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.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int account_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->poll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head poll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
402 /* NOTE: kiocb has the file as the first member, so don't do it here */
410 struct sockaddr __user *addr;
417 struct user_msghdr __user *msg;
423 struct io_buffer *kbuf;
429 struct filename *filename;
431 unsigned long nofile;
434 struct io_files_update {
460 struct epoll_event event;
464 struct file *file_out;
465 struct file *file_in;
472 struct io_provide_buf {
486 const char __user *filename;
487 struct statx __user *buffer;
490 struct io_async_connect {
491 struct sockaddr_storage address;
494 struct io_async_msghdr {
495 struct iovec fast_iov[UIO_FASTIOV];
497 struct sockaddr __user *uaddr;
499 struct sockaddr_storage addr;
503 struct iovec fast_iov[UIO_FASTIOV];
509 struct io_async_ctx {
511 struct io_async_rw rw;
512 struct io_async_msghdr msg;
513 struct io_async_connect connect;
514 struct io_timeout_data timeout;
519 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
520 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
521 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
522 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
523 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
524 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
532 REQ_F_IOPOLL_COMPLETED_BIT,
533 REQ_F_LINK_TIMEOUT_BIT,
537 REQ_F_TIMEOUT_NOSEQ_BIT,
538 REQ_F_COMP_LOCKED_BIT,
539 REQ_F_NEED_CLEANUP_BIT,
542 REQ_F_BUFFER_SELECTED_BIT,
543 REQ_F_NO_FILE_TABLE_BIT,
545 /* not a real bit, just to check we're not overflowing the space */
551 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
552 /* drain existing IO first */
553 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
555 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
556 /* doesn't sever on completion < 0 */
557 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
559 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
560 /* IOSQE_BUFFER_SELECT */
561 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
564 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
565 /* already grabbed next link */
566 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
567 /* fail rest of links */
568 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
569 /* on inflight list */
570 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
571 /* read/write uses file position */
572 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
573 /* must not punt to workers */
574 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
575 /* polled IO has completed */
576 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
577 /* has linked timeout */
578 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
579 /* timeout request */
580 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
582 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
583 /* must be punted even for NONBLOCK */
584 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
585 /* no timeout sequence */
586 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
587 /* completion under lock */
588 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
590 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
591 /* in overflow list */
592 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
593 /* already went through poll handler */
594 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
595 /* buffer already selected */
596 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
597 /* doesn't need file table for this request */
598 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
602 struct io_poll_iocb poll;
603 struct io_wq_work work;
607 * NOTE! Each of the iocb union members has the file pointer
608 * as the first entry in their struct definition. So you can
609 * access the file pointer through any of the sub-structs,
610 * or directly as just 'ki_filp' in this struct.
616 struct io_poll_iocb poll;
617 struct io_accept accept;
619 struct io_cancel cancel;
620 struct io_timeout timeout;
621 struct io_connect connect;
622 struct io_sr_msg sr_msg;
624 struct io_close close;
625 struct io_files_update files_update;
626 struct io_fadvise fadvise;
627 struct io_madvise madvise;
628 struct io_epoll epoll;
629 struct io_splice splice;
630 struct io_provide_buf pbuf;
631 struct io_statx statx;
634 struct io_async_ctx *io;
640 struct io_ring_ctx *ctx;
641 struct list_head list;
644 struct task_struct *task;
650 struct list_head link_list;
652 struct list_head inflight_entry;
654 struct percpu_ref *fixed_file_refs;
658 * Only commands that never go async can use the below fields,
659 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
660 * async armed poll handlers for regular commands. The latter
661 * restore the work, if needed.
664 struct callback_head task_work;
665 struct hlist_node hash_node;
666 struct async_poll *apoll;
668 struct io_wq_work work;
672 #define IO_PLUG_THRESHOLD 2
673 #define IO_IOPOLL_BATCH 8
675 struct io_submit_state {
676 struct blk_plug plug;
679 * io_kiocb alloc cache
681 void *reqs[IO_IOPOLL_BATCH];
682 unsigned int free_reqs;
685 * File reference cache
689 unsigned int has_refs;
690 unsigned int used_refs;
691 unsigned int ios_left;
695 /* needs req->io allocated for deferral/async */
696 unsigned async_ctx : 1;
697 /* needs current->mm setup, does mm access */
698 unsigned needs_mm : 1;
699 /* needs req->file assigned */
700 unsigned needs_file : 1;
701 /* hash wq insertion if file is a regular file */
702 unsigned hash_reg_file : 1;
703 /* unbound wq insertion if file is a non-regular file */
704 unsigned unbound_nonreg_file : 1;
705 /* opcode is not supported by this kernel */
706 unsigned not_supported : 1;
707 /* needs file table */
708 unsigned file_table : 1;
710 unsigned needs_fs : 1;
711 /* set if opcode supports polled "wait" */
713 unsigned pollout : 1;
714 /* op supports buffer selection */
715 unsigned buffer_select : 1;
718 static const struct io_op_def io_op_defs[] = {
719 [IORING_OP_NOP] = {},
720 [IORING_OP_READV] = {
724 .unbound_nonreg_file = 1,
728 [IORING_OP_WRITEV] = {
733 .unbound_nonreg_file = 1,
736 [IORING_OP_FSYNC] = {
739 [IORING_OP_READ_FIXED] = {
741 .unbound_nonreg_file = 1,
744 [IORING_OP_WRITE_FIXED] = {
747 .unbound_nonreg_file = 1,
750 [IORING_OP_POLL_ADD] = {
752 .unbound_nonreg_file = 1,
754 [IORING_OP_POLL_REMOVE] = {},
755 [IORING_OP_SYNC_FILE_RANGE] = {
758 [IORING_OP_SENDMSG] = {
762 .unbound_nonreg_file = 1,
766 [IORING_OP_RECVMSG] = {
770 .unbound_nonreg_file = 1,
775 [IORING_OP_TIMEOUT] = {
779 [IORING_OP_TIMEOUT_REMOVE] = {},
780 [IORING_OP_ACCEPT] = {
783 .unbound_nonreg_file = 1,
787 [IORING_OP_ASYNC_CANCEL] = {},
788 [IORING_OP_LINK_TIMEOUT] = {
792 [IORING_OP_CONNECT] = {
796 .unbound_nonreg_file = 1,
799 [IORING_OP_FALLOCATE] = {
802 [IORING_OP_OPENAT] = {
806 [IORING_OP_CLOSE] = {
809 [IORING_OP_FILES_UPDATE] = {
813 [IORING_OP_STATX] = {
821 .unbound_nonreg_file = 1,
825 [IORING_OP_WRITE] = {
828 .unbound_nonreg_file = 1,
831 [IORING_OP_FADVISE] = {
834 [IORING_OP_MADVISE] = {
840 .unbound_nonreg_file = 1,
846 .unbound_nonreg_file = 1,
850 [IORING_OP_OPENAT2] = {
854 [IORING_OP_EPOLL_CTL] = {
855 .unbound_nonreg_file = 1,
858 [IORING_OP_SPLICE] = {
861 .unbound_nonreg_file = 1,
863 [IORING_OP_PROVIDE_BUFFERS] = {},
864 [IORING_OP_REMOVE_BUFFERS] = {},
868 .unbound_nonreg_file = 1,
872 static void io_wq_submit_work(struct io_wq_work **workptr);
873 static void io_cqring_fill_event(struct io_kiocb *req, long res);
874 static void io_put_req(struct io_kiocb *req);
875 static void __io_double_put_req(struct io_kiocb *req);
876 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
877 static void io_queue_linked_timeout(struct io_kiocb *req);
878 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
879 struct io_uring_files_update *ip,
881 static int io_grab_files(struct io_kiocb *req);
882 static void io_cleanup_req(struct io_kiocb *req);
883 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
884 int fd, struct file **out_file, bool fixed);
885 static void __io_queue_sqe(struct io_kiocb *req,
886 const struct io_uring_sqe *sqe);
888 static struct kmem_cache *req_cachep;
890 static const struct file_operations io_uring_fops;
892 struct sock *io_uring_get_socket(struct file *file)
894 #if defined(CONFIG_UNIX)
895 if (file->f_op == &io_uring_fops) {
896 struct io_ring_ctx *ctx = file->private_data;
898 return ctx->ring_sock->sk;
903 EXPORT_SYMBOL(io_uring_get_socket);
905 static void io_file_put_work(struct work_struct *work);
907 static inline bool io_async_submit(struct io_ring_ctx *ctx)
909 return ctx->flags & IORING_SETUP_SQPOLL;
912 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
914 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
916 complete(&ctx->ref_comp);
919 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
921 struct io_ring_ctx *ctx;
924 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
928 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
929 if (!ctx->fallback_req)
933 * Use 5 bits less than the max cq entries, that should give us around
934 * 32 entries per hash list if totally full and uniformly spread.
936 hash_bits = ilog2(p->cq_entries);
940 ctx->cancel_hash_bits = hash_bits;
941 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
943 if (!ctx->cancel_hash)
945 __hash_init(ctx->cancel_hash, 1U << hash_bits);
947 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
948 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
951 ctx->flags = p->flags;
952 init_waitqueue_head(&ctx->sqo_wait);
953 init_waitqueue_head(&ctx->cq_wait);
954 INIT_LIST_HEAD(&ctx->cq_overflow_list);
955 init_completion(&ctx->ref_comp);
956 init_completion(&ctx->sq_thread_comp);
957 idr_init(&ctx->io_buffer_idr);
958 idr_init(&ctx->personality_idr);
959 mutex_init(&ctx->uring_lock);
960 init_waitqueue_head(&ctx->wait);
961 spin_lock_init(&ctx->completion_lock);
962 INIT_LIST_HEAD(&ctx->poll_list);
963 INIT_LIST_HEAD(&ctx->defer_list);
964 INIT_LIST_HEAD(&ctx->timeout_list);
965 init_waitqueue_head(&ctx->inflight_wait);
966 spin_lock_init(&ctx->inflight_lock);
967 INIT_LIST_HEAD(&ctx->inflight_list);
968 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
969 init_llist_head(&ctx->file_put_llist);
972 if (ctx->fallback_req)
973 kmem_cache_free(req_cachep, ctx->fallback_req);
974 kfree(ctx->cancel_hash);
979 static inline bool __req_need_defer(struct io_kiocb *req)
981 struct io_ring_ctx *ctx = req->ctx;
983 return req->sequence != ctx->cached_cq_tail
984 + atomic_read(&ctx->cached_cq_overflow);
987 static inline bool req_need_defer(struct io_kiocb *req)
989 if (unlikely(req->flags & REQ_F_IO_DRAIN))
990 return __req_need_defer(req);
995 static void __io_commit_cqring(struct io_ring_ctx *ctx)
997 struct io_rings *rings = ctx->rings;
999 /* order cqe stores with ring update */
1000 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1002 if (wq_has_sleeper(&ctx->cq_wait)) {
1003 wake_up_interruptible(&ctx->cq_wait);
1004 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1008 static inline void io_req_work_grab_env(struct io_kiocb *req,
1009 const struct io_op_def *def)
1011 if (!req->work.mm && def->needs_mm) {
1012 mmgrab(current->mm);
1013 req->work.mm = current->mm;
1015 if (!req->work.creds)
1016 req->work.creds = get_current_cred();
1017 if (!req->work.fs && def->needs_fs) {
1018 spin_lock(¤t->fs->lock);
1019 if (!current->fs->in_exec) {
1020 req->work.fs = current->fs;
1021 req->work.fs->users++;
1023 req->work.flags |= IO_WQ_WORK_CANCEL;
1025 spin_unlock(¤t->fs->lock);
1027 if (!req->work.task_pid)
1028 req->work.task_pid = task_pid_vnr(current);
1031 static inline void io_req_work_drop_env(struct io_kiocb *req)
1034 mmdrop(req->work.mm);
1035 req->work.mm = NULL;
1037 if (req->work.creds) {
1038 put_cred(req->work.creds);
1039 req->work.creds = NULL;
1042 struct fs_struct *fs = req->work.fs;
1044 spin_lock(&req->work.fs->lock);
1047 spin_unlock(&req->work.fs->lock);
1053 static inline void io_prep_async_work(struct io_kiocb *req,
1054 struct io_kiocb **link)
1056 const struct io_op_def *def = &io_op_defs[req->opcode];
1058 if (req->flags & REQ_F_ISREG) {
1059 if (def->hash_reg_file)
1060 io_wq_hash_work(&req->work, file_inode(req->file));
1062 if (def->unbound_nonreg_file)
1063 req->work.flags |= IO_WQ_WORK_UNBOUND;
1066 io_req_work_grab_env(req, def);
1068 *link = io_prep_linked_timeout(req);
1071 static inline void io_queue_async_work(struct io_kiocb *req)
1073 struct io_ring_ctx *ctx = req->ctx;
1074 struct io_kiocb *link;
1076 io_prep_async_work(req, &link);
1078 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1079 &req->work, req->flags);
1080 io_wq_enqueue(ctx->io_wq, &req->work);
1083 io_queue_linked_timeout(link);
1086 static void io_kill_timeout(struct io_kiocb *req)
1090 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1092 atomic_inc(&req->ctx->cq_timeouts);
1093 list_del_init(&req->list);
1094 req->flags |= REQ_F_COMP_LOCKED;
1095 io_cqring_fill_event(req, 0);
1100 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1102 struct io_kiocb *req, *tmp;
1104 spin_lock_irq(&ctx->completion_lock);
1105 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1106 io_kill_timeout(req);
1107 spin_unlock_irq(&ctx->completion_lock);
1110 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1113 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1114 struct io_kiocb, list);
1116 if (req_need_defer(req))
1118 list_del_init(&req->list);
1119 io_queue_async_work(req);
1120 } while (!list_empty(&ctx->defer_list));
1123 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1125 while (!list_empty(&ctx->timeout_list)) {
1126 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1127 struct io_kiocb, list);
1129 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
1131 if (req->timeout.target_seq != ctx->cached_cq_tail
1132 - atomic_read(&ctx->cq_timeouts))
1135 list_del_init(&req->list);
1136 io_kill_timeout(req);
1140 static void io_commit_cqring(struct io_ring_ctx *ctx)
1142 io_flush_timeouts(ctx);
1143 __io_commit_cqring(ctx);
1145 if (unlikely(!list_empty(&ctx->defer_list)))
1146 __io_queue_deferred(ctx);
1149 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1151 struct io_rings *rings = ctx->rings;
1154 tail = ctx->cached_cq_tail;
1156 * writes to the cq entry need to come after reading head; the
1157 * control dependency is enough as we're using WRITE_ONCE to
1160 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1163 ctx->cached_cq_tail++;
1164 return &rings->cqes[tail & ctx->cq_mask];
1167 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1171 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1173 if (!ctx->eventfd_async)
1175 return io_wq_current_is_worker();
1178 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1180 if (waitqueue_active(&ctx->wait))
1181 wake_up(&ctx->wait);
1182 if (waitqueue_active(&ctx->sqo_wait))
1183 wake_up(&ctx->sqo_wait);
1184 if (io_should_trigger_evfd(ctx))
1185 eventfd_signal(ctx->cq_ev_fd, 1);
1188 /* Returns true if there are no backlogged entries after the flush */
1189 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1191 struct io_rings *rings = ctx->rings;
1192 struct io_uring_cqe *cqe;
1193 struct io_kiocb *req;
1194 unsigned long flags;
1198 if (list_empty_careful(&ctx->cq_overflow_list))
1200 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1201 rings->cq_ring_entries))
1205 spin_lock_irqsave(&ctx->completion_lock, flags);
1207 /* if force is set, the ring is going away. always drop after that */
1209 ctx->cq_overflow_flushed = 1;
1212 while (!list_empty(&ctx->cq_overflow_list)) {
1213 cqe = io_get_cqring(ctx);
1217 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1219 list_move(&req->list, &list);
1220 req->flags &= ~REQ_F_OVERFLOW;
1222 WRITE_ONCE(cqe->user_data, req->user_data);
1223 WRITE_ONCE(cqe->res, req->result);
1224 WRITE_ONCE(cqe->flags, req->cflags);
1226 WRITE_ONCE(ctx->rings->cq_overflow,
1227 atomic_inc_return(&ctx->cached_cq_overflow));
1231 io_commit_cqring(ctx);
1233 clear_bit(0, &ctx->sq_check_overflow);
1234 clear_bit(0, &ctx->cq_check_overflow);
1236 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1237 io_cqring_ev_posted(ctx);
1239 while (!list_empty(&list)) {
1240 req = list_first_entry(&list, struct io_kiocb, list);
1241 list_del(&req->list);
1248 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1250 struct io_ring_ctx *ctx = req->ctx;
1251 struct io_uring_cqe *cqe;
1253 trace_io_uring_complete(ctx, req->user_data, res);
1256 * If we can't get a cq entry, userspace overflowed the
1257 * submission (by quite a lot). Increment the overflow count in
1260 cqe = io_get_cqring(ctx);
1262 WRITE_ONCE(cqe->user_data, req->user_data);
1263 WRITE_ONCE(cqe->res, res);
1264 WRITE_ONCE(cqe->flags, cflags);
1265 } else if (ctx->cq_overflow_flushed) {
1266 WRITE_ONCE(ctx->rings->cq_overflow,
1267 atomic_inc_return(&ctx->cached_cq_overflow));
1269 if (list_empty(&ctx->cq_overflow_list)) {
1270 set_bit(0, &ctx->sq_check_overflow);
1271 set_bit(0, &ctx->cq_check_overflow);
1273 req->flags |= REQ_F_OVERFLOW;
1274 refcount_inc(&req->refs);
1276 req->cflags = cflags;
1277 list_add_tail(&req->list, &ctx->cq_overflow_list);
1281 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1283 __io_cqring_fill_event(req, res, 0);
1286 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1288 struct io_ring_ctx *ctx = req->ctx;
1289 unsigned long flags;
1291 spin_lock_irqsave(&ctx->completion_lock, flags);
1292 __io_cqring_fill_event(req, res, cflags);
1293 io_commit_cqring(ctx);
1294 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1296 io_cqring_ev_posted(ctx);
1299 static void io_cqring_add_event(struct io_kiocb *req, long res)
1301 __io_cqring_add_event(req, res, 0);
1304 static inline bool io_is_fallback_req(struct io_kiocb *req)
1306 return req == (struct io_kiocb *)
1307 ((unsigned long) req->ctx->fallback_req & ~1UL);
1310 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1312 struct io_kiocb *req;
1314 req = ctx->fallback_req;
1315 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1321 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1322 struct io_submit_state *state)
1324 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1325 struct io_kiocb *req;
1328 req = kmem_cache_alloc(req_cachep, gfp);
1331 } else if (!state->free_reqs) {
1335 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1336 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1339 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1340 * retry single alloc to be on the safe side.
