1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/mmu_context.h>
59 #include <linux/percpu.h>
60 #include <linux/slab.h>
61 #include <linux/kthread.h>
62 #include <linux/blkdev.h>
63 #include <linux/bvec.h>
64 #include <linux/net.h>
66 #include <net/af_unix.h>
68 #include <linux/anon_inodes.h>
69 #include <linux/sched/mm.h>
70 #include <linux/uaccess.h>
71 #include <linux/nospec.h>
72 #include <linux/sizes.h>
73 #include <linux/hugetlb.h>
74 #include <linux/highmem.h>
75 #include <linux/namei.h>
76 #include <linux/fsnotify.h>
77 #include <linux/fadvise.h>
78 #include <linux/eventpoll.h>
79 #include <linux/fs_struct.h>
80 #include <linux/splice.h>
81 #include <linux/task_work.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
155 * Number of completion events lost because the queue was full;
156 * this should be avoided by the application by making sure
157 * there are not more requests pending than there is space in
158 * the completion queue.
160 * Written by the kernel, shouldn't be modified by the
161 * application (i.e. get number of "new events" by comparing to
164 * As completion events come in out of order this counter is not
165 * ordered with any other data.
169 * Ring buffer of completion events.
171 * The kernel writes completion events fresh every time they are
172 * produced, so the application is allowed to modify pending
175 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
178 struct io_mapped_ubuf {
181 struct bio_vec *bvec;
182 unsigned int nr_bvecs;
185 struct fixed_file_table {
189 struct fixed_file_ref_node {
190 struct percpu_ref refs;
191 struct list_head node;
192 struct list_head file_list;
193 struct fixed_file_data *file_data;
194 struct work_struct work;
197 struct fixed_file_data {
198 struct fixed_file_table *table;
199 struct io_ring_ctx *ctx;
201 struct percpu_ref *cur_refs;
202 struct percpu_ref refs;
203 struct completion done;
204 struct list_head ref_list;
209 struct list_head list;
217 struct percpu_ref refs;
218 } ____cacheline_aligned_in_smp;
222 unsigned int compat: 1;
223 unsigned int account_mem: 1;
224 unsigned int cq_overflow_flushed: 1;
225 unsigned int drain_next: 1;
226 unsigned int eventfd_async: 1;
229 * Ring buffer of indices into array of io_uring_sqe, which is
230 * mmapped by the application using the IORING_OFF_SQES offset.
232 * This indirection could e.g. be used to assign fixed
233 * io_uring_sqe entries to operations and only submit them to
234 * the queue when needed.
236 * The kernel modifies neither the indices array nor the entries
240 unsigned cached_sq_head;
243 unsigned sq_thread_idle;
244 unsigned cached_sq_dropped;
245 atomic_t cached_cq_overflow;
246 unsigned long sq_check_overflow;
248 struct list_head defer_list;
249 struct list_head timeout_list;
250 struct list_head cq_overflow_list;
252 wait_queue_head_t inflight_wait;
253 struct io_uring_sqe *sq_sqes;
254 } ____cacheline_aligned_in_smp;
256 struct io_rings *rings;
260 struct task_struct *sqo_thread; /* if using sq thread polling */
261 struct mm_struct *sqo_mm;
262 wait_queue_head_t sqo_wait;
265 * If used, fixed file set. Writers must ensure that ->refs is dead,
266 * readers must ensure that ->refs is alive as long as the file* is
267 * used. Only updated through io_uring_register(2).
269 struct fixed_file_data *file_data;
270 unsigned nr_user_files;
272 struct file *ring_file;
274 /* if used, fixed mapped user buffers */
275 unsigned nr_user_bufs;
276 struct io_mapped_ubuf *user_bufs;
278 struct user_struct *user;
280 const struct cred *creds;
282 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
283 struct completion *completions;
285 /* if all else fails... */
286 struct io_kiocb *fallback_req;
288 #if defined(CONFIG_UNIX)
289 struct socket *ring_sock;
292 struct idr io_buffer_idr;
294 struct idr personality_idr;
297 unsigned cached_cq_tail;
300 atomic_t cq_timeouts;
301 unsigned long cq_check_overflow;
302 struct wait_queue_head cq_wait;
303 struct fasync_struct *cq_fasync;
304 struct eventfd_ctx *cq_ev_fd;
305 } ____cacheline_aligned_in_smp;
308 struct mutex uring_lock;
309 wait_queue_head_t wait;
310 } ____cacheline_aligned_in_smp;
313 spinlock_t completion_lock;
316 * ->poll_list is protected by the ctx->uring_lock for
317 * io_uring instances that don't use IORING_SETUP_SQPOLL.
318 * For SQPOLL, only the single threaded io_sq_thread() will
319 * manipulate the list, hence no extra locking is needed there.
321 struct list_head poll_list;
322 struct hlist_head *cancel_hash;
323 unsigned cancel_hash_bits;
324 bool poll_multi_file;
326 spinlock_t inflight_lock;
327 struct list_head inflight_list;
328 } ____cacheline_aligned_in_smp;
330 struct work_struct exit_work;
334 * First field must be the file pointer in all the
335 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
337 struct io_poll_iocb {
340 struct wait_queue_head *head;
346 struct wait_queue_entry wait;
351 struct file *put_file;
355 struct io_timeout_data {
356 struct io_kiocb *req;
357 struct hrtimer timer;
358 struct timespec64 ts;
359 enum hrtimer_mode mode;
364 struct sockaddr __user *addr;
365 int __user *addr_len;
367 unsigned long nofile;
391 /* NOTE: kiocb has the file as the first member, so don't do it here */
399 struct sockaddr __user *addr;
406 struct user_msghdr __user *msg;
412 struct io_buffer *kbuf;
421 struct filename *filename;
422 struct statx __user *buffer;
424 unsigned long nofile;
427 struct io_files_update {
453 struct epoll_event event;
457 struct file *file_out;
458 struct file *file_in;
465 struct io_provide_buf {
474 struct io_async_connect {
475 struct sockaddr_storage address;
478 struct io_async_msghdr {
479 struct iovec fast_iov[UIO_FASTIOV];
481 struct sockaddr __user *uaddr;
483 struct sockaddr_storage addr;
487 struct iovec fast_iov[UIO_FASTIOV];
493 struct io_async_ctx {
495 struct io_async_rw rw;
496 struct io_async_msghdr msg;
497 struct io_async_connect connect;
498 struct io_timeout_data timeout;
503 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
504 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
505 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
506 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
507 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
508 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
516 REQ_F_IOPOLL_COMPLETED_BIT,
517 REQ_F_LINK_TIMEOUT_BIT,
521 REQ_F_TIMEOUT_NOSEQ_BIT,
522 REQ_F_COMP_LOCKED_BIT,
523 REQ_F_NEED_CLEANUP_BIT,
526 REQ_F_BUFFER_SELECTED_BIT,
527 REQ_F_NO_FILE_TABLE_BIT,
529 /* not a real bit, just to check we're not overflowing the space */
535 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
536 /* drain existing IO first */
537 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
539 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
540 /* doesn't sever on completion < 0 */
541 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
543 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
544 /* IOSQE_BUFFER_SELECT */
545 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
548 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
549 /* already grabbed next link */
550 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
551 /* fail rest of links */
552 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
553 /* on inflight list */
554 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
555 /* read/write uses file position */
556 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
557 /* must not punt to workers */
558 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
559 /* polled IO has completed */
560 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
561 /* has linked timeout */
562 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
563 /* timeout request */
564 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
566 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
567 /* must be punted even for NONBLOCK */
568 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
569 /* no timeout sequence */
570 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
571 /* completion under lock */
572 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
574 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
575 /* in overflow list */
576 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
577 /* already went through poll handler */
578 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
579 /* buffer already selected */
580 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
581 /* doesn't need file table for this request */
582 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
586 struct io_poll_iocb poll;
587 struct io_wq_work work;
591 * NOTE! Each of the iocb union members has the file pointer
592 * as the first entry in their struct definition. So you can
593 * access the file pointer through any of the sub-structs,
594 * or directly as just 'ki_filp' in this struct.
600 struct io_poll_iocb poll;
601 struct io_accept accept;
603 struct io_cancel cancel;
604 struct io_timeout timeout;
605 struct io_connect connect;
606 struct io_sr_msg sr_msg;
608 struct io_close close;
609 struct io_files_update files_update;
610 struct io_fadvise fadvise;
611 struct io_madvise madvise;
612 struct io_epoll epoll;
613 struct io_splice splice;
614 struct io_provide_buf pbuf;
617 struct io_async_ctx *io;
619 bool needs_fixed_file;
622 struct io_ring_ctx *ctx;
623 struct list_head list;
626 struct task_struct *task;
632 struct list_head link_list;
634 struct list_head inflight_entry;
636 struct percpu_ref *fixed_file_refs;
640 * Only commands that never go async can use the below fields,
641 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
642 * async armed poll handlers for regular commands. The latter
643 * restore the work, if needed.
646 struct callback_head task_work;
647 struct hlist_node hash_node;
648 struct async_poll *apoll;
650 struct io_wq_work work;
654 #define IO_PLUG_THRESHOLD 2
655 #define IO_IOPOLL_BATCH 8
657 struct io_submit_state {
658 struct blk_plug plug;
661 * io_kiocb alloc cache
663 void *reqs[IO_IOPOLL_BATCH];
664 unsigned int free_reqs;
667 * File reference cache
671 unsigned int has_refs;
672 unsigned int used_refs;
673 unsigned int ios_left;
677 /* needs req->io allocated for deferral/async */
678 unsigned async_ctx : 1;
679 /* needs current->mm setup, does mm access */
680 unsigned needs_mm : 1;
681 /* needs req->file assigned */
682 unsigned needs_file : 1;
683 /* needs req->file assigned IFF fd is >= 0 */
684 unsigned fd_non_neg : 1;
685 /* hash wq insertion if file is a regular file */
686 unsigned hash_reg_file : 1;
687 /* unbound wq insertion if file is a non-regular file */
688 unsigned unbound_nonreg_file : 1;
689 /* opcode is not supported by this kernel */
690 unsigned not_supported : 1;
691 /* needs file table */
692 unsigned file_table : 1;
694 unsigned needs_fs : 1;
695 /* set if opcode supports polled "wait" */
697 unsigned pollout : 1;
698 /* op supports buffer selection */
699 unsigned buffer_select : 1;
702 static const struct io_op_def io_op_defs[] = {
703 [IORING_OP_NOP] = {},
704 [IORING_OP_READV] = {
708 .unbound_nonreg_file = 1,
712 [IORING_OP_WRITEV] = {
717 .unbound_nonreg_file = 1,
720 [IORING_OP_FSYNC] = {
723 [IORING_OP_READ_FIXED] = {
725 .unbound_nonreg_file = 1,
728 [IORING_OP_WRITE_FIXED] = {
731 .unbound_nonreg_file = 1,
734 [IORING_OP_POLL_ADD] = {
736 .unbound_nonreg_file = 1,
738 [IORING_OP_POLL_REMOVE] = {},
739 [IORING_OP_SYNC_FILE_RANGE] = {
742 [IORING_OP_SENDMSG] = {
746 .unbound_nonreg_file = 1,
750 [IORING_OP_RECVMSG] = {
754 .unbound_nonreg_file = 1,
759 [IORING_OP_TIMEOUT] = {
763 [IORING_OP_TIMEOUT_REMOVE] = {},
764 [IORING_OP_ACCEPT] = {
767 .unbound_nonreg_file = 1,
771 [IORING_OP_ASYNC_CANCEL] = {},
772 [IORING_OP_LINK_TIMEOUT] = {
776 [IORING_OP_CONNECT] = {
780 .unbound_nonreg_file = 1,
783 [IORING_OP_FALLOCATE] = {
786 [IORING_OP_OPENAT] = {
792 [IORING_OP_CLOSE] = {
796 [IORING_OP_FILES_UPDATE] = {
800 [IORING_OP_STATX] = {
810 .unbound_nonreg_file = 1,
814 [IORING_OP_WRITE] = {
817 .unbound_nonreg_file = 1,
820 [IORING_OP_FADVISE] = {
823 [IORING_OP_MADVISE] = {
829 .unbound_nonreg_file = 1,
835 .unbound_nonreg_file = 1,
839 [IORING_OP_OPENAT2] = {
845 [IORING_OP_EPOLL_CTL] = {
846 .unbound_nonreg_file = 1,
849 [IORING_OP_SPLICE] = {
852 .unbound_nonreg_file = 1,
854 [IORING_OP_PROVIDE_BUFFERS] = {},
855 [IORING_OP_REMOVE_BUFFERS] = {},
858 static void io_wq_submit_work(struct io_wq_work **workptr);
859 static void io_cqring_fill_event(struct io_kiocb *req, long res);
860 static void io_put_req(struct io_kiocb *req);
861 static void __io_double_put_req(struct io_kiocb *req);
862 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
863 static void io_queue_linked_timeout(struct io_kiocb *req);
864 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
865 struct io_uring_files_update *ip,
867 static int io_grab_files(struct io_kiocb *req);
868 static void io_cleanup_req(struct io_kiocb *req);
869 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
870 int fd, struct file **out_file, bool fixed);
871 static void __io_queue_sqe(struct io_kiocb *req,
872 const struct io_uring_sqe *sqe);
874 static struct kmem_cache *req_cachep;
876 static const struct file_operations io_uring_fops;
878 struct sock *io_uring_get_socket(struct file *file)
880 #if defined(CONFIG_UNIX)
881 if (file->f_op == &io_uring_fops) {
882 struct io_ring_ctx *ctx = file->private_data;
884 return ctx->ring_sock->sk;
889 EXPORT_SYMBOL(io_uring_get_socket);
891 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
893 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
895 complete(&ctx->completions[0]);
898 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
900 struct io_ring_ctx *ctx;
903 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
907 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
908 if (!ctx->fallback_req)
911 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
912 if (!ctx->completions)
916 * Use 5 bits less than the max cq entries, that should give us around
917 * 32 entries per hash list if totally full and uniformly spread.
