1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/mmu_context.h>
59 #include <linux/percpu.h>
60 #include <linux/slab.h>
61 #include <linux/kthread.h>
62 #include <linux/blkdev.h>
63 #include <linux/bvec.h>
64 #include <linux/net.h>
66 #include <net/af_unix.h>
68 #include <linux/anon_inodes.h>
69 #include <linux/sched/mm.h>
70 #include <linux/uaccess.h>
71 #include <linux/nospec.h>
72 #include <linux/sizes.h>
73 #include <linux/hugetlb.h>
74 #include <linux/highmem.h>
75 #include <linux/namei.h>
76 #include <linux/fsnotify.h>
77 #include <linux/fadvise.h>
78 #include <linux/eventpoll.h>
79 #include <linux/fs_struct.h>
80 #include <linux/splice.h>
81 #include <linux/task_work.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
155 * Number of completion events lost because the queue was full;
156 * this should be avoided by the application by making sure
157 * there are not more requests pending than there is space in
158 * the completion queue.
160 * Written by the kernel, shouldn't be modified by the
161 * application (i.e. get number of "new events" by comparing to
164 * As completion events come in out of order this counter is not
165 * ordered with any other data.
169 * Ring buffer of completion events.
171 * The kernel writes completion events fresh every time they are
172 * produced, so the application is allowed to modify pending
175 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
178 struct io_mapped_ubuf {
181 struct bio_vec *bvec;
182 unsigned int nr_bvecs;
185 struct fixed_file_table {
189 struct fixed_file_ref_node {
190 struct percpu_ref refs;
191 struct list_head node;
192 struct list_head file_list;
193 struct fixed_file_data *file_data;
194 struct work_struct work;
197 struct fixed_file_data {
198 struct fixed_file_table *table;
199 struct io_ring_ctx *ctx;
201 struct percpu_ref *cur_refs;
202 struct percpu_ref refs;
203 struct completion done;
204 struct list_head ref_list;
209 struct list_head list;
217 struct percpu_ref refs;
218 } ____cacheline_aligned_in_smp;
222 unsigned int compat: 1;
223 unsigned int account_mem: 1;
224 unsigned int cq_overflow_flushed: 1;
225 unsigned int drain_next: 1;
226 unsigned int eventfd_async: 1;
229 * Ring buffer of indices into array of io_uring_sqe, which is
230 * mmapped by the application using the IORING_OFF_SQES offset.
232 * This indirection could e.g. be used to assign fixed
233 * io_uring_sqe entries to operations and only submit them to
234 * the queue when needed.
236 * The kernel modifies neither the indices array nor the entries
240 unsigned cached_sq_head;
243 unsigned sq_thread_idle;
244 unsigned cached_sq_dropped;
245 atomic_t cached_cq_overflow;
246 unsigned long sq_check_overflow;
248 struct list_head defer_list;
249 struct list_head timeout_list;
250 struct list_head cq_overflow_list;
252 wait_queue_head_t inflight_wait;
253 struct io_uring_sqe *sq_sqes;
254 } ____cacheline_aligned_in_smp;
256 struct io_rings *rings;
260 struct task_struct *sqo_thread; /* if using sq thread polling */
261 struct mm_struct *sqo_mm;
262 wait_queue_head_t sqo_wait;
265 * If used, fixed file set. Writers must ensure that ->refs is dead,
266 * readers must ensure that ->refs is alive as long as the file* is
267 * used. Only updated through io_uring_register(2).
269 struct fixed_file_data *file_data;
270 unsigned nr_user_files;
272 struct file *ring_file;
274 /* if used, fixed mapped user buffers */
275 unsigned nr_user_bufs;
276 struct io_mapped_ubuf *user_bufs;
278 struct user_struct *user;
280 const struct cred *creds;
282 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
283 struct completion *completions;
285 /* if all else fails... */
286 struct io_kiocb *fallback_req;
288 #if defined(CONFIG_UNIX)
289 struct socket *ring_sock;
292 struct idr io_buffer_idr;
294 struct idr personality_idr;
297 unsigned cached_cq_tail;
300 atomic_t cq_timeouts;
301 unsigned long cq_check_overflow;
302 struct wait_queue_head cq_wait;
303 struct fasync_struct *cq_fasync;
304 struct eventfd_ctx *cq_ev_fd;
305 } ____cacheline_aligned_in_smp;
308 struct mutex uring_lock;
309 wait_queue_head_t wait;
310 } ____cacheline_aligned_in_smp;
313 spinlock_t completion_lock;
316 * ->poll_list is protected by the ctx->uring_lock for
317 * io_uring instances that don't use IORING_SETUP_SQPOLL.
318 * For SQPOLL, only the single threaded io_sq_thread() will
319 * manipulate the list, hence no extra locking is needed there.
321 struct list_head poll_list;
322 struct hlist_head *cancel_hash;
323 unsigned cancel_hash_bits;
324 bool poll_multi_file;
326 spinlock_t inflight_lock;
327 struct list_head inflight_list;
328 } ____cacheline_aligned_in_smp;
330 struct work_struct exit_work;
334 * First field must be the file pointer in all the
335 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
337 struct io_poll_iocb {
340 struct wait_queue_head *head;
346 struct wait_queue_entry wait;
351 struct file *put_file;
355 struct io_timeout_data {
356 struct io_kiocb *req;
357 struct hrtimer timer;
358 struct timespec64 ts;
359 enum hrtimer_mode mode;
364 struct sockaddr __user *addr;
365 int __user *addr_len;
367 unsigned long nofile;
391 /* NOTE: kiocb has the file as the first member, so don't do it here */
399 struct sockaddr __user *addr;
406 struct user_msghdr __user *msg;
412 struct io_buffer *kbuf;
421 struct filename *filename;
422 struct statx __user *buffer;
424 unsigned long nofile;
427 struct io_files_update {
453 struct epoll_event event;
457 struct file *file_out;
458 struct file *file_in;
465 struct io_provide_buf {
474 struct io_async_connect {
475 struct sockaddr_storage address;
478 struct io_async_msghdr {
479 struct iovec fast_iov[UIO_FASTIOV];
481 struct sockaddr __user *uaddr;
483 struct sockaddr_storage addr;
487 struct iovec fast_iov[UIO_FASTIOV];
493 struct io_async_ctx {
495 struct io_async_rw rw;
496 struct io_async_msghdr msg;
497 struct io_async_connect connect;
498 struct io_timeout_data timeout;
503 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
504 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
505 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
506 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
507 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
508 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
516 REQ_F_IOPOLL_COMPLETED_BIT,
517 REQ_F_LINK_TIMEOUT_BIT,
521 REQ_F_TIMEOUT_NOSEQ_BIT,
522 REQ_F_COMP_LOCKED_BIT,
523 REQ_F_NEED_CLEANUP_BIT,
526 REQ_F_BUFFER_SELECTED_BIT,
527 REQ_F_NO_FILE_TABLE_BIT,
529 /* not a real bit, just to check we're not overflowing the space */
535 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
536 /* drain existing IO first */
537 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
539 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
540 /* doesn't sever on completion < 0 */
541 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
543 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
544 /* IOSQE_BUFFER_SELECT */
545 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
548 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
549 /* already grabbed next link */
550 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
551 /* fail rest of links */
552 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
553 /* on inflight list */
554 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
555 /* read/write uses file position */
556 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
557 /* must not punt to workers */
558 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
559 /* polled IO has completed */
560 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
561 /* has linked timeout */
562 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
563 /* timeout request */
564 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
566 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
567 /* must be punted even for NONBLOCK */
568 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
569 /* no timeout sequence */
570 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
571 /* completion under lock */
572 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
574 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
575 /* in overflow list */
576 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
577 /* already went through poll handler */
578 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
579 /* buffer already selected */
580 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
581 /* doesn't need file table for this request */
582 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
586 struct io_poll_iocb poll;
587 struct io_wq_work work;
591 * NOTE! Each of the iocb union members has the file pointer
592 * as the first entry in their struct definition. So you can
593 * access the file pointer through any of the sub-structs,
594 * or directly as just 'ki_filp' in this struct.
600 struct io_poll_iocb poll;
601 struct io_accept accept;
603 struct io_cancel cancel;
604 struct io_timeout timeout;
605 struct io_connect connect;
606 struct io_sr_msg sr_msg;
608 struct io_close close;
609 struct io_files_update files_update;
610 struct io_fadvise fadvise;
611 struct io_madvise madvise;
612 struct io_epoll epoll;
613 struct io_splice splice;
614 struct io_provide_buf pbuf;
617 struct io_async_ctx *io;
619 bool needs_fixed_file;
624 struct io_ring_ctx *ctx;
625 struct list_head list;
628 struct task_struct *task;
634 struct list_head link_list;
636 struct list_head inflight_entry;
638 struct percpu_ref *fixed_file_refs;
642 * Only commands that never go async can use the below fields,
643 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
644 * async armed poll handlers for regular commands. The latter
645 * restore the work, if needed.
648 struct callback_head task_work;
649 struct hlist_node hash_node;
650 struct async_poll *apoll;
652 struct io_wq_work work;
656 #define IO_PLUG_THRESHOLD 2
657 #define IO_IOPOLL_BATCH 8
659 struct io_submit_state {
660 struct blk_plug plug;
663 * io_kiocb alloc cache
665 void *reqs[IO_IOPOLL_BATCH];
666 unsigned int free_reqs;
669 * File reference cache
673 unsigned int has_refs;
674 unsigned int used_refs;
675 unsigned int ios_left;
679 /* needs req->io allocated for deferral/async */
680 unsigned async_ctx : 1;
681 /* needs current->mm setup, does mm access */
682 unsigned needs_mm : 1;
683 /* needs req->file assigned */
684 unsigned needs_file : 1;
685 /* hash wq insertion if file is a regular file */
686 unsigned hash_reg_file : 1;
687 /* unbound wq insertion if file is a non-regular file */
688 unsigned unbound_nonreg_file : 1;
689 /* opcode is not supported by this kernel */
690 unsigned not_supported : 1;
691 /* needs file table */
692 unsigned file_table : 1;
694 unsigned needs_fs : 1;
695 /* set if opcode supports polled "wait" */
697 unsigned pollout : 1;
698 /* op supports buffer selection */
699 unsigned buffer_select : 1;
702 static const struct io_op_def io_op_defs[] = {
703 [IORING_OP_NOP] = {},
704 [IORING_OP_READV] = {
708 .unbound_nonreg_file = 1,
712 [IORING_OP_WRITEV] = {
717 .unbound_nonreg_file = 1,
720 [IORING_OP_FSYNC] = {
723 [IORING_OP_READ_FIXED] = {
725 .unbound_nonreg_file = 1,
728 [IORING_OP_WRITE_FIXED] = {
731 .unbound_nonreg_file = 1,
734 [IORING_OP_POLL_ADD] = {
736 .unbound_nonreg_file = 1,
738 [IORING_OP_POLL_REMOVE] = {},
739 [IORING_OP_SYNC_FILE_RANGE] = {
742 [IORING_OP_SENDMSG] = {
746 .unbound_nonreg_file = 1,
750 [IORING_OP_RECVMSG] = {
754 .unbound_nonreg_file = 1,
759 [IORING_OP_TIMEOUT] = {
763 [IORING_OP_TIMEOUT_REMOVE] = {},
764 [IORING_OP_ACCEPT] = {
767 .unbound_nonreg_file = 1,
771 [IORING_OP_ASYNC_CANCEL] = {},
772 [IORING_OP_LINK_TIMEOUT] = {
776 [IORING_OP_CONNECT] = {
780 .unbound_nonreg_file = 1,
783 [IORING_OP_FALLOCATE] = {
786 [IORING_OP_OPENAT] = {
790 [IORING_OP_CLOSE] = {
794 [IORING_OP_FILES_UPDATE] = {
798 [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] = {
839 [IORING_OP_EPOLL_CTL] = {
840 .unbound_nonreg_file = 1,
843 [IORING_OP_SPLICE] = {
846 .unbound_nonreg_file = 1,
848 [IORING_OP_PROVIDE_BUFFERS] = {},
849 [IORING_OP_REMOVE_BUFFERS] = {},
852 static void io_wq_submit_work(struct io_wq_work **workptr);
853 static void io_cqring_fill_event(struct io_kiocb *req, long res);
854 static void io_put_req(struct io_kiocb *req);
855 static void __io_double_put_req(struct io_kiocb *req);
856 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
857 static void io_queue_linked_timeout(struct io_kiocb *req);
858 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
859 struct io_uring_files_update *ip,
861 static int io_grab_files(struct io_kiocb *req);
862 static void io_cleanup_req(struct io_kiocb *req);
863 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
864 int fd, struct file **out_file, bool fixed);
865 static void __io_queue_sqe(struct io_kiocb *req,
866 const struct io_uring_sqe *sqe);
868 static struct kmem_cache *req_cachep;
870 static const struct file_operations io_uring_fops;
872 struct sock *io_uring_get_socket(struct file *file)
874 #if defined(CONFIG_UNIX)
875 if (file->f_op == &io_uring_fops) {
876 struct io_ring_ctx *ctx = file->private_data;
878 return ctx->ring_sock->sk;
883 EXPORT_SYMBOL(io_uring_get_socket);
885 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
887 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
889 complete(&ctx->completions[0]);
892 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
894 struct io_ring_ctx *ctx;
897 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
901 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
902 if (!ctx->fallback_req)
905 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
906 if (!ctx->completions)
910 * Use 5 bits less than the max cq entries, that should give us around
911 * 32 entries per hash list if totally full and uniformly spread.
913 hash_bits = ilog2(p->cq_entries);
917 ctx->cancel_hash_bits = hash_bits;
918 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
920 if (!ctx->cancel_hash)
922 __hash_init(ctx->cancel_hash, 1U << hash_bits);
924 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
925 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
928 ctx->flags = p->flags;
929 init_waitqueue_head(&ctx->sqo_wait);
930 init_waitqueue_head(&ctx->cq_wait);
931 INIT_LIST_HEAD(&ctx->cq_overflow_list);
932 init_completion(&ctx->completions[0]);
933 init_completion(&ctx->completions[1]);
934 idr_init(&ctx->io_buffer_idr);
935 idr_init(&ctx->personality_idr);
936 mutex_init(&ctx->uring_lock);
937 init_waitqueue_head(&ctx->wait);
938 spin_lock_init(&ctx->completion_lock);
939 INIT_LIST_HEAD(&ctx->poll_list);
940 INIT_LIST_HEAD(&ctx->defer_list);
941 INIT_LIST_HEAD(&ctx->timeout_list);
942 init_waitqueue_head(&ctx->inflight_wait);
943 spin_lock_init(&ctx->inflight_lock);
944 INIT_LIST_HEAD(&ctx->inflight_list);
947 if (ctx->fallback_req)
948 kmem_cache_free(req_cachep, ctx->fallback_req);
949 kfree(ctx->completions);
950 kfree(ctx->cancel_hash);
955 static inline bool __req_need_defer(struct io_kiocb *req)
957 struct io_ring_ctx *ctx = req->ctx;
959 return req->sequence != ctx->cached_cq_tail
960 + atomic_read(&ctx->cached_cq_overflow);
963 static inline bool req_need_defer(struct io_kiocb *req)
965 if (unlikely(req->flags & REQ_F_IO_DRAIN))
966 return __req_need_defer(req);
971 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
973 struct io_kiocb *req;
975 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
976 if (req && !req_need_defer(req)) {
977 list_del_init(&req->list);
984 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
986 struct io_kiocb *req;
988 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
990 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
992 if (!__req_need_defer(req)) {
993 list_del_init(&req->list);
1001 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1003 struct io_rings *rings = ctx->rings;
1005 /* order cqe stores with ring update */
1006 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1008 if (wq_has_sleeper(&ctx->cq_wait)) {
1009 wake_up_interruptible(&ctx->cq_wait);
1010 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1014 static inline void io_req_work_grab_env(struct io_kiocb *req,
1015 const struct io_op_def *def)
1017 if (!req->work.mm && def->needs_mm) {
1018 mmgrab(current->mm);
1019 req->work.mm = current->mm;
1021 if (!req->work.creds)
1022 req->work.creds = get_current_cred();
1023 if (!req->work.fs && def->needs_fs) {
1024 spin_lock(¤t->fs->lock);
1025 if (!current->fs->in_exec) {
1026 req->work.fs = current->fs;
1027 req->work.fs->users++;
1029 req->work.flags |= IO_WQ_WORK_CANCEL;
1031 spin_unlock(¤t->fs->lock);
1033 if (!req->work.task_pid)
1034 req->work.task_pid = task_pid_vnr(current);
1037 static inline void io_req_work_drop_env(struct io_kiocb *req)
1040 mmdrop(req->work.mm);
1041 req->work.mm = NULL;
1043 if (req->work.creds) {
1044 put_cred(req->work.creds);
1045 req->work.creds = NULL;
1048 struct fs_struct *fs = req->work.fs;
1050 spin_lock(&req->work.fs->lock);
1053 spin_unlock(&req->work.fs->lock);
1059 static inline void io_prep_async_work(struct io_kiocb *req,
1060 struct io_kiocb **link)
1062 const struct io_op_def *def = &io_op_defs[req->opcode];
1064 if (req->flags & REQ_F_ISREG) {
1065 if (def->hash_reg_file)
1066 io_wq_hash_work(&req->work, file_inode(req->file));
1068 if (def->unbound_nonreg_file)
1069 req->work.flags |= IO_WQ_WORK_UNBOUND;
1072 io_req_work_grab_env(req, def);
1074 *link = io_prep_linked_timeout(req);
1077 static inline void io_queue_async_work(struct io_kiocb *req)
1079 struct io_ring_ctx *ctx = req->ctx;
1080 struct io_kiocb *link;
1082 io_prep_async_work(req, &link);
1084 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1085 &req->work, req->flags);
1086 io_wq_enqueue(ctx->io_wq, &req->work);
1089 io_queue_linked_timeout(link);
1092 static void io_kill_timeout(struct io_kiocb *req)
1096 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1098 atomic_inc(&req->ctx->cq_timeouts);
1099 list_del_init(&req->list);
1100 req->flags |= REQ_F_COMP_LOCKED;
1101 io_cqring_fill_event(req, 0);
1106 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1108 struct io_kiocb *req, *tmp;
1110 spin_lock_irq(&ctx->completion_lock);
1111 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1112 io_kill_timeout(req);
1113 spin_unlock_irq(&ctx->completion_lock);
1116 static void io_commit_cqring(struct io_ring_ctx *ctx)
1118 struct io_kiocb *req;
1120 while ((req = io_get_timeout_req(ctx)) != NULL)
1121 io_kill_timeout(req);
1123 __io_commit_cqring(ctx);
1125 while ((req = io_get_deferred_req(ctx)) != NULL)
1126 io_queue_async_work(req);
1129 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1131 struct io_rings *rings = ctx->rings;
1134 tail = ctx->cached_cq_tail;
1136 * writes to the cq entry need to come after reading head; the
1137 * control dependency is enough as we're using WRITE_ONCE to
1140 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1143 ctx->cached_cq_tail++;
1144 return &rings->cqes[tail & ctx->cq_mask];
1147 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1151 if (!ctx->eventfd_async)
1153 return io_wq_current_is_worker();
1156 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1158 if (waitqueue_active(&ctx->wait))
1159 wake_up(&ctx->wait);
1160 if (waitqueue_active(&ctx->sqo_wait))
1161 wake_up(&ctx->sqo_wait);
1162 if (io_should_trigger_evfd(ctx))
1163 eventfd_signal(ctx->cq_ev_fd, 1);
1166 /* Returns true if there are no backlogged entries after the flush */
1167 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1169 struct io_rings *rings = ctx->rings;
1170 struct io_uring_cqe *cqe;
1171 struct io_kiocb *req;
1172 unsigned long flags;
1176 if (list_empty_careful(&ctx->cq_overflow_list))
1178 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1179 rings->cq_ring_entries))
1183 spin_lock_irqsave(&ctx->completion_lock, flags);
1185 /* if force is set, the ring is going away. always drop after that */
1187 ctx->cq_overflow_flushed = 1;
1190 while (!list_empty(&ctx->cq_overflow_list)) {
1191 cqe = io_get_cqring(ctx);
1195 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1197 list_move(&req->list, &list);
1198 req->flags &= ~REQ_F_OVERFLOW;
1200 WRITE_ONCE(cqe->user_data, req->user_data);
1201 WRITE_ONCE(cqe->res, req->result);
1202 WRITE_ONCE(cqe->flags, req->cflags);
1204 WRITE_ONCE(ctx->rings->cq_overflow,
1205 atomic_inc_return(&ctx->cached_cq_overflow));
1209 io_commit_cqring(ctx);
1211 clear_bit(0, &ctx->sq_check_overflow);
1212 clear_bit(0, &ctx->cq_check_overflow);
1214 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1215 io_cqring_ev_posted(ctx);
1217 while (!list_empty(&list)) {
1218 req = list_first_entry(&list, struct io_kiocb, list);
1219 list_del(&req->list);
1226 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1228 struct io_ring_ctx *ctx = req->ctx;
1229 struct io_uring_cqe *cqe;
1231 trace_io_uring_complete(ctx, req->user_data, res);
1234 * If we can't get a cq entry, userspace overflowed the
1235 * submission (by quite a lot). Increment the overflow count in
1238 cqe = io_get_cqring(ctx);
1240 WRITE_ONCE(cqe->user_data, req->user_data);
1241 WRITE_ONCE(cqe->res, res);
1242 WRITE_ONCE(cqe->flags, cflags);
1243 } else if (ctx->cq_overflow_flushed) {
1244 WRITE_ONCE(ctx->rings->cq_overflow,
1245 atomic_inc_return(&ctx->cached_cq_overflow));
1247 if (list_empty(&ctx->cq_overflow_list)) {
1248 set_bit(0, &ctx->sq_check_overflow);
1249 set_bit(0, &ctx->cq_check_overflow);
1251 req->flags |= REQ_F_OVERFLOW;
1252 refcount_inc(&req->refs);
1254 req->cflags = cflags;
1255 list_add_tail(&req->list, &ctx->cq_overflow_list);
1259 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1261 __io_cqring_fill_event(req, res, 0);
1264 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1266 struct io_ring_ctx *ctx = req->ctx;
1267 unsigned long flags;
1269 spin_lock_irqsave(&ctx->completion_lock, flags);
1270 __io_cqring_fill_event(req, res, cflags);
1271 io_commit_cqring(ctx);
1272 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1274 io_cqring_ev_posted(ctx);
1277 static void io_cqring_add_event(struct io_kiocb *req, long res)
1279 __io_cqring_add_event(req, res, 0);
1282 static inline bool io_is_fallback_req(struct io_kiocb *req)
1284 return req == (struct io_kiocb *)
1285 ((unsigned long) req->ctx->fallback_req & ~1UL);
1288 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1290 struct io_kiocb *req;
1292 req = ctx->fallback_req;
1293 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1299 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1300 struct io_submit_state *state)
1302 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1303 struct io_kiocb *req;
1306 req = kmem_cache_alloc(req_cachep, gfp);
1309 } else if (!state->free_reqs) {
1313 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1314 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1317 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1318 * retry single alloc to be on the safe side.