1342 if (unlikely(ret <= 0)) {
1343 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1344 if (!state->reqs[0])
1348 state->free_reqs = ret - 1;
1349 req = state->reqs[ret - 1];
1352 req = state->reqs[state->free_reqs];
1357 return io_get_fallback_req(ctx);
1360 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1364 percpu_ref_put(req->fixed_file_refs);
1369 static void __io_req_aux_free(struct io_kiocb *req)
1371 if (req->flags & REQ_F_NEED_CLEANUP)
1372 io_cleanup_req(req);
1376 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1378 put_task_struct(req->task);
1380 io_req_work_drop_env(req);
1383 static void __io_free_req(struct io_kiocb *req)
1385 __io_req_aux_free(req);
1387 if (req->flags & REQ_F_INFLIGHT) {
1388 struct io_ring_ctx *ctx = req->ctx;
1389 unsigned long flags;
1391 spin_lock_irqsave(&ctx->inflight_lock, flags);
1392 list_del(&req->inflight_entry);
1393 if (waitqueue_active(&ctx->inflight_wait))
1394 wake_up(&ctx->inflight_wait);
1395 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1398 percpu_ref_put(&req->ctx->refs);
1399 if (likely(!io_is_fallback_req(req)))
1400 kmem_cache_free(req_cachep, req);
1402 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1406 void *reqs[IO_IOPOLL_BATCH];
1411 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1415 if (rb->need_iter) {
1416 int i, inflight = 0;
1417 unsigned long flags;
1419 for (i = 0; i < rb->to_free; i++) {
1420 struct io_kiocb *req = rb->reqs[i];
1422 if (req->flags & REQ_F_INFLIGHT)
1424 __io_req_aux_free(req);
1429 spin_lock_irqsave(&ctx->inflight_lock, flags);
1430 for (i = 0; i < rb->to_free; i++) {
1431 struct io_kiocb *req = rb->reqs[i];
1433 if (req->flags & REQ_F_INFLIGHT) {
1434 list_del(&req->inflight_entry);
1439 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1441 if (waitqueue_active(&ctx->inflight_wait))
1442 wake_up(&ctx->inflight_wait);
1445 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1446 percpu_ref_put_many(&ctx->refs, rb->to_free);
1447 rb->to_free = rb->need_iter = 0;
1450 static bool io_link_cancel_timeout(struct io_kiocb *req)
1452 struct io_ring_ctx *ctx = req->ctx;
1455 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1457 io_cqring_fill_event(req, -ECANCELED);
1458 io_commit_cqring(ctx);
1459 req->flags &= ~REQ_F_LINK_HEAD;
1467 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1469 struct io_ring_ctx *ctx = req->ctx;
1470 bool wake_ev = false;
1472 /* Already got next link */
1473 if (req->flags & REQ_F_LINK_NEXT)
1477 * The list should never be empty when we are called here. But could
1478 * potentially happen if the chain is messed up, check to be on the
1481 while (!list_empty(&req->link_list)) {
1482 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1483 struct io_kiocb, link_list);
1485 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1486 (nxt->flags & REQ_F_TIMEOUT))) {
1487 list_del_init(&nxt->link_list);
1488 wake_ev |= io_link_cancel_timeout(nxt);
1489 req->flags &= ~REQ_F_LINK_TIMEOUT;
1493 list_del_init(&req->link_list);
1494 if (!list_empty(&nxt->link_list))
1495 nxt->flags |= REQ_F_LINK_HEAD;
1500 req->flags |= REQ_F_LINK_NEXT;
1502 io_cqring_ev_posted(ctx);
1506 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1508 static void io_fail_links(struct io_kiocb *req)
1510 struct io_ring_ctx *ctx = req->ctx;
1511 unsigned long flags;
1513 spin_lock_irqsave(&ctx->completion_lock, flags);
1515 while (!list_empty(&req->link_list)) {
1516 struct io_kiocb *link = list_first_entry(&req->link_list,
1517 struct io_kiocb, link_list);
1519 list_del_init(&link->link_list);
1520 trace_io_uring_fail_link(req, link);
1522 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1523 link->opcode == IORING_OP_LINK_TIMEOUT) {
1524 io_link_cancel_timeout(link);
1526 io_cqring_fill_event(link, -ECANCELED);
1527 __io_double_put_req(link);
1529 req->flags &= ~REQ_F_LINK_TIMEOUT;
1532 io_commit_cqring(ctx);
1533 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1534 io_cqring_ev_posted(ctx);
1537 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1539 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1543 * If LINK is set, we have dependent requests in this chain. If we
1544 * didn't fail this request, queue the first one up, moving any other
1545 * dependencies to the next request. In case of failure, fail the rest
1548 if (req->flags & REQ_F_FAIL_LINK) {
1550 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1551 REQ_F_LINK_TIMEOUT) {
1552 struct io_ring_ctx *ctx = req->ctx;
1553 unsigned long flags;
1556 * If this is a timeout link, we could be racing with the
1557 * timeout timer. Grab the completion lock for this case to
1558 * protect against that.
1560 spin_lock_irqsave(&ctx->completion_lock, flags);
1561 io_req_link_next(req, nxt);
1562 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1564 io_req_link_next(req, nxt);
1568 static void io_free_req(struct io_kiocb *req)
1570 struct io_kiocb *nxt = NULL;
1572 io_req_find_next(req, &nxt);
1576 io_queue_async_work(nxt);
1579 static void io_link_work_cb(struct io_wq_work **workptr)
1581 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1582 struct io_kiocb *link;
1584 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1585 io_queue_linked_timeout(link);
1586 io_wq_submit_work(workptr);
1589 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1591 struct io_kiocb *link;
1592 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1594 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1595 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1597 *workptr = &nxt->work;
1598 link = io_prep_linked_timeout(nxt);
1600 nxt->work.func = io_link_work_cb;
1604 * Drop reference to request, return next in chain (if there is one) if this
1605 * was the last reference to this request.
1607 __attribute__((nonnull))
1608 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1610 if (refcount_dec_and_test(&req->refs)) {
1611 io_req_find_next(req, nxtptr);
1616 static void io_put_req(struct io_kiocb *req)
1618 if (refcount_dec_and_test(&req->refs))
1622 static void io_steal_work(struct io_kiocb *req,
1623 struct io_wq_work **workptr)
1626 * It's in an io-wq worker, so there always should be at least
1627 * one reference, which will be dropped in io_put_work() just
1628 * after the current handler returns.
1630 * It also means, that if the counter dropped to 1, then there is
1631 * no asynchronous users left, so it's safe to steal the next work.
1633 if (refcount_read(&req->refs) == 1) {
1634 struct io_kiocb *nxt = NULL;
1636 io_req_find_next(req, &nxt);
1638 io_wq_assign_next(workptr, nxt);
1643 * Must only be used if we don't need to care about links, usually from
1644 * within the completion handling itself.
1646 static void __io_double_put_req(struct io_kiocb *req)
1648 /* drop both submit and complete references */
1649 if (refcount_sub_and_test(2, &req->refs))
1653 static void io_double_put_req(struct io_kiocb *req)
1655 /* drop both submit and complete references */
1656 if (refcount_sub_and_test(2, &req->refs))
1660 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1662 struct io_rings *rings = ctx->rings;
1664 if (test_bit(0, &ctx->cq_check_overflow)) {
1666 * noflush == true is from the waitqueue handler, just ensure
1667 * we wake up the task, and the next invocation will flush the
1668 * entries. We cannot safely to it from here.
1670 if (noflush && !list_empty(&ctx->cq_overflow_list))
1673 io_cqring_overflow_flush(ctx, false);
1676 /* See comment at the top of this file */
1678 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1681 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1683 struct io_rings *rings = ctx->rings;
1685 /* make sure SQ entry isn't read before tail */
1686 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1689 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1691 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1694 if (req->file || req->io)
1697 rb->reqs[rb->to_free++] = req;
1698 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1699 io_free_req_many(req->ctx, rb);
1703 static int io_put_kbuf(struct io_kiocb *req)
1705 struct io_buffer *kbuf;
1708 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1709 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1710 cflags |= IORING_CQE_F_BUFFER;
1717 * Find and free completed poll iocbs
1719 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1720 struct list_head *done)
1722 struct req_batch rb;
1723 struct io_kiocb *req;
1725 rb.to_free = rb.need_iter = 0;
1726 while (!list_empty(done)) {
1729 req = list_first_entry(done, struct io_kiocb, list);
1730 list_del(&req->list);
1732 if (req->flags & REQ_F_BUFFER_SELECTED)
1733 cflags = io_put_kbuf(req);
1735 __io_cqring_fill_event(req, req->result, cflags);
1738 if (refcount_dec_and_test(&req->refs) &&
1739 !io_req_multi_free(&rb, req))
1743 io_commit_cqring(ctx);
1744 if (ctx->flags & IORING_SETUP_SQPOLL)
1745 io_cqring_ev_posted(ctx);
1746 io_free_req_many(ctx, &rb);
1749 static void io_iopoll_queue(struct list_head *again)
1751 struct io_kiocb *req;
1754 req = list_first_entry(again, struct io_kiocb, list);
1755 list_del(&req->list);
1756 refcount_inc(&req->refs);
1757 io_queue_async_work(req);
1758 } while (!list_empty(again));
1761 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1764 struct io_kiocb *req, *tmp;
1771 * Only spin for completions if we don't have multiple devices hanging
1772 * off our complete list, and we're under the requested amount.
1774 spin = !ctx->poll_multi_file && *nr_events < min;
1777 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1778 struct kiocb *kiocb = &req->rw.kiocb;
1781 * Move completed and retryable entries to our local lists.
1782 * If we find a request that requires polling, break out
1783 * and complete those lists first, if we have entries there.
1785 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1786 list_move_tail(&req->list, &done);
1789 if (!list_empty(&done))
1792 if (req->result == -EAGAIN) {
1793 list_move_tail(&req->list, &again);
1796 if (!list_empty(&again))
1799 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1808 if (!list_empty(&done))
1809 io_iopoll_complete(ctx, nr_events, &done);
1811 if (!list_empty(&again))
1812 io_iopoll_queue(&again);
1818 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1819 * non-spinning poll check - we'll still enter the driver poll loop, but only
1820 * as a non-spinning completion check.
1822 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1825 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1828 ret = io_do_iopoll(ctx, nr_events, min);
1831 if (!min || *nr_events >= min)
1839 * We can't just wait for polled events to come to us, we have to actively
1840 * find and complete them.
1842 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1844 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1847 mutex_lock(&ctx->uring_lock);
1848 while (!list_empty(&ctx->poll_list)) {
1849 unsigned int nr_events = 0;
1851 io_iopoll_getevents(ctx, &nr_events, 1);
1854 * Ensure we allow local-to-the-cpu processing to take place,
1855 * in this case we need to ensure that we reap all events.
1859 mutex_unlock(&ctx->uring_lock);
1862 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1865 int iters = 0, ret = 0;
1868 * We disallow the app entering submit/complete with polling, but we
1869 * still need to lock the ring to prevent racing with polled issue
1870 * that got punted to a workqueue.
1872 mutex_lock(&ctx->uring_lock);
1877 * Don't enter poll loop if we already have events pending.
1878 * If we do, we can potentially be spinning for commands that
1879 * already triggered a CQE (eg in error).
1881 if (io_cqring_events(ctx, false))
1885 * If a submit got punted to a workqueue, we can have the
1886 * application entering polling for a command before it gets
1887 * issued. That app will hold the uring_lock for the duration
1888 * of the poll right here, so we need to take a breather every
1889 * now and then to ensure that the issue has a chance to add
1890 * the poll to the issued list. Otherwise we can spin here
1891 * forever, while the workqueue is stuck trying to acquire the
1894 if (!(++iters & 7)) {
1895 mutex_unlock(&ctx->uring_lock);
1896 mutex_lock(&ctx->uring_lock);
1899 if (*nr_events < min)
1900 tmin = min - *nr_events;
1902 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1906 } while (min && !*nr_events && !need_resched());
1908 mutex_unlock(&ctx->uring_lock);
1912 static void kiocb_end_write(struct io_kiocb *req)
1915 * Tell lockdep we inherited freeze protection from submission
1918 if (req->flags & REQ_F_ISREG) {
1919 struct inode *inode = file_inode(req->file);
1921 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1923 file_end_write(req->file);
1926 static inline void req_set_fail_links(struct io_kiocb *req)
1928 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1929 req->flags |= REQ_F_FAIL_LINK;
1932 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1934 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);
1942 if (req->flags & REQ_F_BUFFER_SELECTED)
1943 cflags = io_put_kbuf(req);
1944 __io_cqring_add_event(req, res, cflags);
1947 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1949 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1951 io_complete_rw_common(kiocb, res);
1955 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1957 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1959 if (kiocb->ki_flags & IOCB_WRITE)
1960 kiocb_end_write(req);
1962 if (res != req->result)
1963 req_set_fail_links(req);
1966 req->flags |= REQ_F_IOPOLL_COMPLETED;
1970 * After the iocb has been issued, it's safe to be found on the poll list.
1971 * Adding the kiocb to the list AFTER submission ensures that we don't
1972 * find it from a io_iopoll_getevents() thread before the issuer is done
1973 * accessing the kiocb cookie.
1975 static void io_iopoll_req_issued(struct io_kiocb *req)
1977 struct io_ring_ctx *ctx = req->ctx;
1980 * Track whether we have multiple files in our lists. This will impact
1981 * how we do polling eventually, not spinning if we're on potentially
1982 * different devices.
1984 if (list_empty(&ctx->poll_list)) {
1985 ctx->poll_multi_file = false;
1986 } else if (!ctx->poll_multi_file) {
1987 struct io_kiocb *list_req;
1989 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1991 if (list_req->file != req->file)
1992 ctx->poll_multi_file = true;
1996 * For fast devices, IO may have already completed. If it has, add
1997 * it to the front so we find it first.
1999 if (req->flags & REQ_F_IOPOLL_COMPLETED)
2000 list_add(&req->list, &ctx->poll_list);
2002 list_add_tail(&req->list, &ctx->poll_list);
2004 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2005 wq_has_sleeper(&ctx->sqo_wait))
2006 wake_up(&ctx->sqo_wait);
2009 static void __io_state_file_put(struct io_submit_state *state)
2011 int diff = state->has_refs - state->used_refs;
2014 fput_many(state->file, diff);
2018 static inline void io_state_file_put(struct io_submit_state *state)
2021 __io_state_file_put(state);
2025 * Get as many references to a file as we have IOs left in this submission,
2026 * assuming most submissions are for one file, or at least that each file
2027 * has more than one submission.
2029 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2035 if (state->fd == fd) {
2040 __io_state_file_put(state);
2042 state->file = fget_many(fd, state->ios_left);
2047 state->has_refs = state->ios_left;
2048 state->used_refs = 1;
2054 * If we tracked the file through the SCM inflight mechanism, we could support
2055 * any file. For now, just ensure that anything potentially problematic is done
2058 static bool io_file_supports_async(struct file *file, int rw)
2060 umode_t mode = file_inode(file)->i_mode;
2062 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2064 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2067 if (!(file->f_mode & FMODE_NOWAIT))
2071 return file->f_op->read_iter != NULL;
2073 return file->f_op->write_iter != NULL;
2076 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2077 bool force_nonblock)
2079 struct io_ring_ctx *ctx = req->ctx;
2080 struct kiocb *kiocb = &req->rw.kiocb;
2084 if (S_ISREG(file_inode(req->file)->i_mode))
2085 req->flags |= REQ_F_ISREG;
2087 kiocb->ki_pos = READ_ONCE(sqe->off);
2088 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2089 req->flags |= REQ_F_CUR_POS;
2090 kiocb->ki_pos = req->file->f_pos;
2092 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2093 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2094 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2098 ioprio = READ_ONCE(sqe->ioprio);
2100 ret = ioprio_check_cap(ioprio);
2104 kiocb->ki_ioprio = ioprio;
2106 kiocb->ki_ioprio = get_current_ioprio();
2108 /* don't allow async punt if RWF_NOWAIT was requested */
2109 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2110 (req->file->f_flags & O_NONBLOCK))
2111 req->flags |= REQ_F_NOWAIT;
2114 kiocb->ki_flags |= IOCB_NOWAIT;
2116 if (ctx->flags & IORING_SETUP_IOPOLL) {
2117 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2118 !kiocb->ki_filp->f_op->iopoll)
2121 kiocb->ki_flags |= IOCB_HIPRI;
2122 kiocb->ki_complete = io_complete_rw_iopoll;
2125 if (kiocb->ki_flags & IOCB_HIPRI)
2127 kiocb->ki_complete = io_complete_rw;
2130 req->rw.addr = READ_ONCE(sqe->addr);
2131 req->rw.len = READ_ONCE(sqe->len);
2132 req->buf_index = READ_ONCE(sqe->buf_index);
2136 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2142 case -ERESTARTNOINTR:
2143 case -ERESTARTNOHAND:
2144 case -ERESTART_RESTARTBLOCK:
2146 * We can't just restart the syscall, since previously
2147 * submitted sqes may already be in progress. Just fail this
2153 kiocb->ki_complete(kiocb, ret, 0);
2157 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2159 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2161 if (req->flags & REQ_F_CUR_POS)
2162 req->file->f_pos = kiocb->ki_pos;
2163 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2164 io_complete_rw(kiocb, ret, 0);
2166 io_rw_done(kiocb, ret);
2169 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2170 struct iov_iter *iter)
2172 struct io_ring_ctx *ctx = req->ctx;
2173 size_t len = req->rw.len;
2174 struct io_mapped_ubuf *imu;
2175 u16 index, buf_index;
2179 /* attempt to use fixed buffers without having provided iovecs */
2180 if (unlikely(!ctx->user_bufs))
2183 buf_index = req->buf_index;
2184 if (unlikely(buf_index >= ctx->nr_user_bufs))
2187 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2188 imu = &ctx->user_bufs[index];
2189 buf_addr = req->rw.addr;
2192 if (buf_addr + len < buf_addr)
2194 /* not inside the mapped region */
2195 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2199 * May not be a start of buffer, set size appropriately
2200 * and advance us to the beginning.
2202 offset = buf_addr - imu->ubuf;
2203 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2207 * Don't use iov_iter_advance() here, as it's really slow for
2208 * using the latter parts of a big fixed buffer - it iterates
2209 * over each segment manually. We can cheat a bit here, because
2212 * 1) it's a BVEC iter, we set it up
2213 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2214 * first and last bvec
2216 * So just find our index, and adjust the iterator afterwards.
2217 * If the offset is within the first bvec (or the whole first
2218 * bvec, just use iov_iter_advance(). This makes it easier
2219 * since we can just skip the first segment, which may not
2220 * be PAGE_SIZE aligned.
2222 const struct bio_vec *bvec = imu->bvec;
2224 if (offset <= bvec->bv_len) {
2225 iov_iter_advance(iter, offset);
2227 unsigned long seg_skip;
2229 /* skip first vec */
2230 offset -= bvec->bv_len;
2231 seg_skip = 1 + (offset >> PAGE_SHIFT);
2233 iter->bvec = bvec + seg_skip;
2234 iter->nr_segs -= seg_skip;
2235 iter->count -= bvec->bv_len + offset;
2236 iter->iov_offset = offset & ~PAGE_MASK;
2243 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2246 mutex_unlock(&ctx->uring_lock);
2249 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2252 * "Normal" inline submissions always hold the uring_lock, since we
2253 * grab it from the system call. Same is true for the SQPOLL offload.
2254 * The only exception is when we've detached the request and issue it
2255 * from an async worker thread, grab the lock for that case.