919 hash_bits = ilog2(p->cq_entries);
923 ctx->cancel_hash_bits = hash_bits;
924 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
926 if (!ctx->cancel_hash)
928 __hash_init(ctx->cancel_hash, 1U << hash_bits);
930 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
931 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
934 ctx->flags = p->flags;
935 init_waitqueue_head(&ctx->cq_wait);
936 INIT_LIST_HEAD(&ctx->cq_overflow_list);
937 init_completion(&ctx->completions[0]);
938 init_completion(&ctx->completions[1]);
939 idr_init(&ctx->io_buffer_idr);
940 idr_init(&ctx->personality_idr);
941 mutex_init(&ctx->uring_lock);
942 init_waitqueue_head(&ctx->wait);
943 spin_lock_init(&ctx->completion_lock);
944 INIT_LIST_HEAD(&ctx->poll_list);
945 INIT_LIST_HEAD(&ctx->defer_list);
946 INIT_LIST_HEAD(&ctx->timeout_list);
947 init_waitqueue_head(&ctx->inflight_wait);
948 spin_lock_init(&ctx->inflight_lock);
949 INIT_LIST_HEAD(&ctx->inflight_list);
952 if (ctx->fallback_req)
953 kmem_cache_free(req_cachep, ctx->fallback_req);
954 kfree(ctx->completions);
955 kfree(ctx->cancel_hash);
960 static inline bool __req_need_defer(struct io_kiocb *req)
962 struct io_ring_ctx *ctx = req->ctx;
964 return req->sequence != ctx->cached_cq_tail
965 + atomic_read(&ctx->cached_cq_overflow);
968 static inline bool req_need_defer(struct io_kiocb *req)
970 if (unlikely(req->flags & REQ_F_IO_DRAIN))
971 return __req_need_defer(req);
976 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
978 struct io_kiocb *req;
980 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
981 if (req && !req_need_defer(req)) {
982 list_del_init(&req->list);
989 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
991 struct io_kiocb *req;
993 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
995 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
997 if (!__req_need_defer(req)) {
998 list_del_init(&req->list);
1006 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1008 struct io_rings *rings = ctx->rings;
1010 /* order cqe stores with ring update */
1011 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1013 if (wq_has_sleeper(&ctx->cq_wait)) {
1014 wake_up_interruptible(&ctx->cq_wait);
1015 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1019 static inline void io_req_work_grab_env(struct io_kiocb *req,
1020 const struct io_op_def *def)
1022 if (!req->work.mm && def->needs_mm) {
1023 mmgrab(current->mm);
1024 req->work.mm = current->mm;
1026 if (!req->work.creds)
1027 req->work.creds = get_current_cred();
1028 if (!req->work.fs && def->needs_fs) {
1029 spin_lock(¤t->fs->lock);
1030 if (!current->fs->in_exec) {
1031 req->work.fs = current->fs;
1032 req->work.fs->users++;
1034 req->work.flags |= IO_WQ_WORK_CANCEL;
1036 spin_unlock(¤t->fs->lock);
1038 if (!req->work.task_pid)
1039 req->work.task_pid = task_pid_vnr(current);
1042 static inline void io_req_work_drop_env(struct io_kiocb *req)
1045 mmdrop(req->work.mm);
1046 req->work.mm = NULL;
1048 if (req->work.creds) {
1049 put_cred(req->work.creds);
1050 req->work.creds = NULL;
1053 struct fs_struct *fs = req->work.fs;
1055 spin_lock(&req->work.fs->lock);
1058 spin_unlock(&req->work.fs->lock);
1064 static inline void io_prep_async_work(struct io_kiocb *req,
1065 struct io_kiocb **link)
1067 const struct io_op_def *def = &io_op_defs[req->opcode];
1069 if (req->flags & REQ_F_ISREG) {
1070 if (def->hash_reg_file)
1071 io_wq_hash_work(&req->work, file_inode(req->file));
1073 if (def->unbound_nonreg_file)
1074 req->work.flags |= IO_WQ_WORK_UNBOUND;
1077 io_req_work_grab_env(req, def);
1079 *link = io_prep_linked_timeout(req);
1082 static inline void io_queue_async_work(struct io_kiocb *req)
1084 struct io_ring_ctx *ctx = req->ctx;
1085 struct io_kiocb *link;
1087 io_prep_async_work(req, &link);
1089 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1090 &req->work, req->flags);
1091 io_wq_enqueue(ctx->io_wq, &req->work);
1094 io_queue_linked_timeout(link);
1097 static void io_kill_timeout(struct io_kiocb *req)
1101 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1103 atomic_inc(&req->ctx->cq_timeouts);
1104 list_del_init(&req->list);
1105 req->flags |= REQ_F_COMP_LOCKED;
1106 io_cqring_fill_event(req, 0);
1111 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1113 struct io_kiocb *req, *tmp;
1115 spin_lock_irq(&ctx->completion_lock);
1116 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1117 io_kill_timeout(req);
1118 spin_unlock_irq(&ctx->completion_lock);
1121 static void io_commit_cqring(struct io_ring_ctx *ctx)
1123 struct io_kiocb *req;
1125 while ((req = io_get_timeout_req(ctx)) != NULL)
1126 io_kill_timeout(req);
1128 __io_commit_cqring(ctx);
1130 while ((req = io_get_deferred_req(ctx)) != NULL)
1131 io_queue_async_work(req);
1134 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1136 struct io_rings *rings = ctx->rings;
1139 tail = ctx->cached_cq_tail;
1141 * writes to the cq entry need to come after reading head; the
1142 * control dependency is enough as we're using WRITE_ONCE to
1145 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1148 ctx->cached_cq_tail++;
1149 return &rings->cqes[tail & ctx->cq_mask];
1152 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1156 if (!ctx->eventfd_async)
1158 return io_wq_current_is_worker();
1161 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1163 if (waitqueue_active(&ctx->wait))
1164 wake_up(&ctx->wait);
1165 if (waitqueue_active(&ctx->sqo_wait))
1166 wake_up(&ctx->sqo_wait);
1167 if (io_should_trigger_evfd(ctx))
1168 eventfd_signal(ctx->cq_ev_fd, 1);
1171 /* Returns true if there are no backlogged entries after the flush */
1172 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1174 struct io_rings *rings = ctx->rings;
1175 struct io_uring_cqe *cqe;
1176 struct io_kiocb *req;
1177 unsigned long flags;
1181 if (list_empty_careful(&ctx->cq_overflow_list))
1183 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1184 rings->cq_ring_entries))
1188 spin_lock_irqsave(&ctx->completion_lock, flags);
1190 /* if force is set, the ring is going away. always drop after that */
1192 ctx->cq_overflow_flushed = 1;
1195 while (!list_empty(&ctx->cq_overflow_list)) {
1196 cqe = io_get_cqring(ctx);
1200 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1202 list_move(&req->list, &list);
1203 req->flags &= ~REQ_F_OVERFLOW;
1205 WRITE_ONCE(cqe->user_data, req->user_data);
1206 WRITE_ONCE(cqe->res, req->result);
1207 WRITE_ONCE(cqe->flags, req->cflags);
1209 WRITE_ONCE(ctx->rings->cq_overflow,
1210 atomic_inc_return(&ctx->cached_cq_overflow));
1214 io_commit_cqring(ctx);
1216 clear_bit(0, &ctx->sq_check_overflow);
1217 clear_bit(0, &ctx->cq_check_overflow);
1219 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1220 io_cqring_ev_posted(ctx);
1222 while (!list_empty(&list)) {
1223 req = list_first_entry(&list, struct io_kiocb, list);
1224 list_del(&req->list);
1231 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1233 struct io_ring_ctx *ctx = req->ctx;
1234 struct io_uring_cqe *cqe;
1236 trace_io_uring_complete(ctx, req->user_data, res);
1239 * If we can't get a cq entry, userspace overflowed the
1240 * submission (by quite a lot). Increment the overflow count in
1243 cqe = io_get_cqring(ctx);
1245 WRITE_ONCE(cqe->user_data, req->user_data);
1246 WRITE_ONCE(cqe->res, res);
1247 WRITE_ONCE(cqe->flags, cflags);
1248 } else if (ctx->cq_overflow_flushed) {
1249 WRITE_ONCE(ctx->rings->cq_overflow,
1250 atomic_inc_return(&ctx->cached_cq_overflow));
1252 if (list_empty(&ctx->cq_overflow_list)) {
1253 set_bit(0, &ctx->sq_check_overflow);
1254 set_bit(0, &ctx->cq_check_overflow);
1256 req->flags |= REQ_F_OVERFLOW;
1257 refcount_inc(&req->refs);
1259 req->cflags = cflags;
1260 list_add_tail(&req->list, &ctx->cq_overflow_list);
1264 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1266 __io_cqring_fill_event(req, res, 0);
1269 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1271 struct io_ring_ctx *ctx = req->ctx;
1272 unsigned long flags;
1274 spin_lock_irqsave(&ctx->completion_lock, flags);
1275 __io_cqring_fill_event(req, res, cflags);
1276 io_commit_cqring(ctx);
1277 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1279 io_cqring_ev_posted(ctx);
1282 static void io_cqring_add_event(struct io_kiocb *req, long res)
1284 __io_cqring_add_event(req, res, 0);
1287 static inline bool io_is_fallback_req(struct io_kiocb *req)
1289 return req == (struct io_kiocb *)
1290 ((unsigned long) req->ctx->fallback_req & ~1UL);
1293 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1295 struct io_kiocb *req;
1297 req = ctx->fallback_req;
1298 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1304 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1305 struct io_submit_state *state)
1307 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1308 struct io_kiocb *req;
1311 req = kmem_cache_alloc(req_cachep, gfp);
1314 } else if (!state->free_reqs) {
1318 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1319 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1322 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1323 * retry single alloc to be on the safe side.