1320 if (unlikely(ret <= 0)) {
1321 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1322 if (!state->reqs[0])
1326 state->free_reqs = ret - 1;
1327 req = state->reqs[ret - 1];
1330 req = state->reqs[state->free_reqs];
1335 return io_get_fallback_req(ctx);
1338 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1342 percpu_ref_put(req->fixed_file_refs);
1347 static void __io_req_aux_free(struct io_kiocb *req)
1349 if (req->flags & REQ_F_NEED_CLEANUP)
1350 io_cleanup_req(req);
1354 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1356 put_task_struct(req->task);
1358 io_req_work_drop_env(req);
1361 static void __io_free_req(struct io_kiocb *req)
1363 __io_req_aux_free(req);
1365 if (req->flags & REQ_F_INFLIGHT) {
1366 struct io_ring_ctx *ctx = req->ctx;
1367 unsigned long flags;
1369 spin_lock_irqsave(&ctx->inflight_lock, flags);
1370 list_del(&req->inflight_entry);
1371 if (waitqueue_active(&ctx->inflight_wait))
1372 wake_up(&ctx->inflight_wait);
1373 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1376 percpu_ref_put(&req->ctx->refs);
1377 if (likely(!io_is_fallback_req(req)))
1378 kmem_cache_free(req_cachep, req);
1380 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1384 void *reqs[IO_IOPOLL_BATCH];
1389 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1393 if (rb->need_iter) {
1394 int i, inflight = 0;
1395 unsigned long flags;
1397 for (i = 0; i < rb->to_free; i++) {
1398 struct io_kiocb *req = rb->reqs[i];
1400 if (req->flags & REQ_F_INFLIGHT)
1402 __io_req_aux_free(req);
1407 spin_lock_irqsave(&ctx->inflight_lock, flags);
1408 for (i = 0; i < rb->to_free; i++) {
1409 struct io_kiocb *req = rb->reqs[i];
1411 if (req->flags & REQ_F_INFLIGHT) {
1412 list_del(&req->inflight_entry);
1417 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1419 if (waitqueue_active(&ctx->inflight_wait))
1420 wake_up(&ctx->inflight_wait);
1423 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1424 percpu_ref_put_many(&ctx->refs, rb->to_free);
1425 rb->to_free = rb->need_iter = 0;
1428 static bool io_link_cancel_timeout(struct io_kiocb *req)
1430 struct io_ring_ctx *ctx = req->ctx;
1433 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1435 io_cqring_fill_event(req, -ECANCELED);
1436 io_commit_cqring(ctx);
1437 req->flags &= ~REQ_F_LINK_HEAD;
1445 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1447 struct io_ring_ctx *ctx = req->ctx;
1448 bool wake_ev = false;
1450 /* Already got next link */
1451 if (req->flags & REQ_F_LINK_NEXT)
1455 * The list should never be empty when we are called here. But could
1456 * potentially happen if the chain is messed up, check to be on the
1459 while (!list_empty(&req->link_list)) {
1460 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1461 struct io_kiocb, link_list);
1463 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1464 (nxt->flags & REQ_F_TIMEOUT))) {
1465 list_del_init(&nxt->link_list);
1466 wake_ev |= io_link_cancel_timeout(nxt);
1467 req->flags &= ~REQ_F_LINK_TIMEOUT;
1471 list_del_init(&req->link_list);
1472 if (!list_empty(&nxt->link_list))
1473 nxt->flags |= REQ_F_LINK_HEAD;
1478 req->flags |= REQ_F_LINK_NEXT;
1480 io_cqring_ev_posted(ctx);
1484 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1486 static void io_fail_links(struct io_kiocb *req)
1488 struct io_ring_ctx *ctx = req->ctx;
1489 unsigned long flags;
1491 spin_lock_irqsave(&ctx->completion_lock, flags);
1493 while (!list_empty(&req->link_list)) {
1494 struct io_kiocb *link = list_first_entry(&req->link_list,
1495 struct io_kiocb, link_list);
1497 list_del_init(&link->link_list);
1498 trace_io_uring_fail_link(req, link);
1500 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1501 link->opcode == IORING_OP_LINK_TIMEOUT) {
1502 io_link_cancel_timeout(link);
1504 io_cqring_fill_event(link, -ECANCELED);
1505 __io_double_put_req(link);
1507 req->flags &= ~REQ_F_LINK_TIMEOUT;
1510 io_commit_cqring(ctx);
1511 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1512 io_cqring_ev_posted(ctx);
1515 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1517 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1521 * If LINK is set, we have dependent requests in this chain. If we
1522 * didn't fail this request, queue the first one up, moving any other
1523 * dependencies to the next request. In case of failure, fail the rest
1526 if (req->flags & REQ_F_FAIL_LINK) {
1528 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1529 REQ_F_LINK_TIMEOUT) {
1530 struct io_ring_ctx *ctx = req->ctx;
1531 unsigned long flags;
1534 * If this is a timeout link, we could be racing with the
1535 * timeout timer. Grab the completion lock for this case to
1536 * protect against that.
1538 spin_lock_irqsave(&ctx->completion_lock, flags);
1539 io_req_link_next(req, nxt);
1540 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1542 io_req_link_next(req, nxt);
1546 static void io_free_req(struct io_kiocb *req)
1548 struct io_kiocb *nxt = NULL;
1550 io_req_find_next(req, &nxt);
1554 io_queue_async_work(nxt);
1557 static void io_link_work_cb(struct io_wq_work **workptr)
1559 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1560 struct io_kiocb *link;
1562 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1563 io_queue_linked_timeout(link);
1564 io_wq_submit_work(workptr);
1567 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1569 struct io_kiocb *link;
1570 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1572 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1573 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1575 *workptr = &nxt->work;
1576 link = io_prep_linked_timeout(nxt);
1578 nxt->work.func = io_link_work_cb;
1582 * Drop reference to request, return next in chain (if there is one) if this
1583 * was the last reference to this request.
1585 __attribute__((nonnull))
1586 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1588 if (refcount_dec_and_test(&req->refs)) {
1589 io_req_find_next(req, nxtptr);
1594 static void io_put_req(struct io_kiocb *req)
1596 if (refcount_dec_and_test(&req->refs))
1600 static void io_steal_work(struct io_kiocb *req,
1601 struct io_wq_work **workptr)
1604 * It's in an io-wq worker, so there always should be at least
1605 * one reference, which will be dropped in io_put_work() just
1606 * after the current handler returns.
1608 * It also means, that if the counter dropped to 1, then there is
1609 * no asynchronous users left, so it's safe to steal the next work.
1611 if (refcount_read(&req->refs) == 1) {
1612 struct io_kiocb *nxt = NULL;
1614 io_req_find_next(req, &nxt);
1616 io_wq_assign_next(workptr, nxt);
1621 * Must only be used if we don't need to care about links, usually from
1622 * within the completion handling itself.
1624 static void __io_double_put_req(struct io_kiocb *req)
1626 /* drop both submit and complete references */
1627 if (refcount_sub_and_test(2, &req->refs))
1631 static void io_double_put_req(struct io_kiocb *req)
1633 /* drop both submit and complete references */
1634 if (refcount_sub_and_test(2, &req->refs))
1638 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1640 struct io_rings *rings = ctx->rings;
1642 if (test_bit(0, &ctx->cq_check_overflow)) {
1644 * noflush == true is from the waitqueue handler, just ensure
1645 * we wake up the task, and the next invocation will flush the
1646 * entries. We cannot safely to it from here.
1648 if (noflush && !list_empty(&ctx->cq_overflow_list))
1651 io_cqring_overflow_flush(ctx, false);
1654 /* See comment at the top of this file */
1656 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1659 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1661 struct io_rings *rings = ctx->rings;
1663 /* make sure SQ entry isn't read before tail */
1664 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1667 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1669 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1672 if (req->file || req->io)
1675 rb->reqs[rb->to_free++] = req;
1676 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1677 io_free_req_many(req->ctx, rb);
1681 static int io_put_kbuf(struct io_kiocb *req)
1683 struct io_buffer *kbuf;
1686 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1687 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1688 cflags |= IORING_CQE_F_BUFFER;
1695 * Find and free completed poll iocbs
1697 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1698 struct list_head *done)
1700 struct req_batch rb;
1701 struct io_kiocb *req;
1703 rb.to_free = rb.need_iter = 0;
1704 while (!list_empty(done)) {
1707 req = list_first_entry(done, struct io_kiocb, list);
1708 list_del(&req->list);
1710 if (req->flags & REQ_F_BUFFER_SELECTED)
1711 cflags = io_put_kbuf(req);
1713 __io_cqring_fill_event(req, req->result, cflags);
1716 if (refcount_dec_and_test(&req->refs) &&
1717 !io_req_multi_free(&rb, req))
1721 io_commit_cqring(ctx);
1722 if (ctx->flags & IORING_SETUP_SQPOLL)
1723 io_cqring_ev_posted(ctx);
1724 io_free_req_many(ctx, &rb);
1727 static void io_iopoll_queue(struct list_head *again)
1729 struct io_kiocb *req;
1732 req = list_first_entry(again, struct io_kiocb, list);
1733 list_del(&req->list);
1734 refcount_inc(&req->refs);
1735 io_queue_async_work(req);
1736 } while (!list_empty(again));
1739 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1742 struct io_kiocb *req, *tmp;
1749 * Only spin for completions if we don't have multiple devices hanging
1750 * off our complete list, and we're under the requested amount.
1752 spin = !ctx->poll_multi_file && *nr_events < min;
1755 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1756 struct kiocb *kiocb = &req->rw.kiocb;
1759 * Move completed and retryable entries to our local lists.
1760 * If we find a request that requires polling, break out
1761 * and complete those lists first, if we have entries there.
1763 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1764 list_move_tail(&req->list, &done);
1767 if (!list_empty(&done))
1770 if (req->result == -EAGAIN) {
1771 list_move_tail(&req->list, &again);
1774 if (!list_empty(&again))
1777 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1786 if (!list_empty(&done))
1787 io_iopoll_complete(ctx, nr_events, &done);
1789 if (!list_empty(&again))
1790 io_iopoll_queue(&again);
1796 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1797 * non-spinning poll check - we'll still enter the driver poll loop, but only
1798 * as a non-spinning completion check.
1800 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1803 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1806 ret = io_do_iopoll(ctx, nr_events, min);
1809 if (!min || *nr_events >= min)
1817 * We can't just wait for polled events to come to us, we have to actively
1818 * find and complete them.
1820 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1822 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1825 mutex_lock(&ctx->uring_lock);
1826 while (!list_empty(&ctx->poll_list)) {
1827 unsigned int nr_events = 0;
1829 io_iopoll_getevents(ctx, &nr_events, 1);
1832 * Ensure we allow local-to-the-cpu processing to take place,
1833 * in this case we need to ensure that we reap all events.
1837 mutex_unlock(&ctx->uring_lock);
1840 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1843 int iters = 0, ret = 0;
1846 * We disallow the app entering submit/complete with polling, but we
1847 * still need to lock the ring to prevent racing with polled issue
1848 * that got punted to a workqueue.
1850 mutex_lock(&ctx->uring_lock);
1855 * Don't enter poll loop if we already have events pending.
1856 * If we do, we can potentially be spinning for commands that
1857 * already triggered a CQE (eg in error).
1859 if (io_cqring_events(ctx, false))
1863 * If a submit got punted to a workqueue, we can have the
1864 * application entering polling for a command before it gets
1865 * issued. That app will hold the uring_lock for the duration
1866 * of the poll right here, so we need to take a breather every
1867 * now and then to ensure that the issue has a chance to add
1868 * the poll to the issued list. Otherwise we can spin here
1869 * forever, while the workqueue is stuck trying to acquire the
1872 if (!(++iters & 7)) {
1873 mutex_unlock(&ctx->uring_lock);
1874 mutex_lock(&ctx->uring_lock);
1877 if (*nr_events < min)
1878 tmin = min - *nr_events;
1880 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1884 } while (min && !*nr_events && !need_resched());
1886 mutex_unlock(&ctx->uring_lock);
1890 static void kiocb_end_write(struct io_kiocb *req)
1893 * Tell lockdep we inherited freeze protection from submission
1896 if (req->flags & REQ_F_ISREG) {
1897 struct inode *inode = file_inode(req->file);
1899 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1901 file_end_write(req->file);
1904 static inline void req_set_fail_links(struct io_kiocb *req)
1906 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1907 req->flags |= REQ_F_FAIL_LINK;
1910 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1912 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1915 if (kiocb->ki_flags & IOCB_WRITE)
1916 kiocb_end_write(req);
1918 if (res != req->result)
1919 req_set_fail_links(req);
1920 if (req->flags & REQ_F_BUFFER_SELECTED)
1921 cflags = io_put_kbuf(req);
1922 __io_cqring_add_event(req, res, cflags);
1925 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1927 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1929 io_complete_rw_common(kiocb, res);
1933 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1935 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1937 if (kiocb->ki_flags & IOCB_WRITE)
1938 kiocb_end_write(req);
1940 if (res != req->result)
1941 req_set_fail_links(req);
1944 req->flags |= REQ_F_IOPOLL_COMPLETED;
1948 * After the iocb has been issued, it's safe to be found on the poll list.
1949 * Adding the kiocb to the list AFTER submission ensures that we don't
1950 * find it from a io_iopoll_getevents() thread before the issuer is done
1951 * accessing the kiocb cookie.
1953 static void io_iopoll_req_issued(struct io_kiocb *req)
1955 struct io_ring_ctx *ctx = req->ctx;
1958 * Track whether we have multiple files in our lists. This will impact
1959 * how we do polling eventually, not spinning if we're on potentially
1960 * different devices.
1962 if (list_empty(&ctx->poll_list)) {
1963 ctx->poll_multi_file = false;
1964 } else if (!ctx->poll_multi_file) {
1965 struct io_kiocb *list_req;
1967 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1969 if (list_req->file != req->file)
1970 ctx->poll_multi_file = true;
1974 * For fast devices, IO may have already completed. If it has, add
1975 * it to the front so we find it first.
1977 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1978 list_add(&req->list, &ctx->poll_list);
1980 list_add_tail(&req->list, &ctx->poll_list);
1982 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1983 wq_has_sleeper(&ctx->sqo_wait))
1984 wake_up(&ctx->sqo_wait);
1987 static void io_file_put(struct io_submit_state *state)
1990 int diff = state->has_refs - state->used_refs;
1993 fput_many(state->file, diff);
1999 * Get as many references to a file as we have IOs left in this submission,
2000 * assuming most submissions are for one file, or at least that each file
2001 * has more than one submission.
2003 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2009 if (state->fd == fd) {
2016 state->file = fget_many(fd, state->ios_left);
2021 state->has_refs = state->ios_left;
2022 state->used_refs = 1;
2028 * If we tracked the file through the SCM inflight mechanism, we could support
2029 * any file. For now, just ensure that anything potentially problematic is done
2032 static bool io_file_supports_async(struct file *file, int rw)
2034 umode_t mode = file_inode(file)->i_mode;
2036 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2038 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2041 if (!(file->f_mode & FMODE_NOWAIT))
2045 return file->f_op->read_iter != NULL;
2047 return file->f_op->write_iter != NULL;
2050 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2051 bool force_nonblock)
2053 struct io_ring_ctx *ctx = req->ctx;
2054 struct kiocb *kiocb = &req->rw.kiocb;
2058 if (S_ISREG(file_inode(req->file)->i_mode))
2059 req->flags |= REQ_F_ISREG;
2061 kiocb->ki_pos = READ_ONCE(sqe->off);
2062 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2063 req->flags |= REQ_F_CUR_POS;
2064 kiocb->ki_pos = req->file->f_pos;
2066 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2067 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2068 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2072 ioprio = READ_ONCE(sqe->ioprio);
2074 ret = ioprio_check_cap(ioprio);
2078 kiocb->ki_ioprio = ioprio;
2080 kiocb->ki_ioprio = get_current_ioprio();
2082 /* don't allow async punt if RWF_NOWAIT was requested */
2083 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2084 (req->file->f_flags & O_NONBLOCK))
2085 req->flags |= REQ_F_NOWAIT;
2088 kiocb->ki_flags |= IOCB_NOWAIT;
2090 if (ctx->flags & IORING_SETUP_IOPOLL) {
2091 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2092 !kiocb->ki_filp->f_op->iopoll)
2095 kiocb->ki_flags |= IOCB_HIPRI;
2096 kiocb->ki_complete = io_complete_rw_iopoll;
2099 if (kiocb->ki_flags & IOCB_HIPRI)
2101 kiocb->ki_complete = io_complete_rw;
2104 req->rw.addr = READ_ONCE(sqe->addr);
2105 req->rw.len = READ_ONCE(sqe->len);
2106 req->buf_index = READ_ONCE(sqe->buf_index);
2110 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2116 case -ERESTARTNOINTR:
2117 case -ERESTARTNOHAND:
2118 case -ERESTART_RESTARTBLOCK:
2120 * We can't just restart the syscall, since previously
2121 * submitted sqes may already be in progress. Just fail this
2127 kiocb->ki_complete(kiocb, ret, 0);
2131 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2133 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2135 if (req->flags & REQ_F_CUR_POS)
2136 req->file->f_pos = kiocb->ki_pos;
2137 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2138 io_complete_rw(kiocb, ret, 0);
2140 io_rw_done(kiocb, ret);
2143 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2144 struct iov_iter *iter)
2146 struct io_ring_ctx *ctx = req->ctx;
2147 size_t len = req->rw.len;
2148 struct io_mapped_ubuf *imu;
2149 u16 index, buf_index;
2153 /* attempt to use fixed buffers without having provided iovecs */
2154 if (unlikely(!ctx->user_bufs))
2157 buf_index = req->buf_index;
2158 if (unlikely(buf_index >= ctx->nr_user_bufs))
2161 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2162 imu = &ctx->user_bufs[index];
2163 buf_addr = req->rw.addr;
2166 if (buf_addr + len < buf_addr)
2168 /* not inside the mapped region */
2169 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2173 * May not be a start of buffer, set size appropriately
2174 * and advance us to the beginning.