2258 mutex_lock(&ctx->uring_lock);
2261 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2262 int bgid, struct io_buffer *kbuf,
2265 struct io_buffer *head;
2267 if (req->flags & REQ_F_BUFFER_SELECTED)
2270 io_ring_submit_lock(req->ctx, needs_lock);
2272 lockdep_assert_held(&req->ctx->uring_lock);
2274 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2276 if (!list_empty(&head->list)) {
2277 kbuf = list_last_entry(&head->list, struct io_buffer,
2279 list_del(&kbuf->list);
2282 idr_remove(&req->ctx->io_buffer_idr, bgid);
2284 if (*len > kbuf->len)
2287 kbuf = ERR_PTR(-ENOBUFS);
2290 io_ring_submit_unlock(req->ctx, needs_lock);
2295 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2298 struct io_buffer *kbuf;
2301 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2302 bgid = req->buf_index;
2303 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2306 req->rw.addr = (u64) (unsigned long) kbuf;
2307 req->flags |= REQ_F_BUFFER_SELECTED;
2308 return u64_to_user_ptr(kbuf->addr);
2311 #ifdef CONFIG_COMPAT
2312 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2315 struct compat_iovec __user *uiov;
2316 compat_ssize_t clen;
2320 uiov = u64_to_user_ptr(req->rw.addr);
2321 if (!access_ok(uiov, sizeof(*uiov)))
2323 if (__get_user(clen, &uiov->iov_len))
2329 buf = io_rw_buffer_select(req, &len, needs_lock);
2331 return PTR_ERR(buf);
2332 iov[0].iov_base = buf;
2333 iov[0].iov_len = (compat_size_t) len;
2338 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2341 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2345 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2348 len = iov[0].iov_len;
2351 buf = io_rw_buffer_select(req, &len, needs_lock);
2353 return PTR_ERR(buf);
2354 iov[0].iov_base = buf;
2355 iov[0].iov_len = len;
2359 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2362 if (req->flags & REQ_F_BUFFER_SELECTED)
2366 else if (req->rw.len > 1)
2369 #ifdef CONFIG_COMPAT
2370 if (req->ctx->compat)
2371 return io_compat_import(req, iov, needs_lock);
2374 return __io_iov_buffer_select(req, iov, needs_lock);
2377 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2378 struct iovec **iovec, struct iov_iter *iter,
2381 void __user *buf = u64_to_user_ptr(req->rw.addr);
2382 size_t sqe_len = req->rw.len;
2386 opcode = req->opcode;
2387 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2389 return io_import_fixed(req, rw, iter);
2392 /* buffer index only valid with fixed read/write, or buffer select */
2393 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2396 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2397 if (req->flags & REQ_F_BUFFER_SELECT) {
2398 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2401 return PTR_ERR(buf);
2403 req->rw.len = sqe_len;
2406 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2408 return ret < 0 ? ret : sqe_len;
2412 struct io_async_rw *iorw = &req->io->rw;
2415 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2416 if (iorw->iov == iorw->fast_iov)
2421 if (req->flags & REQ_F_BUFFER_SELECT) {
2422 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2424 ret = (*iovec)->iov_len;
2425 iov_iter_init(iter, rw, *iovec, 1, ret);
2431 #ifdef CONFIG_COMPAT
2432 if (req->ctx->compat)
2433 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2437 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2441 * For files that don't have ->read_iter() and ->write_iter(), handle them
2442 * by looping over ->read() or ->write() manually.
2444 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2445 struct iov_iter *iter)
2450 * Don't support polled IO through this interface, and we can't
2451 * support non-blocking either. For the latter, this just causes
2452 * the kiocb to be handled from an async context.
2454 if (kiocb->ki_flags & IOCB_HIPRI)
2456 if (kiocb->ki_flags & IOCB_NOWAIT)
2459 while (iov_iter_count(iter)) {
2463 if (!iov_iter_is_bvec(iter)) {
2464 iovec = iov_iter_iovec(iter);
2466 /* fixed buffers import bvec */
2467 iovec.iov_base = kmap(iter->bvec->bv_page)
2469 iovec.iov_len = min(iter->count,
2470 iter->bvec->bv_len - iter->iov_offset);
2474 nr = file->f_op->read(file, iovec.iov_base,
2475 iovec.iov_len, &kiocb->ki_pos);
2477 nr = file->f_op->write(file, iovec.iov_base,
2478 iovec.iov_len, &kiocb->ki_pos);
2481 if (iov_iter_is_bvec(iter))
2482 kunmap(iter->bvec->bv_page);
2490 if (nr != iovec.iov_len)
2492 iov_iter_advance(iter, nr);
2498 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2499 struct iovec *iovec, struct iovec *fast_iov,
2500 struct iov_iter *iter)
2502 req->io->rw.nr_segs = iter->nr_segs;
2503 req->io->rw.size = io_size;
2504 req->io->rw.iov = iovec;
2505 if (!req->io->rw.iov) {
2506 req->io->rw.iov = req->io->rw.fast_iov;
2507 if (req->io->rw.iov != fast_iov)
2508 memcpy(req->io->rw.iov, fast_iov,
2509 sizeof(struct iovec) * iter->nr_segs);
2511 req->flags |= REQ_F_NEED_CLEANUP;
2515 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2517 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2518 return req->io == NULL;
2521 static int io_alloc_async_ctx(struct io_kiocb *req)
2523 if (!io_op_defs[req->opcode].async_ctx)
2526 return __io_alloc_async_ctx(req);
2529 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2530 struct iovec *iovec, struct iovec *fast_iov,
2531 struct iov_iter *iter)
2533 if (!io_op_defs[req->opcode].async_ctx)
2536 if (__io_alloc_async_ctx(req))
2539 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2544 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2545 bool force_nonblock)
2547 struct io_async_ctx *io;
2548 struct iov_iter iter;
2551 ret = io_prep_rw(req, sqe, force_nonblock);
2555 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2558 /* either don't need iovec imported or already have it */
2559 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2563 io->rw.iov = io->rw.fast_iov;
2565 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2570 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2574 static int io_read(struct io_kiocb *req, bool force_nonblock)
2576 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2577 struct kiocb *kiocb = &req->rw.kiocb;
2578 struct iov_iter iter;
2580 ssize_t io_size, ret;
2582 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2586 /* Ensure we clear previously set non-block flag */
2587 if (!force_nonblock)
2588 kiocb->ki_flags &= ~IOCB_NOWAIT;
2592 if (req->flags & REQ_F_LINK_HEAD)
2593 req->result = io_size;
2596 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2597 * we know to async punt it even if it was opened O_NONBLOCK
2599 if (force_nonblock && !io_file_supports_async(req->file, READ))
2602 iov_count = iov_iter_count(&iter);
2603 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2607 if (req->file->f_op->read_iter)
2608 ret2 = call_read_iter(req->file, kiocb, &iter);
2610 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2612 /* Catch -EAGAIN return for forced non-blocking submission */
2613 if (!force_nonblock || ret2 != -EAGAIN) {
2614 kiocb_done(kiocb, ret2);
2617 ret = io_setup_async_rw(req, io_size, iovec,
2618 inline_vecs, &iter);
2621 /* any defer here is final, must blocking retry */
2622 if (!(req->flags & REQ_F_NOWAIT) &&
2623 !file_can_poll(req->file))
2624 req->flags |= REQ_F_MUST_PUNT;
2630 req->flags &= ~REQ_F_NEED_CLEANUP;
2634 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2635 bool force_nonblock)
2637 struct io_async_ctx *io;
2638 struct iov_iter iter;
2641 ret = io_prep_rw(req, sqe, force_nonblock);
2645 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2648 req->fsize = rlimit(RLIMIT_FSIZE);
2650 /* either don't need iovec imported or already have it */
2651 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2655 io->rw.iov = io->rw.fast_iov;
2657 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2662 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2666 static int io_write(struct io_kiocb *req, bool force_nonblock)
2668 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2669 struct kiocb *kiocb = &req->rw.kiocb;
2670 struct iov_iter iter;
2672 ssize_t ret, io_size;
2674 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2678 /* Ensure we clear previously set non-block flag */
2679 if (!force_nonblock)
2680 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2684 if (req->flags & REQ_F_LINK_HEAD)
2685 req->result = io_size;
2688 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2689 * we know to async punt it even if it was opened O_NONBLOCK
2691 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2694 /* file path doesn't support NOWAIT for non-direct_IO */
2695 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2696 (req->flags & REQ_F_ISREG))
2699 iov_count = iov_iter_count(&iter);
2700 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2705 * Open-code file_start_write here to grab freeze protection,
2706 * which will be released by another thread in
2707 * io_complete_rw(). Fool lockdep by telling it the lock got
2708 * released so that it doesn't complain about the held lock when
2709 * we return to userspace.
2711 if (req->flags & REQ_F_ISREG) {
2712 __sb_start_write(file_inode(req->file)->i_sb,
2713 SB_FREEZE_WRITE, true);
2714 __sb_writers_release(file_inode(req->file)->i_sb,
2717 kiocb->ki_flags |= IOCB_WRITE;
2719 if (!force_nonblock)
2720 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2722 if (req->file->f_op->write_iter)
2723 ret2 = call_write_iter(req->file, kiocb, &iter);
2725 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2727 if (!force_nonblock)
2728 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2731 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2732 * retry them without IOCB_NOWAIT.
2734 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2736 if (!force_nonblock || ret2 != -EAGAIN) {
2737 kiocb_done(kiocb, ret2);
2740 ret = io_setup_async_rw(req, io_size, iovec,
2741 inline_vecs, &iter);
2744 /* any defer here is final, must blocking retry */
2745 if (!file_can_poll(req->file))
2746 req->flags |= REQ_F_MUST_PUNT;
2751 req->flags &= ~REQ_F_NEED_CLEANUP;
2756 static int __io_splice_prep(struct io_kiocb *req,
2757 const struct io_uring_sqe *sqe)
2759 struct io_splice* sp = &req->splice;
2760 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2763 if (req->flags & REQ_F_NEED_CLEANUP)
2767 sp->len = READ_ONCE(sqe->len);
2768 sp->flags = READ_ONCE(sqe->splice_flags);
2770 if (unlikely(sp->flags & ~valid_flags))
2773 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2774 (sp->flags & SPLICE_F_FD_IN_FIXED));
2777 req->flags |= REQ_F_NEED_CLEANUP;
2779 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2780 req->work.flags |= IO_WQ_WORK_UNBOUND;
2785 static int io_tee_prep(struct io_kiocb *req,
2786 const struct io_uring_sqe *sqe)
2788 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
2790 return __io_splice_prep(req, sqe);
2793 static int io_tee(struct io_kiocb *req, bool force_nonblock)
2795 struct io_splice *sp = &req->splice;
2796 struct file *in = sp->file_in;
2797 struct file *out = sp->file_out;
2798 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2804 ret = do_tee(in, out, sp->len, flags);
2806 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2807 req->flags &= ~REQ_F_NEED_CLEANUP;
2809 io_cqring_add_event(req, ret);
2811 req_set_fail_links(req);
2816 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2818 struct io_splice* sp = &req->splice;
2820 sp->off_in = READ_ONCE(sqe->splice_off_in);
2821 sp->off_out = READ_ONCE(sqe->off);
2822 return __io_splice_prep(req, sqe);
2825 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2827 struct io_splice *sp = &req->splice;
2828 struct file *in = sp->file_in;
2829 struct file *out = sp->file_out;
2830 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2831 loff_t *poff_in, *poff_out;
2837 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2838 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2841 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2843 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2844 req->flags &= ~REQ_F_NEED_CLEANUP;
2846 io_cqring_add_event(req, ret);
2848 req_set_fail_links(req);
2854 * IORING_OP_NOP just posts a completion event, nothing else.
2856 static int io_nop(struct io_kiocb *req)
2858 struct io_ring_ctx *ctx = req->ctx;
2860 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2863 io_cqring_add_event(req, 0);
2868 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2870 struct io_ring_ctx *ctx = req->ctx;
2875 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2877 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2880 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2881 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2884 req->sync.off = READ_ONCE(sqe->off);
2885 req->sync.len = READ_ONCE(sqe->len);
2889 static bool io_req_cancelled(struct io_kiocb *req)
2891 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2892 req_set_fail_links(req);
2893 io_cqring_add_event(req, -ECANCELED);
2901 static void __io_fsync(struct io_kiocb *req)
2903 loff_t end = req->sync.off + req->sync.len;
2906 ret = vfs_fsync_range(req->file, req->sync.off,
2907 end > 0 ? end : LLONG_MAX,
2908 req->sync.flags & IORING_FSYNC_DATASYNC);
2910 req_set_fail_links(req);
2911 io_cqring_add_event(req, ret);
2915 static void io_fsync_finish(struct io_wq_work **workptr)
2917 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2919 if (io_req_cancelled(req))
2922 io_steal_work(req, workptr);
2925 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2927 /* fsync always requires a blocking context */
2928 if (force_nonblock) {
2929 req->work.func = io_fsync_finish;
2936 static void __io_fallocate(struct io_kiocb *req)
2940 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2941 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2943 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2945 req_set_fail_links(req);
2946 io_cqring_add_event(req, ret);
2950 static void io_fallocate_finish(struct io_wq_work **workptr)
2952 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2954 if (io_req_cancelled(req))
2956 __io_fallocate(req);
2957 io_steal_work(req, workptr);
2960 static int io_fallocate_prep(struct io_kiocb *req,
2961 const struct io_uring_sqe *sqe)
2963 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2966 req->sync.off = READ_ONCE(sqe->off);
2967 req->sync.len = READ_ONCE(sqe->addr);
2968 req->sync.mode = READ_ONCE(sqe->len);
2969 req->fsize = rlimit(RLIMIT_FSIZE);
2973 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2975 /* fallocate always requiring blocking context */
2976 if (force_nonblock) {
2977 req->work.func = io_fallocate_finish;
2981 __io_fallocate(req);
2985 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2987 const char __user *fname;
2990 if (sqe->ioprio || sqe->buf_index)
2992 if (req->flags & REQ_F_FIXED_FILE)
2994 if (req->flags & REQ_F_NEED_CLEANUP)
2997 req->open.dfd = READ_ONCE(sqe->fd);
2998 req->open.how.mode = READ_ONCE(sqe->len);
2999 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3000 req->open.how.flags = READ_ONCE(sqe->open_flags);
3001 if (force_o_largefile())
3002 req->open.how.flags |= O_LARGEFILE;
3004 req->open.filename = getname(fname);
3005 if (IS_ERR(req->open.filename)) {
3006 ret = PTR_ERR(req->open.filename);
3007 req->open.filename = NULL;
3011 req->open.nofile = rlimit(RLIMIT_NOFILE);
3012 req->flags |= REQ_F_NEED_CLEANUP;
3016 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3018 struct open_how __user *how;
3019 const char __user *fname;
3023 if (sqe->ioprio || sqe->buf_index)
3025 if (req->flags & REQ_F_FIXED_FILE)
3027 if (req->flags & REQ_F_NEED_CLEANUP)
3030 req->open.dfd = READ_ONCE(sqe->fd);
3031 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3032 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3033 len = READ_ONCE(sqe->len);
3035 if (len < OPEN_HOW_SIZE_VER0)
3038 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3043 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3044 req->open.how.flags |= O_LARGEFILE;
3046 req->open.filename = getname(fname);
3047 if (IS_ERR(req->open.filename)) {
3048 ret = PTR_ERR(req->open.filename);
3049 req->open.filename = NULL;
3053 req->open.nofile = rlimit(RLIMIT_NOFILE);
3054 req->flags |= REQ_F_NEED_CLEANUP;
3058 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3060 struct open_flags op;
3067 ret = build_open_flags(&req->open.how, &op);
3071 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3075 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3078 ret = PTR_ERR(file);
3080 fsnotify_open(file);
3081 fd_install(ret, file);
3084 putname(req->open.filename);
3085 req->flags &= ~REQ_F_NEED_CLEANUP;
3087 req_set_fail_links(req);
3088 io_cqring_add_event(req, ret);
3093 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3095 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3096 return io_openat2(req, force_nonblock);
3099 static int io_remove_buffers_prep(struct io_kiocb *req,
3100 const struct io_uring_sqe *sqe)
3102 struct io_provide_buf *p = &req->pbuf;
3105 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3108 tmp = READ_ONCE(sqe->fd);
3109 if (!tmp || tmp > USHRT_MAX)
3112 memset(p, 0, sizeof(*p));
3114 p->bgid = READ_ONCE(sqe->buf_group);
3118 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3119 int bgid, unsigned nbufs)
3123 /* shouldn't happen */
3127 /* the head kbuf is the list itself */
3128 while (!list_empty(&buf->list)) {
3129 struct io_buffer *nxt;
3131 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3132 list_del(&nxt->list);
3139 idr_remove(&ctx->io_buffer_idr, bgid);
3144 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3146 struct io_provide_buf *p = &req->pbuf;
3147 struct io_ring_ctx *ctx = req->ctx;
3148 struct io_buffer *head;
3151 io_ring_submit_lock(ctx, !force_nonblock);
3153 lockdep_assert_held(&ctx->uring_lock);
3156 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3158 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3160 io_ring_submit_lock(ctx, !force_nonblock);
3162 req_set_fail_links(req);
3163 io_cqring_add_event(req, ret);
3168 static int io_provide_buffers_prep(struct io_kiocb *req,
3169 const struct io_uring_sqe *sqe)
3171 struct io_provide_buf *p = &req->pbuf;
3174 if (sqe->ioprio || sqe->rw_flags)
3177 tmp = READ_ONCE(sqe->fd);
3178 if (!tmp || tmp > USHRT_MAX)
3181 p->addr = READ_ONCE(sqe->addr);
3182 p->len = READ_ONCE(sqe->len);
3184 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3187 p->bgid = READ_ONCE(sqe->buf_group);
3188 tmp = READ_ONCE(sqe->off);
3189 if (tmp > USHRT_MAX)
3195 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3197 struct io_buffer *buf;
3198 u64 addr = pbuf->addr;
3199 int i, bid = pbuf->bid;
3201 for (i = 0; i < pbuf->nbufs; i++) {
3202 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3207 buf->len = pbuf->len;
3212 INIT_LIST_HEAD(&buf->list);
3215 list_add_tail(&buf->list, &(*head)->list);
3219 return i ? i : -ENOMEM;
3222 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3224 struct io_provide_buf *p = &req->pbuf;
3225 struct io_ring_ctx *ctx = req->ctx;
3226 struct io_buffer *head, *list;
3229 io_ring_submit_lock(ctx, !force_nonblock);
3231 lockdep_assert_held(&ctx->uring_lock);
3233 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3235 ret = io_add_buffers(p, &head);
3240 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3243 __io_remove_buffers(ctx, head, p->bgid, -1U);
3248 io_ring_submit_unlock(ctx, !force_nonblock);
3250 req_set_fail_links(req);
3251 io_cqring_add_event(req, ret);
3256 static int io_epoll_ctl_prep(struct io_kiocb *req,
3257 const struct io_uring_sqe *sqe)
3259 #if defined(CONFIG_EPOLL)
3260 if (sqe->ioprio || sqe->buf_index)
3263 req->epoll.epfd = READ_ONCE(sqe->fd);
3264 req->epoll.op = READ_ONCE(sqe->len);
3265 req->epoll.fd = READ_ONCE(sqe->off);
3267 if (ep_op_has_event(req->epoll.op)) {
3268 struct epoll_event __user *ev;
3270 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3271 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3281 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3283 #if defined(CONFIG_EPOLL)
3284 struct io_epoll *ie = &req->epoll;
3287 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3288 if (force_nonblock && ret == -EAGAIN)
3292 req_set_fail_links(req);
3293 io_cqring_add_event(req, ret);
3301 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3303 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3304 if (sqe->ioprio || sqe->buf_index || sqe->off)
3307 req->madvise.addr = READ_ONCE(sqe->addr);
3308 req->madvise.len = READ_ONCE(sqe->len);
3309 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3316 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3318 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3319 struct io_madvise *ma = &req->madvise;
3325 ret = do_madvise(ma->addr, ma->len, ma->advice);
3327 req_set_fail_links(req);
3328 io_cqring_add_event(req, ret);
3336 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3338 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3341 req->fadvise.offset = READ_ONCE(sqe->off);
3342 req->fadvise.len = READ_ONCE(sqe->len);
3343 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3347 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3349 struct io_fadvise *fa = &req->fadvise;
3352 if (force_nonblock) {
3353 switch (fa->advice) {
3354 case POSIX_FADV_NORMAL:
3355 case POSIX_FADV_RANDOM:
3356 case POSIX_FADV_SEQUENTIAL:
3363 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3365 req_set_fail_links(req);
3366 io_cqring_add_event(req, ret);
3371 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3373 if (sqe->ioprio || sqe->buf_index)
3375 if (req->flags & REQ_F_FIXED_FILE)
3378 req->statx.dfd = READ_ONCE(sqe->fd);
3379 req->statx.mask = READ_ONCE(sqe->len);
3380 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3381 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3382 req->statx.flags = READ_ONCE(sqe->statx_flags);
3387 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3389 struct io_statx *ctx = &req->statx;
3392 if (force_nonblock) {
3393 /* only need file table for an actual valid fd */
3394 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3395 req->flags |= REQ_F_NO_FILE_TABLE;
3399 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3403 req_set_fail_links(req);
3404 io_cqring_add_event(req, ret);
3409 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3412 * If we queue this for async, it must not be cancellable. That would
3413 * leave the 'file' in an undeterminate state.