1325 if (unlikely(ret <= 0)) {
1326 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1327 if (!state->reqs[0])
1331 state->free_reqs = ret - 1;
1332 req = state->reqs[ret - 1];
1335 req = state->reqs[state->free_reqs];
1340 return io_get_fallback_req(ctx);
1343 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1347 percpu_ref_put(req->fixed_file_refs);
1352 static void __io_req_aux_free(struct io_kiocb *req)
1354 if (req->flags & REQ_F_NEED_CLEANUP)
1355 io_cleanup_req(req);
1359 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1361 put_task_struct(req->task);
1363 io_req_work_drop_env(req);
1366 static void __io_free_req(struct io_kiocb *req)
1368 __io_req_aux_free(req);
1370 if (req->flags & REQ_F_INFLIGHT) {
1371 struct io_ring_ctx *ctx = req->ctx;
1372 unsigned long flags;
1374 spin_lock_irqsave(&ctx->inflight_lock, flags);
1375 list_del(&req->inflight_entry);
1376 if (waitqueue_active(&ctx->inflight_wait))
1377 wake_up(&ctx->inflight_wait);
1378 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1381 percpu_ref_put(&req->ctx->refs);
1382 if (likely(!io_is_fallback_req(req)))
1383 kmem_cache_free(req_cachep, req);
1385 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1389 void *reqs[IO_IOPOLL_BATCH];
1394 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1398 if (rb->need_iter) {
1399 int i, inflight = 0;
1400 unsigned long flags;
1402 for (i = 0; i < rb->to_free; i++) {
1403 struct io_kiocb *req = rb->reqs[i];
1405 if (req->flags & REQ_F_FIXED_FILE) {
1407 percpu_ref_put(req->fixed_file_refs);
1409 if (req->flags & REQ_F_INFLIGHT)
1411 __io_req_aux_free(req);
1416 spin_lock_irqsave(&ctx->inflight_lock, flags);
1417 for (i = 0; i < rb->to_free; i++) {
1418 struct io_kiocb *req = rb->reqs[i];
1420 if (req->flags & REQ_F_INFLIGHT) {
1421 list_del(&req->inflight_entry);
1426 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1428 if (waitqueue_active(&ctx->inflight_wait))
1429 wake_up(&ctx->inflight_wait);
1432 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1433 percpu_ref_put_many(&ctx->refs, rb->to_free);
1434 rb->to_free = rb->need_iter = 0;
1437 static bool io_link_cancel_timeout(struct io_kiocb *req)
1439 struct io_ring_ctx *ctx = req->ctx;
1442 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1444 io_cqring_fill_event(req, -ECANCELED);
1445 io_commit_cqring(ctx);
1446 req->flags &= ~REQ_F_LINK_HEAD;
1454 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1456 struct io_ring_ctx *ctx = req->ctx;
1457 bool wake_ev = false;
1459 /* Already got next link */
1460 if (req->flags & REQ_F_LINK_NEXT)
1464 * The list should never be empty when we are called here. But could
1465 * potentially happen if the chain is messed up, check to be on the
1468 while (!list_empty(&req->link_list)) {
1469 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1470 struct io_kiocb, link_list);
1472 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1473 (nxt->flags & REQ_F_TIMEOUT))) {
1474 list_del_init(&nxt->link_list);
1475 wake_ev |= io_link_cancel_timeout(nxt);
1476 req->flags &= ~REQ_F_LINK_TIMEOUT;
1480 list_del_init(&req->link_list);
1481 if (!list_empty(&nxt->link_list))
1482 nxt->flags |= REQ_F_LINK_HEAD;
1487 req->flags |= REQ_F_LINK_NEXT;
1489 io_cqring_ev_posted(ctx);
1493 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1495 static void io_fail_links(struct io_kiocb *req)
1497 struct io_ring_ctx *ctx = req->ctx;
1498 unsigned long flags;
1500 spin_lock_irqsave(&ctx->completion_lock, flags);
1502 while (!list_empty(&req->link_list)) {
1503 struct io_kiocb *link = list_first_entry(&req->link_list,
1504 struct io_kiocb, link_list);
1506 list_del_init(&link->link_list);
1507 trace_io_uring_fail_link(req, link);
1509 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1510 link->opcode == IORING_OP_LINK_TIMEOUT) {
1511 io_link_cancel_timeout(link);
1513 io_cqring_fill_event(link, -ECANCELED);
1514 __io_double_put_req(link);
1516 req->flags &= ~REQ_F_LINK_TIMEOUT;
1519 io_commit_cqring(ctx);
1520 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1521 io_cqring_ev_posted(ctx);
1524 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1526 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1530 * If LINK is set, we have dependent requests in this chain. If we
1531 * didn't fail this request, queue the first one up, moving any other
1532 * dependencies to the next request. In case of failure, fail the rest
1535 if (req->flags & REQ_F_FAIL_LINK) {
1537 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1538 REQ_F_LINK_TIMEOUT) {
1539 struct io_ring_ctx *ctx = req->ctx;
1540 unsigned long flags;
1543 * If this is a timeout link, we could be racing with the
1544 * timeout timer. Grab the completion lock for this case to
1545 * protect against that.
1547 spin_lock_irqsave(&ctx->completion_lock, flags);
1548 io_req_link_next(req, nxt);
1549 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1551 io_req_link_next(req, nxt);
1555 static void io_free_req(struct io_kiocb *req)
1557 struct io_kiocb *nxt = NULL;
1559 io_req_find_next(req, &nxt);
1563 io_queue_async_work(nxt);
1566 static void io_link_work_cb(struct io_wq_work **workptr)
1568 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1569 struct io_kiocb *link;
1571 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1572 io_queue_linked_timeout(link);
1573 io_wq_submit_work(workptr);
1576 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1578 struct io_kiocb *link;
1579 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1581 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1582 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1584 *workptr = &nxt->work;
1585 link = io_prep_linked_timeout(nxt);
1587 nxt->work.func = io_link_work_cb;
1591 * Drop reference to request, return next in chain (if there is one) if this
1592 * was the last reference to this request.
1594 __attribute__((nonnull))
1595 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1597 if (refcount_dec_and_test(&req->refs)) {
1598 io_req_find_next(req, nxtptr);
1603 static void io_put_req(struct io_kiocb *req)
1605 if (refcount_dec_and_test(&req->refs))
1609 static void io_steal_work(struct io_kiocb *req,
1610 struct io_wq_work **workptr)
1613 * It's in an io-wq worker, so there always should be at least
1614 * one reference, which will be dropped in io_put_work() just
1615 * after the current handler returns.
1617 * It also means, that if the counter dropped to 1, then there is
1618 * no asynchronous users left, so it's safe to steal the next work.
1620 if (refcount_read(&req->refs) == 1) {
1621 struct io_kiocb *nxt = NULL;
1623 io_req_find_next(req, &nxt);
1625 io_wq_assign_next(workptr, nxt);
1630 * Must only be used if we don't need to care about links, usually from
1631 * within the completion handling itself.
1633 static void __io_double_put_req(struct io_kiocb *req)
1635 /* drop both submit and complete references */
1636 if (refcount_sub_and_test(2, &req->refs))
1640 static void io_double_put_req(struct io_kiocb *req)
1642 /* drop both submit and complete references */
1643 if (refcount_sub_and_test(2, &req->refs))
1647 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1649 struct io_rings *rings = ctx->rings;
1651 if (test_bit(0, &ctx->cq_check_overflow)) {
1653 * noflush == true is from the waitqueue handler, just ensure
1654 * we wake up the task, and the next invocation will flush the
1655 * entries. We cannot safely to it from here.
1657 if (noflush && !list_empty(&ctx->cq_overflow_list))
1660 io_cqring_overflow_flush(ctx, false);
1663 /* See comment at the top of this file */
1665 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1668 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1670 struct io_rings *rings = ctx->rings;
1672 /* make sure SQ entry isn't read before tail */
1673 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1676 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1678 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1681 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1684 rb->reqs[rb->to_free++] = req;
1685 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1686 io_free_req_many(req->ctx, rb);
1690 static int io_put_kbuf(struct io_kiocb *req)
1692 struct io_buffer *kbuf;
1695 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1696 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1697 cflags |= IORING_CQE_F_BUFFER;
1704 * Find and free completed poll iocbs
1706 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1707 struct list_head *done)
1709 struct req_batch rb;
1710 struct io_kiocb *req;
1712 rb.to_free = rb.need_iter = 0;
1713 while (!list_empty(done)) {
1716 req = list_first_entry(done, struct io_kiocb, list);
1717 list_del(&req->list);
1719 if (req->flags & REQ_F_BUFFER_SELECTED)
1720 cflags = io_put_kbuf(req);
1722 __io_cqring_fill_event(req, req->result, cflags);
1725 if (refcount_dec_and_test(&req->refs) &&
1726 !io_req_multi_free(&rb, req))
1730 io_commit_cqring(ctx);
1731 if (ctx->flags & IORING_SETUP_SQPOLL)
1732 io_cqring_ev_posted(ctx);
1733 io_free_req_many(ctx, &rb);
1736 static void io_iopoll_queue(struct list_head *again)
1738 struct io_kiocb *req;
1741 req = list_first_entry(again, struct io_kiocb, list);
1742 list_del(&req->list);
1743 refcount_inc(&req->refs);
1744 io_queue_async_work(req);
1745 } while (!list_empty(again));
1748 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1751 struct io_kiocb *req, *tmp;
1758 * Only spin for completions if we don't have multiple devices hanging
1759 * off our complete list, and we're under the requested amount.
1761 spin = !ctx->poll_multi_file && *nr_events < min;
1764 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1765 struct kiocb *kiocb = &req->rw.kiocb;
1768 * Move completed and retryable entries to our local lists.
1769 * If we find a request that requires polling, break out
1770 * and complete those lists first, if we have entries there.
1772 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1773 list_move_tail(&req->list, &done);
1776 if (!list_empty(&done))
1779 if (req->result == -EAGAIN) {
1780 list_move_tail(&req->list, &again);
1783 if (!list_empty(&again))
1786 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1795 if (!list_empty(&done))
1796 io_iopoll_complete(ctx, nr_events, &done);
1798 if (!list_empty(&again))
1799 io_iopoll_queue(&again);
1805 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1806 * non-spinning poll check - we'll still enter the driver poll loop, but only
1807 * as a non-spinning completion check.
1809 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1812 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1815 ret = io_do_iopoll(ctx, nr_events, min);
1818 if (!min || *nr_events >= min)
1826 * We can't just wait for polled events to come to us, we have to actively
1827 * find and complete them.
1829 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1831 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1834 mutex_lock(&ctx->uring_lock);
1835 while (!list_empty(&ctx->poll_list)) {
1836 unsigned int nr_events = 0;
1838 io_iopoll_getevents(ctx, &nr_events, 1);
1841 * Ensure we allow local-to-the-cpu processing to take place,
1842 * in this case we need to ensure that we reap all events.
1846 mutex_unlock(&ctx->uring_lock);
1849 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1852 int iters = 0, ret = 0;
1855 * We disallow the app entering submit/complete with polling, but we
1856 * still need to lock the ring to prevent racing with polled issue
1857 * that got punted to a workqueue.
1859 mutex_lock(&ctx->uring_lock);
1864 * Don't enter poll loop if we already have events pending.
1865 * If we do, we can potentially be spinning for commands that
1866 * already triggered a CQE (eg in error).