2176 offset = buf_addr - imu->ubuf;
2177 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2181 * Don't use iov_iter_advance() here, as it's really slow for
2182 * using the latter parts of a big fixed buffer - it iterates
2183 * over each segment manually. We can cheat a bit here, because
2186 * 1) it's a BVEC iter, we set it up
2187 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2188 * first and last bvec
2190 * So just find our index, and adjust the iterator afterwards.
2191 * If the offset is within the first bvec (or the whole first
2192 * bvec, just use iov_iter_advance(). This makes it easier
2193 * since we can just skip the first segment, which may not
2194 * be PAGE_SIZE aligned.
2196 const struct bio_vec *bvec = imu->bvec;
2198 if (offset <= bvec->bv_len) {
2199 iov_iter_advance(iter, offset);
2201 unsigned long seg_skip;
2203 /* skip first vec */
2204 offset -= bvec->bv_len;
2205 seg_skip = 1 + (offset >> PAGE_SHIFT);
2207 iter->bvec = bvec + seg_skip;
2208 iter->nr_segs -= seg_skip;
2209 iter->count -= bvec->bv_len + offset;
2210 iter->iov_offset = offset & ~PAGE_MASK;
2217 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2220 mutex_unlock(&ctx->uring_lock);
2223 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2226 * "Normal" inline submissions always hold the uring_lock, since we
2227 * grab it from the system call. Same is true for the SQPOLL offload.
2228 * The only exception is when we've detached the request and issue it
2229 * from an async worker thread, grab the lock for that case.
2232 mutex_lock(&ctx->uring_lock);
2235 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2236 int bgid, struct io_buffer *kbuf,
2239 struct io_buffer *head;
2241 if (req->flags & REQ_F_BUFFER_SELECTED)
2244 io_ring_submit_lock(req->ctx, needs_lock);
2246 lockdep_assert_held(&req->ctx->uring_lock);
2248 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2250 if (!list_empty(&head->list)) {
2251 kbuf = list_last_entry(&head->list, struct io_buffer,
2253 list_del(&kbuf->list);
2256 idr_remove(&req->ctx->io_buffer_idr, bgid);
2258 if (*len > kbuf->len)
2261 kbuf = ERR_PTR(-ENOBUFS);
2264 io_ring_submit_unlock(req->ctx, needs_lock);
2269 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2272 struct io_buffer *kbuf;
2275 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2276 bgid = req->buf_index;
2277 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2280 req->rw.addr = (u64) (unsigned long) kbuf;
2281 req->flags |= REQ_F_BUFFER_SELECTED;
2282 return u64_to_user_ptr(kbuf->addr);
2285 #ifdef CONFIG_COMPAT
2286 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2289 struct compat_iovec __user *uiov;
2290 compat_ssize_t clen;
2294 uiov = u64_to_user_ptr(req->rw.addr);
2295 if (!access_ok(uiov, sizeof(*uiov)))
2297 if (__get_user(clen, &uiov->iov_len))
2303 buf = io_rw_buffer_select(req, &len, needs_lock);
2305 return PTR_ERR(buf);
2306 iov[0].iov_base = buf;
2307 iov[0].iov_len = (compat_size_t) len;
2312 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2315 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2319 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2322 len = iov[0].iov_len;
2325 buf = io_rw_buffer_select(req, &len, needs_lock);
2327 return PTR_ERR(buf);
2328 iov[0].iov_base = buf;
2329 iov[0].iov_len = len;
2333 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2336 if (req->flags & REQ_F_BUFFER_SELECTED)
2340 else if (req->rw.len > 1)
2343 #ifdef CONFIG_COMPAT
2344 if (req->ctx->compat)
2345 return io_compat_import(req, iov, needs_lock);
2348 return __io_iov_buffer_select(req, iov, needs_lock);
2351 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2352 struct iovec **iovec, struct iov_iter *iter,
2355 void __user *buf = u64_to_user_ptr(req->rw.addr);
2356 size_t sqe_len = req->rw.len;
2360 opcode = req->opcode;
2361 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2363 return io_import_fixed(req, rw, iter);
2366 /* buffer index only valid with fixed read/write, or buffer select */
2367 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2370 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2371 if (req->flags & REQ_F_BUFFER_SELECT) {
2372 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2375 return PTR_ERR(buf);
2377 req->rw.len = sqe_len;
2380 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2382 return ret < 0 ? ret : sqe_len;
2386 struct io_async_rw *iorw = &req->io->rw;
2389 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2390 if (iorw->iov == iorw->fast_iov)
2395 if (req->flags & REQ_F_BUFFER_SELECT) {
2396 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2398 ret = (*iovec)->iov_len;
2399 iov_iter_init(iter, rw, *iovec, 1, ret);
2405 #ifdef CONFIG_COMPAT
2406 if (req->ctx->compat)
2407 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2411 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2415 * For files that don't have ->read_iter() and ->write_iter(), handle them
2416 * by looping over ->read() or ->write() manually.
2418 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2419 struct iov_iter *iter)
2424 * Don't support polled IO through this interface, and we can't
2425 * support non-blocking either. For the latter, this just causes
2426 * the kiocb to be handled from an async context.
2428 if (kiocb->ki_flags & IOCB_HIPRI)
2430 if (kiocb->ki_flags & IOCB_NOWAIT)
2433 while (iov_iter_count(iter)) {
2437 if (!iov_iter_is_bvec(iter)) {
2438 iovec = iov_iter_iovec(iter);
2440 /* fixed buffers import bvec */
2441 iovec.iov_base = kmap(iter->bvec->bv_page)
2443 iovec.iov_len = min(iter->count,
2444 iter->bvec->bv_len - iter->iov_offset);
2448 nr = file->f_op->read(file, iovec.iov_base,
2449 iovec.iov_len, &kiocb->ki_pos);
2451 nr = file->f_op->write(file, iovec.iov_base,
2452 iovec.iov_len, &kiocb->ki_pos);
2455 if (iov_iter_is_bvec(iter))
2456 kunmap(iter->bvec->bv_page);
2464 if (nr != iovec.iov_len)
2466 iov_iter_advance(iter, nr);
2472 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2473 struct iovec *iovec, struct iovec *fast_iov,
2474 struct iov_iter *iter)
2476 req->io->rw.nr_segs = iter->nr_segs;
2477 req->io->rw.size = io_size;
2478 req->io->rw.iov = iovec;
2479 if (!req->io->rw.iov) {
2480 req->io->rw.iov = req->io->rw.fast_iov;
2481 if (req->io->rw.iov != fast_iov)
2482 memcpy(req->io->rw.iov, fast_iov,
2483 sizeof(struct iovec) * iter->nr_segs);
2485 req->flags |= REQ_F_NEED_CLEANUP;
2489 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2491 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2492 return req->io == NULL;
2495 static int io_alloc_async_ctx(struct io_kiocb *req)
2497 if (!io_op_defs[req->opcode].async_ctx)
2500 return __io_alloc_async_ctx(req);
2503 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2504 struct iovec *iovec, struct iovec *fast_iov,
2505 struct iov_iter *iter)
2507 if (!io_op_defs[req->opcode].async_ctx)
2510 if (__io_alloc_async_ctx(req))
2513 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2518 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2519 bool force_nonblock)
2521 struct io_async_ctx *io;
2522 struct iov_iter iter;
2525 ret = io_prep_rw(req, sqe, force_nonblock);
2529 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2532 /* either don't need iovec imported or already have it */
2533 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2537 io->rw.iov = io->rw.fast_iov;
2539 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2544 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2548 static int io_read(struct io_kiocb *req, bool force_nonblock)
2550 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2551 struct kiocb *kiocb = &req->rw.kiocb;
2552 struct iov_iter iter;
2554 ssize_t io_size, ret;
2556 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2560 /* Ensure we clear previously set non-block flag */
2561 if (!force_nonblock)
2562 kiocb->ki_flags &= ~IOCB_NOWAIT;
2566 if (req->flags & REQ_F_LINK_HEAD)
2567 req->result = io_size;
2570 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2571 * we know to async punt it even if it was opened O_NONBLOCK
2573 if (force_nonblock && !io_file_supports_async(req->file, READ))
2576 iov_count = iov_iter_count(&iter);
2577 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2581 if (req->file->f_op->read_iter)
2582 ret2 = call_read_iter(req->file, kiocb, &iter);
2584 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2586 /* Catch -EAGAIN return for forced non-blocking submission */
2587 if (!force_nonblock || ret2 != -EAGAIN) {
2588 kiocb_done(kiocb, ret2);
2591 ret = io_setup_async_rw(req, io_size, iovec,
2592 inline_vecs, &iter);
2595 /* any defer here is final, must blocking retry */
2596 if (!(req->flags & REQ_F_NOWAIT) &&
2597 !file_can_poll(req->file))
2598 req->flags |= REQ_F_MUST_PUNT;
2604 req->flags &= ~REQ_F_NEED_CLEANUP;
2608 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2609 bool force_nonblock)
2611 struct io_async_ctx *io;
2612 struct iov_iter iter;
2615 ret = io_prep_rw(req, sqe, force_nonblock);
2619 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2622 req->fsize = rlimit(RLIMIT_FSIZE);
2624 /* either don't need iovec imported or already have it */
2625 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2629 io->rw.iov = io->rw.fast_iov;
2631 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2636 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2640 static int io_write(struct io_kiocb *req, bool force_nonblock)
2642 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2643 struct kiocb *kiocb = &req->rw.kiocb;
2644 struct iov_iter iter;
2646 ssize_t ret, io_size;
2648 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2652 /* Ensure we clear previously set non-block flag */
2653 if (!force_nonblock)
2654 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2658 if (req->flags & REQ_F_LINK_HEAD)
2659 req->result = io_size;
2662 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2663 * we know to async punt it even if it was opened O_NONBLOCK
2665 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2668 /* file path doesn't support NOWAIT for non-direct_IO */
2669 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2670 (req->flags & REQ_F_ISREG))
2673 iov_count = iov_iter_count(&iter);
2674 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2679 * Open-code file_start_write here to grab freeze protection,
2680 * which will be released by another thread in
2681 * io_complete_rw(). Fool lockdep by telling it the lock got
2682 * released so that it doesn't complain about the held lock when
2683 * we return to userspace.
2685 if (req->flags & REQ_F_ISREG) {
2686 __sb_start_write(file_inode(req->file)->i_sb,
2687 SB_FREEZE_WRITE, true);
2688 __sb_writers_release(file_inode(req->file)->i_sb,
2691 kiocb->ki_flags |= IOCB_WRITE;
2693 if (!force_nonblock)
2694 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2696 if (req->file->f_op->write_iter)
2697 ret2 = call_write_iter(req->file, kiocb, &iter);
2699 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2701 if (!force_nonblock)
2702 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2705 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2706 * retry them without IOCB_NOWAIT.
2708 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2710 if (!force_nonblock || ret2 != -EAGAIN) {
2711 kiocb_done(kiocb, ret2);
2714 ret = io_setup_async_rw(req, io_size, iovec,
2715 inline_vecs, &iter);
2718 /* any defer here is final, must blocking retry */
2719 if (!file_can_poll(req->file))
2720 req->flags |= REQ_F_MUST_PUNT;
2725 req->flags &= ~REQ_F_NEED_CLEANUP;
2730 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2732 struct io_splice* sp = &req->splice;
2733 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2736 if (req->flags & REQ_F_NEED_CLEANUP)
2740 sp->off_in = READ_ONCE(sqe->splice_off_in);
2741 sp->off_out = READ_ONCE(sqe->off);
2742 sp->len = READ_ONCE(sqe->len);
2743 sp->flags = READ_ONCE(sqe->splice_flags);
2745 if (unlikely(sp->flags & ~valid_flags))
2748 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2749 (sp->flags & SPLICE_F_FD_IN_FIXED));
2752 req->flags |= REQ_F_NEED_CLEANUP;
2754 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2755 req->work.flags |= IO_WQ_WORK_UNBOUND;
2760 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2762 struct io_splice *sp = &req->splice;
2763 struct file *in = sp->file_in;
2764 struct file *out = sp->file_out;
2765 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2766 loff_t *poff_in, *poff_out;
2772 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2773 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2776 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2778 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2779 req->flags &= ~REQ_F_NEED_CLEANUP;
2781 io_cqring_add_event(req, ret);
2783 req_set_fail_links(req);
2789 * IORING_OP_NOP just posts a completion event, nothing else.
2791 static int io_nop(struct io_kiocb *req)
2793 struct io_ring_ctx *ctx = req->ctx;
2795 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2798 io_cqring_add_event(req, 0);
2803 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2805 struct io_ring_ctx *ctx = req->ctx;
2810 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2812 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2815 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2816 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2819 req->sync.off = READ_ONCE(sqe->off);
2820 req->sync.len = READ_ONCE(sqe->len);
2824 static bool io_req_cancelled(struct io_kiocb *req)
2826 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2827 req_set_fail_links(req);
2828 io_cqring_add_event(req, -ECANCELED);
2836 static void __io_fsync(struct io_kiocb *req)
2838 loff_t end = req->sync.off + req->sync.len;
2841 ret = vfs_fsync_range(req->file, req->sync.off,
2842 end > 0 ? end : LLONG_MAX,
2843 req->sync.flags & IORING_FSYNC_DATASYNC);
2845 req_set_fail_links(req);
2846 io_cqring_add_event(req, ret);
2850 static void io_fsync_finish(struct io_wq_work **workptr)
2852 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2854 if (io_req_cancelled(req))
2857 io_steal_work(req, workptr);
2860 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2862 /* fsync always requires a blocking context */
2863 if (force_nonblock) {
2864 req->work.func = io_fsync_finish;
2871 static void __io_fallocate(struct io_kiocb *req)
2875 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2876 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2878 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2880 req_set_fail_links(req);
2881 io_cqring_add_event(req, ret);
2885 static void io_fallocate_finish(struct io_wq_work **workptr)
2887 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2889 if (io_req_cancelled(req))
2891 __io_fallocate(req);
2892 io_steal_work(req, workptr);
2895 static int io_fallocate_prep(struct io_kiocb *req,
2896 const struct io_uring_sqe *sqe)
2898 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2901 req->sync.off = READ_ONCE(sqe->off);
2902 req->sync.len = READ_ONCE(sqe->addr);
2903 req->sync.mode = READ_ONCE(sqe->len);
2904 req->fsize = rlimit(RLIMIT_FSIZE);
2908 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2910 /* fallocate always requiring blocking context */
2911 if (force_nonblock) {
2912 req->work.func = io_fallocate_finish;
2916 __io_fallocate(req);
2920 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2922 const char __user *fname;
2925 if (sqe->ioprio || sqe->buf_index)
2927 if (req->flags & REQ_F_FIXED_FILE)
2929 if (req->flags & REQ_F_NEED_CLEANUP)
2932 req->open.dfd = READ_ONCE(sqe->fd);
2933 req->open.how.mode = READ_ONCE(sqe->len);
2934 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2935 req->open.how.flags = READ_ONCE(sqe->open_flags);
2936 if (force_o_largefile())
2937 req->open.how.flags |= O_LARGEFILE;
2939 req->open.filename = getname(fname);
2940 if (IS_ERR(req->open.filename)) {
2941 ret = PTR_ERR(req->open.filename);
2942 req->open.filename = NULL;
2946 req->open.nofile = rlimit(RLIMIT_NOFILE);
2947 req->flags |= REQ_F_NEED_CLEANUP;
2951 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2953 struct open_how __user *how;
2954 const char __user *fname;
2958 if (sqe->ioprio || sqe->buf_index)
2960 if (req->flags & REQ_F_FIXED_FILE)
2962 if (req->flags & REQ_F_NEED_CLEANUP)
2965 req->open.dfd = READ_ONCE(sqe->fd);
2966 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2967 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2968 len = READ_ONCE(sqe->len);
2970 if (len < OPEN_HOW_SIZE_VER0)
2973 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2978 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2979 req->open.how.flags |= O_LARGEFILE;
2981 req->open.filename = getname(fname);
2982 if (IS_ERR(req->open.filename)) {
2983 ret = PTR_ERR(req->open.filename);
2984 req->open.filename = NULL;
2988 req->open.nofile = rlimit(RLIMIT_NOFILE);
2989 req->flags |= REQ_F_NEED_CLEANUP;
2993 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
2995 struct open_flags op;
3002 ret = build_open_flags(&req->open.how, &op);
3006 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3010 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3013 ret = PTR_ERR(file);
3015 fsnotify_open(file);
3016 fd_install(ret, file);
3019 putname(req->open.filename);
3020 req->flags &= ~REQ_F_NEED_CLEANUP;
3022 req_set_fail_links(req);
3023 io_cqring_add_event(req, ret);
3028 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3030 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3031 return io_openat2(req, force_nonblock);
3034 static int io_remove_buffers_prep(struct io_kiocb *req,
3035 const struct io_uring_sqe *sqe)
3037 struct io_provide_buf *p = &req->pbuf;
3040 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3043 tmp = READ_ONCE(sqe->fd);
3044 if (!tmp || tmp > USHRT_MAX)
3047 memset(p, 0, sizeof(*p));
3049 p->bgid = READ_ONCE(sqe->buf_group);
3053 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3054 int bgid, unsigned nbufs)
3058 /* shouldn't happen */
3062 /* the head kbuf is the list itself */
3063 while (!list_empty(&buf->list)) {
3064 struct io_buffer *nxt;
3066 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3067 list_del(&nxt->list);
3074 idr_remove(&ctx->io_buffer_idr, bgid);
3079 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3081 struct io_provide_buf *p = &req->pbuf;
3082 struct io_ring_ctx *ctx = req->ctx;
3083 struct io_buffer *head;
3086 io_ring_submit_lock(ctx, !force_nonblock);
3088 lockdep_assert_held(&ctx->uring_lock);
3091 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3093 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3095 io_ring_submit_lock(ctx, !force_nonblock);
3097 req_set_fail_links(req);
3098 io_cqring_add_event(req, ret);
3103 static int io_provide_buffers_prep(struct io_kiocb *req,
3104 const struct io_uring_sqe *sqe)
3106 struct io_provide_buf *p = &req->pbuf;
3109 if (sqe->ioprio || sqe->rw_flags)
3112 tmp = READ_ONCE(sqe->fd);
3113 if (!tmp || tmp > USHRT_MAX)
3116 p->addr = READ_ONCE(sqe->addr);
3117 p->len = READ_ONCE(sqe->len);
3119 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3122 p->bgid = READ_ONCE(sqe->buf_group);
3123 tmp = READ_ONCE(sqe->off);
3124 if (tmp > USHRT_MAX)
3130 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3132 struct io_buffer *buf;
3133 u64 addr = pbuf->addr;
3134 int i, bid = pbuf->bid;
3136 for (i = 0; i < pbuf->nbufs; i++) {
3137 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3142 buf->len = pbuf->len;
3147 INIT_LIST_HEAD(&buf->list);
3150 list_add_tail(&buf->list, &(*head)->list);
3154 return i ? i : -ENOMEM;
3157 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3159 struct io_provide_buf *p = &req->pbuf;
3160 struct io_ring_ctx *ctx = req->ctx;
3161 struct io_buffer *head, *list;
3164 io_ring_submit_lock(ctx, !force_nonblock);
3166 lockdep_assert_held(&ctx->uring_lock);
3168 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3170 ret = io_add_buffers(p, &head);
3175 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3178 __io_remove_buffers(ctx, head, p->bgid, -1U);
3183 io_ring_submit_unlock(ctx, !force_nonblock);
3185 req_set_fail_links(req);
3186 io_cqring_add_event(req, ret);
3191 static int io_epoll_ctl_prep(struct io_kiocb *req,
3192 const struct io_uring_sqe *sqe)
3194 #if defined(CONFIG_EPOLL)
3195 if (sqe->ioprio || sqe->buf_index)
3198 req->epoll.epfd = READ_ONCE(sqe->fd);
3199 req->epoll.op = READ_ONCE(sqe->len);
3200 req->epoll.fd = READ_ONCE(sqe->off);
3202 if (ep_op_has_event(req->epoll.op)) {
3203 struct epoll_event __user *ev;
3205 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3206 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3216 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3218 #if defined(CONFIG_EPOLL)
3219 struct io_epoll *ie = &req->epoll;
3222 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3223 if (force_nonblock && ret == -EAGAIN)
3227 req_set_fail_links(req);
3228 io_cqring_add_event(req, ret);
3236 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3238 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3239 if (sqe->ioprio || sqe->buf_index || sqe->off)
3242 req->madvise.addr = READ_ONCE(sqe->addr);
3243 req->madvise.len = READ_ONCE(sqe->len);
3244 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3251 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3253 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3254 struct io_madvise *ma = &req->madvise;
3260 ret = do_madvise(ma->addr, ma->len, ma->advice);
3262 req_set_fail_links(req);
3263 io_cqring_add_event(req, ret);
3271 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3273 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3276 req->fadvise.offset = READ_ONCE(sqe->off);
3277 req->fadvise.len = READ_ONCE(sqe->len);
3278 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3282 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3284 struct io_fadvise *fa = &req->fadvise;
3287 if (force_nonblock) {
3288 switch (fa->advice) {
3289 case POSIX_FADV_NORMAL:
3290 case POSIX_FADV_RANDOM:
3291 case POSIX_FADV_SEQUENTIAL:
3298 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3300 req_set_fail_links(req);
3301 io_cqring_add_event(req, ret);
3306 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3308 const char __user *fname;
3309 unsigned lookup_flags;
3312 if (sqe->ioprio || sqe->buf_index)
3314 if (req->flags & REQ_F_FIXED_FILE)
3316 if (req->flags & REQ_F_NEED_CLEANUP)
3319 req->open.dfd = READ_ONCE(sqe->fd);
3320 req->open.mask = READ_ONCE(sqe->len);
3321 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3322 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3323 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3325 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3328 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3329 if (IS_ERR(req->open.filename)) {
3330 ret = PTR_ERR(req->open.filename);
3331 req->open.filename = NULL;
3335 req->flags |= REQ_F_NEED_CLEANUP;
3339 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3341 struct io_open *ctx = &req->open;
3342 unsigned lookup_flags;
3347 if (force_nonblock) {
3348 /* only need file table for an actual valid fd */
3349 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3350 req->flags |= REQ_F_NO_FILE_TABLE;
3354 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3358 /* filename_lookup() drops it, keep a reference */
3359 ctx->filename->refcnt++;
3361 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3366 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3368 if (retry_estale(ret, lookup_flags)) {
3369 lookup_flags |= LOOKUP_REVAL;
3373 ret = cp_statx(&stat, ctx->buffer);
3375 putname(ctx->filename);
3376 req->flags &= ~REQ_F_NEED_CLEANUP;
3378 req_set_fail_links(req);
3379 io_cqring_add_event(req, ret);
3384 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3387 * If we queue this for async, it must not be cancellable. That would
3388 * leave the 'file' in an undeterminate state.