3415 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3417 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3418 sqe->rw_flags || sqe->buf_index)
3420 if (req->flags & REQ_F_FIXED_FILE)
3423 req->close.fd = READ_ONCE(sqe->fd);
3427 /* only called when __close_fd_get_file() is done */
3428 static void __io_close_finish(struct io_kiocb *req)
3432 ret = filp_close(req->close.put_file, req->work.files);
3434 req_set_fail_links(req);
3435 io_cqring_add_event(req, ret);
3436 fput(req->close.put_file);
3440 static void io_close_finish(struct io_wq_work **workptr)
3442 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3444 /* not cancellable, don't do io_req_cancelled() */
3445 __io_close_finish(req);
3446 io_steal_work(req, workptr);
3449 static int io_close(struct io_kiocb *req, bool force_nonblock)
3453 req->close.put_file = NULL;
3454 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3456 return (ret == -ENOENT) ? -EBADF : ret;
3458 /* if the file has a flush method, be safe and punt to async */
3459 if (req->close.put_file->f_op->flush && force_nonblock) {
3460 /* avoid grabbing files - we don't need the files */
3461 req->flags |= REQ_F_NO_FILE_TABLE | REQ_F_MUST_PUNT;
3462 req->work.func = io_close_finish;
3467 * No ->flush(), safely close from here and just punt the
3468 * fput() to async context.
3470 __io_close_finish(req);
3474 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3476 struct io_ring_ctx *ctx = req->ctx;
3481 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3483 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3486 req->sync.off = READ_ONCE(sqe->off);
3487 req->sync.len = READ_ONCE(sqe->len);
3488 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3492 static void __io_sync_file_range(struct io_kiocb *req)
3496 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3499 req_set_fail_links(req);
3500 io_cqring_add_event(req, ret);
3505 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3507 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3509 if (io_req_cancelled(req))
3511 __io_sync_file_range(req);
3512 io_steal_work(req, workptr);
3515 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3517 /* sync_file_range always requires a blocking context */
3518 if (force_nonblock) {
3519 req->work.func = io_sync_file_range_finish;
3523 __io_sync_file_range(req);
3527 #if defined(CONFIG_NET)
3528 static int io_setup_async_msg(struct io_kiocb *req,
3529 struct io_async_msghdr *kmsg)
3533 if (io_alloc_async_ctx(req)) {
3534 if (kmsg->iov != kmsg->fast_iov)
3538 req->flags |= REQ_F_NEED_CLEANUP;
3539 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3543 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3545 struct io_sr_msg *sr = &req->sr_msg;
3546 struct io_async_ctx *io = req->io;
3549 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3550 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3551 sr->len = READ_ONCE(sqe->len);
3553 #ifdef CONFIG_COMPAT
3554 if (req->ctx->compat)
3555 sr->msg_flags |= MSG_CMSG_COMPAT;
3558 if (!io || req->opcode == IORING_OP_SEND)
3560 /* iovec is already imported */
3561 if (req->flags & REQ_F_NEED_CLEANUP)
3564 io->msg.iov = io->msg.fast_iov;
3565 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3568 req->flags |= REQ_F_NEED_CLEANUP;
3572 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3574 struct io_async_msghdr *kmsg = NULL;
3575 struct socket *sock;
3578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3581 sock = sock_from_file(req->file, &ret);
3583 struct io_async_ctx io;
3587 kmsg = &req->io->msg;
3588 kmsg->msg.msg_name = &req->io->msg.addr;
3589 /* if iov is set, it's allocated already */
3591 kmsg->iov = kmsg->fast_iov;
3592 kmsg->msg.msg_iter.iov = kmsg->iov;
3594 struct io_sr_msg *sr = &req->sr_msg;
3597 kmsg->msg.msg_name = &io.msg.addr;
3599 io.msg.iov = io.msg.fast_iov;
3600 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3601 sr->msg_flags, &io.msg.iov);
3606 flags = req->sr_msg.msg_flags;
3607 if (flags & MSG_DONTWAIT)
3608 req->flags |= REQ_F_NOWAIT;
3609 else if (force_nonblock)
3610 flags |= MSG_DONTWAIT;
3612 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3613 if (force_nonblock && ret == -EAGAIN)
3614 return io_setup_async_msg(req, kmsg);
3615 if (ret == -ERESTARTSYS)
3619 if (kmsg && kmsg->iov != kmsg->fast_iov)
3621 req->flags &= ~REQ_F_NEED_CLEANUP;
3622 io_cqring_add_event(req, ret);
3624 req_set_fail_links(req);
3629 static int io_send(struct io_kiocb *req, bool force_nonblock)
3631 struct socket *sock;
3634 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3637 sock = sock_from_file(req->file, &ret);
3639 struct io_sr_msg *sr = &req->sr_msg;
3644 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3649 msg.msg_name = NULL;
3650 msg.msg_control = NULL;
3651 msg.msg_controllen = 0;
3652 msg.msg_namelen = 0;
3654 flags = req->sr_msg.msg_flags;
3655 if (flags & MSG_DONTWAIT)
3656 req->flags |= REQ_F_NOWAIT;
3657 else if (force_nonblock)
3658 flags |= MSG_DONTWAIT;
3660 msg.msg_flags = flags;
3661 ret = sock_sendmsg(sock, &msg);
3662 if (force_nonblock && ret == -EAGAIN)
3664 if (ret == -ERESTARTSYS)
3668 io_cqring_add_event(req, ret);
3670 req_set_fail_links(req);
3675 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3677 struct io_sr_msg *sr = &req->sr_msg;
3678 struct iovec __user *uiov;
3682 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3687 if (req->flags & REQ_F_BUFFER_SELECT) {
3690 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3692 sr->len = io->msg.iov[0].iov_len;
3693 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3697 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3698 &io->msg.iov, &io->msg.msg.msg_iter);
3706 #ifdef CONFIG_COMPAT
3707 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3708 struct io_async_ctx *io)
3710 struct compat_msghdr __user *msg_compat;
3711 struct io_sr_msg *sr = &req->sr_msg;
3712 struct compat_iovec __user *uiov;
3717 msg_compat = (struct compat_msghdr __user *) sr->msg;
3718 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3723 uiov = compat_ptr(ptr);
3724 if (req->flags & REQ_F_BUFFER_SELECT) {
3725 compat_ssize_t clen;
3729 if (!access_ok(uiov, sizeof(*uiov)))
3731 if (__get_user(clen, &uiov->iov_len))
3735 sr->len = io->msg.iov[0].iov_len;
3738 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3740 &io->msg.msg.msg_iter);
3749 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3751 io->msg.iov = io->msg.fast_iov;
3753 #ifdef CONFIG_COMPAT
3754 if (req->ctx->compat)
3755 return __io_compat_recvmsg_copy_hdr(req, io);
3758 return __io_recvmsg_copy_hdr(req, io);
3761 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3762 int *cflags, bool needs_lock)
3764 struct io_sr_msg *sr = &req->sr_msg;
3765 struct io_buffer *kbuf;
3767 if (!(req->flags & REQ_F_BUFFER_SELECT))
3770 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3775 req->flags |= REQ_F_BUFFER_SELECTED;
3777 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3778 *cflags |= IORING_CQE_F_BUFFER;
3782 static int io_recvmsg_prep(struct io_kiocb *req,
3783 const struct io_uring_sqe *sqe)
3785 struct io_sr_msg *sr = &req->sr_msg;
3786 struct io_async_ctx *io = req->io;
3789 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3790 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3791 sr->len = READ_ONCE(sqe->len);
3792 sr->bgid = READ_ONCE(sqe->buf_group);
3794 #ifdef CONFIG_COMPAT
3795 if (req->ctx->compat)
3796 sr->msg_flags |= MSG_CMSG_COMPAT;
3799 if (!io || req->opcode == IORING_OP_RECV)
3801 /* iovec is already imported */
3802 if (req->flags & REQ_F_NEED_CLEANUP)
3805 ret = io_recvmsg_copy_hdr(req, io);
3807 req->flags |= REQ_F_NEED_CLEANUP;
3811 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3813 struct io_async_msghdr *kmsg = NULL;
3814 struct socket *sock;
3815 int ret, cflags = 0;
3817 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3820 sock = sock_from_file(req->file, &ret);
3822 struct io_buffer *kbuf;
3823 struct io_async_ctx io;
3827 kmsg = &req->io->msg;
3828 kmsg->msg.msg_name = &req->io->msg.addr;
3829 /* if iov is set, it's allocated already */
3831 kmsg->iov = kmsg->fast_iov;
3832 kmsg->msg.msg_iter.iov = kmsg->iov;
3835 kmsg->msg.msg_name = &io.msg.addr;
3837 ret = io_recvmsg_copy_hdr(req, &io);
3842 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3844 return PTR_ERR(kbuf);
3846 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3847 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3848 1, req->sr_msg.len);
3851 flags = req->sr_msg.msg_flags;
3852 if (flags & MSG_DONTWAIT)
3853 req->flags |= REQ_F_NOWAIT;
3854 else if (force_nonblock)
3855 flags |= MSG_DONTWAIT;
3857 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3858 kmsg->uaddr, flags);
3859 if (force_nonblock && ret == -EAGAIN)
3860 return io_setup_async_msg(req, kmsg);
3861 if (ret == -ERESTARTSYS)
3865 if (kmsg && kmsg->iov != kmsg->fast_iov)
3867 req->flags &= ~REQ_F_NEED_CLEANUP;
3868 __io_cqring_add_event(req, ret, cflags);
3870 req_set_fail_links(req);
3875 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3877 struct io_buffer *kbuf = NULL;
3878 struct socket *sock;
3879 int ret, cflags = 0;
3881 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3884 sock = sock_from_file(req->file, &ret);
3886 struct io_sr_msg *sr = &req->sr_msg;
3887 void __user *buf = sr->buf;
3892 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3894 return PTR_ERR(kbuf);
3896 buf = u64_to_user_ptr(kbuf->addr);
3898 ret = import_single_range(READ, buf, sr->len, &iov,
3905 req->flags |= REQ_F_NEED_CLEANUP;
3906 msg.msg_name = NULL;
3907 msg.msg_control = NULL;
3908 msg.msg_controllen = 0;
3909 msg.msg_namelen = 0;
3910 msg.msg_iocb = NULL;
3913 flags = req->sr_msg.msg_flags;
3914 if (flags & MSG_DONTWAIT)
3915 req->flags |= REQ_F_NOWAIT;
3916 else if (force_nonblock)
3917 flags |= MSG_DONTWAIT;
3919 ret = sock_recvmsg(sock, &msg, flags);
3920 if (force_nonblock && ret == -EAGAIN)
3922 if (ret == -ERESTARTSYS)
3927 req->flags &= ~REQ_F_NEED_CLEANUP;
3928 __io_cqring_add_event(req, ret, cflags);
3930 req_set_fail_links(req);
3935 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3937 struct io_accept *accept = &req->accept;
3939 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3941 if (sqe->ioprio || sqe->len || sqe->buf_index)
3944 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3945 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3946 accept->flags = READ_ONCE(sqe->accept_flags);
3947 accept->nofile = rlimit(RLIMIT_NOFILE);
3951 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3953 struct io_accept *accept = &req->accept;
3954 unsigned file_flags;
3957 file_flags = force_nonblock ? O_NONBLOCK : 0;
3958 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3959 accept->addr_len, accept->flags,
3961 if (ret == -EAGAIN && force_nonblock)
3963 if (ret == -ERESTARTSYS)
3966 req_set_fail_links(req);
3967 io_cqring_add_event(req, ret);
3972 static void io_accept_finish(struct io_wq_work **workptr)
3974 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3976 if (io_req_cancelled(req))
3978 __io_accept(req, false);
3979 io_steal_work(req, workptr);
3982 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3986 ret = __io_accept(req, force_nonblock);
3987 if (ret == -EAGAIN && force_nonblock) {
3988 req->work.func = io_accept_finish;
3994 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3996 struct io_connect *conn = &req->connect;
3997 struct io_async_ctx *io = req->io;
3999 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4001 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4004 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4005 conn->addr_len = READ_ONCE(sqe->addr2);
4010 return move_addr_to_kernel(conn->addr, conn->addr_len,
4011 &io->connect.address);
4014 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4016 struct io_async_ctx __io, *io;
4017 unsigned file_flags;
4023 ret = move_addr_to_kernel(req->connect.addr,
4024 req->connect.addr_len,
4025 &__io.connect.address);
4031 file_flags = force_nonblock ? O_NONBLOCK : 0;
4033 ret = __sys_connect_file(req->file, &io->connect.address,
4034 req->connect.addr_len, file_flags);
4035 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4038 if (io_alloc_async_ctx(req)) {
4042 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4045 if (ret == -ERESTARTSYS)
4049 req_set_fail_links(req);
4050 io_cqring_add_event(req, ret);
4054 #else /* !CONFIG_NET */
4055 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4060 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4065 static int io_send(struct io_kiocb *req, bool force_nonblock)
4070 static int io_recvmsg_prep(struct io_kiocb *req,
4071 const struct io_uring_sqe *sqe)
4076 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4081 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4086 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4091 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4096 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4101 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4105 #endif /* CONFIG_NET */
4107 struct io_poll_table {
4108 struct poll_table_struct pt;
4109 struct io_kiocb *req;
4113 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4114 __poll_t mask, task_work_func_t func)
4116 struct task_struct *tsk;
4119 /* for instances that support it check for an event match first: */
4120 if (mask && !(mask & poll->events))
4123 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4125 list_del_init(&poll->wait.entry);
4129 init_task_work(&req->task_work, func);
4131 * If this fails, then the task is exiting. When a task exits, the
4132 * work gets canceled, so just cancel this request as well instead
4133 * of executing it. We can't safely execute it anyway, as we may not
4134 * have the needed state needed for it anyway.
4136 ret = task_work_add(tsk, &req->task_work, true);
4137 if (unlikely(ret)) {
4138 WRITE_ONCE(poll->canceled, true);
4139 tsk = io_wq_get_task(req->ctx->io_wq);
4140 task_work_add(tsk, &req->task_work, true);
4142 wake_up_process(tsk);
4146 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4147 __acquires(&req->ctx->completion_lock)
4149 struct io_ring_ctx *ctx = req->ctx;
4151 if (!req->result && !READ_ONCE(poll->canceled)) {
4152 struct poll_table_struct pt = { ._key = poll->events };
4154 req->result = vfs_poll(req->file, &pt) & poll->events;
4157 spin_lock_irq(&ctx->completion_lock);
4158 if (!req->result && !READ_ONCE(poll->canceled)) {
4159 add_wait_queue(poll->head, &poll->wait);
4166 static void io_poll_remove_double(struct io_kiocb *req)
4168 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4170 lockdep_assert_held(&req->ctx->completion_lock);
4172 if (poll && poll->head) {
4173 struct wait_queue_head *head = poll->head;
4175 spin_lock(&head->lock);
4176 list_del_init(&poll->wait.entry);
4177 if (poll->wait.private)
4178 refcount_dec(&req->refs);
4180 spin_unlock(&head->lock);
4184 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4186 struct io_ring_ctx *ctx = req->ctx;
4188 io_poll_remove_double(req);
4189 req->poll.done = true;
4190 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4191 io_commit_cqring(ctx);
4194 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4196 struct io_ring_ctx *ctx = req->ctx;
4198 if (io_poll_rewait(req, &req->poll)) {
4199 spin_unlock_irq(&ctx->completion_lock);
4203 hash_del(&req->hash_node);
4204 io_poll_complete(req, req->result, 0);
4205 req->flags |= REQ_F_COMP_LOCKED;
4206 io_put_req_find_next(req, nxt);
4207 spin_unlock_irq(&ctx->completion_lock);
4209 io_cqring_ev_posted(ctx);
4212 static void io_poll_task_func(struct callback_head *cb)
4214 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4215 struct io_kiocb *nxt = NULL;
4217 io_poll_task_handler(req, &nxt);
4219 struct io_ring_ctx *ctx = nxt->ctx;
4221 mutex_lock(&ctx->uring_lock);
4222 __io_queue_sqe(nxt, NULL);
4223 mutex_unlock(&ctx->uring_lock);
4227 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4228 int sync, void *key)
4230 struct io_kiocb *req = wait->private;
4231 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4232 __poll_t mask = key_to_poll(key);
4234 /* for instances that support it check for an event match first: */
4235 if (mask && !(mask & poll->events))
4238 if (req->poll.head) {
4241 spin_lock(&req->poll.head->lock);
4242 done = list_empty(&req->poll.wait.entry);
4244 list_del_init(&req->poll.wait.entry);
4245 spin_unlock(&req->poll.head->lock);
4247 __io_async_wake(req, poll, mask, io_poll_task_func);
4249 refcount_dec(&req->refs);
4253 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4254 wait_queue_func_t wake_func)
4258 poll->canceled = false;
4259 poll->events = events;
4260 INIT_LIST_HEAD(&poll->wait.entry);
4261 init_waitqueue_func_entry(&poll->wait, wake_func);
4264 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4265 struct wait_queue_head *head)
4267 struct io_kiocb *req = pt->req;
4270 * If poll->head is already set, it's because the file being polled
4271 * uses multiple waitqueues for poll handling (eg one for read, one
4272 * for write). Setup a separate io_poll_iocb if this happens.