1868 if (io_cqring_events(ctx, false))
1872 * If a submit got punted to a workqueue, we can have the
1873 * application entering polling for a command before it gets
1874 * issued. That app will hold the uring_lock for the duration
1875 * of the poll right here, so we need to take a breather every
1876 * now and then to ensure that the issue has a chance to add
1877 * the poll to the issued list. Otherwise we can spin here
1878 * forever, while the workqueue is stuck trying to acquire the
1881 if (!(++iters & 7)) {
1882 mutex_unlock(&ctx->uring_lock);
1883 mutex_lock(&ctx->uring_lock);
1886 if (*nr_events < min)
1887 tmin = min - *nr_events;
1889 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1893 } while (min && !*nr_events && !need_resched());
1895 mutex_unlock(&ctx->uring_lock);
1899 static void kiocb_end_write(struct io_kiocb *req)
1902 * Tell lockdep we inherited freeze protection from submission
1905 if (req->flags & REQ_F_ISREG) {
1906 struct inode *inode = file_inode(req->file);
1908 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1910 file_end_write(req->file);
1913 static inline void req_set_fail_links(struct io_kiocb *req)
1915 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1916 req->flags |= REQ_F_FAIL_LINK;
1919 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1921 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1924 if (kiocb->ki_flags & IOCB_WRITE)
1925 kiocb_end_write(req);
1927 if (res != req->result)
1928 req_set_fail_links(req);
1929 if (req->flags & REQ_F_BUFFER_SELECTED)
1930 cflags = io_put_kbuf(req);
1931 __io_cqring_add_event(req, res, cflags);
1934 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1936 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1938 io_complete_rw_common(kiocb, res);
1942 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1944 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1946 if (kiocb->ki_flags & IOCB_WRITE)
1947 kiocb_end_write(req);
1949 if (res != req->result)
1950 req_set_fail_links(req);
1953 req->flags |= REQ_F_IOPOLL_COMPLETED;
1957 * After the iocb has been issued, it's safe to be found on the poll list.
1958 * Adding the kiocb to the list AFTER submission ensures that we don't
1959 * find it from a io_iopoll_getevents() thread before the issuer is done
1960 * accessing the kiocb cookie.
1962 static void io_iopoll_req_issued(struct io_kiocb *req)
1964 struct io_ring_ctx *ctx = req->ctx;
1967 * Track whether we have multiple files in our lists. This will impact
1968 * how we do polling eventually, not spinning if we're on potentially
1969 * different devices.
1971 if (list_empty(&ctx->poll_list)) {
1972 ctx->poll_multi_file = false;
1973 } else if (!ctx->poll_multi_file) {
1974 struct io_kiocb *list_req;
1976 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1978 if (list_req->file != req->file)
1979 ctx->poll_multi_file = true;
1983 * For fast devices, IO may have already completed. If it has, add
1984 * it to the front so we find it first.
1986 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1987 list_add(&req->list, &ctx->poll_list);
1989 list_add_tail(&req->list, &ctx->poll_list);
1991 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1992 wq_has_sleeper(&ctx->sqo_wait))
1993 wake_up(&ctx->sqo_wait);
1996 static void io_file_put(struct io_submit_state *state)
1999 int diff = state->has_refs - state->used_refs;
2002 fput_many(state->file, diff);
2008 * Get as many references to a file as we have IOs left in this submission,
2009 * assuming most submissions are for one file, or at least that each file
2010 * has more than one submission.
2012 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2018 if (state->fd == fd) {
2025 state->file = fget_many(fd, state->ios_left);
2030 state->has_refs = state->ios_left;
2031 state->used_refs = 1;
2037 * If we tracked the file through the SCM inflight mechanism, we could support
2038 * any file. For now, just ensure that anything potentially problematic is done
2041 static bool io_file_supports_async(struct file *file, int rw)
2043 umode_t mode = file_inode(file)->i_mode;
2045 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2047 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2050 if (!(file->f_mode & FMODE_NOWAIT))
2054 return file->f_op->read_iter != NULL;
2056 return file->f_op->write_iter != NULL;
2059 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2060 bool force_nonblock)
2062 struct io_ring_ctx *ctx = req->ctx;
2063 struct kiocb *kiocb = &req->rw.kiocb;
2067 if (S_ISREG(file_inode(req->file)->i_mode))
2068 req->flags |= REQ_F_ISREG;
2070 kiocb->ki_pos = READ_ONCE(sqe->off);
2071 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2072 req->flags |= REQ_F_CUR_POS;
2073 kiocb->ki_pos = req->file->f_pos;
2075 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2076 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2077 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2081 ioprio = READ_ONCE(sqe->ioprio);
2083 ret = ioprio_check_cap(ioprio);
2087 kiocb->ki_ioprio = ioprio;
2089 kiocb->ki_ioprio = get_current_ioprio();
2091 /* don't allow async punt if RWF_NOWAIT was requested */
2092 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2093 (req->file->f_flags & O_NONBLOCK))
2094 req->flags |= REQ_F_NOWAIT;
2097 kiocb->ki_flags |= IOCB_NOWAIT;
2099 if (ctx->flags & IORING_SETUP_IOPOLL) {
2100 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2101 !kiocb->ki_filp->f_op->iopoll)
2104 kiocb->ki_flags |= IOCB_HIPRI;
2105 kiocb->ki_complete = io_complete_rw_iopoll;
2108 if (kiocb->ki_flags & IOCB_HIPRI)
2110 kiocb->ki_complete = io_complete_rw;
2113 req->rw.addr = READ_ONCE(sqe->addr);
2114 req->rw.len = READ_ONCE(sqe->len);
2115 /* we own ->private, reuse it for the buffer index / buffer ID */
2116 req->rw.kiocb.private = (void *) (unsigned long)
2117 READ_ONCE(sqe->buf_index);
2121 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2127 case -ERESTARTNOINTR:
2128 case -ERESTARTNOHAND:
2129 case -ERESTART_RESTARTBLOCK:
2131 * We can't just restart the syscall, since previously
2132 * submitted sqes may already be in progress. Just fail this
2138 kiocb->ki_complete(kiocb, ret, 0);
2142 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2144 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2146 if (req->flags & REQ_F_CUR_POS)
2147 req->file->f_pos = kiocb->ki_pos;
2148 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2149 io_complete_rw(kiocb, ret, 0);
2151 io_rw_done(kiocb, ret);
2154 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2155 struct iov_iter *iter)
2157 struct io_ring_ctx *ctx = req->ctx;
2158 size_t len = req->rw.len;
2159 struct io_mapped_ubuf *imu;
2160 unsigned index, buf_index;
2164 /* attempt to use fixed buffers without having provided iovecs */
2165 if (unlikely(!ctx->user_bufs))
2168 buf_index = (unsigned long) req->rw.kiocb.private;
2169 if (unlikely(buf_index >= ctx->nr_user_bufs))
2172 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2173 imu = &ctx->user_bufs[index];
2174 buf_addr = req->rw.addr;
2177 if (buf_addr + len < buf_addr)
2179 /* not inside the mapped region */
2180 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2184 * May not be a start of buffer, set size appropriately
2185 * and advance us to the beginning.
2187 offset = buf_addr - imu->ubuf;
2188 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2192 * Don't use iov_iter_advance() here, as it's really slow for
2193 * using the latter parts of a big fixed buffer - it iterates
2194 * over each segment manually. We can cheat a bit here, because
2197 * 1) it's a BVEC iter, we set it up
2198 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2199 * first and last bvec
2201 * So just find our index, and adjust the iterator afterwards.
2202 * If the offset is within the first bvec (or the whole first
2203 * bvec, just use iov_iter_advance(). This makes it easier
2204 * since we can just skip the first segment, which may not
2205 * be PAGE_SIZE aligned.
2207 const struct bio_vec *bvec = imu->bvec;
2209 if (offset <= bvec->bv_len) {
2210 iov_iter_advance(iter, offset);
2212 unsigned long seg_skip;
2214 /* skip first vec */
2215 offset -= bvec->bv_len;
2216 seg_skip = 1 + (offset >> PAGE_SHIFT);
2218 iter->bvec = bvec + seg_skip;
2219 iter->nr_segs -= seg_skip;
2220 iter->count -= bvec->bv_len + offset;
2221 iter->iov_offset = offset & ~PAGE_MASK;
2228 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2231 mutex_unlock(&ctx->uring_lock);
2234 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2237 * "Normal" inline submissions always hold the uring_lock, since we
2238 * grab it from the system call. Same is true for the SQPOLL offload.
2239 * The only exception is when we've detached the request and issue it
2240 * from an async worker thread, grab the lock for that case.
2243 mutex_lock(&ctx->uring_lock);
2246 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2247 int bgid, struct io_buffer *kbuf,
2250 struct io_buffer *head;
2252 if (req->flags & REQ_F_BUFFER_SELECTED)
2255 io_ring_submit_lock(req->ctx, needs_lock);
2257 lockdep_assert_held(&req->ctx->uring_lock);
2259 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2261 if (!list_empty(&head->list)) {
2262 kbuf = list_last_entry(&head->list, struct io_buffer,
2264 list_del(&kbuf->list);
2267 idr_remove(&req->ctx->io_buffer_idr, bgid);
2269 if (*len > kbuf->len)
2272 kbuf = ERR_PTR(-ENOBUFS);
2275 io_ring_submit_unlock(req->ctx, needs_lock);
2280 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2283 struct io_buffer *kbuf;
2286 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2287 bgid = (int) (unsigned long) req->rw.kiocb.private;
2288 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2291 req->rw.addr = (u64) (unsigned long) kbuf;
2292 req->flags |= REQ_F_BUFFER_SELECTED;
2293 return u64_to_user_ptr(kbuf->addr);
2296 #ifdef CONFIG_COMPAT
2297 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2300 struct compat_iovec __user *uiov;
2301 compat_ssize_t clen;
2305 uiov = u64_to_user_ptr(req->rw.addr);
2306 if (!access_ok(uiov, sizeof(*uiov)))
2308 if (__get_user(clen, &uiov->iov_len))
2314 buf = io_rw_buffer_select(req, &len, needs_lock);
2316 return PTR_ERR(buf);
2317 iov[0].iov_base = buf;
2318 iov[0].iov_len = (compat_size_t) len;
2323 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2326 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2330 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2333 len = iov[0].iov_len;
2336 buf = io_rw_buffer_select(req, &len, needs_lock);
2338 return PTR_ERR(buf);
2339 iov[0].iov_base = buf;
2340 iov[0].iov_len = len;
2344 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2347 if (req->flags & REQ_F_BUFFER_SELECTED)
2351 else if (req->rw.len > 1)
2354 #ifdef CONFIG_COMPAT
2355 if (req->ctx->compat)
2356 return io_compat_import(req, iov, needs_lock);
2359 return __io_iov_buffer_select(req, iov, needs_lock);
2362 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2363 struct iovec **iovec, struct iov_iter *iter,
2366 void __user *buf = u64_to_user_ptr(req->rw.addr);
2367 size_t sqe_len = req->rw.len;
2371 opcode = req->opcode;
2372 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2374 return io_import_fixed(req, rw, iter);
2377 /* buffer index only valid with fixed read/write, or buffer select */
2378 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2381 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2382 if (req->flags & REQ_F_BUFFER_SELECT) {
2383 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2386 return PTR_ERR(buf);
2388 req->rw.len = sqe_len;
2391 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2393 return ret < 0 ? ret : sqe_len;
2397 struct io_async_rw *iorw = &req->io->rw;
2400 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2401 if (iorw->iov == iorw->fast_iov)
2406 if (req->flags & REQ_F_BUFFER_SELECT) {
2407 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2409 ret = (*iovec)->iov_len;
2410 iov_iter_init(iter, rw, *iovec, 1, ret);
2416 #ifdef CONFIG_COMPAT
2417 if (req->ctx->compat)
2418 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2422 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2426 * For files that don't have ->read_iter() and ->write_iter(), handle them
2427 * by looping over ->read() or ->write() manually.
2429 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2430 struct iov_iter *iter)
2435 * Don't support polled IO through this interface, and we can't
2436 * support non-blocking either. For the latter, this just causes
2437 * the kiocb to be handled from an async context.