3390 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3392 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3393 sqe->rw_flags || sqe->buf_index)
3395 if (req->flags & REQ_F_FIXED_FILE)
3398 req->close.fd = READ_ONCE(sqe->fd);
3399 if (req->file->f_op == &io_uring_fops ||
3400 req->close.fd == req->ctx->ring_fd)
3406 /* only called when __close_fd_get_file() is done */
3407 static void __io_close_finish(struct io_kiocb *req)
3411 ret = filp_close(req->close.put_file, req->work.files);
3413 req_set_fail_links(req);
3414 io_cqring_add_event(req, ret);
3415 fput(req->close.put_file);
3419 static void io_close_finish(struct io_wq_work **workptr)
3421 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3423 /* not cancellable, don't do io_req_cancelled() */
3424 __io_close_finish(req);
3425 io_steal_work(req, workptr);
3428 static int io_close(struct io_kiocb *req, bool force_nonblock)
3432 req->close.put_file = NULL;
3433 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3437 /* if the file has a flush method, be safe and punt to async */
3438 if (req->close.put_file->f_op->flush && force_nonblock) {
3439 /* submission ref will be dropped, take it for async */
3440 refcount_inc(&req->refs);
3442 req->work.func = io_close_finish;
3444 * Do manual async queue here to avoid grabbing files - we don't
3445 * need the files, and it'll cause io_close_finish() to close
3446 * the file again and cause a double CQE entry for this request
3448 io_queue_async_work(req);
3453 * No ->flush(), safely close from here and just punt the
3454 * fput() to async context.
3456 __io_close_finish(req);
3460 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3462 struct io_ring_ctx *ctx = req->ctx;
3467 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3469 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3472 req->sync.off = READ_ONCE(sqe->off);
3473 req->sync.len = READ_ONCE(sqe->len);
3474 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3478 static void __io_sync_file_range(struct io_kiocb *req)
3482 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3485 req_set_fail_links(req);
3486 io_cqring_add_event(req, ret);
3491 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3493 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3495 if (io_req_cancelled(req))
3497 __io_sync_file_range(req);
3498 io_steal_work(req, workptr);
3501 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3503 /* sync_file_range always requires a blocking context */
3504 if (force_nonblock) {
3505 req->work.func = io_sync_file_range_finish;
3509 __io_sync_file_range(req);
3513 #if defined(CONFIG_NET)
3514 static int io_setup_async_msg(struct io_kiocb *req,
3515 struct io_async_msghdr *kmsg)
3519 if (io_alloc_async_ctx(req)) {
3520 if (kmsg->iov != kmsg->fast_iov)
3524 req->flags |= REQ_F_NEED_CLEANUP;
3525 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3529 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3531 struct io_sr_msg *sr = &req->sr_msg;
3532 struct io_async_ctx *io = req->io;
3535 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3536 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3537 sr->len = READ_ONCE(sqe->len);
3539 #ifdef CONFIG_COMPAT
3540 if (req->ctx->compat)
3541 sr->msg_flags |= MSG_CMSG_COMPAT;
3544 if (!io || req->opcode == IORING_OP_SEND)
3546 /* iovec is already imported */
3547 if (req->flags & REQ_F_NEED_CLEANUP)
3550 io->msg.iov = io->msg.fast_iov;
3551 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3554 req->flags |= REQ_F_NEED_CLEANUP;
3558 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3560 struct io_async_msghdr *kmsg = NULL;
3561 struct socket *sock;
3564 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3567 sock = sock_from_file(req->file, &ret);
3569 struct io_async_ctx io;
3573 kmsg = &req->io->msg;
3574 kmsg->msg.msg_name = &req->io->msg.addr;
3575 /* if iov is set, it's allocated already */
3577 kmsg->iov = kmsg->fast_iov;
3578 kmsg->msg.msg_iter.iov = kmsg->iov;
3580 struct io_sr_msg *sr = &req->sr_msg;
3583 kmsg->msg.msg_name = &io.msg.addr;
3585 io.msg.iov = io.msg.fast_iov;
3586 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3587 sr->msg_flags, &io.msg.iov);
3592 flags = req->sr_msg.msg_flags;
3593 if (flags & MSG_DONTWAIT)
3594 req->flags |= REQ_F_NOWAIT;
3595 else if (force_nonblock)
3596 flags |= MSG_DONTWAIT;
3598 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3599 if (force_nonblock && ret == -EAGAIN)
3600 return io_setup_async_msg(req, kmsg);
3601 if (ret == -ERESTARTSYS)
3605 if (kmsg && kmsg->iov != kmsg->fast_iov)
3607 req->flags &= ~REQ_F_NEED_CLEANUP;
3608 io_cqring_add_event(req, ret);
3610 req_set_fail_links(req);
3615 static int io_send(struct io_kiocb *req, bool force_nonblock)
3617 struct socket *sock;
3620 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3623 sock = sock_from_file(req->file, &ret);
3625 struct io_sr_msg *sr = &req->sr_msg;
3630 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3635 msg.msg_name = NULL;
3636 msg.msg_control = NULL;
3637 msg.msg_controllen = 0;
3638 msg.msg_namelen = 0;
3640 flags = req->sr_msg.msg_flags;
3641 if (flags & MSG_DONTWAIT)
3642 req->flags |= REQ_F_NOWAIT;
3643 else if (force_nonblock)
3644 flags |= MSG_DONTWAIT;
3646 msg.msg_flags = flags;
3647 ret = sock_sendmsg(sock, &msg);
3648 if (force_nonblock && ret == -EAGAIN)
3650 if (ret == -ERESTARTSYS)
3654 io_cqring_add_event(req, ret);
3656 req_set_fail_links(req);
3661 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3663 struct io_sr_msg *sr = &req->sr_msg;
3664 struct iovec __user *uiov;
3668 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3673 if (req->flags & REQ_F_BUFFER_SELECT) {
3676 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3678 sr->len = io->msg.iov[0].iov_len;
3679 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3683 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3684 &io->msg.iov, &io->msg.msg.msg_iter);
3692 #ifdef CONFIG_COMPAT
3693 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3694 struct io_async_ctx *io)
3696 struct compat_msghdr __user *msg_compat;
3697 struct io_sr_msg *sr = &req->sr_msg;
3698 struct compat_iovec __user *uiov;
3703 msg_compat = (struct compat_msghdr __user *) sr->msg;
3704 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3709 uiov = compat_ptr(ptr);
3710 if (req->flags & REQ_F_BUFFER_SELECT) {
3711 compat_ssize_t clen;
3715 if (!access_ok(uiov, sizeof(*uiov)))
3717 if (__get_user(clen, &uiov->iov_len))
3721 sr->len = io->msg.iov[0].iov_len;
3724 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3726 &io->msg.msg.msg_iter);
3735 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3737 io->msg.iov = io->msg.fast_iov;
3739 #ifdef CONFIG_COMPAT
3740 if (req->ctx->compat)
3741 return __io_compat_recvmsg_copy_hdr(req, io);
3744 return __io_recvmsg_copy_hdr(req, io);
3747 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3748 int *cflags, bool needs_lock)
3750 struct io_sr_msg *sr = &req->sr_msg;
3751 struct io_buffer *kbuf;
3753 if (!(req->flags & REQ_F_BUFFER_SELECT))
3756 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3761 req->flags |= REQ_F_BUFFER_SELECTED;
3763 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3764 *cflags |= IORING_CQE_F_BUFFER;
3768 static int io_recvmsg_prep(struct io_kiocb *req,
3769 const struct io_uring_sqe *sqe)
3771 struct io_sr_msg *sr = &req->sr_msg;
3772 struct io_async_ctx *io = req->io;
3775 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3776 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3777 sr->len = READ_ONCE(sqe->len);
3778 sr->bgid = READ_ONCE(sqe->buf_group);
3780 #ifdef CONFIG_COMPAT
3781 if (req->ctx->compat)
3782 sr->msg_flags |= MSG_CMSG_COMPAT;
3785 if (!io || req->opcode == IORING_OP_RECV)
3787 /* iovec is already imported */
3788 if (req->flags & REQ_F_NEED_CLEANUP)
3791 ret = io_recvmsg_copy_hdr(req, io);
3793 req->flags |= REQ_F_NEED_CLEANUP;
3797 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3799 struct io_async_msghdr *kmsg = NULL;
3800 struct socket *sock;
3801 int ret, cflags = 0;
3803 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3806 sock = sock_from_file(req->file, &ret);
3808 struct io_buffer *kbuf;
3809 struct io_async_ctx io;
3813 kmsg = &req->io->msg;
3814 kmsg->msg.msg_name = &req->io->msg.addr;
3815 /* if iov is set, it's allocated already */
3817 kmsg->iov = kmsg->fast_iov;
3818 kmsg->msg.msg_iter.iov = kmsg->iov;
3821 kmsg->msg.msg_name = &io.msg.addr;
3823 ret = io_recvmsg_copy_hdr(req, &io);
3828 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3830 return PTR_ERR(kbuf);
3832 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3833 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3834 1, req->sr_msg.len);
3837 flags = req->sr_msg.msg_flags;
3838 if (flags & MSG_DONTWAIT)
3839 req->flags |= REQ_F_NOWAIT;
3840 else if (force_nonblock)
3841 flags |= MSG_DONTWAIT;
3843 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3844 kmsg->uaddr, flags);
3845 if (force_nonblock && ret == -EAGAIN)
3846 return io_setup_async_msg(req, kmsg);
3847 if (ret == -ERESTARTSYS)
3851 if (kmsg && kmsg->iov != kmsg->fast_iov)
3853 req->flags &= ~REQ_F_NEED_CLEANUP;
3854 __io_cqring_add_event(req, ret, cflags);
3856 req_set_fail_links(req);
3861 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3863 struct io_buffer *kbuf = NULL;
3864 struct socket *sock;
3865 int ret, cflags = 0;
3867 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3870 sock = sock_from_file(req->file, &ret);
3872 struct io_sr_msg *sr = &req->sr_msg;
3873 void __user *buf = sr->buf;
3878 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3880 return PTR_ERR(kbuf);
3882 buf = u64_to_user_ptr(kbuf->addr);
3884 ret = import_single_range(READ, buf, sr->len, &iov,
3891 req->flags |= REQ_F_NEED_CLEANUP;
3892 msg.msg_name = NULL;
3893 msg.msg_control = NULL;
3894 msg.msg_controllen = 0;
3895 msg.msg_namelen = 0;
3896 msg.msg_iocb = NULL;
3899 flags = req->sr_msg.msg_flags;
3900 if (flags & MSG_DONTWAIT)
3901 req->flags |= REQ_F_NOWAIT;
3902 else if (force_nonblock)
3903 flags |= MSG_DONTWAIT;
3905 ret = sock_recvmsg(sock, &msg, flags);
3906 if (force_nonblock && ret == -EAGAIN)
3908 if (ret == -ERESTARTSYS)
3913 req->flags &= ~REQ_F_NEED_CLEANUP;
3914 __io_cqring_add_event(req, ret, cflags);
3916 req_set_fail_links(req);
3921 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3923 struct io_accept *accept = &req->accept;
3925 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3927 if (sqe->ioprio || sqe->len || sqe->buf_index)
3930 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3931 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3932 accept->flags = READ_ONCE(sqe->accept_flags);
3933 accept->nofile = rlimit(RLIMIT_NOFILE);
3937 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3939 struct io_accept *accept = &req->accept;
3940 unsigned file_flags;
3943 file_flags = force_nonblock ? O_NONBLOCK : 0;
3944 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3945 accept->addr_len, accept->flags,
3947 if (ret == -EAGAIN && force_nonblock)
3949 if (ret == -ERESTARTSYS)
3952 req_set_fail_links(req);
3953 io_cqring_add_event(req, ret);
3958 static void io_accept_finish(struct io_wq_work **workptr)
3960 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3962 if (io_req_cancelled(req))
3964 __io_accept(req, false);
3965 io_steal_work(req, workptr);
3968 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3972 ret = __io_accept(req, force_nonblock);
3973 if (ret == -EAGAIN && force_nonblock) {
3974 req->work.func = io_accept_finish;
3980 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3982 struct io_connect *conn = &req->connect;
3983 struct io_async_ctx *io = req->io;
3985 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3987 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3990 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3991 conn->addr_len = READ_ONCE(sqe->addr2);
3996 return move_addr_to_kernel(conn->addr, conn->addr_len,
3997 &io->connect.address);
4000 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4002 struct io_async_ctx __io, *io;
4003 unsigned file_flags;
4009 ret = move_addr_to_kernel(req->connect.addr,
4010 req->connect.addr_len,
4011 &__io.connect.address);
4017 file_flags = force_nonblock ? O_NONBLOCK : 0;
4019 ret = __sys_connect_file(req->file, &io->connect.address,
4020 req->connect.addr_len, file_flags);
4021 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4024 if (io_alloc_async_ctx(req)) {
4028 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4031 if (ret == -ERESTARTSYS)
4035 req_set_fail_links(req);
4036 io_cqring_add_event(req, ret);
4040 #else /* !CONFIG_NET */
4041 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4046 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4051 static int io_send(struct io_kiocb *req, bool force_nonblock)
4056 static int io_recvmsg_prep(struct io_kiocb *req,
4057 const struct io_uring_sqe *sqe)
4062 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4067 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4072 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4077 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4082 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4087 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4091 #endif /* CONFIG_NET */
4093 struct io_poll_table {
4094 struct poll_table_struct pt;
4095 struct io_kiocb *req;
4099 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4100 struct wait_queue_head *head)
4102 if (unlikely(poll->head)) {
4103 pt->error = -EINVAL;
4109 add_wait_queue(head, &poll->wait);
4112 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4113 struct poll_table_struct *p)
4115 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4117 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4120 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4121 __poll_t mask, task_work_func_t func)
4123 struct task_struct *tsk;
4126 /* for instances that support it check for an event match first: */
4127 if (mask && !(mask & poll->events))
4130 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4132 list_del_init(&poll->wait.entry);
4136 init_task_work(&req->task_work, func);
4138 * If this fails, then the task is exiting. When a task exits, the
4139 * work gets canceled, so just cancel this request as well instead
4140 * of executing it. We can't safely execute it anyway, as we may not
4141 * have the needed state needed for it anyway.