4274 if (unlikely(poll->head)) {
4275 /* already have a 2nd entry, fail a third attempt */
4277 pt->error = -EINVAL;
4280 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4282 pt->error = -ENOMEM;
4285 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4286 refcount_inc(&req->refs);
4287 poll->wait.private = req;
4288 req->io = (void *) poll;
4293 add_wait_queue(head, &poll->wait);
4296 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4297 struct poll_table_struct *p)
4299 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4301 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4304 static void io_async_task_func(struct callback_head *cb)
4306 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4307 struct async_poll *apoll = req->apoll;
4308 struct io_ring_ctx *ctx = req->ctx;
4309 bool canceled = false;
4311 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4313 if (io_poll_rewait(req, &apoll->poll)) {
4314 spin_unlock_irq(&ctx->completion_lock);
4318 /* If req is still hashed, it cannot have been canceled. Don't check. */
4319 if (hash_hashed(&req->hash_node)) {
4320 hash_del(&req->hash_node);
4322 canceled = READ_ONCE(apoll->poll.canceled);
4324 io_cqring_fill_event(req, -ECANCELED);
4325 io_commit_cqring(ctx);
4329 spin_unlock_irq(&ctx->completion_lock);
4331 /* restore ->work in case we need to retry again */
4332 memcpy(&req->work, &apoll->work, sizeof(req->work));
4336 __set_current_state(TASK_RUNNING);
4337 mutex_lock(&ctx->uring_lock);
4338 __io_queue_sqe(req, NULL);
4339 mutex_unlock(&ctx->uring_lock);
4341 io_cqring_ev_posted(ctx);
4342 req_set_fail_links(req);
4343 io_double_put_req(req);
4347 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4350 struct io_kiocb *req = wait->private;
4351 struct io_poll_iocb *poll = &req->apoll->poll;
4353 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4356 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4359 static void io_poll_req_insert(struct io_kiocb *req)
4361 struct io_ring_ctx *ctx = req->ctx;
4362 struct hlist_head *list;
4364 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4365 hlist_add_head(&req->hash_node, list);
4368 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4369 struct io_poll_iocb *poll,
4370 struct io_poll_table *ipt, __poll_t mask,
4371 wait_queue_func_t wake_func)
4372 __acquires(&ctx->completion_lock)
4374 struct io_ring_ctx *ctx = req->ctx;
4375 bool cancel = false;
4377 poll->file = req->file;
4378 io_init_poll_iocb(poll, mask, wake_func);
4379 poll->wait.private = req;
4381 ipt->pt._key = mask;
4383 ipt->error = -EINVAL;
4385 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4387 spin_lock_irq(&ctx->completion_lock);
4388 if (likely(poll->head)) {
4389 spin_lock(&poll->head->lock);
4390 if (unlikely(list_empty(&poll->wait.entry))) {
4396 if (mask || ipt->error)
4397 list_del_init(&poll->wait.entry);
4399 WRITE_ONCE(poll->canceled, true);
4400 else if (!poll->done) /* actually waiting for an event */
4401 io_poll_req_insert(req);
4402 spin_unlock(&poll->head->lock);
4408 static bool io_arm_poll_handler(struct io_kiocb *req)
4410 const struct io_op_def *def = &io_op_defs[req->opcode];
4411 struct io_ring_ctx *ctx = req->ctx;
4412 struct async_poll *apoll;
4413 struct io_poll_table ipt;
4417 if (!req->file || !file_can_poll(req->file))
4419 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4421 if (!def->pollin && !def->pollout)
4424 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4425 if (unlikely(!apoll))
4428 req->flags |= REQ_F_POLLED;
4429 memcpy(&apoll->work, &req->work, sizeof(req->work));
4430 had_io = req->io != NULL;
4432 get_task_struct(current);
4433 req->task = current;
4435 INIT_HLIST_NODE(&req->hash_node);
4439 mask |= POLLIN | POLLRDNORM;
4441 mask |= POLLOUT | POLLWRNORM;
4442 mask |= POLLERR | POLLPRI;
4444 ipt.pt._qproc = io_async_queue_proc;
4446 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4450 /* only remove double add if we did it here */
4452 io_poll_remove_double(req);
4453 spin_unlock_irq(&ctx->completion_lock);
4454 memcpy(&req->work, &apoll->work, sizeof(req->work));
4458 spin_unlock_irq(&ctx->completion_lock);
4459 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4460 apoll->poll.events);
4464 static bool __io_poll_remove_one(struct io_kiocb *req,
4465 struct io_poll_iocb *poll)
4467 bool do_complete = false;
4469 spin_lock(&poll->head->lock);
4470 WRITE_ONCE(poll->canceled, true);
4471 if (!list_empty(&poll->wait.entry)) {
4472 list_del_init(&poll->wait.entry);
4475 spin_unlock(&poll->head->lock);
4476 hash_del(&req->hash_node);
4480 static bool io_poll_remove_one(struct io_kiocb *req)
4484 if (req->opcode == IORING_OP_POLL_ADD) {
4485 io_poll_remove_double(req);
4486 do_complete = __io_poll_remove_one(req, &req->poll);
4488 struct async_poll *apoll = req->apoll;
4490 /* non-poll requests have submit ref still */
4491 do_complete = __io_poll_remove_one(req, &apoll->poll);
4495 * restore ->work because we will call
4496 * io_req_work_drop_env below when dropping the
4499 memcpy(&req->work, &apoll->work, sizeof(req->work));
4505 io_cqring_fill_event(req, -ECANCELED);
4506 io_commit_cqring(req->ctx);
4507 req->flags |= REQ_F_COMP_LOCKED;
4514 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4516 struct hlist_node *tmp;
4517 struct io_kiocb *req;
4520 spin_lock_irq(&ctx->completion_lock);
4521 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4522 struct hlist_head *list;
4524 list = &ctx->cancel_hash[i];
4525 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4526 posted += io_poll_remove_one(req);
4528 spin_unlock_irq(&ctx->completion_lock);
4531 io_cqring_ev_posted(ctx);
4534 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4536 struct hlist_head *list;
4537 struct io_kiocb *req;
4539 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4540 hlist_for_each_entry(req, list, hash_node) {
4541 if (sqe_addr != req->user_data)
4543 if (io_poll_remove_one(req))
4551 static int io_poll_remove_prep(struct io_kiocb *req,
4552 const struct io_uring_sqe *sqe)
4554 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4556 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4560 req->poll.addr = READ_ONCE(sqe->addr);
4565 * Find a running poll command that matches one specified in sqe->addr,
4566 * and remove it if found.
4568 static int io_poll_remove(struct io_kiocb *req)
4570 struct io_ring_ctx *ctx = req->ctx;
4574 addr = req->poll.addr;
4575 spin_lock_irq(&ctx->completion_lock);
4576 ret = io_poll_cancel(ctx, addr);
4577 spin_unlock_irq(&ctx->completion_lock);
4579 io_cqring_add_event(req, ret);
4581 req_set_fail_links(req);
4586 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4589 struct io_kiocb *req = wait->private;
4590 struct io_poll_iocb *poll = &req->poll;
4592 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4595 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4596 struct poll_table_struct *p)
4598 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4600 __io_queue_proc(&pt->req->poll, pt, head);
4603 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4605 struct io_poll_iocb *poll = &req->poll;
4608 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4610 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4615 events = READ_ONCE(sqe->poll_events);
4616 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4618 get_task_struct(current);
4619 req->task = current;
4623 static int io_poll_add(struct io_kiocb *req)
4625 struct io_poll_iocb *poll = &req->poll;
4626 struct io_ring_ctx *ctx = req->ctx;
4627 struct io_poll_table ipt;
4630 INIT_HLIST_NODE(&req->hash_node);
4631 INIT_LIST_HEAD(&req->list);
4632 ipt.pt._qproc = io_poll_queue_proc;
4634 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4637 if (mask) { /* no async, we'd stolen it */
4639 io_poll_complete(req, mask, 0);
4641 spin_unlock_irq(&ctx->completion_lock);
4644 io_cqring_ev_posted(ctx);
4650 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4652 struct io_timeout_data *data = container_of(timer,
4653 struct io_timeout_data, timer);
4654 struct io_kiocb *req = data->req;
4655 struct io_ring_ctx *ctx = req->ctx;
4656 unsigned long flags;
4658 atomic_inc(&ctx->cq_timeouts);
4660 spin_lock_irqsave(&ctx->completion_lock, flags);
4662 * We could be racing with timeout deletion. If the list is empty,
4663 * then timeout lookup already found it and will be handling it.
4665 if (!list_empty(&req->list))
4666 list_del_init(&req->list);
4668 io_cqring_fill_event(req, -ETIME);
4669 io_commit_cqring(ctx);
4670 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4672 io_cqring_ev_posted(ctx);
4673 req_set_fail_links(req);
4675 return HRTIMER_NORESTART;
4678 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4680 struct io_kiocb *req;
4683 list_for_each_entry(req, &ctx->timeout_list, list) {
4684 if (user_data == req->user_data) {
4685 list_del_init(&req->list);
4694 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4698 req_set_fail_links(req);
4699 io_cqring_fill_event(req, -ECANCELED);
4704 static int io_timeout_remove_prep(struct io_kiocb *req,
4705 const struct io_uring_sqe *sqe)
4707 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4709 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4712 req->timeout.addr = READ_ONCE(sqe->addr);
4713 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4714 if (req->timeout.flags)
4721 * Remove or update an existing timeout command
4723 static int io_timeout_remove(struct io_kiocb *req)
4725 struct io_ring_ctx *ctx = req->ctx;
4728 spin_lock_irq(&ctx->completion_lock);
4729 ret = io_timeout_cancel(ctx, req->timeout.addr);
4731 io_cqring_fill_event(req, ret);
4732 io_commit_cqring(ctx);
4733 spin_unlock_irq(&ctx->completion_lock);
4734 io_cqring_ev_posted(ctx);
4736 req_set_fail_links(req);
4741 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4742 bool is_timeout_link)
4744 struct io_timeout_data *data;
4746 u32 off = READ_ONCE(sqe->off);
4748 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4750 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4752 if (off && is_timeout_link)
4754 flags = READ_ONCE(sqe->timeout_flags);
4755 if (flags & ~IORING_TIMEOUT_ABS)
4758 req->timeout.off = off;
4760 if (!req->io && io_alloc_async_ctx(req))
4763 data = &req->io->timeout;
4765 req->flags |= REQ_F_TIMEOUT;
4767 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4770 if (flags & IORING_TIMEOUT_ABS)
4771 data->mode = HRTIMER_MODE_ABS;
4773 data->mode = HRTIMER_MODE_REL;
4775 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4779 static int io_timeout(struct io_kiocb *req)
4781 struct io_ring_ctx *ctx = req->ctx;
4782 struct io_timeout_data *data = &req->io->timeout;
4783 struct list_head *entry;
4784 u32 tail, off = req->timeout.off;
4786 spin_lock_irq(&ctx->completion_lock);
4789 * sqe->off holds how many events that need to occur for this
4790 * timeout event to be satisfied. If it isn't set, then this is
4791 * a pure timeout request, sequence isn't used.
4794 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4795 entry = ctx->timeout_list.prev;
4799 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
4800 req->timeout.target_seq = tail + off;
4803 * Insertion sort, ensuring the first entry in the list is always
4804 * the one we need first.
4806 list_for_each_prev(entry, &ctx->timeout_list) {
4807 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4809 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4811 /* nxt.seq is behind @tail, otherwise would've been completed */
4812 if (off >= nxt->timeout.target_seq - tail)
4816 list_add(&req->list, entry);
4817 data->timer.function = io_timeout_fn;
4818 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4819 spin_unlock_irq(&ctx->completion_lock);
4823 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4825 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4827 return req->user_data == (unsigned long) data;
4830 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4832 enum io_wq_cancel cancel_ret;
4835 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4836 switch (cancel_ret) {
4837 case IO_WQ_CANCEL_OK:
4840 case IO_WQ_CANCEL_RUNNING:
4843 case IO_WQ_CANCEL_NOTFOUND:
4851 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4852 struct io_kiocb *req, __u64 sqe_addr,
4855 unsigned long flags;
4858 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4859 if (ret != -ENOENT) {
4860 spin_lock_irqsave(&ctx->completion_lock, flags);
4864 spin_lock_irqsave(&ctx->completion_lock, flags);
4865 ret = io_timeout_cancel(ctx, sqe_addr);
4868 ret = io_poll_cancel(ctx, sqe_addr);
4872 io_cqring_fill_event(req, ret);
4873 io_commit_cqring(ctx);
4874 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4875 io_cqring_ev_posted(ctx);
4878 req_set_fail_links(req);
4882 static int io_async_cancel_prep(struct io_kiocb *req,
4883 const struct io_uring_sqe *sqe)
4885 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4887 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4891 req->cancel.addr = READ_ONCE(sqe->addr);
4895 static int io_async_cancel(struct io_kiocb *req)
4897 struct io_ring_ctx *ctx = req->ctx;
4899 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4903 static int io_files_update_prep(struct io_kiocb *req,
4904 const struct io_uring_sqe *sqe)
4906 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4909 req->files_update.offset = READ_ONCE(sqe->off);
4910 req->files_update.nr_args = READ_ONCE(sqe->len);
4911 if (!req->files_update.nr_args)
4913 req->files_update.arg = READ_ONCE(sqe->addr);
4917 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4919 struct io_ring_ctx *ctx = req->ctx;
4920 struct io_uring_files_update up;
4926 up.offset = req->files_update.offset;
4927 up.fds = req->files_update.arg;
4929 mutex_lock(&ctx->uring_lock);
4930 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4931 mutex_unlock(&ctx->uring_lock);
4934 req_set_fail_links(req);
4935 io_cqring_add_event(req, ret);
4940 static int io_req_defer_prep(struct io_kiocb *req,
4941 const struct io_uring_sqe *sqe)
4948 if (io_op_defs[req->opcode].file_table) {
4949 ret = io_grab_files(req);
4954 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4956 switch (req->opcode) {
4959 case IORING_OP_READV:
4960 case IORING_OP_READ_FIXED:
4961 case IORING_OP_READ:
4962 ret = io_read_prep(req, sqe, true);
4964 case IORING_OP_WRITEV:
4965 case IORING_OP_WRITE_FIXED:
4966 case IORING_OP_WRITE:
4967 ret = io_write_prep(req, sqe, true);
4969 case IORING_OP_POLL_ADD:
4970 ret = io_poll_add_prep(req, sqe);
4972 case IORING_OP_POLL_REMOVE:
4973 ret = io_poll_remove_prep(req, sqe);
4975 case IORING_OP_FSYNC:
4976 ret = io_prep_fsync(req, sqe);
4978 case IORING_OP_SYNC_FILE_RANGE:
4979 ret = io_prep_sfr(req, sqe);
4981 case IORING_OP_SENDMSG:
4982 case IORING_OP_SEND:
4983 ret = io_sendmsg_prep(req, sqe);
4985 case IORING_OP_RECVMSG:
4986 case IORING_OP_RECV:
4987 ret = io_recvmsg_prep(req, sqe);
4989 case IORING_OP_CONNECT:
4990 ret = io_connect_prep(req, sqe);
4992 case IORING_OP_TIMEOUT:
4993 ret = io_timeout_prep(req, sqe, false);
4995 case IORING_OP_TIMEOUT_REMOVE:
4996 ret = io_timeout_remove_prep(req, sqe);
4998 case IORING_OP_ASYNC_CANCEL:
4999 ret = io_async_cancel_prep(req, sqe);
5001 case IORING_OP_LINK_TIMEOUT:
5002 ret = io_timeout_prep(req, sqe, true);
5004 case IORING_OP_ACCEPT:
5005 ret = io_accept_prep(req, sqe);
5007 case IORING_OP_FALLOCATE:
5008 ret = io_fallocate_prep(req, sqe);
5010 case IORING_OP_OPENAT:
5011 ret = io_openat_prep(req, sqe);
5013 case IORING_OP_CLOSE:
5014 ret = io_close_prep(req, sqe);
5016 case IORING_OP_FILES_UPDATE:
5017 ret = io_files_update_prep(req, sqe);
5019 case IORING_OP_STATX:
5020 ret = io_statx_prep(req, sqe);
5022 case IORING_OP_FADVISE:
5023 ret = io_fadvise_prep(req, sqe);
5025 case IORING_OP_MADVISE:
5026 ret = io_madvise_prep(req, sqe);
5028 case IORING_OP_OPENAT2:
5029 ret = io_openat2_prep(req, sqe);
5031 case IORING_OP_EPOLL_CTL:
5032 ret = io_epoll_ctl_prep(req, sqe);
5034 case IORING_OP_SPLICE:
5035 ret = io_splice_prep(req, sqe);
5037 case IORING_OP_PROVIDE_BUFFERS:
5038 ret = io_provide_buffers_prep(req, sqe);
5040 case IORING_OP_REMOVE_BUFFERS:
5041 ret = io_remove_buffers_prep(req, sqe);
5044 ret = io_tee_prep(req, sqe);
5047 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5056 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5058 struct io_ring_ctx *ctx = req->ctx;
5061 /* Still need defer if there is pending req in defer list. */
5062 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5066 if (io_alloc_async_ctx(req))
5068 ret = io_req_defer_prep(req, sqe);
5073 spin_lock_irq(&ctx->completion_lock);
5074 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5075 spin_unlock_irq(&ctx->completion_lock);
5079 trace_io_uring_defer(ctx, req, req->user_data);
5080 list_add_tail(&req->list, &ctx->defer_list);
5081 spin_unlock_irq(&ctx->completion_lock);
5082 return -EIOCBQUEUED;
5085 static void io_cleanup_req(struct io_kiocb *req)
5087 struct io_async_ctx *io = req->io;
5089 switch (req->opcode) {
5090 case IORING_OP_READV:
5091 case IORING_OP_READ_FIXED:
5092 case IORING_OP_READ:
5093 if (req->flags & REQ_F_BUFFER_SELECTED)
5094 kfree((void *)(unsigned long)req->rw.addr);
5096 case IORING_OP_WRITEV:
5097 case IORING_OP_WRITE_FIXED:
5098 case IORING_OP_WRITE:
5099 if (io->rw.iov != io->rw.fast_iov)
5102 case IORING_OP_RECVMSG:
5103 if (req->flags & REQ_F_BUFFER_SELECTED)
5104 kfree(req->sr_msg.kbuf);
5106 case IORING_OP_SENDMSG:
5107 if (io->msg.iov != io->msg.fast_iov)
5110 case IORING_OP_RECV:
5111 if (req->flags & REQ_F_BUFFER_SELECTED)
5112 kfree(req->sr_msg.kbuf);
5114 case IORING_OP_OPENAT:
5115 case IORING_OP_OPENAT2:
5117 case IORING_OP_SPLICE:
5119 io_put_file(req, req->splice.file_in,
5120 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5124 req->flags &= ~REQ_F_NEED_CLEANUP;
5127 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5128 bool force_nonblock)
5130 struct io_ring_ctx *ctx = req->ctx;
5133 switch (req->opcode) {
5137 case IORING_OP_READV:
5138 case IORING_OP_READ_FIXED:
5139 case IORING_OP_READ:
5141 ret = io_read_prep(req, sqe, force_nonblock);
5145 ret = io_read(req, force_nonblock);
5147 case IORING_OP_WRITEV:
5148 case IORING_OP_WRITE_FIXED:
5149 case IORING_OP_WRITE:
5151 ret = io_write_prep(req, sqe, force_nonblock);
5155 ret = io_write(req, force_nonblock);
5157 case IORING_OP_FSYNC:
5159 ret = io_prep_fsync(req, sqe);
5163 ret = io_fsync(req, force_nonblock);
5165 case IORING_OP_POLL_ADD:
5167 ret = io_poll_add_prep(req, sqe);
5171 ret = io_poll_add(req);
5173 case IORING_OP_POLL_REMOVE:
5175 ret = io_poll_remove_prep(req, sqe);
5179 ret = io_poll_remove(req);
5181 case IORING_OP_SYNC_FILE_RANGE:
5183 ret = io_prep_sfr(req, sqe);
5187 ret = io_sync_file_range(req, force_nonblock);
5189 case IORING_OP_SENDMSG:
5190 case IORING_OP_SEND:
5192 ret = io_sendmsg_prep(req, sqe);
5196 if (req->opcode == IORING_OP_SENDMSG)
5197 ret = io_sendmsg(req, force_nonblock);
5199 ret = io_send(req, force_nonblock);
5201 case IORING_OP_RECVMSG:
5202 case IORING_OP_RECV:
5204 ret = io_recvmsg_prep(req, sqe);
5208 if (req->opcode == IORING_OP_RECVMSG)
5209 ret = io_recvmsg(req, force_nonblock);
5211 ret = io_recv(req, force_nonblock);
5213 case IORING_OP_TIMEOUT:
5215 ret = io_timeout_prep(req, sqe, false);
5219 ret = io_timeout(req);
5221 case IORING_OP_TIMEOUT_REMOVE:
5223 ret = io_timeout_remove_prep(req, sqe);
5227 ret = io_timeout_remove(req);
5229 case IORING_OP_ACCEPT:
5231 ret = io_accept_prep(req, sqe);
5235 ret = io_accept(req, force_nonblock);
5237 case IORING_OP_CONNECT:
5239 ret = io_connect_prep(req, sqe);
5243 ret = io_connect(req, force_nonblock);
5245 case IORING_OP_ASYNC_CANCEL:
5247 ret = io_async_cancel_prep(req, sqe);
5251 ret = io_async_cancel(req);
5253 case IORING_OP_FALLOCATE:
5255 ret = io_fallocate_prep(req, sqe);
5259 ret = io_fallocate(req, force_nonblock);
5261 case IORING_OP_OPENAT:
5263 ret = io_openat_prep(req, sqe);
5267 ret = io_openat(req, force_nonblock);
5269 case IORING_OP_CLOSE:
5271 ret = io_close_prep(req, sqe);
5275 ret = io_close(req, force_nonblock);
5277 case IORING_OP_FILES_UPDATE:
5279 ret = io_files_update_prep(req, sqe);
5283 ret = io_files_update(req, force_nonblock);
5285 case IORING_OP_STATX:
5287 ret = io_statx_prep(req, sqe);
5291 ret = io_statx(req, force_nonblock);
5293 case IORING_OP_FADVISE:
5295 ret = io_fadvise_prep(req, sqe);
5299 ret = io_fadvise(req, force_nonblock);
5301 case IORING_OP_MADVISE:
5303 ret = io_madvise_prep(req, sqe);
5307 ret = io_madvise(req, force_nonblock);
5309 case IORING_OP_OPENAT2:
5311 ret = io_openat2_prep(req, sqe);
5315 ret = io_openat2(req, force_nonblock);
5317 case IORING_OP_EPOLL_CTL:
5319 ret = io_epoll_ctl_prep(req, sqe);
5323 ret = io_epoll_ctl(req, force_nonblock);
5325 case IORING_OP_SPLICE:
5327 ret = io_splice_prep(req, sqe);
5331 ret = io_splice(req, force_nonblock);
5333 case IORING_OP_PROVIDE_BUFFERS:
5335 ret = io_provide_buffers_prep(req, sqe);
5339 ret = io_provide_buffers(req, force_nonblock);
5341 case IORING_OP_REMOVE_BUFFERS:
5343 ret = io_remove_buffers_prep(req, sqe);
5347 ret = io_remove_buffers(req, force_nonblock);
5351 ret = io_tee_prep(req, sqe);
5355 ret = io_tee(req, force_nonblock);
5365 /* If the op doesn't have a file, we're not polling for it */
5366 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5367 const bool in_async = io_wq_current_is_worker();
5369 if (req->result == -EAGAIN)
5372 /* workqueue context doesn't hold uring_lock, grab it now */
5374 mutex_lock(&ctx->uring_lock);
5376 io_iopoll_req_issued(req);
5379 mutex_unlock(&ctx->uring_lock);
5385 static void io_wq_submit_work(struct io_wq_work **workptr)
5387 struct io_wq_work *work = *workptr;
5388 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5391 /* if NO_CANCEL is set, we must still run the work */
5392 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5393 IO_WQ_WORK_CANCEL) {
5399 ret = io_issue_sqe(req, NULL, false);
5401 * We can get EAGAIN for polled IO even though we're
5402 * forcing a sync submission from here, since we can't
5403 * wait for request slots on the block side.