2439 if (kiocb->ki_flags & IOCB_HIPRI)
2441 if (kiocb->ki_flags & IOCB_NOWAIT)
2444 while (iov_iter_count(iter)) {
2448 if (!iov_iter_is_bvec(iter)) {
2449 iovec = iov_iter_iovec(iter);
2451 /* fixed buffers import bvec */
2452 iovec.iov_base = kmap(iter->bvec->bv_page)
2454 iovec.iov_len = min(iter->count,
2455 iter->bvec->bv_len - iter->iov_offset);
2459 nr = file->f_op->read(file, iovec.iov_base,
2460 iovec.iov_len, &kiocb->ki_pos);
2462 nr = file->f_op->write(file, iovec.iov_base,
2463 iovec.iov_len, &kiocb->ki_pos);
2466 if (iov_iter_is_bvec(iter))
2467 kunmap(iter->bvec->bv_page);
2475 if (nr != iovec.iov_len)
2477 iov_iter_advance(iter, nr);
2483 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2484 struct iovec *iovec, struct iovec *fast_iov,
2485 struct iov_iter *iter)
2487 req->io->rw.nr_segs = iter->nr_segs;
2488 req->io->rw.size = io_size;
2489 req->io->rw.iov = iovec;
2490 if (!req->io->rw.iov) {
2491 req->io->rw.iov = req->io->rw.fast_iov;
2492 if (req->io->rw.iov != fast_iov)
2493 memcpy(req->io->rw.iov, fast_iov,
2494 sizeof(struct iovec) * iter->nr_segs);
2496 req->flags |= REQ_F_NEED_CLEANUP;
2500 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2502 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2503 return req->io == NULL;
2506 static int io_alloc_async_ctx(struct io_kiocb *req)
2508 if (!io_op_defs[req->opcode].async_ctx)
2511 return __io_alloc_async_ctx(req);
2514 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2515 struct iovec *iovec, struct iovec *fast_iov,
2516 struct iov_iter *iter)
2518 if (!io_op_defs[req->opcode].async_ctx)
2521 if (__io_alloc_async_ctx(req))
2524 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2529 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2530 bool force_nonblock)
2532 struct io_async_ctx *io;
2533 struct iov_iter iter;
2536 ret = io_prep_rw(req, sqe, force_nonblock);
2540 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2543 /* either don't need iovec imported or already have it */
2544 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2548 io->rw.iov = io->rw.fast_iov;
2550 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2555 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2559 static int io_read(struct io_kiocb *req, bool force_nonblock)
2561 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2562 struct kiocb *kiocb = &req->rw.kiocb;
2563 struct iov_iter iter;
2565 ssize_t io_size, ret;
2567 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2571 /* Ensure we clear previously set non-block flag */
2572 if (!force_nonblock)
2573 kiocb->ki_flags &= ~IOCB_NOWAIT;
2577 if (req->flags & REQ_F_LINK_HEAD)
2578 req->result = io_size;
2581 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2582 * we know to async punt it even if it was opened O_NONBLOCK
2584 if (force_nonblock && !io_file_supports_async(req->file, READ))
2587 iov_count = iov_iter_count(&iter);
2588 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2592 if (req->file->f_op->read_iter)
2593 ret2 = call_read_iter(req->file, kiocb, &iter);
2595 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2597 /* Catch -EAGAIN return for forced non-blocking submission */
2598 if (!force_nonblock || ret2 != -EAGAIN) {
2599 kiocb_done(kiocb, ret2);
2602 ret = io_setup_async_rw(req, io_size, iovec,
2603 inline_vecs, &iter);
2606 /* any defer here is final, must blocking retry */
2607 if (!(req->flags & REQ_F_NOWAIT))
2608 req->flags |= REQ_F_MUST_PUNT;
2614 req->flags &= ~REQ_F_NEED_CLEANUP;
2618 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2619 bool force_nonblock)
2621 struct io_async_ctx *io;
2622 struct iov_iter iter;
2625 ret = io_prep_rw(req, sqe, force_nonblock);
2629 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2632 req->fsize = rlimit(RLIMIT_FSIZE);
2634 /* either don't need iovec imported or already have it */
2635 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2639 io->rw.iov = io->rw.fast_iov;
2641 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2646 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2650 static int io_write(struct io_kiocb *req, bool force_nonblock)
2652 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2653 struct kiocb *kiocb = &req->rw.kiocb;
2654 struct iov_iter iter;
2656 ssize_t ret, io_size;
2658 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2662 /* Ensure we clear previously set non-block flag */
2663 if (!force_nonblock)
2664 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2668 if (req->flags & REQ_F_LINK_HEAD)
2669 req->result = io_size;
2672 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2673 * we know to async punt it even if it was opened O_NONBLOCK
2675 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2678 /* file path doesn't support NOWAIT for non-direct_IO */
2679 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2680 (req->flags & REQ_F_ISREG))
2683 iov_count = iov_iter_count(&iter);
2684 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2689 * Open-code file_start_write here to grab freeze protection,
2690 * which will be released by another thread in
2691 * io_complete_rw(). Fool lockdep by telling it the lock got
2692 * released so that it doesn't complain about the held lock when
2693 * we return to userspace.
2695 if (req->flags & REQ_F_ISREG) {
2696 __sb_start_write(file_inode(req->file)->i_sb,
2697 SB_FREEZE_WRITE, true);
2698 __sb_writers_release(file_inode(req->file)->i_sb,
2701 kiocb->ki_flags |= IOCB_WRITE;
2703 if (!force_nonblock)
2704 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2706 if (req->file->f_op->write_iter)
2707 ret2 = call_write_iter(req->file, kiocb, &iter);
2709 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2711 if (!force_nonblock)
2712 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2715 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2716 * retry them without IOCB_NOWAIT.
2718 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2720 if (!force_nonblock || ret2 != -EAGAIN) {
2721 kiocb_done(kiocb, ret2);
2724 ret = io_setup_async_rw(req, io_size, iovec,
2725 inline_vecs, &iter);
2728 /* any defer here is final, must blocking retry */
2729 req->flags |= REQ_F_MUST_PUNT;
2734 req->flags &= ~REQ_F_NEED_CLEANUP;
2739 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2741 struct io_splice* sp = &req->splice;
2742 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2745 if (req->flags & REQ_F_NEED_CLEANUP)
2749 sp->off_in = READ_ONCE(sqe->splice_off_in);
2750 sp->off_out = READ_ONCE(sqe->off);
2751 sp->len = READ_ONCE(sqe->len);
2752 sp->flags = READ_ONCE(sqe->splice_flags);
2754 if (unlikely(sp->flags & ~valid_flags))
2757 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2758 (sp->flags & SPLICE_F_FD_IN_FIXED));
2761 req->flags |= REQ_F_NEED_CLEANUP;
2763 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2764 req->work.flags |= IO_WQ_WORK_UNBOUND;
2769 static bool io_splice_punt(struct file *file, int rw)
2771 if (get_pipe_info(file))
2773 if (!io_file_supports_async(file, rw))
2775 return !(file->f_flags & O_NONBLOCK);
2778 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2780 struct io_splice *sp = &req->splice;
2781 struct file *in = sp->file_in;
2782 struct file *out = sp->file_out;
2783 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2784 loff_t *poff_in, *poff_out;
2787 if (force_nonblock) {
2788 if (io_splice_punt(in, READ) || io_splice_punt(out, WRITE))
2790 flags |= SPLICE_F_NONBLOCK;
2793 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2794 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2795 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2796 if (force_nonblock && ret == -EAGAIN)
2799 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2800 req->flags &= ~REQ_F_NEED_CLEANUP;
2802 io_cqring_add_event(req, ret);
2804 req_set_fail_links(req);
2810 * IORING_OP_NOP just posts a completion event, nothing else.
2812 static int io_nop(struct io_kiocb *req)
2814 struct io_ring_ctx *ctx = req->ctx;
2816 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2819 io_cqring_add_event(req, 0);
2824 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2826 struct io_ring_ctx *ctx = req->ctx;
2831 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2833 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2836 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2837 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2840 req->sync.off = READ_ONCE(sqe->off);
2841 req->sync.len = READ_ONCE(sqe->len);
2845 static bool io_req_cancelled(struct io_kiocb *req)
2847 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2848 req_set_fail_links(req);
2849 io_cqring_add_event(req, -ECANCELED);
2857 static void __io_fsync(struct io_kiocb *req)
2859 loff_t end = req->sync.off + req->sync.len;
2862 ret = vfs_fsync_range(req->file, req->sync.off,
2863 end > 0 ? end : LLONG_MAX,
2864 req->sync.flags & IORING_FSYNC_DATASYNC);
2866 req_set_fail_links(req);
2867 io_cqring_add_event(req, ret);
2871 static void io_fsync_finish(struct io_wq_work **workptr)
2873 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2875 if (io_req_cancelled(req))
2878 io_steal_work(req, workptr);
2881 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2883 /* fsync always requires a blocking context */
2884 if (force_nonblock) {
2885 req->work.func = io_fsync_finish;
2892 static void __io_fallocate(struct io_kiocb *req)
2896 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2897 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2899 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2901 req_set_fail_links(req);
2902 io_cqring_add_event(req, ret);
2906 static void io_fallocate_finish(struct io_wq_work **workptr)
2908 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2910 if (io_req_cancelled(req))
2912 __io_fallocate(req);
2913 io_steal_work(req, workptr);
2916 static int io_fallocate_prep(struct io_kiocb *req,
2917 const struct io_uring_sqe *sqe)
2919 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2922 req->sync.off = READ_ONCE(sqe->off);
2923 req->sync.len = READ_ONCE(sqe->addr);
2924 req->sync.mode = READ_ONCE(sqe->len);
2925 req->fsize = rlimit(RLIMIT_FSIZE);
2929 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2931 /* fallocate always requiring blocking context */
2932 if (force_nonblock) {
2933 req->work.func = io_fallocate_finish;
2937 __io_fallocate(req);
2941 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2943 const char __user *fname;
2946 if (sqe->ioprio || sqe->buf_index)
2948 if (req->flags & REQ_F_FIXED_FILE)
2950 if (req->flags & REQ_F_NEED_CLEANUP)
2953 req->open.dfd = READ_ONCE(sqe->fd);
2954 req->open.how.mode = READ_ONCE(sqe->len);
2955 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2956 req->open.how.flags = READ_ONCE(sqe->open_flags);
2957 if (force_o_largefile())
2958 req->open.how.flags |= O_LARGEFILE;
2960 req->open.filename = getname(fname);
2961 if (IS_ERR(req->open.filename)) {
2962 ret = PTR_ERR(req->open.filename);
2963 req->open.filename = NULL;
2967 req->open.nofile = rlimit(RLIMIT_NOFILE);
2968 req->flags |= REQ_F_NEED_CLEANUP;
2972 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2974 struct open_how __user *how;
2975 const char __user *fname;
2979 if (sqe->ioprio || sqe->buf_index)
2981 if (req->flags & REQ_F_FIXED_FILE)
2983 if (req->flags & REQ_F_NEED_CLEANUP)
2986 req->open.dfd = READ_ONCE(sqe->fd);
2987 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2988 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2989 len = READ_ONCE(sqe->len);
2991 if (len < OPEN_HOW_SIZE_VER0)
2994 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2999 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3000 req->open.how.flags |= O_LARGEFILE;
3002 req->open.filename = getname(fname);
3003 if (IS_ERR(req->open.filename)) {
3004 ret = PTR_ERR(req->open.filename);
3005 req->open.filename = NULL;
3009 req->open.nofile = rlimit(RLIMIT_NOFILE);
3010 req->flags |= REQ_F_NEED_CLEANUP;
3014 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3016 struct open_flags op;
3023 ret = build_open_flags(&req->open.how, &op);
3027 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3031 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3034 ret = PTR_ERR(file);
3036 fsnotify_open(file);
3037 fd_install(ret, file);
3040 putname(req->open.filename);
3041 req->flags &= ~REQ_F_NEED_CLEANUP;
3043 req_set_fail_links(req);
3044 io_cqring_add_event(req, ret);
3049 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3051 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3052 return io_openat2(req, force_nonblock);