4143 ret = task_work_add(tsk, &req->task_work, true);
4144 if (unlikely(ret)) {
4145 WRITE_ONCE(poll->canceled, true);
4146 tsk = io_wq_get_task(req->ctx->io_wq);
4147 task_work_add(tsk, &req->task_work, true);
4149 wake_up_process(tsk);
4153 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4154 __acquires(&req->ctx->completion_lock)
4156 struct io_ring_ctx *ctx = req->ctx;
4158 if (!req->result && !READ_ONCE(poll->canceled)) {
4159 struct poll_table_struct pt = { ._key = poll->events };
4161 req->result = vfs_poll(req->file, &pt) & poll->events;
4164 spin_lock_irq(&ctx->completion_lock);
4165 if (!req->result && !READ_ONCE(poll->canceled)) {
4166 add_wait_queue(poll->head, &poll->wait);
4173 static void io_async_task_func(struct callback_head *cb)
4175 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4176 struct async_poll *apoll = req->apoll;
4177 struct io_ring_ctx *ctx = req->ctx;
4180 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4182 if (io_poll_rewait(req, &apoll->poll)) {
4183 spin_unlock_irq(&ctx->completion_lock);
4187 if (hash_hashed(&req->hash_node))
4188 hash_del(&req->hash_node);
4190 canceled = READ_ONCE(apoll->poll.canceled);
4192 io_cqring_fill_event(req, -ECANCELED);
4193 io_commit_cqring(ctx);
4196 spin_unlock_irq(&ctx->completion_lock);
4198 /* restore ->work in case we need to retry again */
4199 memcpy(&req->work, &apoll->work, sizeof(req->work));
4203 io_cqring_ev_posted(ctx);
4204 req_set_fail_links(req);
4205 io_double_put_req(req);
4209 __set_current_state(TASK_RUNNING);
4210 mutex_lock(&ctx->uring_lock);
4211 __io_queue_sqe(req, NULL);
4212 mutex_unlock(&ctx->uring_lock);
4217 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4220 struct io_kiocb *req = wait->private;
4221 struct io_poll_iocb *poll = &req->apoll->poll;
4223 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4226 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4229 static void io_poll_req_insert(struct io_kiocb *req)
4231 struct io_ring_ctx *ctx = req->ctx;
4232 struct hlist_head *list;
4234 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4235 hlist_add_head(&req->hash_node, list);
4238 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4239 struct io_poll_iocb *poll,
4240 struct io_poll_table *ipt, __poll_t mask,
4241 wait_queue_func_t wake_func)
4242 __acquires(&ctx->completion_lock)
4244 struct io_ring_ctx *ctx = req->ctx;
4245 bool cancel = false;
4247 poll->file = req->file;
4249 poll->done = poll->canceled = false;
4250 poll->events = mask;
4252 ipt->pt._key = mask;
4254 ipt->error = -EINVAL;
4256 INIT_LIST_HEAD(&poll->wait.entry);
4257 init_waitqueue_func_entry(&poll->wait, wake_func);
4258 poll->wait.private = req;
4260 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4262 spin_lock_irq(&ctx->completion_lock);
4263 if (likely(poll->head)) {
4264 spin_lock(&poll->head->lock);
4265 if (unlikely(list_empty(&poll->wait.entry))) {
4271 if (mask || ipt->error)
4272 list_del_init(&poll->wait.entry);
4274 WRITE_ONCE(poll->canceled, true);
4275 else if (!poll->done) /* actually waiting for an event */
4276 io_poll_req_insert(req);
4277 spin_unlock(&poll->head->lock);
4283 static bool io_arm_poll_handler(struct io_kiocb *req)
4285 const struct io_op_def *def = &io_op_defs[req->opcode];
4286 struct io_ring_ctx *ctx = req->ctx;
4287 struct async_poll *apoll;
4288 struct io_poll_table ipt;
4291 if (!req->file || !file_can_poll(req->file))
4293 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4295 if (!def->pollin && !def->pollout)
4298 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4299 if (unlikely(!apoll))
4302 req->flags |= REQ_F_POLLED;
4303 memcpy(&apoll->work, &req->work, sizeof(req->work));
4305 get_task_struct(current);
4306 req->task = current;
4308 INIT_HLIST_NODE(&req->hash_node);
4312 mask |= POLLIN | POLLRDNORM;
4314 mask |= POLLOUT | POLLWRNORM;
4315 mask |= POLLERR | POLLPRI;
4317 ipt.pt._qproc = io_async_queue_proc;
4319 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4323 apoll->poll.done = true;
4324 spin_unlock_irq(&ctx->completion_lock);
4325 memcpy(&req->work, &apoll->work, sizeof(req->work));
4329 spin_unlock_irq(&ctx->completion_lock);
4330 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4331 apoll->poll.events);
4335 static bool __io_poll_remove_one(struct io_kiocb *req,
4336 struct io_poll_iocb *poll)
4338 bool do_complete = false;
4340 spin_lock(&poll->head->lock);
4341 WRITE_ONCE(poll->canceled, true);
4342 if (!list_empty(&poll->wait.entry)) {
4343 list_del_init(&poll->wait.entry);
4346 spin_unlock(&poll->head->lock);
4350 static bool io_poll_remove_one(struct io_kiocb *req)
4352 struct async_poll *apoll = NULL;
4355 if (req->opcode == IORING_OP_POLL_ADD) {
4356 do_complete = __io_poll_remove_one(req, &req->poll);
4359 /* non-poll requests have submit ref still */
4360 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4365 hash_del(&req->hash_node);
4367 if (do_complete && apoll) {
4369 * restore ->work because we need to call io_req_work_drop_env.
4371 memcpy(&req->work, &apoll->work, sizeof(req->work));
4376 io_cqring_fill_event(req, -ECANCELED);
4377 io_commit_cqring(req->ctx);
4378 req->flags |= REQ_F_COMP_LOCKED;
4385 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4387 struct hlist_node *tmp;
4388 struct io_kiocb *req;
4391 spin_lock_irq(&ctx->completion_lock);
4392 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4393 struct hlist_head *list;
4395 list = &ctx->cancel_hash[i];
4396 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4397 posted += io_poll_remove_one(req);
4399 spin_unlock_irq(&ctx->completion_lock);
4402 io_cqring_ev_posted(ctx);
4405 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4407 struct hlist_head *list;
4408 struct io_kiocb *req;
4410 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4411 hlist_for_each_entry(req, list, hash_node) {
4412 if (sqe_addr != req->user_data)
4414 if (io_poll_remove_one(req))
4422 static int io_poll_remove_prep(struct io_kiocb *req,
4423 const struct io_uring_sqe *sqe)
4425 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4427 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4431 req->poll.addr = READ_ONCE(sqe->addr);
4436 * Find a running poll command that matches one specified in sqe->addr,
4437 * and remove it if found.
4439 static int io_poll_remove(struct io_kiocb *req)
4441 struct io_ring_ctx *ctx = req->ctx;
4445 addr = req->poll.addr;
4446 spin_lock_irq(&ctx->completion_lock);
4447 ret = io_poll_cancel(ctx, addr);
4448 spin_unlock_irq(&ctx->completion_lock);
4450 io_cqring_add_event(req, ret);
4452 req_set_fail_links(req);
4457 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4459 struct io_ring_ctx *ctx = req->ctx;
4461 req->poll.done = true;
4462 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4463 io_commit_cqring(ctx);
4466 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4468 struct io_ring_ctx *ctx = req->ctx;
4469 struct io_poll_iocb *poll = &req->poll;
4471 if (io_poll_rewait(req, poll)) {
4472 spin_unlock_irq(&ctx->completion_lock);
4476 hash_del(&req->hash_node);
4477 io_poll_complete(req, req->result, 0);
4478 req->flags |= REQ_F_COMP_LOCKED;
4479 io_put_req_find_next(req, nxt);
4480 spin_unlock_irq(&ctx->completion_lock);
4482 io_cqring_ev_posted(ctx);
4485 static void io_poll_task_func(struct callback_head *cb)
4487 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4488 struct io_kiocb *nxt = NULL;
4490 io_poll_task_handler(req, &nxt);
4492 struct io_ring_ctx *ctx = nxt->ctx;
4494 mutex_lock(&ctx->uring_lock);
4495 __io_queue_sqe(nxt, NULL);
4496 mutex_unlock(&ctx->uring_lock);
4500 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4503 struct io_kiocb *req = wait->private;
4504 struct io_poll_iocb *poll = &req->poll;
4506 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4509 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4510 struct poll_table_struct *p)
4512 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4514 __io_queue_proc(&pt->req->poll, pt, head);
4517 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4519 struct io_poll_iocb *poll = &req->poll;
4522 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4524 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4529 events = READ_ONCE(sqe->poll_events);
4530 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4532 get_task_struct(current);
4533 req->task = current;
4537 static int io_poll_add(struct io_kiocb *req)
4539 struct io_poll_iocb *poll = &req->poll;
4540 struct io_ring_ctx *ctx = req->ctx;
4541 struct io_poll_table ipt;
4544 INIT_HLIST_NODE(&req->hash_node);
4545 INIT_LIST_HEAD(&req->list);
4546 ipt.pt._qproc = io_poll_queue_proc;
4548 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4551 if (mask) { /* no async, we'd stolen it */
4553 io_poll_complete(req, mask, 0);
4555 spin_unlock_irq(&ctx->completion_lock);
4558 io_cqring_ev_posted(ctx);
4564 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4566 struct io_timeout_data *data = container_of(timer,
4567 struct io_timeout_data, timer);
4568 struct io_kiocb *req = data->req;
4569 struct io_ring_ctx *ctx = req->ctx;
4570 unsigned long flags;
4572 atomic_inc(&ctx->cq_timeouts);
4574 spin_lock_irqsave(&ctx->completion_lock, flags);
4576 * We could be racing with timeout deletion. If the list is empty,
4577 * then timeout lookup already found it and will be handling it.
4579 if (!list_empty(&req->list)) {
4580 struct io_kiocb *prev;
4583 * Adjust the reqs sequence before the current one because it
4584 * will consume a slot in the cq_ring and the cq_tail
4585 * pointer will be increased, otherwise other timeout reqs may
4586 * return in advance without waiting for enough wait_nr.
4589 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4591 list_del_init(&req->list);
4594 io_cqring_fill_event(req, -ETIME);
4595 io_commit_cqring(ctx);
4596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4598 io_cqring_ev_posted(ctx);
4599 req_set_fail_links(req);
4601 return HRTIMER_NORESTART;
4604 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4606 struct io_kiocb *req;
4609 list_for_each_entry(req, &ctx->timeout_list, list) {
4610 if (user_data == req->user_data) {
4611 list_del_init(&req->list);
4620 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4624 req_set_fail_links(req);
4625 io_cqring_fill_event(req, -ECANCELED);
4630 static int io_timeout_remove_prep(struct io_kiocb *req,
4631 const struct io_uring_sqe *sqe)
4633 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4635 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4638 req->timeout.addr = READ_ONCE(sqe->addr);
4639 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4640 if (req->timeout.flags)
4647 * Remove or update an existing timeout command
4649 static int io_timeout_remove(struct io_kiocb *req)
4651 struct io_ring_ctx *ctx = req->ctx;
4654 spin_lock_irq(&ctx->completion_lock);
4655 ret = io_timeout_cancel(ctx, req->timeout.addr);
4657 io_cqring_fill_event(req, ret);
4658 io_commit_cqring(ctx);
4659 spin_unlock_irq(&ctx->completion_lock);
4660 io_cqring_ev_posted(ctx);
4662 req_set_fail_links(req);
4667 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4668 bool is_timeout_link)
4670 struct io_timeout_data *data;
4673 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4675 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4677 if (sqe->off && is_timeout_link)
4679 flags = READ_ONCE(sqe->timeout_flags);
4680 if (flags & ~IORING_TIMEOUT_ABS)
4683 req->timeout.count = READ_ONCE(sqe->off);
4685 if (!req->io && io_alloc_async_ctx(req))
4688 data = &req->io->timeout;
4690 req->flags |= REQ_F_TIMEOUT;
4692 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4695 if (flags & IORING_TIMEOUT_ABS)
4696 data->mode = HRTIMER_MODE_ABS;
4698 data->mode = HRTIMER_MODE_REL;
4700 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4704 static int io_timeout(struct io_kiocb *req)
4706 struct io_ring_ctx *ctx = req->ctx;
4707 struct io_timeout_data *data;
4708 struct list_head *entry;
4710 u32 count = req->timeout.count;
4711 u32 seq = req->sequence;
4713 data = &req->io->timeout;
4716 * sqe->off holds how many events that need to occur for this
4717 * timeout event to be satisfied. If it isn't set, then this is
4718 * a pure timeout request, sequence isn't used.
4721 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4722 spin_lock_irq(&ctx->completion_lock);
4723 entry = ctx->timeout_list.prev;
4727 req->sequence = seq + count;
4730 * Insertion sort, ensuring the first entry in the list is always
4731 * the one we need first.
4733 spin_lock_irq(&ctx->completion_lock);
4734 list_for_each_prev(entry, &ctx->timeout_list) {
4735 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4737 long long tmp, tmp_nxt;
4738 u32 nxt_offset = nxt->timeout.count;
4740 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4744 * Since seq + count can overflow, use type long
4747 tmp = (long long)seq + count;
4748 nxt_seq = nxt->sequence - nxt_offset;
4749 tmp_nxt = (long long)nxt_seq + nxt_offset;
4752 * cached_sq_head may overflow, and it will never overflow twice
4753 * once there is some timeout req still be valid.
4762 * Sequence of reqs after the insert one and itself should
4763 * be adjusted because each timeout req consumes a slot.
4768 req->sequence -= span;
4770 list_add(&req->list, entry);
4771 data->timer.function = io_timeout_fn;
4772 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4773 spin_unlock_irq(&ctx->completion_lock);
4777 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4779 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4781 return req->user_data == (unsigned long) data;
4784 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4786 enum io_wq_cancel cancel_ret;
4789 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4790 switch (cancel_ret) {
4791 case IO_WQ_CANCEL_OK:
4794 case IO_WQ_CANCEL_RUNNING:
4797 case IO_WQ_CANCEL_NOTFOUND:
4805 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4806 struct io_kiocb *req, __u64 sqe_addr,
4809 unsigned long flags;
4812 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4813 if (ret != -ENOENT) {
4814 spin_lock_irqsave(&ctx->completion_lock, flags);
4818 spin_lock_irqsave(&ctx->completion_lock, flags);
4819 ret = io_timeout_cancel(ctx, sqe_addr);
4822 ret = io_poll_cancel(ctx, sqe_addr);
4826 io_cqring_fill_event(req, ret);
4827 io_commit_cqring(ctx);
4828 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4829 io_cqring_ev_posted(ctx);
4832 req_set_fail_links(req);
4836 static int io_async_cancel_prep(struct io_kiocb *req,
4837 const struct io_uring_sqe *sqe)
4839 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4841 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4845 req->cancel.addr = READ_ONCE(sqe->addr);
4849 static int io_async_cancel(struct io_kiocb *req)
4851 struct io_ring_ctx *ctx = req->ctx;
4853 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4857 static int io_files_update_prep(struct io_kiocb *req,
4858 const struct io_uring_sqe *sqe)
4860 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4863 req->files_update.offset = READ_ONCE(sqe->off);
4864 req->files_update.nr_args = READ_ONCE(sqe->len);
4865 if (!req->files_update.nr_args)
4867 req->files_update.arg = READ_ONCE(sqe->addr);
4871 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4873 struct io_ring_ctx *ctx = req->ctx;
4874 struct io_uring_files_update up;
4880 up.offset = req->files_update.offset;
4881 up.fds = req->files_update.arg;
4883 mutex_lock(&ctx->uring_lock);
4884 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4885 mutex_unlock(&ctx->uring_lock);
4888 req_set_fail_links(req);
4889 io_cqring_add_event(req, ret);
4894 static int io_req_defer_prep(struct io_kiocb *req,
4895 const struct io_uring_sqe *sqe)
4902 if (io_op_defs[req->opcode].file_table) {
4903 ret = io_grab_files(req);
4908 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4910 switch (req->opcode) {
4913 case IORING_OP_READV:
4914 case IORING_OP_READ_FIXED:
4915 case IORING_OP_READ:
4916 ret = io_read_prep(req, sqe, true);
4918 case IORING_OP_WRITEV:
4919 case IORING_OP_WRITE_FIXED:
4920 case IORING_OP_WRITE:
4921 ret = io_write_prep(req, sqe, true);
4923 case IORING_OP_POLL_ADD:
4924 ret = io_poll_add_prep(req, sqe);
4926 case IORING_OP_POLL_REMOVE:
4927 ret = io_poll_remove_prep(req, sqe);
4929 case IORING_OP_FSYNC:
4930 ret = io_prep_fsync(req, sqe);
4932 case IORING_OP_SYNC_FILE_RANGE:
4933 ret = io_prep_sfr(req, sqe);
4935 case IORING_OP_SENDMSG:
4936 case IORING_OP_SEND:
4937 ret = io_sendmsg_prep(req, sqe);
4939 case IORING_OP_RECVMSG:
4940 case IORING_OP_RECV:
4941 ret = io_recvmsg_prep(req, sqe);
4943 case IORING_OP_CONNECT:
4944 ret = io_connect_prep(req, sqe);
4946 case IORING_OP_TIMEOUT:
4947 ret = io_timeout_prep(req, sqe, false);
4949 case IORING_OP_TIMEOUT_REMOVE:
4950 ret = io_timeout_remove_prep(req, sqe);
4952 case IORING_OP_ASYNC_CANCEL:
4953 ret = io_async_cancel_prep(req, sqe);
4955 case IORING_OP_LINK_TIMEOUT:
4956 ret = io_timeout_prep(req, sqe, true);
4958 case IORING_OP_ACCEPT:
4959 ret = io_accept_prep(req, sqe);
4961 case IORING_OP_FALLOCATE:
4962 ret = io_fallocate_prep(req, sqe);
4964 case IORING_OP_OPENAT:
4965 ret = io_openat_prep(req, sqe);
4967 case IORING_OP_CLOSE:
4968 ret = io_close_prep(req, sqe);
4970 case IORING_OP_FILES_UPDATE:
4971 ret = io_files_update_prep(req, sqe);
4973 case IORING_OP_STATX:
4974 ret = io_statx_prep(req, sqe);
4976 case IORING_OP_FADVISE:
4977 ret = io_fadvise_prep(req, sqe);
4979 case IORING_OP_MADVISE:
4980 ret = io_madvise_prep(req, sqe);
4982 case IORING_OP_OPENAT2:
4983 ret = io_openat2_prep(req, sqe);
4985 case IORING_OP_EPOLL_CTL:
4986 ret = io_epoll_ctl_prep(req, sqe);
4988 case IORING_OP_SPLICE:
4989 ret = io_splice_prep(req, sqe);
4991 case IORING_OP_PROVIDE_BUFFERS:
4992 ret = io_provide_buffers_prep(req, sqe);
4994 case IORING_OP_REMOVE_BUFFERS:
4995 ret = io_remove_buffers_prep(req, sqe);
4998 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5007 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5009 struct io_ring_ctx *ctx = req->ctx;
5012 /* Still need defer if there is pending req in defer list. */
5013 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5017 if (io_alloc_async_ctx(req))
5019 ret = io_req_defer_prep(req, sqe);
5024 spin_lock_irq(&ctx->completion_lock);
5025 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5026 spin_unlock_irq(&ctx->completion_lock);
5030 trace_io_uring_defer(ctx, req, req->user_data);
5031 list_add_tail(&req->list, &ctx->defer_list);
5032 spin_unlock_irq(&ctx->completion_lock);
5033 return -EIOCBQUEUED;
5036 static void io_cleanup_req(struct io_kiocb *req)
5038 struct io_async_ctx *io = req->io;
5040 switch (req->opcode) {
5041 case IORING_OP_READV:
5042 case IORING_OP_READ_FIXED:
5043 case IORING_OP_READ:
5044 if (req->flags & REQ_F_BUFFER_SELECTED)
5045 kfree((void *)(unsigned long)req->rw.addr);
5047 case IORING_OP_WRITEV:
5048 case IORING_OP_WRITE_FIXED:
5049 case IORING_OP_WRITE:
5050 if (io->rw.iov != io->rw.fast_iov)
5053 case IORING_OP_RECVMSG:
5054 if (req->flags & REQ_F_BUFFER_SELECTED)
5055 kfree(req->sr_msg.kbuf);
5057 case IORING_OP_SENDMSG:
5058 if (io->msg.iov != io->msg.fast_iov)
5061 case IORING_OP_RECV:
5062 if (req->flags & REQ_F_BUFFER_SELECTED)
5063 kfree(req->sr_msg.kbuf);
5065 case IORING_OP_OPENAT:
5066 case IORING_OP_OPENAT2:
5067 case IORING_OP_STATX:
5068 putname(req->open.filename);
5070 case IORING_OP_SPLICE:
5071 io_put_file(req, req->splice.file_in,
5072 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5076 req->flags &= ~REQ_F_NEED_CLEANUP;
5079 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5080 bool force_nonblock)
5082 struct io_ring_ctx *ctx = req->ctx;
5085 switch (req->opcode) {
5089 case IORING_OP_READV:
5090 case IORING_OP_READ_FIXED:
5091 case IORING_OP_READ:
5093 ret = io_read_prep(req, sqe, force_nonblock);
5097 ret = io_read(req, force_nonblock);
5099 case IORING_OP_WRITEV:
5100 case IORING_OP_WRITE_FIXED:
5101 case IORING_OP_WRITE:
5103 ret = io_write_prep(req, sqe, force_nonblock);
5107 ret = io_write(req, force_nonblock);
5109 case IORING_OP_FSYNC:
5111 ret = io_prep_fsync(req, sqe);
5115 ret = io_fsync(req, force_nonblock);
5117 case IORING_OP_POLL_ADD:
5119 ret = io_poll_add_prep(req, sqe);
5123 ret = io_poll_add(req);
5125 case IORING_OP_POLL_REMOVE:
5127 ret = io_poll_remove_prep(req, sqe);
5131 ret = io_poll_remove(req);
5133 case IORING_OP_SYNC_FILE_RANGE:
5135 ret = io_prep_sfr(req, sqe);
5139 ret = io_sync_file_range(req, force_nonblock);
5141 case IORING_OP_SENDMSG:
5142 case IORING_OP_SEND:
5144 ret = io_sendmsg_prep(req, sqe);
5148 if (req->opcode == IORING_OP_SENDMSG)
5149 ret = io_sendmsg(req, force_nonblock);
5151 ret = io_send(req, force_nonblock);
5153 case IORING_OP_RECVMSG:
5154 case IORING_OP_RECV:
5156 ret = io_recvmsg_prep(req, sqe);
5160 if (req->opcode == IORING_OP_RECVMSG)
5161 ret = io_recvmsg(req, force_nonblock);
5163 ret = io_recv(req, force_nonblock);
5165 case IORING_OP_TIMEOUT:
5167 ret = io_timeout_prep(req, sqe, false);
5171 ret = io_timeout(req);
5173 case IORING_OP_TIMEOUT_REMOVE:
5175 ret = io_timeout_remove_prep(req, sqe);
5179 ret = io_timeout_remove(req);
5181 case IORING_OP_ACCEPT:
5183 ret = io_accept_prep(req, sqe);
5187 ret = io_accept(req, force_nonblock);
5189 case IORING_OP_CONNECT:
5191 ret = io_connect_prep(req, sqe);
5195 ret = io_connect(req, force_nonblock);
5197 case IORING_OP_ASYNC_CANCEL:
5199 ret = io_async_cancel_prep(req, sqe);
5203 ret = io_async_cancel(req);
5205 case IORING_OP_FALLOCATE:
5207 ret = io_fallocate_prep(req, sqe);
5211 ret = io_fallocate(req, force_nonblock);
5213 case IORING_OP_OPENAT:
5215 ret = io_openat_prep(req, sqe);
5219 ret = io_openat(req, force_nonblock);
5221 case IORING_OP_CLOSE:
5223 ret = io_close_prep(req, sqe);
5227 ret = io_close(req, force_nonblock);
5229 case IORING_OP_FILES_UPDATE:
5231 ret = io_files_update_prep(req, sqe);
5235 ret = io_files_update(req, force_nonblock);
5237 case IORING_OP_STATX:
5239 ret = io_statx_prep(req, sqe);
5243 ret = io_statx(req, force_nonblock);
5245 case IORING_OP_FADVISE:
5247 ret = io_fadvise_prep(req, sqe);
5251 ret = io_fadvise(req, force_nonblock);
5253 case IORING_OP_MADVISE:
5255 ret = io_madvise_prep(req, sqe);
5259 ret = io_madvise(req, force_nonblock);
5261 case IORING_OP_OPENAT2:
5263 ret = io_openat2_prep(req, sqe);
5267 ret = io_openat2(req, force_nonblock);
5269 case IORING_OP_EPOLL_CTL:
5271 ret = io_epoll_ctl_prep(req, sqe);
5275 ret = io_epoll_ctl(req, force_nonblock);
5277 case IORING_OP_SPLICE:
5279 ret = io_splice_prep(req, sqe);
5283 ret = io_splice(req, force_nonblock);
5285 case IORING_OP_PROVIDE_BUFFERS:
5287 ret = io_provide_buffers_prep(req, sqe);
5291 ret = io_provide_buffers(req, force_nonblock);
5293 case IORING_OP_REMOVE_BUFFERS:
5295 ret = io_remove_buffers_prep(req, sqe);
5299 ret = io_remove_buffers(req, force_nonblock);
5309 /* If the op doesn't have a file, we're not polling for it */
5310 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5311 const bool in_async = io_wq_current_is_worker();
5313 if (req->result == -EAGAIN)
5316 /* workqueue context doesn't hold uring_lock, grab it now */
5318 mutex_lock(&ctx->uring_lock);
5320 io_iopoll_req_issued(req);
5323 mutex_unlock(&ctx->uring_lock);
5329 static void io_wq_submit_work(struct io_wq_work **workptr)
5331 struct io_wq_work *work = *workptr;
5332 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5335 /* if NO_CANCEL is set, we must still run the work */
5336 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5337 IO_WQ_WORK_CANCEL) {
5343 ret = io_issue_sqe(req, NULL, false);
5345 * We can get EAGAIN for polled IO even though we're
5346 * forcing a sync submission from here, since we can't
5347 * wait for request slots on the block side.