5412 req_set_fail_links(req);
5413 io_cqring_add_event(req, ret);
5417 io_steal_work(req, workptr);
5420 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5423 struct fixed_file_table *table;
5425 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5426 return table->files[index & IORING_FILE_TABLE_MASK];
5429 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5430 int fd, struct file **out_file, bool fixed)
5432 struct io_ring_ctx *ctx = req->ctx;
5436 if (unlikely(!ctx->file_data ||
5437 (unsigned) fd >= ctx->nr_user_files))
5439 fd = array_index_nospec(fd, ctx->nr_user_files);
5440 file = io_file_from_index(ctx, fd);
5443 req->fixed_file_refs = ctx->file_data->cur_refs;
5444 percpu_ref_get(req->fixed_file_refs);
5446 trace_io_uring_file_get(ctx, fd);
5447 file = __io_file_get(state, fd);
5448 if (unlikely(!file))
5456 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5461 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5462 if (unlikely(!fixed && io_async_submit(req->ctx)))
5465 return io_file_get(state, req, fd, &req->file, fixed);
5468 static int io_grab_files(struct io_kiocb *req)
5471 struct io_ring_ctx *ctx = req->ctx;
5473 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5475 if (!ctx->ring_file)
5479 spin_lock_irq(&ctx->inflight_lock);
5481 * We use the f_ops->flush() handler to ensure that we can flush
5482 * out work accessing these files if the fd is closed. Check if
5483 * the fd has changed since we started down this path, and disallow
5484 * this operation if it has.
5486 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5487 list_add(&req->inflight_entry, &ctx->inflight_list);
5488 req->flags |= REQ_F_INFLIGHT;
5489 req->work.files = current->files;
5492 spin_unlock_irq(&ctx->inflight_lock);
5498 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5500 struct io_timeout_data *data = container_of(timer,
5501 struct io_timeout_data, timer);
5502 struct io_kiocb *req = data->req;
5503 struct io_ring_ctx *ctx = req->ctx;
5504 struct io_kiocb *prev = NULL;
5505 unsigned long flags;
5507 spin_lock_irqsave(&ctx->completion_lock, flags);
5510 * We don't expect the list to be empty, that will only happen if we
5511 * race with the completion of the linked work.
5513 if (!list_empty(&req->link_list)) {
5514 prev = list_entry(req->link_list.prev, struct io_kiocb,
5516 if (refcount_inc_not_zero(&prev->refs)) {
5517 list_del_init(&req->link_list);
5518 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5523 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5526 req_set_fail_links(prev);
5527 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5530 io_cqring_add_event(req, -ETIME);
5533 return HRTIMER_NORESTART;
5536 static void io_queue_linked_timeout(struct io_kiocb *req)
5538 struct io_ring_ctx *ctx = req->ctx;
5541 * If the list is now empty, then our linked request finished before
5542 * we got a chance to setup the timer
5544 spin_lock_irq(&ctx->completion_lock);
5545 if (!list_empty(&req->link_list)) {
5546 struct io_timeout_data *data = &req->io->timeout;
5548 data->timer.function = io_link_timeout_fn;
5549 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5552 spin_unlock_irq(&ctx->completion_lock);
5554 /* drop submission reference */
5558 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5560 struct io_kiocb *nxt;
5562 if (!(req->flags & REQ_F_LINK_HEAD))
5564 /* for polled retry, if flag is set, we already went through here */
5565 if (req->flags & REQ_F_POLLED)
5568 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5570 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5573 req->flags |= REQ_F_LINK_TIMEOUT;
5577 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5579 struct io_kiocb *linked_timeout;
5580 struct io_kiocb *nxt;
5581 const struct cred *old_creds = NULL;
5585 linked_timeout = io_prep_linked_timeout(req);
5587 if (req->work.creds && req->work.creds != current_cred()) {
5589 revert_creds(old_creds);
5590 if (old_creds == req->work.creds)
5591 old_creds = NULL; /* restored original creds */
5593 old_creds = override_creds(req->work.creds);
5596 ret = io_issue_sqe(req, sqe, true);
5599 * We async punt it if the file wasn't marked NOWAIT, or if the file
5600 * doesn't support non-blocking read/write attempts
5602 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5603 (req->flags & REQ_F_MUST_PUNT))) {
5604 if (io_arm_poll_handler(req)) {
5606 io_queue_linked_timeout(linked_timeout);
5610 if (io_op_defs[req->opcode].file_table) {
5611 ret = io_grab_files(req);
5617 * Queued up for async execution, worker will release
5618 * submit reference when the iocb is actually submitted.
5620 io_queue_async_work(req);
5626 /* drop submission reference */
5627 io_put_req_find_next(req, &nxt);
5629 if (linked_timeout) {
5631 io_queue_linked_timeout(linked_timeout);
5633 io_put_req(linked_timeout);
5636 /* and drop final reference, if we failed */
5638 io_cqring_add_event(req, ret);
5639 req_set_fail_links(req);
5645 if (req->flags & REQ_F_FORCE_ASYNC)
5651 revert_creds(old_creds);
5654 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5658 ret = io_req_defer(req, sqe);
5660 if (ret != -EIOCBQUEUED) {
5662 io_cqring_add_event(req, ret);
5663 req_set_fail_links(req);
5664 io_double_put_req(req);
5666 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5669 if (io_alloc_async_ctx(req))
5671 ret = io_req_defer_prep(req, sqe);
5672 if (unlikely(ret < 0))
5677 * Never try inline submit of IOSQE_ASYNC is set, go straight
5678 * to async execution.
5680 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5681 io_queue_async_work(req);
5683 __io_queue_sqe(req, sqe);
5687 static inline void io_queue_link_head(struct io_kiocb *req)
5689 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5690 io_cqring_add_event(req, -ECANCELED);
5691 io_double_put_req(req);
5693 io_queue_sqe(req, NULL);
5696 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5697 struct io_kiocb **link)
5699 struct io_ring_ctx *ctx = req->ctx;
5703 * If we already have a head request, queue this one for async
5704 * submittal once the head completes. If we don't have a head but
5705 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5706 * submitted sync once the chain is complete. If none of those
5707 * conditions are true (normal request), then just queue it.
5710 struct io_kiocb *head = *link;
5713 * Taking sequential execution of a link, draining both sides
5714 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5715 * requests in the link. So, it drains the head and the
5716 * next after the link request. The last one is done via
5717 * drain_next flag to persist the effect across calls.
5719 if (req->flags & REQ_F_IO_DRAIN) {
5720 head->flags |= REQ_F_IO_DRAIN;
5721 ctx->drain_next = 1;
5723 if (io_alloc_async_ctx(req))
5726 ret = io_req_defer_prep(req, sqe);
5728 /* fail even hard links since we don't submit */
5729 head->flags |= REQ_F_FAIL_LINK;
5732 trace_io_uring_link(ctx, req, head);
5733 list_add_tail(&req->link_list, &head->link_list);
5735 /* last request of a link, enqueue the link */
5736 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5737 io_queue_link_head(head);
5741 if (unlikely(ctx->drain_next)) {
5742 req->flags |= REQ_F_IO_DRAIN;
5743 ctx->drain_next = 0;
5745 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5746 req->flags |= REQ_F_LINK_HEAD;
5747 INIT_LIST_HEAD(&req->link_list);
5749 if (io_alloc_async_ctx(req))
5752 ret = io_req_defer_prep(req, sqe);
5754 req->flags |= REQ_F_FAIL_LINK;
5757 io_queue_sqe(req, sqe);
5765 * Batched submission is done, ensure local IO is flushed out.
5767 static void io_submit_state_end(struct io_submit_state *state)
5769 blk_finish_plug(&state->plug);
5770 io_state_file_put(state);
5771 if (state->free_reqs)
5772 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5776 * Start submission side cache.
5778 static void io_submit_state_start(struct io_submit_state *state,
5779 unsigned int max_ios)
5781 blk_start_plug(&state->plug);
5782 state->free_reqs = 0;
5784 state->ios_left = max_ios;
5787 static void io_commit_sqring(struct io_ring_ctx *ctx)
5789 struct io_rings *rings = ctx->rings;
5792 * Ensure any loads from the SQEs are done at this point,
5793 * since once we write the new head, the application could
5794 * write new data to them.
5796 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5800 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5801 * that is mapped by userspace. This means that care needs to be taken to
5802 * ensure that reads are stable, as we cannot rely on userspace always
5803 * being a good citizen. If members of the sqe are validated and then later
5804 * used, it's important that those reads are done through READ_ONCE() to
5805 * prevent a re-load down the line.
5807 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5809 u32 *sq_array = ctx->sq_array;
5813 * The cached sq head (or cq tail) serves two purposes:
5815 * 1) allows us to batch the cost of updating the user visible
5817 * 2) allows the kernel side to track the head on its own, even
5818 * though the application is the one updating it.
5820 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5821 if (likely(head < ctx->sq_entries))
5822 return &ctx->sq_sqes[head];
5824 /* drop invalid entries */
5825 ctx->cached_sq_dropped++;
5826 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5830 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5832 ctx->cached_sq_head++;
5835 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5836 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5837 IOSQE_BUFFER_SELECT)
5839 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5840 const struct io_uring_sqe *sqe,
5841 struct io_submit_state *state)
5843 unsigned int sqe_flags;
5847 * All io need record the previous position, if LINK vs DARIN,
5848 * it can be used to mark the position of the first IO in the
5851 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5852 req->opcode = READ_ONCE(sqe->opcode);
5853 req->user_data = READ_ONCE(sqe->user_data);
5858 /* one is dropped after submission, the other at completion */
5859 refcount_set(&req->refs, 2);
5862 INIT_IO_WORK(&req->work, io_wq_submit_work);
5864 if (unlikely(req->opcode >= IORING_OP_LAST))
5867 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5868 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5870 use_mm(ctx->sqo_mm);
5873 sqe_flags = READ_ONCE(sqe->flags);
5874 /* enforce forwards compatibility on users */
5875 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5878 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5879 !io_op_defs[req->opcode].buffer_select)
5882 id = READ_ONCE(sqe->personality);
5884 req->work.creds = idr_find(&ctx->personality_idr, id);
5885 if (unlikely(!req->work.creds))
5887 get_cred(req->work.creds);
5890 /* same numerical values with corresponding REQ_F_*, safe to copy */
5891 req->flags |= sqe_flags;
5893 if (!io_op_defs[req->opcode].needs_file)
5896 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
5899 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5900 struct file *ring_file, int ring_fd)
5902 struct io_submit_state state, *statep = NULL;
5903 struct io_kiocb *link = NULL;
5904 int i, submitted = 0;
5906 /* if we have a backlog and couldn't flush it all, return BUSY */
5907 if (test_bit(0, &ctx->sq_check_overflow)) {
5908 if (!list_empty(&ctx->cq_overflow_list) &&
5909 !io_cqring_overflow_flush(ctx, false))
5913 /* make sure SQ entry isn't read before tail */
5914 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5916 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5919 if (nr > IO_PLUG_THRESHOLD) {
5920 io_submit_state_start(&state, nr);
5924 ctx->ring_fd = ring_fd;
5925 ctx->ring_file = ring_file;
5927 for (i = 0; i < nr; i++) {
5928 const struct io_uring_sqe *sqe;
5929 struct io_kiocb *req;
5932 sqe = io_get_sqe(ctx);
5933 if (unlikely(!sqe)) {
5934 io_consume_sqe(ctx);
5937 req = io_alloc_req(ctx, statep);
5938 if (unlikely(!req)) {
5940 submitted = -EAGAIN;
5944 err = io_init_req(ctx, req, sqe, statep);
5945 io_consume_sqe(ctx);
5946 /* will complete beyond this point, count as submitted */
5949 if (unlikely(err)) {
5951 io_cqring_add_event(req, err);
5952 io_double_put_req(req);
5956 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5957 true, io_async_submit(ctx));
5958 err = io_submit_sqe(req, sqe, &link);
5963 if (unlikely(submitted != nr)) {
5964 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5966 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5969 io_queue_link_head(link);
5971 io_submit_state_end(&state);
5973 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5974 io_commit_sqring(ctx);
5979 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5981 struct mm_struct *mm = current->mm;
5989 static int io_sq_thread(void *data)
5991 struct io_ring_ctx *ctx = data;
5992 const struct cred *old_cred;
5993 mm_segment_t old_fs;
5995 unsigned long timeout;
5998 complete(&ctx->sq_thread_comp);
6002 old_cred = override_creds(ctx->creds);
6004 timeout = jiffies + ctx->sq_thread_idle;
6005 while (!kthread_should_park()) {
6006 unsigned int to_submit;
6008 if (!list_empty(&ctx->poll_list)) {
6009 unsigned nr_events = 0;
6011 mutex_lock(&ctx->uring_lock);
6012 if (!list_empty(&ctx->poll_list))
6013 io_iopoll_getevents(ctx, &nr_events, 0);
6015 timeout = jiffies + ctx->sq_thread_idle;
6016 mutex_unlock(&ctx->uring_lock);
6019 to_submit = io_sqring_entries(ctx);
6022 * If submit got -EBUSY, flag us as needing the application
6023 * to enter the kernel to reap and flush events.
6025 if (!to_submit || ret == -EBUSY) {
6027 * Drop cur_mm before scheduling, we can't hold it for
6028 * long periods (or over schedule()). Do this before
6029 * adding ourselves to the waitqueue, as the unuse/drop
6032 io_sq_thread_drop_mm(ctx);
6035 * We're polling. If we're within the defined idle
6036 * period, then let us spin without work before going
6037 * to sleep. The exception is if we got EBUSY doing
6038 * more IO, we should wait for the application to
6039 * reap events and wake us up.
6041 if (!list_empty(&ctx->poll_list) ||
6042 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6043 !percpu_ref_is_dying(&ctx->refs))) {
6044 if (current->task_works)
6050 prepare_to_wait(&ctx->sqo_wait, &wait,
6051 TASK_INTERRUPTIBLE);
6054 * While doing polled IO, before going to sleep, we need
6055 * to check if there are new reqs added to poll_list, it
6056 * is because reqs may have been punted to io worker and
6057 * will be added to poll_list later, hence check the
6060 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6061 !list_empty_careful(&ctx->poll_list)) {
6062 finish_wait(&ctx->sqo_wait, &wait);
6066 /* Tell userspace we may need a wakeup call */
6067 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6068 /* make sure to read SQ tail after writing flags */
6071 to_submit = io_sqring_entries(ctx);
6072 if (!to_submit || ret == -EBUSY) {
6073 if (kthread_should_park()) {
6074 finish_wait(&ctx->sqo_wait, &wait);
6077 if (current->task_works) {
6079 finish_wait(&ctx->sqo_wait, &wait);
6082 if (signal_pending(current))
6083 flush_signals(current);
6085 finish_wait(&ctx->sqo_wait, &wait);
6087 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6091 finish_wait(&ctx->sqo_wait, &wait);
6093 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6096 mutex_lock(&ctx->uring_lock);
6097 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6098 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6099 mutex_unlock(&ctx->uring_lock);
6100 timeout = jiffies + ctx->sq_thread_idle;
6103 if (current->task_works)
6107 io_sq_thread_drop_mm(ctx);
6108 revert_creds(old_cred);
6115 struct io_wait_queue {
6116 struct wait_queue_entry wq;
6117 struct io_ring_ctx *ctx;
6119 unsigned nr_timeouts;
6122 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6124 struct io_ring_ctx *ctx = iowq->ctx;
6127 * Wake up if we have enough events, or if a timeout occurred since we
6128 * started waiting. For timeouts, we always want to return to userspace,
6129 * regardless of event count.
6131 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6132 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6135 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6136 int wake_flags, void *key)
6138 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6141 /* use noflush == true, as we can't safely rely on locking context */
6142 if (!io_should_wake(iowq, true))
6145 return autoremove_wake_function(curr, mode, wake_flags, key);
6149 * Wait until events become available, if we don't already have some. The
6150 * application must reap them itself, as they reside on the shared cq ring.