3055 static int io_remove_buffers_prep(struct io_kiocb *req,
3056 const struct io_uring_sqe *sqe)
3058 struct io_provide_buf *p = &req->pbuf;
3061 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3064 tmp = READ_ONCE(sqe->fd);
3065 if (!tmp || tmp > USHRT_MAX)
3068 memset(p, 0, sizeof(*p));
3070 p->bgid = READ_ONCE(sqe->buf_group);
3074 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3075 int bgid, unsigned nbufs)
3079 /* shouldn't happen */
3083 /* the head kbuf is the list itself */
3084 while (!list_empty(&buf->list)) {
3085 struct io_buffer *nxt;
3087 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3088 list_del(&nxt->list);
3095 idr_remove(&ctx->io_buffer_idr, bgid);
3100 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3102 struct io_provide_buf *p = &req->pbuf;
3103 struct io_ring_ctx *ctx = req->ctx;
3104 struct io_buffer *head;
3107 io_ring_submit_lock(ctx, !force_nonblock);
3109 lockdep_assert_held(&ctx->uring_lock);
3112 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3114 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3116 io_ring_submit_lock(ctx, !force_nonblock);
3118 req_set_fail_links(req);
3119 io_cqring_add_event(req, ret);
3124 static int io_provide_buffers_prep(struct io_kiocb *req,
3125 const struct io_uring_sqe *sqe)
3127 struct io_provide_buf *p = &req->pbuf;
3130 if (sqe->ioprio || sqe->rw_flags)
3133 tmp = READ_ONCE(sqe->fd);
3134 if (!tmp || tmp > USHRT_MAX)
3137 p->addr = READ_ONCE(sqe->addr);
3138 p->len = READ_ONCE(sqe->len);
3140 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3143 p->bgid = READ_ONCE(sqe->buf_group);
3144 tmp = READ_ONCE(sqe->off);
3145 if (tmp > USHRT_MAX)
3151 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3153 struct io_buffer *buf;
3154 u64 addr = pbuf->addr;
3155 int i, bid = pbuf->bid;
3157 for (i = 0; i < pbuf->nbufs; i++) {
3158 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3163 buf->len = pbuf->len;
3168 INIT_LIST_HEAD(&buf->list);
3171 list_add_tail(&buf->list, &(*head)->list);
3175 return i ? i : -ENOMEM;
3178 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3180 struct io_provide_buf *p = &req->pbuf;
3181 struct io_ring_ctx *ctx = req->ctx;
3182 struct io_buffer *head, *list;
3185 io_ring_submit_lock(ctx, !force_nonblock);
3187 lockdep_assert_held(&ctx->uring_lock);
3189 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3191 ret = io_add_buffers(p, &head);
3196 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3199 __io_remove_buffers(ctx, head, p->bgid, -1U);
3204 io_ring_submit_unlock(ctx, !force_nonblock);
3206 req_set_fail_links(req);
3207 io_cqring_add_event(req, ret);
3212 static int io_epoll_ctl_prep(struct io_kiocb *req,
3213 const struct io_uring_sqe *sqe)
3215 #if defined(CONFIG_EPOLL)
3216 if (sqe->ioprio || sqe->buf_index)
3219 req->epoll.epfd = READ_ONCE(sqe->fd);
3220 req->epoll.op = READ_ONCE(sqe->len);
3221 req->epoll.fd = READ_ONCE(sqe->off);
3223 if (ep_op_has_event(req->epoll.op)) {
3224 struct epoll_event __user *ev;
3226 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3227 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3237 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3239 #if defined(CONFIG_EPOLL)
3240 struct io_epoll *ie = &req->epoll;
3243 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3244 if (force_nonblock && ret == -EAGAIN)
3248 req_set_fail_links(req);
3249 io_cqring_add_event(req, ret);
3257 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3259 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3260 if (sqe->ioprio || sqe->buf_index || sqe->off)
3263 req->madvise.addr = READ_ONCE(sqe->addr);
3264 req->madvise.len = READ_ONCE(sqe->len);
3265 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3272 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3274 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3275 struct io_madvise *ma = &req->madvise;
3281 ret = do_madvise(ma->addr, ma->len, ma->advice);
3283 req_set_fail_links(req);
3284 io_cqring_add_event(req, ret);
3292 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3294 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3297 req->fadvise.offset = READ_ONCE(sqe->off);
3298 req->fadvise.len = READ_ONCE(sqe->len);
3299 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3303 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3305 struct io_fadvise *fa = &req->fadvise;
3308 if (force_nonblock) {
3309 switch (fa->advice) {
3310 case POSIX_FADV_NORMAL:
3311 case POSIX_FADV_RANDOM:
3312 case POSIX_FADV_SEQUENTIAL:
3319 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3321 req_set_fail_links(req);
3322 io_cqring_add_event(req, ret);
3327 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3329 const char __user *fname;
3330 unsigned lookup_flags;
3333 if (sqe->ioprio || sqe->buf_index)
3335 if (req->flags & REQ_F_FIXED_FILE)
3337 if (req->flags & REQ_F_NEED_CLEANUP)
3340 req->open.dfd = READ_ONCE(sqe->fd);
3341 req->open.mask = READ_ONCE(sqe->len);
3342 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3343 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3344 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3346 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3349 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3350 if (IS_ERR(req->open.filename)) {
3351 ret = PTR_ERR(req->open.filename);
3352 req->open.filename = NULL;
3356 req->flags |= REQ_F_NEED_CLEANUP;
3360 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3362 struct io_open *ctx = &req->open;
3363 unsigned lookup_flags;
3368 if (force_nonblock) {
3369 /* only need file table for an actual valid fd */
3370 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3371 req->flags |= REQ_F_NO_FILE_TABLE;
3375 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3379 /* filename_lookup() drops it, keep a reference */
3380 ctx->filename->refcnt++;
3382 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3387 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3389 if (retry_estale(ret, lookup_flags)) {
3390 lookup_flags |= LOOKUP_REVAL;
3394 ret = cp_statx(&stat, ctx->buffer);
3396 putname(ctx->filename);
3397 req->flags &= ~REQ_F_NEED_CLEANUP;
3399 req_set_fail_links(req);
3400 io_cqring_add_event(req, ret);
3405 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3408 * If we queue this for async, it must not be cancellable. That would
3409 * leave the 'file' in an undeterminate state.
3411 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3413 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3414 sqe->rw_flags || sqe->buf_index)
3416 if (req->flags & REQ_F_FIXED_FILE)
3419 req->close.fd = READ_ONCE(sqe->fd);
3420 if (req->file->f_op == &io_uring_fops ||
3421 req->close.fd == req->ctx->ring_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);
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 /* submission ref will be dropped, take it for async */
3461 refcount_inc(&req->refs);
3463 req->work.func = io_close_finish;
3465 * Do manual async queue here to avoid grabbing files - we don't
3466 * need the files, and it'll cause io_close_finish() to close
3467 * the file again and cause a double CQE entry for this request
3469 io_queue_async_work(req);
3474 * No ->flush(), safely close from here and just punt the
3475 * fput() to async context.
3477 __io_close_finish(req);
3481 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3483 struct io_ring_ctx *ctx = req->ctx;
3488 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3490 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3493 req->sync.off = READ_ONCE(sqe->off);
3494 req->sync.len = READ_ONCE(sqe->len);
3495 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3499 static void __io_sync_file_range(struct io_kiocb *req)
3503 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3506 req_set_fail_links(req);
3507 io_cqring_add_event(req, ret);
3512 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3514 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3516 if (io_req_cancelled(req))
3518 __io_sync_file_range(req);
3519 io_put_req(req); /* put submission ref */
3522 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3524 /* sync_file_range always requires a blocking context */
3525 if (force_nonblock) {
3526 req->work.func = io_sync_file_range_finish;
3530 __io_sync_file_range(req);
3534 #if defined(CONFIG_NET)
3535 static int io_setup_async_msg(struct io_kiocb *req,
3536 struct io_async_msghdr *kmsg)
3540 if (io_alloc_async_ctx(req)) {
3541 if (kmsg->iov != kmsg->fast_iov)
3545 req->flags |= REQ_F_NEED_CLEANUP;
3546 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3550 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3552 struct io_sr_msg *sr = &req->sr_msg;
3553 struct io_async_ctx *io = req->io;
3556 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3557 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3558 sr->len = READ_ONCE(sqe->len);
3560 #ifdef CONFIG_COMPAT
3561 if (req->ctx->compat)
3562 sr->msg_flags |= MSG_CMSG_COMPAT;
3565 if (!io || req->opcode == IORING_OP_SEND)
3567 /* iovec is already imported */
3568 if (req->flags & REQ_F_NEED_CLEANUP)
3571 io->msg.iov = io->msg.fast_iov;
3572 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3575 req->flags |= REQ_F_NEED_CLEANUP;
3579 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3581 struct io_async_msghdr *kmsg = NULL;
3582 struct socket *sock;
3585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3588 sock = sock_from_file(req->file, &ret);
3590 struct io_async_ctx io;
3594 kmsg = &req->io->msg;
3595 kmsg->msg.msg_name = &req->io->msg.addr;
3596 /* if iov is set, it's allocated already */
3598 kmsg->iov = kmsg->fast_iov;
3599 kmsg->msg.msg_iter.iov = kmsg->iov;
3601 struct io_sr_msg *sr = &req->sr_msg;
3604 kmsg->msg.msg_name = &io.msg.addr;
3606 io.msg.iov = io.msg.fast_iov;
3607 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3608 sr->msg_flags, &io.msg.iov);
3613 flags = req->sr_msg.msg_flags;
3614 if (flags & MSG_DONTWAIT)
3615 req->flags |= REQ_F_NOWAIT;
3616 else if (force_nonblock)
3617 flags |= MSG_DONTWAIT;
3619 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3620 if (force_nonblock && ret == -EAGAIN)
3621 return io_setup_async_msg(req, kmsg);
3622 if (ret == -ERESTARTSYS)
3626 if (kmsg && kmsg->iov != kmsg->fast_iov)
3628 req->flags &= ~REQ_F_NEED_CLEANUP;
3629 io_cqring_add_event(req, ret);
3631 req_set_fail_links(req);
3636 static int io_send(struct io_kiocb *req, bool force_nonblock)
3638 struct socket *sock;
3641 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3644 sock = sock_from_file(req->file, &ret);
3646 struct io_sr_msg *sr = &req->sr_msg;
3651 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3656 msg.msg_name = NULL;
3657 msg.msg_control = NULL;
3658 msg.msg_controllen = 0;
3659 msg.msg_namelen = 0;
3661 flags = req->sr_msg.msg_flags;
3662 if (flags & MSG_DONTWAIT)
3663 req->flags |= REQ_F_NOWAIT;
3664 else if (force_nonblock)
3665 flags |= MSG_DONTWAIT;
3667 msg.msg_flags = flags;
3668 ret = sock_sendmsg(sock, &msg);
3669 if (force_nonblock && ret == -EAGAIN)
3671 if (ret == -ERESTARTSYS)
3675 io_cqring_add_event(req, ret);
3677 req_set_fail_links(req);
3682 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3684 struct io_sr_msg *sr = &req->sr_msg;
3685 struct iovec __user *uiov;
3689 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3694 if (req->flags & REQ_F_BUFFER_SELECT) {
3697 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3699 sr->len = io->msg.iov[0].iov_len;
3700 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3704 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3705 &io->msg.iov, &io->msg.msg.msg_iter);
3713 #ifdef CONFIG_COMPAT
3714 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3715 struct io_async_ctx *io)
3717 struct compat_msghdr __user *msg_compat;
3718 struct io_sr_msg *sr = &req->sr_msg;
3719 struct compat_iovec __user *uiov;
3724 msg_compat = (struct compat_msghdr __user *) sr->msg;
3725 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3730 uiov = compat_ptr(ptr);
3731 if (req->flags & REQ_F_BUFFER_SELECT) {
3732 compat_ssize_t clen;
3736 if (!access_ok(uiov, sizeof(*uiov)))
3738 if (__get_user(clen, &uiov->iov_len))
3742 sr->len = io->msg.iov[0].iov_len;
3745 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3747 &io->msg.msg.msg_iter);
3756 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3758 io->msg.iov = io->msg.fast_iov;
3760 #ifdef CONFIG_COMPAT
3761 if (req->ctx->compat)
3762 return __io_compat_recvmsg_copy_hdr(req, io);