5356 req_set_fail_links(req);
5357 io_cqring_add_event(req, ret);
5361 io_steal_work(req, workptr);
5364 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5367 struct fixed_file_table *table;
5369 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5370 return table->files[index & IORING_FILE_TABLE_MASK];;
5373 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5374 int fd, struct file **out_file, bool fixed)
5376 struct io_ring_ctx *ctx = req->ctx;
5380 if (unlikely(!ctx->file_data ||
5381 (unsigned) fd >= ctx->nr_user_files))
5383 fd = array_index_nospec(fd, ctx->nr_user_files);
5384 file = io_file_from_index(ctx, fd);
5387 req->fixed_file_refs = ctx->file_data->cur_refs;
5388 percpu_ref_get(req->fixed_file_refs);
5390 trace_io_uring_file_get(ctx, fd);
5391 file = __io_file_get(state, fd);
5392 if (unlikely(!file))
5400 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5405 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5406 if (unlikely(!fixed && req->needs_fixed_file))
5409 return io_file_get(state, req, fd, &req->file, fixed);
5412 static int io_grab_files(struct io_kiocb *req)
5415 struct io_ring_ctx *ctx = req->ctx;
5417 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5419 if (!ctx->ring_file)
5423 spin_lock_irq(&ctx->inflight_lock);
5425 * We use the f_ops->flush() handler to ensure that we can flush
5426 * out work accessing these files if the fd is closed. Check if
5427 * the fd has changed since we started down this path, and disallow
5428 * this operation if it has.
5430 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5431 list_add(&req->inflight_entry, &ctx->inflight_list);
5432 req->flags |= REQ_F_INFLIGHT;
5433 req->work.files = current->files;
5436 spin_unlock_irq(&ctx->inflight_lock);
5442 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5444 struct io_timeout_data *data = container_of(timer,
5445 struct io_timeout_data, timer);
5446 struct io_kiocb *req = data->req;
5447 struct io_ring_ctx *ctx = req->ctx;
5448 struct io_kiocb *prev = NULL;
5449 unsigned long flags;
5451 spin_lock_irqsave(&ctx->completion_lock, flags);
5454 * We don't expect the list to be empty, that will only happen if we
5455 * race with the completion of the linked work.
5457 if (!list_empty(&req->link_list)) {
5458 prev = list_entry(req->link_list.prev, struct io_kiocb,
5460 if (refcount_inc_not_zero(&prev->refs)) {
5461 list_del_init(&req->link_list);
5462 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5467 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5470 req_set_fail_links(prev);
5471 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5474 io_cqring_add_event(req, -ETIME);
5477 return HRTIMER_NORESTART;
5480 static void io_queue_linked_timeout(struct io_kiocb *req)
5482 struct io_ring_ctx *ctx = req->ctx;
5485 * If the list is now empty, then our linked request finished before
5486 * we got a chance to setup the timer
5488 spin_lock_irq(&ctx->completion_lock);
5489 if (!list_empty(&req->link_list)) {
5490 struct io_timeout_data *data = &req->io->timeout;
5492 data->timer.function = io_link_timeout_fn;
5493 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5496 spin_unlock_irq(&ctx->completion_lock);
5498 /* drop submission reference */
5502 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5504 struct io_kiocb *nxt;
5506 if (!(req->flags & REQ_F_LINK_HEAD))
5508 /* for polled retry, if flag is set, we already went through here */
5509 if (req->flags & REQ_F_POLLED)
5512 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5514 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5517 req->flags |= REQ_F_LINK_TIMEOUT;
5521 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5523 struct io_kiocb *linked_timeout;
5524 struct io_kiocb *nxt;
5525 const struct cred *old_creds = NULL;
5529 linked_timeout = io_prep_linked_timeout(req);
5531 if (req->work.creds && req->work.creds != current_cred()) {
5533 revert_creds(old_creds);
5534 if (old_creds == req->work.creds)
5535 old_creds = NULL; /* restored original creds */
5537 old_creds = override_creds(req->work.creds);
5540 ret = io_issue_sqe(req, sqe, true);
5543 * We async punt it if the file wasn't marked NOWAIT, or if the file
5544 * doesn't support non-blocking read/write attempts
5546 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5547 (req->flags & REQ_F_MUST_PUNT))) {
5548 if (io_arm_poll_handler(req)) {
5550 io_queue_linked_timeout(linked_timeout);
5554 if (io_op_defs[req->opcode].file_table) {
5555 ret = io_grab_files(req);
5561 * Queued up for async execution, worker will release
5562 * submit reference when the iocb is actually submitted.
5564 io_queue_async_work(req);
5570 /* drop submission reference */
5571 io_put_req_find_next(req, &nxt);
5573 if (linked_timeout) {
5575 io_queue_linked_timeout(linked_timeout);
5577 io_put_req(linked_timeout);
5580 /* and drop final reference, if we failed */
5582 io_cqring_add_event(req, ret);
5583 req_set_fail_links(req);
5589 if (req->flags & REQ_F_FORCE_ASYNC)
5595 revert_creds(old_creds);
5598 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5602 ret = io_req_defer(req, sqe);
5604 if (ret != -EIOCBQUEUED) {
5606 io_cqring_add_event(req, ret);
5607 req_set_fail_links(req);
5608 io_double_put_req(req);
5610 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5613 if (io_alloc_async_ctx(req))
5615 ret = io_req_defer_prep(req, sqe);
5616 if (unlikely(ret < 0))
5621 * Never try inline submit of IOSQE_ASYNC is set, go straight
5622 * to async execution.
5624 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5625 io_queue_async_work(req);
5627 __io_queue_sqe(req, sqe);
5631 static inline void io_queue_link_head(struct io_kiocb *req)
5633 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5634 io_cqring_add_event(req, -ECANCELED);
5635 io_double_put_req(req);
5637 io_queue_sqe(req, NULL);
5640 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5641 struct io_submit_state *state, struct io_kiocb **link)
5643 struct io_ring_ctx *ctx = req->ctx;
5647 * If we already have a head request, queue this one for async
5648 * submittal once the head completes. If we don't have a head but
5649 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5650 * submitted sync once the chain is complete. If none of those
5651 * conditions are true (normal request), then just queue it.
5654 struct io_kiocb *head = *link;
5657 * Taking sequential execution of a link, draining both sides
5658 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5659 * requests in the link. So, it drains the head and the
5660 * next after the link request. The last one is done via
5661 * drain_next flag to persist the effect across calls.
5663 if (req->flags & REQ_F_IO_DRAIN) {
5664 head->flags |= REQ_F_IO_DRAIN;
5665 ctx->drain_next = 1;
5667 if (io_alloc_async_ctx(req))
5670 ret = io_req_defer_prep(req, sqe);
5672 /* fail even hard links since we don't submit */
5673 head->flags |= REQ_F_FAIL_LINK;
5676 trace_io_uring_link(ctx, req, head);
5677 list_add_tail(&req->link_list, &head->link_list);
5679 /* last request of a link, enqueue the link */
5680 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5681 io_queue_link_head(head);
5685 if (unlikely(ctx->drain_next)) {
5686 req->flags |= REQ_F_IO_DRAIN;
5687 ctx->drain_next = 0;
5689 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5690 req->flags |= REQ_F_LINK_HEAD;
5691 INIT_LIST_HEAD(&req->link_list);
5693 if (io_alloc_async_ctx(req))
5696 ret = io_req_defer_prep(req, sqe);
5698 req->flags |= REQ_F_FAIL_LINK;
5701 io_queue_sqe(req, sqe);
5709 * Batched submission is done, ensure local IO is flushed out.
5711 static void io_submit_state_end(struct io_submit_state *state)
5713 blk_finish_plug(&state->plug);
5715 if (state->free_reqs)
5716 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5720 * Start submission side cache.
5722 static void io_submit_state_start(struct io_submit_state *state,
5723 unsigned int max_ios)
5725 blk_start_plug(&state->plug);
5726 state->free_reqs = 0;
5728 state->ios_left = max_ios;
5731 static void io_commit_sqring(struct io_ring_ctx *ctx)
5733 struct io_rings *rings = ctx->rings;
5736 * Ensure any loads from the SQEs are done at this point,
5737 * since once we write the new head, the application could
5738 * write new data to them.
5740 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5744 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5745 * that is mapped by userspace. This means that care needs to be taken to
5746 * ensure that reads are stable, as we cannot rely on userspace always
5747 * being a good citizen. If members of the sqe are validated and then later
5748 * used, it's important that those reads are done through READ_ONCE() to
5749 * prevent a re-load down the line.
5751 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5753 u32 *sq_array = ctx->sq_array;
5757 * The cached sq head (or cq tail) serves two purposes:
5759 * 1) allows us to batch the cost of updating the user visible
5761 * 2) allows the kernel side to track the head on its own, even
5762 * though the application is the one updating it.
5764 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5765 if (likely(head < ctx->sq_entries))
5766 return &ctx->sq_sqes[head];
5768 /* drop invalid entries */
5769 ctx->cached_sq_dropped++;
5770 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5774 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5776 ctx->cached_sq_head++;
5779 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5780 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5781 IOSQE_BUFFER_SELECT)
5783 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5784 const struct io_uring_sqe *sqe,
5785 struct io_submit_state *state, bool async)
5787 unsigned int sqe_flags;
5791 * All io need record the previous position, if LINK vs DARIN,
5792 * it can be used to mark the position of the first IO in the
5795 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5796 req->opcode = READ_ONCE(sqe->opcode);
5797 req->user_data = READ_ONCE(sqe->user_data);
5802 /* one is dropped after submission, the other at completion */
5803 refcount_set(&req->refs, 2);
5806 req->needs_fixed_file = async;
5807 INIT_IO_WORK(&req->work, io_wq_submit_work);
5809 if (unlikely(req->opcode >= IORING_OP_LAST))
5812 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5813 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5815 use_mm(ctx->sqo_mm);
5818 sqe_flags = READ_ONCE(sqe->flags);
5819 /* enforce forwards compatibility on users */
5820 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5823 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5824 !io_op_defs[req->opcode].buffer_select)
5827 id = READ_ONCE(sqe->personality);
5829 req->work.creds = idr_find(&ctx->personality_idr, id);
5830 if (unlikely(!req->work.creds))
5832 get_cred(req->work.creds);
5835 /* same numerical values with corresponding REQ_F_*, safe to copy */
5836 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5837 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5838 IOSQE_BUFFER_SELECT | IOSQE_IO_LINK);
5840 if (!io_op_defs[req->opcode].needs_file)
5843 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
5846 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5847 struct file *ring_file, int ring_fd, bool async)
5849 struct io_submit_state state, *statep = NULL;
5850 struct io_kiocb *link = NULL;
5851 int i, submitted = 0;
5853 /* if we have a backlog and couldn't flush it all, return BUSY */
5854 if (test_bit(0, &ctx->sq_check_overflow)) {
5855 if (!list_empty(&ctx->cq_overflow_list) &&
5856 !io_cqring_overflow_flush(ctx, false))
5860 /* make sure SQ entry isn't read before tail */
5861 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5863 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5866 if (nr > IO_PLUG_THRESHOLD) {
5867 io_submit_state_start(&state, nr);
5871 ctx->ring_fd = ring_fd;
5872 ctx->ring_file = ring_file;
5874 for (i = 0; i < nr; i++) {
5875 const struct io_uring_sqe *sqe;
5876 struct io_kiocb *req;
5879 sqe = io_get_sqe(ctx);
5880 if (unlikely(!sqe)) {
5881 io_consume_sqe(ctx);
5884 req = io_alloc_req(ctx, statep);
5885 if (unlikely(!req)) {
5887 submitted = -EAGAIN;
5891 err = io_init_req(ctx, req, sqe, statep, async);
5892 io_consume_sqe(ctx);
5893 /* will complete beyond this point, count as submitted */
5896 if (unlikely(err)) {
5898 io_cqring_add_event(req, err);
5899 io_double_put_req(req);
5903 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5905 err = io_submit_sqe(req, sqe, statep, &link);
5910 if (unlikely(submitted != nr)) {
5911 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5913 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5916 io_queue_link_head(link);
5918 io_submit_state_end(&state);
5920 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5921 io_commit_sqring(ctx);
5926 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5928 struct mm_struct *mm = current->mm;
5936 static int io_sq_thread(void *data)
5938 struct io_ring_ctx *ctx = data;
5939 const struct cred *old_cred;
5940 mm_segment_t old_fs;
5942 unsigned long timeout;
5945 complete(&ctx->completions[1]);
5949 old_cred = override_creds(ctx->creds);
5951 timeout = jiffies + ctx->sq_thread_idle;
5952 while (!kthread_should_park()) {
5953 unsigned int to_submit;
5955 if (!list_empty(&ctx->poll_list)) {
5956 unsigned nr_events = 0;
5958 mutex_lock(&ctx->uring_lock);
5959 if (!list_empty(&ctx->poll_list))
5960 io_iopoll_getevents(ctx, &nr_events, 0);
5962 timeout = jiffies + ctx->sq_thread_idle;
5963 mutex_unlock(&ctx->uring_lock);
5966 to_submit = io_sqring_entries(ctx);
5969 * If submit got -EBUSY, flag us as needing the application
5970 * to enter the kernel to reap and flush events.
5972 if (!to_submit || ret == -EBUSY) {
5974 * Drop cur_mm before scheduling, we can't hold it for
5975 * long periods (or over schedule()). Do this before
5976 * adding ourselves to the waitqueue, as the unuse/drop
5979 io_sq_thread_drop_mm(ctx);
5982 * We're polling. If we're within the defined idle
5983 * period, then let us spin without work before going
5984 * to sleep. The exception is if we got EBUSY doing
5985 * more IO, we should wait for the application to
5986 * reap events and wake us up.
5988 if (!list_empty(&ctx->poll_list) ||
5989 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5990 !percpu_ref_is_dying(&ctx->refs))) {
5991 if (current->task_works)
5997 prepare_to_wait(&ctx->sqo_wait, &wait,
5998 TASK_INTERRUPTIBLE);
6001 * While doing polled IO, before going to sleep, we need
6002 * to check if there are new reqs added to poll_list, it
6003 * is because reqs may have been punted to io worker and
6004 * will be added to poll_list later, hence check the
6007 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6008 !list_empty_careful(&ctx->poll_list)) {
6009 finish_wait(&ctx->sqo_wait, &wait);
6013 /* Tell userspace we may need a wakeup call */
6014 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6015 /* make sure to read SQ tail after writing flags */
6018 to_submit = io_sqring_entries(ctx);
6019 if (!to_submit || ret == -EBUSY) {
6020 if (kthread_should_park()) {
6021 finish_wait(&ctx->sqo_wait, &wait);
6024 if (current->task_works) {
6026 finish_wait(&ctx->sqo_wait, &wait);
6029 if (signal_pending(current))
6030 flush_signals(current);
6032 finish_wait(&ctx->sqo_wait, &wait);
6034 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6038 finish_wait(&ctx->sqo_wait, &wait);
6040 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6043 mutex_lock(&ctx->uring_lock);
6044 ret = io_submit_sqes(ctx, to_submit, NULL, -1, true);
6045 mutex_unlock(&ctx->uring_lock);
6046 timeout = jiffies + ctx->sq_thread_idle;
6049 if (current->task_works)
6053 io_sq_thread_drop_mm(ctx);
6054 revert_creds(old_cred);
6061 struct io_wait_queue {
6062 struct wait_queue_entry wq;
6063 struct io_ring_ctx *ctx;
6065 unsigned nr_timeouts;
6068 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6070 struct io_ring_ctx *ctx = iowq->ctx;
6073 * Wake up if we have enough events, or if a timeout occurred since we
6074 * started waiting. For timeouts, we always want to return to userspace,
6075 * regardless of event count.
6077 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6078 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6081 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6082 int wake_flags, void *key)
6084 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6087 /* use noflush == true, as we can't safely rely on locking context */
6088 if (!io_should_wake(iowq, true))
6091 return autoremove_wake_function(curr, mode, wake_flags, key);
6095 * Wait until events become available, if we don't already have some. The
6096 * application must reap them itself, as they reside on the shared cq ring.