6152 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6153 const sigset_t __user *sig, size_t sigsz)
6155 struct io_wait_queue iowq = {
6158 .func = io_wake_function,
6159 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6162 .to_wait = min_events,
6164 struct io_rings *rings = ctx->rings;
6168 if (io_cqring_events(ctx, false) >= min_events)
6170 if (!current->task_works)
6176 #ifdef CONFIG_COMPAT
6177 if (in_compat_syscall())
6178 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6182 ret = set_user_sigmask(sig, sigsz);
6188 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6189 trace_io_uring_cqring_wait(ctx, min_events);
6191 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6192 TASK_INTERRUPTIBLE);
6193 if (current->task_works)
6195 if (io_should_wake(&iowq, false))
6198 if (signal_pending(current)) {
6203 finish_wait(&ctx->wait, &iowq.wq);
6205 restore_saved_sigmask_unless(ret == -EINTR);
6207 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6210 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6212 #if defined(CONFIG_UNIX)
6213 if (ctx->ring_sock) {
6214 struct sock *sock = ctx->ring_sock->sk;
6215 struct sk_buff *skb;
6217 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6223 for (i = 0; i < ctx->nr_user_files; i++) {
6226 file = io_file_from_index(ctx, i);
6233 static void io_file_ref_kill(struct percpu_ref *ref)
6235 struct fixed_file_data *data;
6237 data = container_of(ref, struct fixed_file_data, refs);
6238 complete(&data->done);
6241 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6243 struct fixed_file_data *data = ctx->file_data;
6244 struct fixed_file_ref_node *ref_node = NULL;
6245 unsigned nr_tables, i;
6250 spin_lock(&data->lock);
6251 if (!list_empty(&data->ref_list))
6252 ref_node = list_first_entry(&data->ref_list,
6253 struct fixed_file_ref_node, node);
6254 spin_unlock(&data->lock);
6256 percpu_ref_kill(&ref_node->refs);
6258 percpu_ref_kill(&data->refs);
6260 /* wait for all refs nodes to complete */
6261 flush_delayed_work(&ctx->file_put_work);
6262 wait_for_completion(&data->done);
6264 __io_sqe_files_unregister(ctx);
6265 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6266 for (i = 0; i < nr_tables; i++)
6267 kfree(data->table[i].files);
6269 percpu_ref_exit(&data->refs);
6271 ctx->file_data = NULL;
6272 ctx->nr_user_files = 0;
6276 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6278 if (ctx->sqo_thread) {
6279 wait_for_completion(&ctx->sq_thread_comp);
6281 * The park is a bit of a work-around, without it we get
6282 * warning spews on shutdown with SQPOLL set and affinity
6283 * set to a single CPU.
6285 kthread_park(ctx->sqo_thread);
6286 kthread_stop(ctx->sqo_thread);
6287 ctx->sqo_thread = NULL;
6291 static void io_finish_async(struct io_ring_ctx *ctx)
6293 io_sq_thread_stop(ctx);
6296 io_wq_destroy(ctx->io_wq);
6301 #if defined(CONFIG_UNIX)
6303 * Ensure the UNIX gc is aware of our file set, so we are certain that
6304 * the io_uring can be safely unregistered on process exit, even if we have
6305 * loops in the file referencing.
6307 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6309 struct sock *sk = ctx->ring_sock->sk;
6310 struct scm_fp_list *fpl;
6311 struct sk_buff *skb;
6314 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6318 skb = alloc_skb(0, GFP_KERNEL);
6327 fpl->user = get_uid(ctx->user);
6328 for (i = 0; i < nr; i++) {
6329 struct file *file = io_file_from_index(ctx, i + offset);
6333 fpl->fp[nr_files] = get_file(file);
6334 unix_inflight(fpl->user, fpl->fp[nr_files]);
6339 fpl->max = SCM_MAX_FD;
6340 fpl->count = nr_files;
6341 UNIXCB(skb).fp = fpl;
6342 skb->destructor = unix_destruct_scm;
6343 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6344 skb_queue_head(&sk->sk_receive_queue, skb);
6346 for (i = 0; i < nr_files; i++)
6357 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6358 * causes regular reference counting to break down. We rely on the UNIX
6359 * garbage collection to take care of this problem for us.
6361 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6363 unsigned left, total;
6367 left = ctx->nr_user_files;
6369 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6371 ret = __io_sqe_files_scm(ctx, this_files, total);
6375 total += this_files;
6381 while (total < ctx->nr_user_files) {
6382 struct file *file = io_file_from_index(ctx, total);
6392 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6398 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6403 for (i = 0; i < nr_tables; i++) {
6404 struct fixed_file_table *table = &ctx->file_data->table[i];
6405 unsigned this_files;
6407 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6408 table->files = kcalloc(this_files, sizeof(struct file *),
6412 nr_files -= this_files;
6418 for (i = 0; i < nr_tables; i++) {
6419 struct fixed_file_table *table = &ctx->file_data->table[i];
6420 kfree(table->files);
6425 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6427 #if defined(CONFIG_UNIX)
6428 struct sock *sock = ctx->ring_sock->sk;
6429 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6430 struct sk_buff *skb;
6433 __skb_queue_head_init(&list);
6436 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6437 * remove this entry and rearrange the file array.
6439 skb = skb_dequeue(head);
6441 struct scm_fp_list *fp;
6443 fp = UNIXCB(skb).fp;
6444 for (i = 0; i < fp->count; i++) {
6447 if (fp->fp[i] != file)
6450 unix_notinflight(fp->user, fp->fp[i]);
6451 left = fp->count - 1 - i;
6453 memmove(&fp->fp[i], &fp->fp[i + 1],
6454 left * sizeof(struct file *));
6461 __skb_queue_tail(&list, skb);
6471 __skb_queue_tail(&list, skb);
6473 skb = skb_dequeue(head);
6476 if (skb_peek(&list)) {
6477 spin_lock_irq(&head->lock);
6478 while ((skb = __skb_dequeue(&list)) != NULL)
6479 __skb_queue_tail(head, skb);
6480 spin_unlock_irq(&head->lock);
6487 struct io_file_put {
6488 struct list_head list;
6492 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6494 struct fixed_file_data *file_data = ref_node->file_data;
6495 struct io_ring_ctx *ctx = file_data->ctx;
6496 struct io_file_put *pfile, *tmp;
6498 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6499 list_del(&pfile->list);
6500 io_ring_file_put(ctx, pfile->file);
6504 spin_lock(&file_data->lock);
6505 list_del(&ref_node->node);
6506 spin_unlock(&file_data->lock);
6508 percpu_ref_exit(&ref_node->refs);
6510 percpu_ref_put(&file_data->refs);
6513 static void io_file_put_work(struct work_struct *work)
6515 struct io_ring_ctx *ctx;
6516 struct llist_node *node;
6518 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6519 node = llist_del_all(&ctx->file_put_llist);
6522 struct fixed_file_ref_node *ref_node;
6523 struct llist_node *next = node->next;
6525 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6526 __io_file_put_work(ref_node);
6531 static void io_file_data_ref_zero(struct percpu_ref *ref)
6533 struct fixed_file_ref_node *ref_node;
6534 struct io_ring_ctx *ctx;
6538 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6539 ctx = ref_node->file_data->ctx;
6541 if (percpu_ref_is_dying(&ctx->file_data->refs))
6544 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6546 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6548 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6551 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6552 struct io_ring_ctx *ctx)
6554 struct fixed_file_ref_node *ref_node;
6556 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6558 return ERR_PTR(-ENOMEM);
6560 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6563 return ERR_PTR(-ENOMEM);
6565 INIT_LIST_HEAD(&ref_node->node);
6566 INIT_LIST_HEAD(&ref_node->file_list);
6567 ref_node->file_data = ctx->file_data;
6571 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6573 percpu_ref_exit(&ref_node->refs);
6577 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6580 __s32 __user *fds = (__s32 __user *) arg;
6585 struct fixed_file_ref_node *ref_node;
6591 if (nr_args > IORING_MAX_FIXED_FILES)
6594 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6595 if (!ctx->file_data)
6597 ctx->file_data->ctx = ctx;
6598 init_completion(&ctx->file_data->done);
6599 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6600 spin_lock_init(&ctx->file_data->lock);
6602 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6603 ctx->file_data->table = kcalloc(nr_tables,
6604 sizeof(struct fixed_file_table),
6606 if (!ctx->file_data->table) {
6607 kfree(ctx->file_data);
6608 ctx->file_data = NULL;
6612 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6613 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6614 kfree(ctx->file_data->table);
6615 kfree(ctx->file_data);
6616 ctx->file_data = NULL;
6620 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6621 percpu_ref_exit(&ctx->file_data->refs);
6622 kfree(ctx->file_data->table);
6623 kfree(ctx->file_data);
6624 ctx->file_data = NULL;
6628 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6629 struct fixed_file_table *table;
6633 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6635 /* allow sparse sets */
6641 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6642 index = i & IORING_FILE_TABLE_MASK;
6650 * Don't allow io_uring instances to be registered. If UNIX
6651 * isn't enabled, then this causes a reference cycle and this
6652 * instance can never get freed. If UNIX is enabled we'll
6653 * handle it just fine, but there's still no point in allowing
6654 * a ring fd as it doesn't support regular read/write anyway.
6656 if (file->f_op == &io_uring_fops) {
6661 table->files[index] = file;
6665 for (i = 0; i < ctx->nr_user_files; i++) {
6666 file = io_file_from_index(ctx, i);
6670 for (i = 0; i < nr_tables; i++)
6671 kfree(ctx->file_data->table[i].files);
6673 kfree(ctx->file_data->table);
6674 kfree(ctx->file_data);
6675 ctx->file_data = NULL;
6676 ctx->nr_user_files = 0;
6680 ret = io_sqe_files_scm(ctx);
6682 io_sqe_files_unregister(ctx);
6686 ref_node = alloc_fixed_file_ref_node(ctx);
6687 if (IS_ERR(ref_node)) {
6688 io_sqe_files_unregister(ctx);
6689 return PTR_ERR(ref_node);
6692 ctx->file_data->cur_refs = &ref_node->refs;
6693 spin_lock(&ctx->file_data->lock);
6694 list_add(&ref_node->node, &ctx->file_data->ref_list);
6695 spin_unlock(&ctx->file_data->lock);
6696 percpu_ref_get(&ctx->file_data->refs);
6700 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6703 #if defined(CONFIG_UNIX)
6704 struct sock *sock = ctx->ring_sock->sk;
6705 struct sk_buff_head *head = &sock->sk_receive_queue;
6706 struct sk_buff *skb;
6709 * See if we can merge this file into an existing skb SCM_RIGHTS
6710 * file set. If there's no room, fall back to allocating a new skb
6711 * and filling it in.
6713 spin_lock_irq(&head->lock);
6714 skb = skb_peek(head);
6716 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6718 if (fpl->count < SCM_MAX_FD) {
6719 __skb_unlink(skb, head);
6720 spin_unlock_irq(&head->lock);
6721 fpl->fp[fpl->count] = get_file(file);
6722 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6724 spin_lock_irq(&head->lock);
6725 __skb_queue_head(head, skb);
6730 spin_unlock_irq(&head->lock);
6737 return __io_sqe_files_scm(ctx, 1, index);
6743 static int io_queue_file_removal(struct fixed_file_data *data,
6746 struct io_file_put *pfile;
6747 struct percpu_ref *refs = data->cur_refs;
6748 struct fixed_file_ref_node *ref_node;
6750 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6754 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6756 list_add(&pfile->list, &ref_node->file_list);
6761 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6762 struct io_uring_files_update *up,
6765 struct fixed_file_data *data = ctx->file_data;
6766 struct fixed_file_ref_node *ref_node;
6771 bool needs_switch = false;
6773 if (check_add_overflow(up->offset, nr_args, &done))
6775 if (done > ctx->nr_user_files)
6778 ref_node = alloc_fixed_file_ref_node(ctx);
6779 if (IS_ERR(ref_node))
6780 return PTR_ERR(ref_node);
6783 fds = u64_to_user_ptr(up->fds);
6785 struct fixed_file_table *table;
6789 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6793 i = array_index_nospec(up->offset, ctx->nr_user_files);
6794 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6795 index = i & IORING_FILE_TABLE_MASK;
6796 if (table->files[index]) {
6797 file = io_file_from_index(ctx, index);
6798 err = io_queue_file_removal(data, file);
6801 table->files[index] = NULL;
6802 needs_switch = true;
6811 * Don't allow io_uring instances to be registered. If
6812 * UNIX isn't enabled, then this causes a reference
6813 * cycle and this instance can never get freed. If UNIX
6814 * is enabled we'll handle it just fine, but there's
6815 * still no point in allowing a ring fd as it doesn't
6816 * support regular read/write anyway.
6818 if (file->f_op == &io_uring_fops) {
6823 table->files[index] = file;
6824 err = io_sqe_file_register(ctx, file, i);
6834 percpu_ref_kill(data->cur_refs);
6835 spin_lock(&data->lock);
6836 list_add(&ref_node->node, &data->ref_list);
6837 data->cur_refs = &ref_node->refs;
6838 spin_unlock(&data->lock);
6839 percpu_ref_get(&ctx->file_data->refs);
6841 destroy_fixed_file_ref_node(ref_node);
6843 return done ? done : err;
6846 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6849 struct io_uring_files_update up;
6851 if (!ctx->file_data)
6855 if (copy_from_user(&up, arg, sizeof(up)))
6860 return __io_sqe_files_update(ctx, &up, nr_args);
6863 static void io_free_work(struct io_wq_work *work)
6865 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6867 /* Consider that io_steal_work() relies on this ref */
6871 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6872 struct io_uring_params *p)
6874 struct io_wq_data data;
6876 struct io_ring_ctx *ctx_attach;
6877 unsigned int concurrency;
6880 data.user = ctx->user;
6881 data.free_work = io_free_work;
6883 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6884 /* Do QD, or 4 * CPUS, whatever is smallest */
6885 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6887 ctx->io_wq = io_wq_create(concurrency, &data);
6888 if (IS_ERR(ctx->io_wq)) {
6889 ret = PTR_ERR(ctx->io_wq);
6895 f = fdget(p->wq_fd);
6899 if (f.file->f_op != &io_uring_fops) {
6904 ctx_attach = f.file->private_data;
6905 /* @io_wq is protected by holding the fd */
6906 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6911 ctx->io_wq = ctx_attach->io_wq;
6917 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6918 struct io_uring_params *p)
6922 mmgrab(current->mm);
6923 ctx->sqo_mm = current->mm;
6925 if (ctx->flags & IORING_SETUP_SQPOLL) {
6927 if (!capable(CAP_SYS_ADMIN))
6930 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6931 if (!ctx->sq_thread_idle)
6932 ctx->sq_thread_idle = HZ;
6934 if (p->flags & IORING_SETUP_SQ_AFF) {
6935 int cpu = p->sq_thread_cpu;
6938 if (cpu >= nr_cpu_ids)
6940 if (!cpu_online(cpu))
6943 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6947 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6950 if (IS_ERR(ctx->sqo_thread)) {
6951 ret = PTR_ERR(ctx->sqo_thread);
6952 ctx->sqo_thread = NULL;
6955 wake_up_process(ctx->sqo_thread);
6956 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6957 /* Can't have SQ_AFF without SQPOLL */
6962 ret = io_init_wq_offload(ctx, p);
6968 io_finish_async(ctx);
6969 mmdrop(ctx->sqo_mm);
6974 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6976 atomic_long_sub(nr_pages, &user->locked_vm);
6979 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6981 unsigned long page_limit, cur_pages, new_pages;
6983 /* Don't allow more pages than we can safely lock */
6984 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6987 cur_pages = atomic_long_read(&user->locked_vm);
6988 new_pages = cur_pages + nr_pages;
6989 if (new_pages > page_limit)
6991 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6992 new_pages) != cur_pages);
6997 static void io_mem_free(void *ptr)
7004 page = virt_to_head_page(ptr);
7005 if (put_page_testzero(page))
7006 free_compound_page(page);
7009 static void *io_mem_alloc(size_t size)
7011 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7014 return (void *) __get_free_pages(gfp_flags, get_order(size));
7017 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7020 struct io_rings *rings;
7021 size_t off, sq_array_size;
7023 off = struct_size(rings, cqes, cq_entries);
7024 if (off == SIZE_MAX)
7028 off = ALIGN(off, SMP_CACHE_BYTES);
7033 sq_array_size = array_size(sizeof(u32), sq_entries);
7034 if (sq_array_size == SIZE_MAX)
7037 if (check_add_overflow(off, sq_array_size, &off))
7046 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7050 pages = (size_t)1 << get_order(
7051 rings_size(sq_entries, cq_entries, NULL));
7052 pages += (size_t)1 << get_order(
7053 array_size(sizeof(struct io_uring_sqe), sq_entries));
7058 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7062 if (!ctx->user_bufs)
7065 for (i = 0; i < ctx->nr_user_bufs; i++) {
7066 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7068 for (j = 0; j < imu->nr_bvecs; j++)
7069 unpin_user_page(imu->bvec[j].bv_page);
7071 if (ctx->account_mem)
7072 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7077 kfree(ctx->user_bufs);
7078 ctx->user_bufs = NULL;
7079 ctx->nr_user_bufs = 0;
7083 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7084 void __user *arg, unsigned index)
7086 struct iovec __user *src;
7088 #ifdef CONFIG_COMPAT
7090 struct compat_iovec __user *ciovs;
7091 struct compat_iovec ciov;
7093 ciovs = (struct compat_iovec __user *) arg;
7094 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7097 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7098 dst->iov_len = ciov.iov_len;
7102 src = (struct iovec __user *) arg;
7103 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7108 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7111 struct vm_area_struct **vmas = NULL;
7112 struct page **pages = NULL;
7113 int i, j, got_pages = 0;
7118 if (!nr_args || nr_args > UIO_MAXIOV)
7121 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7123 if (!ctx->user_bufs)
7126 for (i = 0; i < nr_args; i++) {
7127 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7128 unsigned long off, start, end, ubuf;
7133 ret = io_copy_iov(ctx, &iov, arg, i);
7138 * Don't impose further limits on the size and buffer
7139 * constraints here, we'll -EINVAL later when IO is
7140 * submitted if they are wrong.