3765 return __io_recvmsg_copy_hdr(req, io);
3768 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3769 int *cflags, bool needs_lock)
3771 struct io_sr_msg *sr = &req->sr_msg;
3772 struct io_buffer *kbuf;
3774 if (!(req->flags & REQ_F_BUFFER_SELECT))
3777 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3782 req->flags |= REQ_F_BUFFER_SELECTED;
3784 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3785 *cflags |= IORING_CQE_F_BUFFER;
3789 static int io_recvmsg_prep(struct io_kiocb *req,
3790 const struct io_uring_sqe *sqe)
3792 struct io_sr_msg *sr = &req->sr_msg;
3793 struct io_async_ctx *io = req->io;
3796 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3797 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3798 sr->len = READ_ONCE(sqe->len);
3799 sr->bgid = READ_ONCE(sqe->buf_group);
3801 #ifdef CONFIG_COMPAT
3802 if (req->ctx->compat)
3803 sr->msg_flags |= MSG_CMSG_COMPAT;
3806 if (!io || req->opcode == IORING_OP_RECV)
3808 /* iovec is already imported */
3809 if (req->flags & REQ_F_NEED_CLEANUP)
3812 ret = io_recvmsg_copy_hdr(req, io);
3814 req->flags |= REQ_F_NEED_CLEANUP;
3818 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3820 struct io_async_msghdr *kmsg = NULL;
3821 struct socket *sock;
3822 int ret, cflags = 0;
3824 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3827 sock = sock_from_file(req->file, &ret);
3829 struct io_buffer *kbuf;
3830 struct io_async_ctx io;
3834 kmsg = &req->io->msg;
3835 kmsg->msg.msg_name = &req->io->msg.addr;
3836 /* if iov is set, it's allocated already */
3838 kmsg->iov = kmsg->fast_iov;
3839 kmsg->msg.msg_iter.iov = kmsg->iov;
3842 kmsg->msg.msg_name = &io.msg.addr;
3844 ret = io_recvmsg_copy_hdr(req, &io);
3849 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3851 return PTR_ERR(kbuf);
3853 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3854 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3855 1, req->sr_msg.len);
3858 flags = req->sr_msg.msg_flags;
3859 if (flags & MSG_DONTWAIT)
3860 req->flags |= REQ_F_NOWAIT;
3861 else if (force_nonblock)
3862 flags |= MSG_DONTWAIT;
3864 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3865 kmsg->uaddr, flags);
3866 if (force_nonblock && ret == -EAGAIN)
3867 return io_setup_async_msg(req, kmsg);
3868 if (ret == -ERESTARTSYS)
3872 if (kmsg && kmsg->iov != kmsg->fast_iov)
3874 req->flags &= ~REQ_F_NEED_CLEANUP;
3875 __io_cqring_add_event(req, ret, cflags);
3877 req_set_fail_links(req);
3882 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3884 struct io_buffer *kbuf = NULL;
3885 struct socket *sock;
3886 int ret, cflags = 0;
3888 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3891 sock = sock_from_file(req->file, &ret);
3893 struct io_sr_msg *sr = &req->sr_msg;
3894 void __user *buf = sr->buf;
3899 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3901 return PTR_ERR(kbuf);
3903 buf = u64_to_user_ptr(kbuf->addr);
3905 ret = import_single_range(READ, buf, sr->len, &iov,
3912 req->flags |= REQ_F_NEED_CLEANUP;
3913 msg.msg_name = NULL;
3914 msg.msg_control = NULL;
3915 msg.msg_controllen = 0;
3916 msg.msg_namelen = 0;
3917 msg.msg_iocb = NULL;
3920 flags = req->sr_msg.msg_flags;
3921 if (flags & MSG_DONTWAIT)
3922 req->flags |= REQ_F_NOWAIT;
3923 else if (force_nonblock)
3924 flags |= MSG_DONTWAIT;
3926 ret = sock_recvmsg(sock, &msg, flags);
3927 if (force_nonblock && ret == -EAGAIN)
3929 if (ret == -ERESTARTSYS)
3934 req->flags &= ~REQ_F_NEED_CLEANUP;
3935 __io_cqring_add_event(req, ret, cflags);
3937 req_set_fail_links(req);
3942 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3944 struct io_accept *accept = &req->accept;
3946 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3948 if (sqe->ioprio || sqe->len || sqe->buf_index)
3951 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3952 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3953 accept->flags = READ_ONCE(sqe->accept_flags);
3954 accept->nofile = rlimit(RLIMIT_NOFILE);
3958 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3960 struct io_accept *accept = &req->accept;
3961 unsigned file_flags;
3964 file_flags = force_nonblock ? O_NONBLOCK : 0;
3965 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3966 accept->addr_len, accept->flags,
3968 if (ret == -EAGAIN && force_nonblock)
3970 if (ret == -ERESTARTSYS)
3973 req_set_fail_links(req);
3974 io_cqring_add_event(req, ret);
3979 static void io_accept_finish(struct io_wq_work **workptr)
3981 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3983 if (io_req_cancelled(req))
3985 __io_accept(req, false);
3986 io_steal_work(req, workptr);
3989 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3993 ret = __io_accept(req, force_nonblock);
3994 if (ret == -EAGAIN && force_nonblock) {
3995 req->work.func = io_accept_finish;
4001 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4003 struct io_connect *conn = &req->connect;
4004 struct io_async_ctx *io = req->io;
4006 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4008 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4011 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4012 conn->addr_len = READ_ONCE(sqe->addr2);
4017 return move_addr_to_kernel(conn->addr, conn->addr_len,
4018 &io->connect.address);
4021 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4023 struct io_async_ctx __io, *io;
4024 unsigned file_flags;
4030 ret = move_addr_to_kernel(req->connect.addr,
4031 req->connect.addr_len,
4032 &__io.connect.address);
4038 file_flags = force_nonblock ? O_NONBLOCK : 0;
4040 ret = __sys_connect_file(req->file, &io->connect.address,
4041 req->connect.addr_len, file_flags);
4042 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4045 if (io_alloc_async_ctx(req)) {
4049 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4052 if (ret == -ERESTARTSYS)
4056 req_set_fail_links(req);
4057 io_cqring_add_event(req, ret);
4061 #else /* !CONFIG_NET */
4062 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4067 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4072 static int io_send(struct io_kiocb *req, bool force_nonblock)
4077 static int io_recvmsg_prep(struct io_kiocb *req,
4078 const struct io_uring_sqe *sqe)
4083 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4088 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4093 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4098 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4103 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4108 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4112 #endif /* CONFIG_NET */
4114 struct io_poll_table {
4115 struct poll_table_struct pt;
4116 struct io_kiocb *req;
4120 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4121 struct wait_queue_head *head)
4123 if (unlikely(poll->head)) {
4124 pt->error = -EINVAL;
4130 add_wait_queue(head, &poll->wait);
4133 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4134 struct poll_table_struct *p)
4136 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4138 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4141 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4142 __poll_t mask, task_work_func_t func)
4144 struct task_struct *tsk;
4147 /* for instances that support it check for an event match first: */
4148 if (mask && !(mask & poll->events))
4151 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4153 list_del_init(&poll->wait.entry);
4157 init_task_work(&req->task_work, func);
4159 * If this fails, then the task is exiting. Punt to one of the io-wq
4160 * threads to ensure the work gets run, we can't always rely on exit
4161 * cancelation taking care of this.
4163 ret = task_work_add(tsk, &req->task_work, true);
4164 if (unlikely(ret)) {
4165 tsk = io_wq_get_task(req->ctx->io_wq);
4166 task_work_add(tsk, &req->task_work, true);
4168 wake_up_process(tsk);
4172 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4173 __acquires(&req->ctx->completion_lock)
4175 struct io_ring_ctx *ctx = req->ctx;
4177 if (!req->result && !READ_ONCE(poll->canceled)) {
4178 struct poll_table_struct pt = { ._key = poll->events };
4180 req->result = vfs_poll(req->file, &pt) & poll->events;
4183 spin_lock_irq(&ctx->completion_lock);
4184 if (!req->result && !READ_ONCE(poll->canceled)) {
4185 add_wait_queue(poll->head, &poll->wait);
4192 static void io_async_task_func(struct callback_head *cb)
4194 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4195 struct async_poll *apoll = req->apoll;
4196 struct io_ring_ctx *ctx = req->ctx;
4199 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4201 if (io_poll_rewait(req, &apoll->poll)) {
4202 spin_unlock_irq(&ctx->completion_lock);
4206 if (hash_hashed(&req->hash_node))
4207 hash_del(&req->hash_node);
4209 canceled = READ_ONCE(apoll->poll.canceled);
4211 io_cqring_fill_event(req, -ECANCELED);
4212 io_commit_cqring(ctx);
4215 spin_unlock_irq(&ctx->completion_lock);
4217 /* restore ->work in case we need to retry again */
4218 memcpy(&req->work, &apoll->work, sizeof(req->work));
4222 io_cqring_ev_posted(ctx);
4223 req_set_fail_links(req);
4224 io_double_put_req(req);
4228 __set_current_state(TASK_RUNNING);
4229 mutex_lock(&ctx->uring_lock);
4230 __io_queue_sqe(req, NULL);
4231 mutex_unlock(&ctx->uring_lock);
4236 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4239 struct io_kiocb *req = wait->private;
4240 struct io_poll_iocb *poll = &req->apoll->poll;
4242 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4245 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4248 static void io_poll_req_insert(struct io_kiocb *req)
4250 struct io_ring_ctx *ctx = req->ctx;
4251 struct hlist_head *list;
4253 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4254 hlist_add_head(&req->hash_node, list);
4257 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4258 struct io_poll_iocb *poll,
4259 struct io_poll_table *ipt, __poll_t mask,
4260 wait_queue_func_t wake_func)
4261 __acquires(&ctx->completion_lock)
4263 struct io_ring_ctx *ctx = req->ctx;
4264 bool cancel = false;
4266 poll->file = req->file;
4268 poll->done = poll->canceled = false;
4269 poll->events = mask;
4271 ipt->pt._key = mask;
4273 ipt->error = -EINVAL;
4275 INIT_LIST_HEAD(&poll->wait.entry);
4276 init_waitqueue_func_entry(&poll->wait, wake_func);
4277 poll->wait.private = req;
4279 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4281 spin_lock_irq(&ctx->completion_lock);
4282 if (likely(poll->head)) {
4283 spin_lock(&poll->head->lock);
4284 if (unlikely(list_empty(&poll->wait.entry))) {
4290 if (mask || ipt->error)
4291 list_del_init(&poll->wait.entry);
4293 WRITE_ONCE(poll->canceled, true);
4294 else if (!poll->done) /* actually waiting for an event */
4295 io_poll_req_insert(req);
4296 spin_unlock(&poll->head->lock);
4302 static bool io_arm_poll_handler(struct io_kiocb *req)
4304 const struct io_op_def *def = &io_op_defs[req->opcode];
4305 struct io_ring_ctx *ctx = req->ctx;
4306 struct async_poll *apoll;
4307 struct io_poll_table ipt;
4310 if (!req->file || !file_can_poll(req->file))
4312 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4314 if (!def->pollin && !def->pollout)
4317 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4318 if (unlikely(!apoll))
4321 req->flags |= REQ_F_POLLED;
4322 memcpy(&apoll->work, &req->work, sizeof(req->work));
4324 get_task_struct(current);
4325 req->task = current;
4327 INIT_HLIST_NODE(&req->hash_node);
4331 mask |= POLLIN | POLLRDNORM;
4333 mask |= POLLOUT | POLLWRNORM;
4334 mask |= POLLERR | POLLPRI;
4336 ipt.pt._qproc = io_async_queue_proc;
4338 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4342 apoll->poll.done = true;
4343 spin_unlock_irq(&ctx->completion_lock);
4344 memcpy(&req->work, &apoll->work, sizeof(req->work));
4348 spin_unlock_irq(&ctx->completion_lock);
4349 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4350 apoll->poll.events);
4354 static bool __io_poll_remove_one(struct io_kiocb *req,
4355 struct io_poll_iocb *poll)
4357 bool do_complete = false;
4359 spin_lock(&poll->head->lock);
4360 WRITE_ONCE(poll->canceled, true);
4361 if (!list_empty(&poll->wait.entry)) {
4362 list_del_init(&poll->wait.entry);
4365 spin_unlock(&poll->head->lock);
4369 static bool io_poll_remove_one(struct io_kiocb *req)
4371 struct async_poll *apoll = NULL;
4374 if (req->opcode == IORING_OP_POLL_ADD) {
4375 do_complete = __io_poll_remove_one(req, &req->poll);
4378 /* non-poll requests have submit ref still */
4379 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4384 hash_del(&req->hash_node);
4386 if (do_complete && apoll) {
4388 * restore ->work because we need to call io_req_work_drop_env.