6098 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6099 const sigset_t __user *sig, size_t sigsz)
6101 struct io_wait_queue iowq = {
6104 .func = io_wake_function,
6105 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6108 .to_wait = min_events,
6110 struct io_rings *rings = ctx->rings;
6114 if (io_cqring_events(ctx, false) >= min_events)
6116 if (!current->task_works)
6122 #ifdef CONFIG_COMPAT
6123 if (in_compat_syscall())
6124 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6128 ret = set_user_sigmask(sig, sigsz);
6134 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6135 trace_io_uring_cqring_wait(ctx, min_events);
6137 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6138 TASK_INTERRUPTIBLE);
6139 if (current->task_works)
6141 if (io_should_wake(&iowq, false))
6144 if (signal_pending(current)) {
6149 finish_wait(&ctx->wait, &iowq.wq);
6151 restore_saved_sigmask_unless(ret == -EINTR);
6153 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6156 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6158 #if defined(CONFIG_UNIX)
6159 if (ctx->ring_sock) {
6160 struct sock *sock = ctx->ring_sock->sk;
6161 struct sk_buff *skb;
6163 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6169 for (i = 0; i < ctx->nr_user_files; i++) {
6172 file = io_file_from_index(ctx, i);
6179 static void io_file_ref_kill(struct percpu_ref *ref)
6181 struct fixed_file_data *data;
6183 data = container_of(ref, struct fixed_file_data, refs);
6184 complete(&data->done);
6187 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6189 struct fixed_file_data *data = ctx->file_data;
6190 struct fixed_file_ref_node *ref_node = NULL;
6191 unsigned nr_tables, i;
6192 unsigned long flags;
6197 spin_lock_irqsave(&data->lock, flags);
6198 if (!list_empty(&data->ref_list))
6199 ref_node = list_first_entry(&data->ref_list,
6200 struct fixed_file_ref_node, node);
6201 spin_unlock_irqrestore(&data->lock, flags);
6203 percpu_ref_kill(&ref_node->refs);
6205 percpu_ref_kill(&data->refs);
6207 /* wait for all refs nodes to complete */
6208 wait_for_completion(&data->done);
6210 __io_sqe_files_unregister(ctx);
6211 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6212 for (i = 0; i < nr_tables; i++)
6213 kfree(data->table[i].files);
6215 percpu_ref_exit(&data->refs);
6217 ctx->file_data = NULL;
6218 ctx->nr_user_files = 0;
6222 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6224 if (ctx->sqo_thread) {
6225 wait_for_completion(&ctx->completions[1]);
6227 * The park is a bit of a work-around, without it we get
6228 * warning spews on shutdown with SQPOLL set and affinity
6229 * set to a single CPU.
6231 kthread_park(ctx->sqo_thread);
6232 kthread_stop(ctx->sqo_thread);
6233 ctx->sqo_thread = NULL;
6237 static void io_finish_async(struct io_ring_ctx *ctx)
6239 io_sq_thread_stop(ctx);
6242 io_wq_destroy(ctx->io_wq);
6247 #if defined(CONFIG_UNIX)
6249 * Ensure the UNIX gc is aware of our file set, so we are certain that
6250 * the io_uring can be safely unregistered on process exit, even if we have
6251 * loops in the file referencing.
6253 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6255 struct sock *sk = ctx->ring_sock->sk;
6256 struct scm_fp_list *fpl;
6257 struct sk_buff *skb;
6260 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6264 skb = alloc_skb(0, GFP_KERNEL);
6273 fpl->user = get_uid(ctx->user);
6274 for (i = 0; i < nr; i++) {
6275 struct file *file = io_file_from_index(ctx, i + offset);
6279 fpl->fp[nr_files] = get_file(file);
6280 unix_inflight(fpl->user, fpl->fp[nr_files]);
6285 fpl->max = SCM_MAX_FD;
6286 fpl->count = nr_files;
6287 UNIXCB(skb).fp = fpl;
6288 skb->destructor = unix_destruct_scm;
6289 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6290 skb_queue_head(&sk->sk_receive_queue, skb);
6292 for (i = 0; i < nr_files; i++)
6303 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6304 * causes regular reference counting to break down. We rely on the UNIX
6305 * garbage collection to take care of this problem for us.
6307 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6309 unsigned left, total;
6313 left = ctx->nr_user_files;
6315 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6317 ret = __io_sqe_files_scm(ctx, this_files, total);
6321 total += this_files;
6327 while (total < ctx->nr_user_files) {
6328 struct file *file = io_file_from_index(ctx, total);
6338 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6344 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6349 for (i = 0; i < nr_tables; i++) {
6350 struct fixed_file_table *table = &ctx->file_data->table[i];
6351 unsigned this_files;
6353 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6354 table->files = kcalloc(this_files, sizeof(struct file *),
6358 nr_files -= this_files;
6364 for (i = 0; i < nr_tables; i++) {
6365 struct fixed_file_table *table = &ctx->file_data->table[i];
6366 kfree(table->files);
6371 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6373 #if defined(CONFIG_UNIX)
6374 struct sock *sock = ctx->ring_sock->sk;
6375 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6376 struct sk_buff *skb;
6379 __skb_queue_head_init(&list);
6382 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6383 * remove this entry and rearrange the file array.
6385 skb = skb_dequeue(head);
6387 struct scm_fp_list *fp;
6389 fp = UNIXCB(skb).fp;
6390 for (i = 0; i < fp->count; i++) {
6393 if (fp->fp[i] != file)
6396 unix_notinflight(fp->user, fp->fp[i]);
6397 left = fp->count - 1 - i;
6399 memmove(&fp->fp[i], &fp->fp[i + 1],
6400 left * sizeof(struct file *));
6407 __skb_queue_tail(&list, skb);
6417 __skb_queue_tail(&list, skb);
6419 skb = skb_dequeue(head);
6422 if (skb_peek(&list)) {
6423 spin_lock_irq(&head->lock);
6424 while ((skb = __skb_dequeue(&list)) != NULL)
6425 __skb_queue_tail(head, skb);
6426 spin_unlock_irq(&head->lock);
6433 struct io_file_put {
6434 struct list_head list;
6438 static void io_file_put_work(struct work_struct *work)
6440 struct fixed_file_ref_node *ref_node;
6441 struct fixed_file_data *file_data;
6442 struct io_ring_ctx *ctx;
6443 struct io_file_put *pfile, *tmp;
6444 unsigned long flags;
6446 ref_node = container_of(work, struct fixed_file_ref_node, work);
6447 file_data = ref_node->file_data;
6448 ctx = file_data->ctx;
6450 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6451 list_del_init(&pfile->list);
6452 io_ring_file_put(ctx, pfile->file);
6456 spin_lock_irqsave(&file_data->lock, flags);
6457 list_del_init(&ref_node->node);
6458 spin_unlock_irqrestore(&file_data->lock, flags);
6460 percpu_ref_exit(&ref_node->refs);
6462 percpu_ref_put(&file_data->refs);
6465 static void io_file_data_ref_zero(struct percpu_ref *ref)
6467 struct fixed_file_ref_node *ref_node;
6469 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6471 queue_work(system_wq, &ref_node->work);
6474 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6475 struct io_ring_ctx *ctx)
6477 struct fixed_file_ref_node *ref_node;
6479 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6481 return ERR_PTR(-ENOMEM);
6483 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6486 return ERR_PTR(-ENOMEM);
6488 INIT_LIST_HEAD(&ref_node->node);
6489 INIT_LIST_HEAD(&ref_node->file_list);
6490 INIT_WORK(&ref_node->work, io_file_put_work);
6491 ref_node->file_data = ctx->file_data;
6496 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6498 percpu_ref_exit(&ref_node->refs);
6502 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6505 __s32 __user *fds = (__s32 __user *) arg;
6510 struct fixed_file_ref_node *ref_node;
6511 unsigned long flags;
6517 if (nr_args > IORING_MAX_FIXED_FILES)
6520 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6521 if (!ctx->file_data)
6523 ctx->file_data->ctx = ctx;
6524 init_completion(&ctx->file_data->done);
6525 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6526 spin_lock_init(&ctx->file_data->lock);
6528 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6529 ctx->file_data->table = kcalloc(nr_tables,
6530 sizeof(struct fixed_file_table),
6532 if (!ctx->file_data->table) {
6533 kfree(ctx->file_data);
6534 ctx->file_data = NULL;
6538 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6539 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6540 kfree(ctx->file_data->table);
6541 kfree(ctx->file_data);
6542 ctx->file_data = NULL;
6546 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6547 percpu_ref_exit(&ctx->file_data->refs);
6548 kfree(ctx->file_data->table);
6549 kfree(ctx->file_data);
6550 ctx->file_data = NULL;
6554 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6555 struct fixed_file_table *table;
6559 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6561 /* allow sparse sets */
6567 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6568 index = i & IORING_FILE_TABLE_MASK;
6576 * Don't allow io_uring instances to be registered. If UNIX
6577 * isn't enabled, then this causes a reference cycle and this
6578 * instance can never get freed. If UNIX is enabled we'll
6579 * handle it just fine, but there's still no point in allowing
6580 * a ring fd as it doesn't support regular read/write anyway.
6582 if (file->f_op == &io_uring_fops) {
6587 table->files[index] = file;
6591 for (i = 0; i < ctx->nr_user_files; i++) {
6592 file = io_file_from_index(ctx, i);
6596 for (i = 0; i < nr_tables; i++)
6597 kfree(ctx->file_data->table[i].files);
6599 kfree(ctx->file_data->table);
6600 kfree(ctx->file_data);
6601 ctx->file_data = NULL;
6602 ctx->nr_user_files = 0;
6606 ret = io_sqe_files_scm(ctx);
6608 io_sqe_files_unregister(ctx);
6612 ref_node = alloc_fixed_file_ref_node(ctx);
6613 if (IS_ERR(ref_node)) {
6614 io_sqe_files_unregister(ctx);
6615 return PTR_ERR(ref_node);
6618 ctx->file_data->cur_refs = &ref_node->refs;
6619 spin_lock_irqsave(&ctx->file_data->lock, flags);
6620 list_add(&ref_node->node, &ctx->file_data->ref_list);
6621 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6622 percpu_ref_get(&ctx->file_data->refs);
6626 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6629 #if defined(CONFIG_UNIX)
6630 struct sock *sock = ctx->ring_sock->sk;
6631 struct sk_buff_head *head = &sock->sk_receive_queue;
6632 struct sk_buff *skb;
6635 * See if we can merge this file into an existing skb SCM_RIGHTS
6636 * file set. If there's no room, fall back to allocating a new skb
6637 * and filling it in.
6639 spin_lock_irq(&head->lock);
6640 skb = skb_peek(head);
6642 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6644 if (fpl->count < SCM_MAX_FD) {
6645 __skb_unlink(skb, head);
6646 spin_unlock_irq(&head->lock);
6647 fpl->fp[fpl->count] = get_file(file);
6648 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6650 spin_lock_irq(&head->lock);
6651 __skb_queue_head(head, skb);
6656 spin_unlock_irq(&head->lock);
6663 return __io_sqe_files_scm(ctx, 1, index);
6669 static int io_queue_file_removal(struct fixed_file_data *data,
6672 struct io_file_put *pfile;
6673 struct percpu_ref *refs = data->cur_refs;
6674 struct fixed_file_ref_node *ref_node;
6676 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6680 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6682 list_add(&pfile->list, &ref_node->file_list);
6687 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6688 struct io_uring_files_update *up,
6691 struct fixed_file_data *data = ctx->file_data;
6692 struct fixed_file_ref_node *ref_node;
6697 unsigned long flags;
6698 bool needs_switch = false;
6700 if (check_add_overflow(up->offset, nr_args, &done))
6702 if (done > ctx->nr_user_files)
6705 ref_node = alloc_fixed_file_ref_node(ctx);
6706 if (IS_ERR(ref_node))
6707 return PTR_ERR(ref_node);
6710 fds = u64_to_user_ptr(up->fds);
6712 struct fixed_file_table *table;
6716 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6720 i = array_index_nospec(up->offset, ctx->nr_user_files);
6721 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6722 index = i & IORING_FILE_TABLE_MASK;
6723 if (table->files[index]) {
6724 file = io_file_from_index(ctx, index);
6725 err = io_queue_file_removal(data, file);
6728 table->files[index] = NULL;
6729 needs_switch = true;
6738 * Don't allow io_uring instances to be registered. If
6739 * UNIX isn't enabled, then this causes a reference
6740 * cycle and this instance can never get freed. If UNIX
6741 * is enabled we'll handle it just fine, but there's
6742 * still no point in allowing a ring fd as it doesn't
6743 * support regular read/write anyway.
6745 if (file->f_op == &io_uring_fops) {
6750 table->files[index] = file;
6751 err = io_sqe_file_register(ctx, file, i);
6761 percpu_ref_kill(data->cur_refs);
6762 spin_lock_irqsave(&data->lock, flags);
6763 list_add(&ref_node->node, &data->ref_list);
6764 data->cur_refs = &ref_node->refs;
6765 spin_unlock_irqrestore(&data->lock, flags);
6766 percpu_ref_get(&ctx->file_data->refs);
6768 destroy_fixed_file_ref_node(ref_node);
6770 return done ? done : err;
6773 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6776 struct io_uring_files_update up;
6778 if (!ctx->file_data)
6782 if (copy_from_user(&up, arg, sizeof(up)))
6787 return __io_sqe_files_update(ctx, &up, nr_args);
6790 static void io_free_work(struct io_wq_work *work)
6792 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6794 /* Consider that io_steal_work() relies on this ref */
6798 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6799 struct io_uring_params *p)
6801 struct io_wq_data data;
6803 struct io_ring_ctx *ctx_attach;
6804 unsigned int concurrency;
6807 data.user = ctx->user;
6808 data.free_work = io_free_work;
6810 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6811 /* Do QD, or 4 * CPUS, whatever is smallest */
6812 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6814 ctx->io_wq = io_wq_create(concurrency, &data);
6815 if (IS_ERR(ctx->io_wq)) {
6816 ret = PTR_ERR(ctx->io_wq);
6822 f = fdget(p->wq_fd);
6826 if (f.file->f_op != &io_uring_fops) {
6831 ctx_attach = f.file->private_data;
6832 /* @io_wq is protected by holding the fd */
6833 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6838 ctx->io_wq = ctx_attach->io_wq;
6844 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6845 struct io_uring_params *p)
6849 mmgrab(current->mm);
6850 ctx->sqo_mm = current->mm;
6852 if (ctx->flags & IORING_SETUP_SQPOLL) {
6854 if (!capable(CAP_SYS_ADMIN))
6857 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6858 if (!ctx->sq_thread_idle)
6859 ctx->sq_thread_idle = HZ;
6861 if (p->flags & IORING_SETUP_SQ_AFF) {
6862 int cpu = p->sq_thread_cpu;
6865 if (cpu >= nr_cpu_ids)
6867 if (!cpu_online(cpu))
6870 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6874 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6877 if (IS_ERR(ctx->sqo_thread)) {
6878 ret = PTR_ERR(ctx->sqo_thread);
6879 ctx->sqo_thread = NULL;
6882 wake_up_process(ctx->sqo_thread);
6883 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6884 /* Can't have SQ_AFF without SQPOLL */
6889 ret = io_init_wq_offload(ctx, p);
6895 io_finish_async(ctx);
6896 mmdrop(ctx->sqo_mm);
6901 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6903 atomic_long_sub(nr_pages, &user->locked_vm);
6906 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6908 unsigned long page_limit, cur_pages, new_pages;
6910 /* Don't allow more pages than we can safely lock */
6911 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6914 cur_pages = atomic_long_read(&user->locked_vm);
6915 new_pages = cur_pages + nr_pages;
6916 if (new_pages > page_limit)
6918 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6919 new_pages) != cur_pages);
6924 static void io_mem_free(void *ptr)
6931 page = virt_to_head_page(ptr);
6932 if (put_page_testzero(page))
6933 free_compound_page(page);
6936 static void *io_mem_alloc(size_t size)
6938 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6941 return (void *) __get_free_pages(gfp_flags, get_order(size));
6944 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6947 struct io_rings *rings;
6948 size_t off, sq_array_size;
6950 off = struct_size(rings, cqes, cq_entries);
6951 if (off == SIZE_MAX)
6955 off = ALIGN(off, SMP_CACHE_BYTES);
6960 sq_array_size = array_size(sizeof(u32), sq_entries);
6961 if (sq_array_size == SIZE_MAX)
6964 if (check_add_overflow(off, sq_array_size, &off))
6973 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6977 pages = (size_t)1 << get_order(
6978 rings_size(sq_entries, cq_entries, NULL));
6979 pages += (size_t)1 << get_order(
6980 array_size(sizeof(struct io_uring_sqe), sq_entries));
6985 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6989 if (!ctx->user_bufs)
6992 for (i = 0; i < ctx->nr_user_bufs; i++) {
6993 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6995 for (j = 0; j < imu->nr_bvecs; j++)
6996 unpin_user_page(imu->bvec[j].bv_page);
6998 if (ctx->account_mem)
6999 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7004 kfree(ctx->user_bufs);
7005 ctx->user_bufs = NULL;
7006 ctx->nr_user_bufs = 0;
7010 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7011 void __user *arg, unsigned index)
7013 struct iovec __user *src;
7015 #ifdef CONFIG_COMPAT
7017 struct compat_iovec __user *ciovs;
7018 struct compat_iovec ciov;
7020 ciovs = (struct compat_iovec __user *) arg;
7021 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7024 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7025 dst->iov_len = ciov.iov_len;
7029 src = (struct iovec __user *) arg;
7030 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7035 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7038 struct vm_area_struct **vmas = NULL;
7039 struct page **pages = NULL;
7040 int i, j, got_pages = 0;
7045 if (!nr_args || nr_args > UIO_MAXIOV)
7048 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7050 if (!ctx->user_bufs)
7053 for (i = 0; i < nr_args; i++) {
7054 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7055 unsigned long off, start, end, ubuf;
7060 ret = io_copy_iov(ctx, &iov, arg, i);
7065 * Don't impose further limits on the size and buffer
7066 * constraints here, we'll -EINVAL later when IO is
7067 * submitted if they are wrong.