7143 if (!iov.iov_base || !iov.iov_len)
7146 /* arbitrary limit, but we need something */
7147 if (iov.iov_len > SZ_1G)
7150 ubuf = (unsigned long) iov.iov_base;
7151 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7152 start = ubuf >> PAGE_SHIFT;
7153 nr_pages = end - start;
7155 if (ctx->account_mem) {
7156 ret = io_account_mem(ctx->user, nr_pages);
7162 if (!pages || nr_pages > got_pages) {
7165 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7167 vmas = kvmalloc_array(nr_pages,
7168 sizeof(struct vm_area_struct *),
7170 if (!pages || !vmas) {
7172 if (ctx->account_mem)
7173 io_unaccount_mem(ctx->user, nr_pages);
7176 got_pages = nr_pages;
7179 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7183 if (ctx->account_mem)
7184 io_unaccount_mem(ctx->user, nr_pages);
7189 mmap_read_lock(current->mm);
7190 pret = pin_user_pages(ubuf, nr_pages,
7191 FOLL_WRITE | FOLL_LONGTERM,
7193 if (pret == nr_pages) {
7194 /* don't support file backed memory */
7195 for (j = 0; j < nr_pages; j++) {
7196 struct vm_area_struct *vma = vmas[j];
7199 !is_file_hugepages(vma->vm_file)) {
7205 ret = pret < 0 ? pret : -EFAULT;
7207 mmap_read_unlock(current->mm);
7210 * if we did partial map, or found file backed vmas,
7211 * release any pages we did get
7214 unpin_user_pages(pages, pret);
7215 if (ctx->account_mem)
7216 io_unaccount_mem(ctx->user, nr_pages);
7221 off = ubuf & ~PAGE_MASK;
7223 for (j = 0; j < nr_pages; j++) {
7226 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7227 imu->bvec[j].bv_page = pages[j];
7228 imu->bvec[j].bv_len = vec_len;
7229 imu->bvec[j].bv_offset = off;
7233 /* store original address for later verification */
7235 imu->len = iov.iov_len;
7236 imu->nr_bvecs = nr_pages;
7238 ctx->nr_user_bufs++;
7246 io_sqe_buffer_unregister(ctx);
7250 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7252 __s32 __user *fds = arg;
7258 if (copy_from_user(&fd, fds, sizeof(*fds)))
7261 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7262 if (IS_ERR(ctx->cq_ev_fd)) {
7263 int ret = PTR_ERR(ctx->cq_ev_fd);
7264 ctx->cq_ev_fd = NULL;
7271 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7273 if (ctx->cq_ev_fd) {
7274 eventfd_ctx_put(ctx->cq_ev_fd);
7275 ctx->cq_ev_fd = NULL;
7282 static int __io_destroy_buffers(int id, void *p, void *data)
7284 struct io_ring_ctx *ctx = data;
7285 struct io_buffer *buf = p;
7287 __io_remove_buffers(ctx, buf, id, -1U);
7291 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7293 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7294 idr_destroy(&ctx->io_buffer_idr);
7297 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7299 io_finish_async(ctx);
7301 mmdrop(ctx->sqo_mm);
7303 io_iopoll_reap_events(ctx);
7304 io_sqe_buffer_unregister(ctx);
7305 io_sqe_files_unregister(ctx);
7306 io_eventfd_unregister(ctx);
7307 io_destroy_buffers(ctx);
7308 idr_destroy(&ctx->personality_idr);
7310 #if defined(CONFIG_UNIX)
7311 if (ctx->ring_sock) {
7312 ctx->ring_sock->file = NULL; /* so that iput() is called */
7313 sock_release(ctx->ring_sock);
7317 io_mem_free(ctx->rings);
7318 io_mem_free(ctx->sq_sqes);
7320 percpu_ref_exit(&ctx->refs);
7321 if (ctx->account_mem)
7322 io_unaccount_mem(ctx->user,
7323 ring_pages(ctx->sq_entries, ctx->cq_entries));
7324 free_uid(ctx->user);
7325 put_cred(ctx->creds);
7326 kfree(ctx->cancel_hash);
7327 kmem_cache_free(req_cachep, ctx->fallback_req);
7331 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7333 struct io_ring_ctx *ctx = file->private_data;
7336 poll_wait(file, &ctx->cq_wait, wait);
7338 * synchronizes with barrier from wq_has_sleeper call in
7342 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7343 ctx->rings->sq_ring_entries)
7344 mask |= EPOLLOUT | EPOLLWRNORM;
7345 if (io_cqring_events(ctx, false))
7346 mask |= EPOLLIN | EPOLLRDNORM;
7351 static int io_uring_fasync(int fd, struct file *file, int on)
7353 struct io_ring_ctx *ctx = file->private_data;
7355 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7358 static int io_remove_personalities(int id, void *p, void *data)
7360 struct io_ring_ctx *ctx = data;
7361 const struct cred *cred;
7363 cred = idr_remove(&ctx->personality_idr, id);
7369 static void io_ring_exit_work(struct work_struct *work)
7371 struct io_ring_ctx *ctx;
7373 ctx = container_of(work, struct io_ring_ctx, exit_work);
7375 io_cqring_overflow_flush(ctx, true);
7377 wait_for_completion(&ctx->ref_comp);
7378 io_ring_ctx_free(ctx);
7381 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7383 mutex_lock(&ctx->uring_lock);
7384 percpu_ref_kill(&ctx->refs);
7385 mutex_unlock(&ctx->uring_lock);
7387 io_kill_timeouts(ctx);
7388 io_poll_remove_all(ctx);
7391 io_wq_cancel_all(ctx->io_wq);
7393 io_iopoll_reap_events(ctx);
7394 /* if we failed setting up the ctx, we might not have any rings */
7396 io_cqring_overflow_flush(ctx, true);
7397 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7398 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7399 queue_work(system_wq, &ctx->exit_work);
7402 static int io_uring_release(struct inode *inode, struct file *file)
7404 struct io_ring_ctx *ctx = file->private_data;
7406 file->private_data = NULL;
7407 io_ring_ctx_wait_and_kill(ctx);
7411 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7412 struct files_struct *files)
7414 while (!list_empty_careful(&ctx->inflight_list)) {
7415 struct io_kiocb *cancel_req = NULL, *req;
7418 spin_lock_irq(&ctx->inflight_lock);
7419 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7420 if (req->work.files != files)
7422 /* req is being completed, ignore */
7423 if (!refcount_inc_not_zero(&req->refs))
7429 prepare_to_wait(&ctx->inflight_wait, &wait,
7430 TASK_UNINTERRUPTIBLE);
7431 spin_unlock_irq(&ctx->inflight_lock);
7433 /* We need to keep going until we don't find a matching req */
7437 if (cancel_req->flags & REQ_F_OVERFLOW) {
7438 spin_lock_irq(&ctx->completion_lock);
7439 list_del(&cancel_req->list);
7440 cancel_req->flags &= ~REQ_F_OVERFLOW;
7441 if (list_empty(&ctx->cq_overflow_list)) {
7442 clear_bit(0, &ctx->sq_check_overflow);
7443 clear_bit(0, &ctx->cq_check_overflow);
7445 spin_unlock_irq(&ctx->completion_lock);
7447 WRITE_ONCE(ctx->rings->cq_overflow,
7448 atomic_inc_return(&ctx->cached_cq_overflow));
7451 * Put inflight ref and overflow ref. If that's
7452 * all we had, then we're done with this request.
7454 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7455 io_free_req(cancel_req);
7456 finish_wait(&ctx->inflight_wait, &wait);
7460 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7461 io_put_req(cancel_req);
7465 finish_wait(&ctx->inflight_wait, &wait);
7469 static int io_uring_flush(struct file *file, void *data)
7471 struct io_ring_ctx *ctx = file->private_data;
7473 io_uring_cancel_files(ctx, data);
7476 * If the task is going away, cancel work it may have pending
7478 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7479 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7484 static void *io_uring_validate_mmap_request(struct file *file,
7485 loff_t pgoff, size_t sz)
7487 struct io_ring_ctx *ctx = file->private_data;
7488 loff_t offset = pgoff << PAGE_SHIFT;
7493 case IORING_OFF_SQ_RING:
7494 case IORING_OFF_CQ_RING:
7497 case IORING_OFF_SQES:
7501 return ERR_PTR(-EINVAL);
7504 page = virt_to_head_page(ptr);
7505 if (sz > page_size(page))
7506 return ERR_PTR(-EINVAL);
7513 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7515 size_t sz = vma->vm_end - vma->vm_start;
7519 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7521 return PTR_ERR(ptr);
7523 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7524 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7527 #else /* !CONFIG_MMU */
7529 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7531 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7534 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7536 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7539 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7540 unsigned long addr, unsigned long len,
7541 unsigned long pgoff, unsigned long flags)
7545 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7547 return PTR_ERR(ptr);
7549 return (unsigned long) ptr;
7552 #endif /* !CONFIG_MMU */
7554 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7555 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7558 struct io_ring_ctx *ctx;
7563 if (current->task_works)
7566 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7574 if (f.file->f_op != &io_uring_fops)
7578 ctx = f.file->private_data;
7579 if (!percpu_ref_tryget(&ctx->refs))
7583 * For SQ polling, the thread will do all submissions and completions.
7584 * Just return the requested submit count, and wake the thread if
7588 if (ctx->flags & IORING_SETUP_SQPOLL) {
7589 if (!list_empty_careful(&ctx->cq_overflow_list))
7590 io_cqring_overflow_flush(ctx, false);
7591 if (flags & IORING_ENTER_SQ_WAKEUP)
7592 wake_up(&ctx->sqo_wait);
7593 submitted = to_submit;
7594 } else if (to_submit) {
7595 mutex_lock(&ctx->uring_lock);
7596 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7597 mutex_unlock(&ctx->uring_lock);
7599 if (submitted != to_submit)
7602 if (flags & IORING_ENTER_GETEVENTS) {
7603 unsigned nr_events = 0;
7605 min_complete = min(min_complete, ctx->cq_entries);
7608 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7609 * space applications don't need to do io completion events
7610 * polling again, they can rely on io_sq_thread to do polling
7611 * work, which can reduce cpu usage and uring_lock contention.
7613 if (ctx->flags & IORING_SETUP_IOPOLL &&
7614 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7615 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7617 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7622 percpu_ref_put(&ctx->refs);
7625 return submitted ? submitted : ret;
7628 #ifdef CONFIG_PROC_FS
7629 static int io_uring_show_cred(int id, void *p, void *data)
7631 const struct cred *cred = p;
7632 struct seq_file *m = data;
7633 struct user_namespace *uns = seq_user_ns(m);
7634 struct group_info *gi;
7639 seq_printf(m, "%5d\n", id);
7640 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7641 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7642 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7643 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7644 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7645 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7646 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7647 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7648 seq_puts(m, "\n\tGroups:\t");
7649 gi = cred->group_info;
7650 for (g = 0; g < gi->ngroups; g++) {
7651 seq_put_decimal_ull(m, g ? " " : "",
7652 from_kgid_munged(uns, gi->gid[g]));
7654 seq_puts(m, "\n\tCapEff:\t");
7655 cap = cred->cap_effective;
7656 CAP_FOR_EACH_U32(__capi)
7657 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7662 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7666 mutex_lock(&ctx->uring_lock);
7667 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7668 for (i = 0; i < ctx->nr_user_files; i++) {
7669 struct fixed_file_table *table;
7672 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7673 f = table->files[i & IORING_FILE_TABLE_MASK];
7675 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7677 seq_printf(m, "%5u: <none>\n", i);
7679 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7680 for (i = 0; i < ctx->nr_user_bufs; i++) {
7681 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7683 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7684 (unsigned int) buf->len);
7686 if (!idr_is_empty(&ctx->personality_idr)) {
7687 seq_printf(m, "Personalities:\n");
7688 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7690 seq_printf(m, "PollList:\n");
7691 spin_lock_irq(&ctx->completion_lock);
7692 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7693 struct hlist_head *list = &ctx->cancel_hash[i];
7694 struct io_kiocb *req;
7696 hlist_for_each_entry(req, list, hash_node)
7697 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7698 req->task->task_works != NULL);
7700 spin_unlock_irq(&ctx->completion_lock);
7701 mutex_unlock(&ctx->uring_lock);
7704 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7706 struct io_ring_ctx *ctx = f->private_data;
7708 if (percpu_ref_tryget(&ctx->refs)) {
7709 __io_uring_show_fdinfo(ctx, m);
7710 percpu_ref_put(&ctx->refs);
7715 static const struct file_operations io_uring_fops = {
7716 .release = io_uring_release,
7717 .flush = io_uring_flush,
7718 .mmap = io_uring_mmap,
7720 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7721 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7723 .poll = io_uring_poll,
7724 .fasync = io_uring_fasync,
7725 #ifdef CONFIG_PROC_FS
7726 .show_fdinfo = io_uring_show_fdinfo,
7730 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7731 struct io_uring_params *p)
7733 struct io_rings *rings;
7734 size_t size, sq_array_offset;
7736 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7737 if (size == SIZE_MAX)
7740 rings = io_mem_alloc(size);
7745 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7746 rings->sq_ring_mask = p->sq_entries - 1;
7747 rings->cq_ring_mask = p->cq_entries - 1;
7748 rings->sq_ring_entries = p->sq_entries;
7749 rings->cq_ring_entries = p->cq_entries;
7750 ctx->sq_mask = rings->sq_ring_mask;
7751 ctx->cq_mask = rings->cq_ring_mask;
7752 ctx->sq_entries = rings->sq_ring_entries;
7753 ctx->cq_entries = rings->cq_ring_entries;
7755 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7756 if (size == SIZE_MAX) {
7757 io_mem_free(ctx->rings);
7762 ctx->sq_sqes = io_mem_alloc(size);
7763 if (!ctx->sq_sqes) {
7764 io_mem_free(ctx->rings);
7773 * Allocate an anonymous fd, this is what constitutes the application
7774 * visible backing of an io_uring instance. The application mmaps this
7775 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7776 * we have to tie this fd to a socket for file garbage collection purposes.
7778 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7783 #if defined(CONFIG_UNIX)
7784 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7790 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7794 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7795 O_RDWR | O_CLOEXEC);
7798 ret = PTR_ERR(file);
7802 #if defined(CONFIG_UNIX)
7803 ctx->ring_sock->file = file;
7805 fd_install(ret, file);
7808 #if defined(CONFIG_UNIX)
7809 sock_release(ctx->ring_sock);
7810 ctx->ring_sock = NULL;
7815 static int io_uring_create(unsigned entries, struct io_uring_params *p,
7816 struct io_uring_params __user *params)
7818 struct user_struct *user = NULL;
7819 struct io_ring_ctx *ctx;
7825 if (entries > IORING_MAX_ENTRIES) {
7826 if (!(p->flags & IORING_SETUP_CLAMP))
7828 entries = IORING_MAX_ENTRIES;
7832 * Use twice as many entries for the CQ ring. It's possible for the
7833 * application to drive a higher depth than the size of the SQ ring,
7834 * since the sqes are only used at submission time. This allows for
7835 * some flexibility in overcommitting a bit. If the application has
7836 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7837 * of CQ ring entries manually.
7839 p->sq_entries = roundup_pow_of_two(entries);
7840 if (p->flags & IORING_SETUP_CQSIZE) {
7842 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7843 * to a power-of-two, if it isn't already. We do NOT impose
7844 * any cq vs sq ring sizing.
7846 if (p->cq_entries < p->sq_entries)
7848 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7849 if (!(p->flags & IORING_SETUP_CLAMP))
7851 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7853 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7855 p->cq_entries = 2 * p->sq_entries;
7858 user = get_uid(current_user());
7859 account_mem = !capable(CAP_IPC_LOCK);
7862 ret = io_account_mem(user,
7863 ring_pages(p->sq_entries, p->cq_entries));
7870 ctx = io_ring_ctx_alloc(p);
7873 io_unaccount_mem(user, ring_pages(p->sq_entries,
7878 ctx->compat = in_compat_syscall();
7879 ctx->account_mem = account_mem;
7881 ctx->creds = get_current_cred();
7883 ret = io_allocate_scq_urings(ctx, p);
7887 ret = io_sq_offload_start(ctx, p);
7891 memset(&p->sq_off, 0, sizeof(p->sq_off));
7892 p->sq_off.head = offsetof(struct io_rings, sq.head);
7893 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7894 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7895 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7896 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7897 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7898 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7900 memset(&p->cq_off, 0, sizeof(p->cq_off));
7901 p->cq_off.head = offsetof(struct io_rings, cq.head);
7902 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7903 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7904 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7905 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7906 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7907 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
7909 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7910 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7911 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7913 if (copy_to_user(params, p, sizeof(*p))) {
7918 * Install ring fd as the very last thing, so we don't risk someone
7919 * having closed it before we finish setup
7921 ret = io_uring_get_fd(ctx);
7925 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7928 io_ring_ctx_wait_and_kill(ctx);
7933 * Sets up an aio uring context, and returns the fd. Applications asks for a
7934 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7935 * params structure passed in.
7937 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7939 struct io_uring_params p;
7942 if (copy_from_user(&p, params, sizeof(p)))
7944 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7949 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7950 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7951 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7954 return io_uring_create(entries, &p, params);
7957 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7958 struct io_uring_params __user *, params)
7960 return io_uring_setup(entries, params);
7963 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7965 struct io_uring_probe *p;
7969 size = struct_size(p, ops, nr_args);
7970 if (size == SIZE_MAX)
7972 p = kzalloc(size, GFP_KERNEL);
7977 if (copy_from_user(p, arg, size))
7980 if (memchr_inv(p, 0, size))
7983 p->last_op = IORING_OP_LAST - 1;
7984 if (nr_args > IORING_OP_LAST)
7985 nr_args = IORING_OP_LAST;
7987 for (i = 0; i < nr_args; i++) {
7989 if (!io_op_defs[i].not_supported)
7990 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7995 if (copy_to_user(arg, p, size))
8002 static int io_register_personality(struct io_ring_ctx *ctx)
8004 const struct cred *creds = get_current_cred();
8007 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8008 USHRT_MAX, GFP_KERNEL);
8014 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8016 const struct cred *old_creds;
8018 old_creds = idr_remove(&ctx->personality_idr, id);
8020 put_cred(old_creds);
8027 static bool io_register_op_must_quiesce(int op)
8030 case IORING_UNREGISTER_FILES:
8031 case IORING_REGISTER_FILES_UPDATE:
8032 case IORING_REGISTER_PROBE:
8033 case IORING_REGISTER_PERSONALITY:
8034 case IORING_UNREGISTER_PERSONALITY:
8041 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8042 void __user *arg, unsigned nr_args)
8043 __releases(ctx->uring_lock)
8044 __acquires(ctx->uring_lock)
8049 * We're inside the ring mutex, if the ref is already dying, then
8050 * someone else killed the ctx or is already going through
8051 * io_uring_register().
8053 if (percpu_ref_is_dying(&ctx->refs))
8056 if (io_register_op_must_quiesce(opcode)) {
8057 percpu_ref_kill(&ctx->refs);
8060 * Drop uring mutex before waiting for references to exit. If
8061 * another thread is currently inside io_uring_enter() it might
8062 * need to grab the uring_lock to make progress. If we hold it
8063 * here across the drain wait, then we can deadlock. It's safe
8064 * to drop the mutex here, since no new references will come in
8065 * after we've killed the percpu ref.
8067 mutex_unlock(&ctx->uring_lock);
8068 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8069 mutex_lock(&ctx->uring_lock);
8071 percpu_ref_resurrect(&ctx->refs);
8078 case IORING_REGISTER_BUFFERS:
8079 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8081 case IORING_UNREGISTER_BUFFERS:
8085 ret = io_sqe_buffer_unregister(ctx);
8087 case IORING_REGISTER_FILES:
8088 ret = io_sqe_files_register(ctx, arg, nr_args);
8090 case IORING_UNREGISTER_FILES:
8094 ret = io_sqe_files_unregister(ctx);
8096 case IORING_REGISTER_FILES_UPDATE:
8097 ret = io_sqe_files_update(ctx, arg, nr_args);
8099 case IORING_REGISTER_EVENTFD:
8100 case IORING_REGISTER_EVENTFD_ASYNC:
8104 ret = io_eventfd_register(ctx, arg);
8107 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8108 ctx->eventfd_async = 1;
8110 ctx->eventfd_async = 0;
8112 case IORING_UNREGISTER_EVENTFD:
8116 ret = io_eventfd_unregister(ctx);
8118 case IORING_REGISTER_PROBE:
8120 if (!arg || nr_args > 256)
8122 ret = io_probe(ctx, arg, nr_args);
8124 case IORING_REGISTER_PERSONALITY:
8128 ret = io_register_personality(ctx);
8130 case IORING_UNREGISTER_PERSONALITY:
8134 ret = io_unregister_personality(ctx, nr_args);
8141 if (io_register_op_must_quiesce(opcode)) {
8142 /* bring the ctx back to life */
8143 percpu_ref_reinit(&ctx->refs);
8145 reinit_completion(&ctx->ref_comp);
8150 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8151 void __user *, arg, unsigned int, nr_args)
8153 struct io_ring_ctx *ctx;
8162 if (f.file->f_op != &io_uring_fops)
8165 ctx = f.file->private_data;
8167 mutex_lock(&ctx->uring_lock);
8168 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8169 mutex_unlock(&ctx->uring_lock);
8170 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8171 ctx->cq_ev_fd != NULL, ret);
8177 static int __init io_uring_init(void)
8179 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8180 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8181 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8184 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8185 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8186 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8187 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8188 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8189 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8190 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8191 BUILD_BUG_SQE_ELEM(8, __u64, off);
8192 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8193 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8194 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8195 BUILD_BUG_SQE_ELEM(24, __u32, len);
8196 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8197 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8198 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8199 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8200 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8201 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8202 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8203 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8204 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8205 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8206 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8207 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8208 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8209 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8210 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8211 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8212 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8213 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8215 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8216 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8217 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8220 __initcall(io_uring_init);