7070 if (!iov.iov_base || !iov.iov_len)
7073 /* arbitrary limit, but we need something */
7074 if (iov.iov_len > SZ_1G)
7077 ubuf = (unsigned long) iov.iov_base;
7078 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7079 start = ubuf >> PAGE_SHIFT;
7080 nr_pages = end - start;
7082 if (ctx->account_mem) {
7083 ret = io_account_mem(ctx->user, nr_pages);
7089 if (!pages || nr_pages > got_pages) {
7092 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7094 vmas = kvmalloc_array(nr_pages,
7095 sizeof(struct vm_area_struct *),
7097 if (!pages || !vmas) {
7099 if (ctx->account_mem)
7100 io_unaccount_mem(ctx->user, nr_pages);
7103 got_pages = nr_pages;
7106 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7110 if (ctx->account_mem)
7111 io_unaccount_mem(ctx->user, nr_pages);
7116 down_read(¤t->mm->mmap_sem);
7117 pret = pin_user_pages(ubuf, nr_pages,
7118 FOLL_WRITE | FOLL_LONGTERM,
7120 if (pret == nr_pages) {
7121 /* don't support file backed memory */
7122 for (j = 0; j < nr_pages; j++) {
7123 struct vm_area_struct *vma = vmas[j];
7126 !is_file_hugepages(vma->vm_file)) {
7132 ret = pret < 0 ? pret : -EFAULT;
7134 up_read(¤t->mm->mmap_sem);
7137 * if we did partial map, or found file backed vmas,
7138 * release any pages we did get
7141 unpin_user_pages(pages, pret);
7142 if (ctx->account_mem)
7143 io_unaccount_mem(ctx->user, nr_pages);
7148 off = ubuf & ~PAGE_MASK;
7150 for (j = 0; j < nr_pages; j++) {
7153 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7154 imu->bvec[j].bv_page = pages[j];
7155 imu->bvec[j].bv_len = vec_len;
7156 imu->bvec[j].bv_offset = off;
7160 /* store original address for later verification */
7162 imu->len = iov.iov_len;
7163 imu->nr_bvecs = nr_pages;
7165 ctx->nr_user_bufs++;
7173 io_sqe_buffer_unregister(ctx);
7177 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7179 __s32 __user *fds = arg;
7185 if (copy_from_user(&fd, fds, sizeof(*fds)))
7188 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7189 if (IS_ERR(ctx->cq_ev_fd)) {
7190 int ret = PTR_ERR(ctx->cq_ev_fd);
7191 ctx->cq_ev_fd = NULL;
7198 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7200 if (ctx->cq_ev_fd) {
7201 eventfd_ctx_put(ctx->cq_ev_fd);
7202 ctx->cq_ev_fd = NULL;
7209 static int __io_destroy_buffers(int id, void *p, void *data)
7211 struct io_ring_ctx *ctx = data;
7212 struct io_buffer *buf = p;
7214 __io_remove_buffers(ctx, buf, id, -1U);
7218 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7220 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7221 idr_destroy(&ctx->io_buffer_idr);
7224 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7226 io_finish_async(ctx);
7228 mmdrop(ctx->sqo_mm);
7230 io_iopoll_reap_events(ctx);
7231 io_sqe_buffer_unregister(ctx);
7232 io_sqe_files_unregister(ctx);
7233 io_eventfd_unregister(ctx);
7234 io_destroy_buffers(ctx);
7235 idr_destroy(&ctx->personality_idr);
7237 #if defined(CONFIG_UNIX)
7238 if (ctx->ring_sock) {
7239 ctx->ring_sock->file = NULL; /* so that iput() is called */
7240 sock_release(ctx->ring_sock);
7244 io_mem_free(ctx->rings);
7245 io_mem_free(ctx->sq_sqes);
7247 percpu_ref_exit(&ctx->refs);
7248 if (ctx->account_mem)
7249 io_unaccount_mem(ctx->user,
7250 ring_pages(ctx->sq_entries, ctx->cq_entries));
7251 free_uid(ctx->user);
7252 put_cred(ctx->creds);
7253 kfree(ctx->completions);
7254 kfree(ctx->cancel_hash);
7255 kmem_cache_free(req_cachep, ctx->fallback_req);
7259 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7261 struct io_ring_ctx *ctx = file->private_data;
7264 poll_wait(file, &ctx->cq_wait, wait);
7266 * synchronizes with barrier from wq_has_sleeper call in
7270 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7271 ctx->rings->sq_ring_entries)
7272 mask |= EPOLLOUT | EPOLLWRNORM;
7273 if (io_cqring_events(ctx, false))
7274 mask |= EPOLLIN | EPOLLRDNORM;
7279 static int io_uring_fasync(int fd, struct file *file, int on)
7281 struct io_ring_ctx *ctx = file->private_data;
7283 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7286 static int io_remove_personalities(int id, void *p, void *data)
7288 struct io_ring_ctx *ctx = data;
7289 const struct cred *cred;
7291 cred = idr_remove(&ctx->personality_idr, id);
7297 static void io_ring_exit_work(struct work_struct *work)
7299 struct io_ring_ctx *ctx;
7301 ctx = container_of(work, struct io_ring_ctx, exit_work);
7303 io_cqring_overflow_flush(ctx, true);
7305 wait_for_completion(&ctx->completions[0]);
7306 io_ring_ctx_free(ctx);
7309 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7311 mutex_lock(&ctx->uring_lock);
7312 percpu_ref_kill(&ctx->refs);
7313 mutex_unlock(&ctx->uring_lock);
7316 * Wait for sq thread to idle, if we have one. It won't spin on new
7317 * work after we've killed the ctx ref above. This is important to do
7318 * before we cancel existing commands, as the thread could otherwise
7319 * be queueing new work post that. If that's work we need to cancel,
7320 * it could cause shutdown to hang.
7322 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7325 io_kill_timeouts(ctx);
7326 io_poll_remove_all(ctx);
7329 io_wq_cancel_all(ctx->io_wq);
7331 io_iopoll_reap_events(ctx);
7332 /* if we failed setting up the ctx, we might not have any rings */
7334 io_cqring_overflow_flush(ctx, true);
7335 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7336 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7337 queue_work(system_wq, &ctx->exit_work);
7340 static int io_uring_release(struct inode *inode, struct file *file)
7342 struct io_ring_ctx *ctx = file->private_data;
7344 file->private_data = NULL;
7345 io_ring_ctx_wait_and_kill(ctx);
7349 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7350 struct files_struct *files)
7352 while (!list_empty_careful(&ctx->inflight_list)) {
7353 struct io_kiocb *cancel_req = NULL, *req;
7356 spin_lock_irq(&ctx->inflight_lock);
7357 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7358 if (req->work.files != files)
7360 /* req is being completed, ignore */
7361 if (!refcount_inc_not_zero(&req->refs))
7367 prepare_to_wait(&ctx->inflight_wait, &wait,
7368 TASK_UNINTERRUPTIBLE);
7369 spin_unlock_irq(&ctx->inflight_lock);
7371 /* We need to keep going until we don't find a matching req */
7375 if (cancel_req->flags & REQ_F_OVERFLOW) {
7376 spin_lock_irq(&ctx->completion_lock);
7377 list_del(&cancel_req->list);
7378 cancel_req->flags &= ~REQ_F_OVERFLOW;
7379 if (list_empty(&ctx->cq_overflow_list)) {
7380 clear_bit(0, &ctx->sq_check_overflow);
7381 clear_bit(0, &ctx->cq_check_overflow);
7383 spin_unlock_irq(&ctx->completion_lock);
7385 WRITE_ONCE(ctx->rings->cq_overflow,
7386 atomic_inc_return(&ctx->cached_cq_overflow));
7389 * Put inflight ref and overflow ref. If that's
7390 * all we had, then we're done with this request.
7392 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7393 io_put_req(cancel_req);
7394 finish_wait(&ctx->inflight_wait, &wait);
7399 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7400 io_put_req(cancel_req);
7402 finish_wait(&ctx->inflight_wait, &wait);
7406 static int io_uring_flush(struct file *file, void *data)
7408 struct io_ring_ctx *ctx = file->private_data;
7410 io_uring_cancel_files(ctx, data);
7413 * If the task is going away, cancel work it may have pending
7415 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7416 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7421 static void *io_uring_validate_mmap_request(struct file *file,
7422 loff_t pgoff, size_t sz)
7424 struct io_ring_ctx *ctx = file->private_data;
7425 loff_t offset = pgoff << PAGE_SHIFT;
7430 case IORING_OFF_SQ_RING:
7431 case IORING_OFF_CQ_RING:
7434 case IORING_OFF_SQES:
7438 return ERR_PTR(-EINVAL);
7441 page = virt_to_head_page(ptr);
7442 if (sz > page_size(page))
7443 return ERR_PTR(-EINVAL);
7450 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7452 size_t sz = vma->vm_end - vma->vm_start;
7456 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7458 return PTR_ERR(ptr);
7460 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7461 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7464 #else /* !CONFIG_MMU */
7466 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7468 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7471 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7473 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7476 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7477 unsigned long addr, unsigned long len,
7478 unsigned long pgoff, unsigned long flags)
7482 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7484 return PTR_ERR(ptr);
7486 return (unsigned long) ptr;
7489 #endif /* !CONFIG_MMU */
7491 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7492 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7495 struct io_ring_ctx *ctx;
7500 if (current->task_works)
7503 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7511 if (f.file->f_op != &io_uring_fops)
7515 ctx = f.file->private_data;
7516 if (!percpu_ref_tryget(&ctx->refs))
7520 * For SQ polling, the thread will do all submissions and completions.
7521 * Just return the requested submit count, and wake the thread if
7525 if (ctx->flags & IORING_SETUP_SQPOLL) {
7526 if (!list_empty_careful(&ctx->cq_overflow_list))
7527 io_cqring_overflow_flush(ctx, false);
7528 if (flags & IORING_ENTER_SQ_WAKEUP)
7529 wake_up(&ctx->sqo_wait);
7530 submitted = to_submit;
7531 } else if (to_submit) {
7532 mutex_lock(&ctx->uring_lock);
7533 submitted = io_submit_sqes(ctx, to_submit, f.file, fd, false);
7534 mutex_unlock(&ctx->uring_lock);
7536 if (submitted != to_submit)
7539 if (flags & IORING_ENTER_GETEVENTS) {
7540 unsigned nr_events = 0;
7542 min_complete = min(min_complete, ctx->cq_entries);
7545 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7546 * space applications don't need to do io completion events
7547 * polling again, they can rely on io_sq_thread to do polling
7548 * work, which can reduce cpu usage and uring_lock contention.
7550 if (ctx->flags & IORING_SETUP_IOPOLL &&
7551 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7552 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7554 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7559 percpu_ref_put(&ctx->refs);
7562 return submitted ? submitted : ret;
7565 #ifdef CONFIG_PROC_FS
7566 static int io_uring_show_cred(int id, void *p, void *data)
7568 const struct cred *cred = p;
7569 struct seq_file *m = data;
7570 struct user_namespace *uns = seq_user_ns(m);
7571 struct group_info *gi;
7576 seq_printf(m, "%5d\n", id);
7577 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7578 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7579 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7580 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7581 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7582 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7583 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7584 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7585 seq_puts(m, "\n\tGroups:\t");
7586 gi = cred->group_info;
7587 for (g = 0; g < gi->ngroups; g++) {
7588 seq_put_decimal_ull(m, g ? " " : "",
7589 from_kgid_munged(uns, gi->gid[g]));
7591 seq_puts(m, "\n\tCapEff:\t");
7592 cap = cred->cap_effective;
7593 CAP_FOR_EACH_U32(__capi)
7594 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7599 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7603 mutex_lock(&ctx->uring_lock);
7604 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7605 for (i = 0; i < ctx->nr_user_files; i++) {
7606 struct fixed_file_table *table;
7609 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7610 f = table->files[i & IORING_FILE_TABLE_MASK];
7612 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7614 seq_printf(m, "%5u: <none>\n", i);
7616 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7617 for (i = 0; i < ctx->nr_user_bufs; i++) {
7618 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7620 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7621 (unsigned int) buf->len);
7623 if (!idr_is_empty(&ctx->personality_idr)) {
7624 seq_printf(m, "Personalities:\n");
7625 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7627 seq_printf(m, "PollList:\n");
7628 spin_lock_irq(&ctx->completion_lock);
7629 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7630 struct hlist_head *list = &ctx->cancel_hash[i];
7631 struct io_kiocb *req;
7633 hlist_for_each_entry(req, list, hash_node)
7634 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7635 req->task->task_works != NULL);
7637 spin_unlock_irq(&ctx->completion_lock);
7638 mutex_unlock(&ctx->uring_lock);
7641 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7643 struct io_ring_ctx *ctx = f->private_data;
7645 if (percpu_ref_tryget(&ctx->refs)) {
7646 __io_uring_show_fdinfo(ctx, m);
7647 percpu_ref_put(&ctx->refs);
7652 static const struct file_operations io_uring_fops = {
7653 .release = io_uring_release,
7654 .flush = io_uring_flush,
7655 .mmap = io_uring_mmap,
7657 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7658 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7660 .poll = io_uring_poll,
7661 .fasync = io_uring_fasync,
7662 #ifdef CONFIG_PROC_FS
7663 .show_fdinfo = io_uring_show_fdinfo,
7667 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7668 struct io_uring_params *p)
7670 struct io_rings *rings;
7671 size_t size, sq_array_offset;
7673 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7674 if (size == SIZE_MAX)
7677 rings = io_mem_alloc(size);
7682 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7683 rings->sq_ring_mask = p->sq_entries - 1;
7684 rings->cq_ring_mask = p->cq_entries - 1;
7685 rings->sq_ring_entries = p->sq_entries;
7686 rings->cq_ring_entries = p->cq_entries;
7687 ctx->sq_mask = rings->sq_ring_mask;
7688 ctx->cq_mask = rings->cq_ring_mask;
7689 ctx->sq_entries = rings->sq_ring_entries;
7690 ctx->cq_entries = rings->cq_ring_entries;
7692 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7693 if (size == SIZE_MAX) {
7694 io_mem_free(ctx->rings);
7699 ctx->sq_sqes = io_mem_alloc(size);
7700 if (!ctx->sq_sqes) {
7701 io_mem_free(ctx->rings);
7710 * Allocate an anonymous fd, this is what constitutes the application
7711 * visible backing of an io_uring instance. The application mmaps this
7712 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7713 * we have to tie this fd to a socket for file garbage collection purposes.
7715 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7720 #if defined(CONFIG_UNIX)
7721 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7727 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7731 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7732 O_RDWR | O_CLOEXEC);
7735 ret = PTR_ERR(file);
7739 #if defined(CONFIG_UNIX)
7740 ctx->ring_sock->file = file;
7742 fd_install(ret, file);
7745 #if defined(CONFIG_UNIX)
7746 sock_release(ctx->ring_sock);
7747 ctx->ring_sock = NULL;
7752 static int io_uring_create(unsigned entries, struct io_uring_params *p,
7753 struct io_uring_params __user *params)
7755 struct user_struct *user = NULL;
7756 struct io_ring_ctx *ctx;
7762 if (entries > IORING_MAX_ENTRIES) {
7763 if (!(p->flags & IORING_SETUP_CLAMP))
7765 entries = IORING_MAX_ENTRIES;
7769 * Use twice as many entries for the CQ ring. It's possible for the
7770 * application to drive a higher depth than the size of the SQ ring,
7771 * since the sqes are only used at submission time. This allows for
7772 * some flexibility in overcommitting a bit. If the application has
7773 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7774 * of CQ ring entries manually.
7776 p->sq_entries = roundup_pow_of_two(entries);
7777 if (p->flags & IORING_SETUP_CQSIZE) {
7779 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7780 * to a power-of-two, if it isn't already. We do NOT impose
7781 * any cq vs sq ring sizing.
7783 if (p->cq_entries < p->sq_entries)
7785 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7786 if (!(p->flags & IORING_SETUP_CLAMP))
7788 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7790 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7792 p->cq_entries = 2 * p->sq_entries;
7795 user = get_uid(current_user());
7796 account_mem = !capable(CAP_IPC_LOCK);
7799 ret = io_account_mem(user,
7800 ring_pages(p->sq_entries, p->cq_entries));
7807 ctx = io_ring_ctx_alloc(p);
7810 io_unaccount_mem(user, ring_pages(p->sq_entries,
7815 ctx->compat = in_compat_syscall();
7816 ctx->account_mem = account_mem;
7818 ctx->creds = get_current_cred();
7820 ret = io_allocate_scq_urings(ctx, p);
7824 ret = io_sq_offload_start(ctx, p);
7828 memset(&p->sq_off, 0, sizeof(p->sq_off));
7829 p->sq_off.head = offsetof(struct io_rings, sq.head);
7830 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7831 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7832 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7833 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7834 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7835 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7837 memset(&p->cq_off, 0, sizeof(p->cq_off));
7838 p->cq_off.head = offsetof(struct io_rings, cq.head);
7839 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7840 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7841 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7842 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7843 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7845 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7846 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7847 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7849 if (copy_to_user(params, p, sizeof(*p))) {
7854 * Install ring fd as the very last thing, so we don't risk someone
7855 * having closed it before we finish setup
7857 ret = io_uring_get_fd(ctx);
7861 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7864 io_ring_ctx_wait_and_kill(ctx);
7869 * Sets up an aio uring context, and returns the fd. Applications asks for a
7870 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7871 * params structure passed in.
7873 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7875 struct io_uring_params p;
7878 if (copy_from_user(&p, params, sizeof(p)))
7880 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7885 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7886 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7887 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7890 return io_uring_create(entries, &p, params);
7893 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7894 struct io_uring_params __user *, params)
7896 return io_uring_setup(entries, params);
7899 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7901 struct io_uring_probe *p;
7905 size = struct_size(p, ops, nr_args);
7906 if (size == SIZE_MAX)
7908 p = kzalloc(size, GFP_KERNEL);
7913 if (copy_from_user(p, arg, size))
7916 if (memchr_inv(p, 0, size))
7919 p->last_op = IORING_OP_LAST - 1;
7920 if (nr_args > IORING_OP_LAST)
7921 nr_args = IORING_OP_LAST;
7923 for (i = 0; i < nr_args; i++) {
7925 if (!io_op_defs[i].not_supported)
7926 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7931 if (copy_to_user(arg, p, size))
7938 static int io_register_personality(struct io_ring_ctx *ctx)
7940 const struct cred *creds = get_current_cred();
7943 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7944 USHRT_MAX, GFP_KERNEL);
7950 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7952 const struct cred *old_creds;
7954 old_creds = idr_remove(&ctx->personality_idr, id);
7956 put_cred(old_creds);
7963 static bool io_register_op_must_quiesce(int op)
7966 case IORING_UNREGISTER_FILES:
7967 case IORING_REGISTER_FILES_UPDATE:
7968 case IORING_REGISTER_PROBE:
7969 case IORING_REGISTER_PERSONALITY:
7970 case IORING_UNREGISTER_PERSONALITY:
7977 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7978 void __user *arg, unsigned nr_args)
7979 __releases(ctx->uring_lock)
7980 __acquires(ctx->uring_lock)
7985 * We're inside the ring mutex, if the ref is already dying, then
7986 * someone else killed the ctx or is already going through
7987 * io_uring_register().
7989 if (percpu_ref_is_dying(&ctx->refs))
7992 if (io_register_op_must_quiesce(opcode)) {
7993 percpu_ref_kill(&ctx->refs);
7996 * Drop uring mutex before waiting for references to exit. If
7997 * another thread is currently inside io_uring_enter() it might
7998 * need to grab the uring_lock to make progress. If we hold it
7999 * here across the drain wait, then we can deadlock. It's safe
8000 * to drop the mutex here, since no new references will come in
8001 * after we've killed the percpu ref.
8003 mutex_unlock(&ctx->uring_lock);
8004 ret = wait_for_completion_interruptible(&ctx->completions[0]);
8005 mutex_lock(&ctx->uring_lock);
8007 percpu_ref_resurrect(&ctx->refs);
8014 case IORING_REGISTER_BUFFERS:
8015 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8017 case IORING_UNREGISTER_BUFFERS:
8021 ret = io_sqe_buffer_unregister(ctx);
8023 case IORING_REGISTER_FILES:
8024 ret = io_sqe_files_register(ctx, arg, nr_args);
8026 case IORING_UNREGISTER_FILES:
8030 ret = io_sqe_files_unregister(ctx);
8032 case IORING_REGISTER_FILES_UPDATE:
8033 ret = io_sqe_files_update(ctx, arg, nr_args);
8035 case IORING_REGISTER_EVENTFD:
8036 case IORING_REGISTER_EVENTFD_ASYNC:
8040 ret = io_eventfd_register(ctx, arg);
8043 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8044 ctx->eventfd_async = 1;
8046 ctx->eventfd_async = 0;
8048 case IORING_UNREGISTER_EVENTFD:
8052 ret = io_eventfd_unregister(ctx);
8054 case IORING_REGISTER_PROBE:
8056 if (!arg || nr_args > 256)
8058 ret = io_probe(ctx, arg, nr_args);
8060 case IORING_REGISTER_PERSONALITY:
8064 ret = io_register_personality(ctx);
8066 case IORING_UNREGISTER_PERSONALITY:
8070 ret = io_unregister_personality(ctx, nr_args);
8077 if (io_register_op_must_quiesce(opcode)) {
8078 /* bring the ctx back to life */
8079 percpu_ref_reinit(&ctx->refs);
8081 reinit_completion(&ctx->completions[0]);
8086 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8087 void __user *, arg, unsigned int, nr_args)
8089 struct io_ring_ctx *ctx;
8098 if (f.file->f_op != &io_uring_fops)
8101 ctx = f.file->private_data;
8103 mutex_lock(&ctx->uring_lock);
8104 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8105 mutex_unlock(&ctx->uring_lock);
8106 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8107 ctx->cq_ev_fd != NULL, ret);
8113 static int __init io_uring_init(void)
8115 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8116 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8117 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8120 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8121 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8122 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8123 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8124 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8125 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8126 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8127 BUILD_BUG_SQE_ELEM(8, __u64, off);
8128 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8129 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8130 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8131 BUILD_BUG_SQE_ELEM(24, __u32, len);
8132 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8133 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8134 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8135 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8136 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8137 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8138 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8139 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8140 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8141 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8142 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8143 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8144 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8145 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8146 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8147 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8148 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8149 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8151 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8152 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8153 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8156 __initcall(io_uring_init);