4390 memcpy(&req->work, &apoll->work, sizeof(req->work));
4395 io_cqring_fill_event(req, -ECANCELED);
4396 io_commit_cqring(req->ctx);
4397 req->flags |= REQ_F_COMP_LOCKED;
4404 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4406 struct hlist_node *tmp;
4407 struct io_kiocb *req;
4410 spin_lock_irq(&ctx->completion_lock);
4411 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4412 struct hlist_head *list;
4414 list = &ctx->cancel_hash[i];
4415 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4416 posted += io_poll_remove_one(req);
4418 spin_unlock_irq(&ctx->completion_lock);
4421 io_cqring_ev_posted(ctx);
4424 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4426 struct hlist_head *list;
4427 struct io_kiocb *req;
4429 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4430 hlist_for_each_entry(req, list, hash_node) {
4431 if (sqe_addr != req->user_data)
4433 if (io_poll_remove_one(req))
4441 static int io_poll_remove_prep(struct io_kiocb *req,
4442 const struct io_uring_sqe *sqe)
4444 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4446 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4450 req->poll.addr = READ_ONCE(sqe->addr);
4455 * Find a running poll command that matches one specified in sqe->addr,
4456 * and remove it if found.
4458 static int io_poll_remove(struct io_kiocb *req)
4460 struct io_ring_ctx *ctx = req->ctx;
4464 addr = req->poll.addr;
4465 spin_lock_irq(&ctx->completion_lock);
4466 ret = io_poll_cancel(ctx, addr);
4467 spin_unlock_irq(&ctx->completion_lock);
4469 io_cqring_add_event(req, ret);
4471 req_set_fail_links(req);
4476 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4478 struct io_ring_ctx *ctx = req->ctx;
4480 req->poll.done = true;
4481 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4482 io_commit_cqring(ctx);
4485 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4487 struct io_ring_ctx *ctx = req->ctx;
4488 struct io_poll_iocb *poll = &req->poll;
4490 if (io_poll_rewait(req, poll)) {
4491 spin_unlock_irq(&ctx->completion_lock);
4495 hash_del(&req->hash_node);
4496 io_poll_complete(req, req->result, 0);
4497 req->flags |= REQ_F_COMP_LOCKED;
4498 io_put_req_find_next(req, nxt);
4499 spin_unlock_irq(&ctx->completion_lock);
4501 io_cqring_ev_posted(ctx);
4504 static void io_poll_task_func(struct callback_head *cb)
4506 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4507 struct io_kiocb *nxt = NULL;
4509 io_poll_task_handler(req, &nxt);
4511 struct io_ring_ctx *ctx = nxt->ctx;
4513 mutex_lock(&ctx->uring_lock);
4514 __io_queue_sqe(nxt, NULL);
4515 mutex_unlock(&ctx->uring_lock);
4519 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4522 struct io_kiocb *req = wait->private;
4523 struct io_poll_iocb *poll = &req->poll;
4525 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4528 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4529 struct poll_table_struct *p)
4531 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4533 __io_queue_proc(&pt->req->poll, pt, head);
4536 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4538 struct io_poll_iocb *poll = &req->poll;
4541 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4543 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4548 events = READ_ONCE(sqe->poll_events);
4549 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4551 get_task_struct(current);
4552 req->task = current;
4556 static int io_poll_add(struct io_kiocb *req)
4558 struct io_poll_iocb *poll = &req->poll;
4559 struct io_ring_ctx *ctx = req->ctx;
4560 struct io_poll_table ipt;
4563 INIT_HLIST_NODE(&req->hash_node);
4564 INIT_LIST_HEAD(&req->list);
4565 ipt.pt._qproc = io_poll_queue_proc;
4567 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4570 if (mask) { /* no async, we'd stolen it */
4572 io_poll_complete(req, mask, 0);
4574 spin_unlock_irq(&ctx->completion_lock);
4577 io_cqring_ev_posted(ctx);
4583 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4585 struct io_timeout_data *data = container_of(timer,
4586 struct io_timeout_data, timer);
4587 struct io_kiocb *req = data->req;
4588 struct io_ring_ctx *ctx = req->ctx;
4589 unsigned long flags;
4591 atomic_inc(&ctx->cq_timeouts);
4593 spin_lock_irqsave(&ctx->completion_lock, flags);
4595 * We could be racing with timeout deletion. If the list is empty,
4596 * then timeout lookup already found it and will be handling it.
4598 if (!list_empty(&req->list)) {
4599 struct io_kiocb *prev;
4602 * Adjust the reqs sequence before the current one because it
4603 * will consume a slot in the cq_ring and the cq_tail
4604 * pointer will be increased, otherwise other timeout reqs may
4605 * return in advance without waiting for enough wait_nr.
4608 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4610 list_del_init(&req->list);
4613 io_cqring_fill_event(req, -ETIME);
4614 io_commit_cqring(ctx);
4615 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4617 io_cqring_ev_posted(ctx);
4618 req_set_fail_links(req);
4620 return HRTIMER_NORESTART;
4623 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4625 struct io_kiocb *req;
4628 list_for_each_entry(req, &ctx->timeout_list, list) {
4629 if (user_data == req->user_data) {
4630 list_del_init(&req->list);
4639 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4643 req_set_fail_links(req);
4644 io_cqring_fill_event(req, -ECANCELED);
4649 static int io_timeout_remove_prep(struct io_kiocb *req,
4650 const struct io_uring_sqe *sqe)
4652 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4654 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4657 req->timeout.addr = READ_ONCE(sqe->addr);
4658 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4659 if (req->timeout.flags)
4666 * Remove or update an existing timeout command
4668 static int io_timeout_remove(struct io_kiocb *req)
4670 struct io_ring_ctx *ctx = req->ctx;
4673 spin_lock_irq(&ctx->completion_lock);
4674 ret = io_timeout_cancel(ctx, req->timeout.addr);
4676 io_cqring_fill_event(req, ret);
4677 io_commit_cqring(ctx);
4678 spin_unlock_irq(&ctx->completion_lock);
4679 io_cqring_ev_posted(ctx);
4681 req_set_fail_links(req);
4686 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4687 bool is_timeout_link)
4689 struct io_timeout_data *data;
4692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4694 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4696 if (sqe->off && is_timeout_link)
4698 flags = READ_ONCE(sqe->timeout_flags);
4699 if (flags & ~IORING_TIMEOUT_ABS)
4702 req->timeout.count = READ_ONCE(sqe->off);
4704 if (!req->io && io_alloc_async_ctx(req))
4707 data = &req->io->timeout;
4709 req->flags |= REQ_F_TIMEOUT;
4711 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4714 if (flags & IORING_TIMEOUT_ABS)
4715 data->mode = HRTIMER_MODE_ABS;
4717 data->mode = HRTIMER_MODE_REL;
4719 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4723 static int io_timeout(struct io_kiocb *req)
4725 struct io_ring_ctx *ctx = req->ctx;
4726 struct io_timeout_data *data;
4727 struct list_head *entry;
4729 u32 count = req->timeout.count;
4730 u32 seq = req->sequence;
4732 data = &req->io->timeout;
4735 * sqe->off holds how many events that need to occur for this
4736 * timeout event to be satisfied. If it isn't set, then this is
4737 * a pure timeout request, sequence isn't used.
4740 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4741 spin_lock_irq(&ctx->completion_lock);
4742 entry = ctx->timeout_list.prev;
4746 req->sequence = seq + count;
4749 * Insertion sort, ensuring the first entry in the list is always
4750 * the one we need first.
4752 spin_lock_irq(&ctx->completion_lock);
4753 list_for_each_prev(entry, &ctx->timeout_list) {
4754 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4756 long long tmp, tmp_nxt;
4757 u32 nxt_offset = nxt->timeout.count;
4759 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4763 * Since seq + count can overflow, use type long
4766 tmp = (long long)seq + count;
4767 nxt_seq = nxt->sequence - nxt_offset;
4768 tmp_nxt = (long long)nxt_seq + nxt_offset;
4771 * cached_sq_head may overflow, and it will never overflow twice
4772 * once there is some timeout req still be valid.
4781 * Sequence of reqs after the insert one and itself should
4782 * be adjusted because each timeout req consumes a slot.
4787 req->sequence -= span;
4789 list_add(&req->list, entry);
4790 data->timer.function = io_timeout_fn;
4791 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4792 spin_unlock_irq(&ctx->completion_lock);
4796 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4798 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4800 return req->user_data == (unsigned long) data;
4803 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4805 enum io_wq_cancel cancel_ret;
4808 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4809 switch (cancel_ret) {
4810 case IO_WQ_CANCEL_OK:
4813 case IO_WQ_CANCEL_RUNNING:
4816 case IO_WQ_CANCEL_NOTFOUND:
4824 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4825 struct io_kiocb *req, __u64 sqe_addr,
4828 unsigned long flags;
4831 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4832 if (ret != -ENOENT) {
4833 spin_lock_irqsave(&ctx->completion_lock, flags);
4837 spin_lock_irqsave(&ctx->completion_lock, flags);
4838 ret = io_timeout_cancel(ctx, sqe_addr);
4841 ret = io_poll_cancel(ctx, sqe_addr);
4845 io_cqring_fill_event(req, ret);
4846 io_commit_cqring(ctx);
4847 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4848 io_cqring_ev_posted(ctx);
4851 req_set_fail_links(req);
4855 static int io_async_cancel_prep(struct io_kiocb *req,
4856 const struct io_uring_sqe *sqe)
4858 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4860 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4864 req->cancel.addr = READ_ONCE(sqe->addr);
4868 static int io_async_cancel(struct io_kiocb *req)
4870 struct io_ring_ctx *ctx = req->ctx;
4872 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4876 static int io_files_update_prep(struct io_kiocb *req,
4877 const struct io_uring_sqe *sqe)
4879 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4882 req->files_update.offset = READ_ONCE(sqe->off);
4883 req->files_update.nr_args = READ_ONCE(sqe->len);
4884 if (!req->files_update.nr_args)
4886 req->files_update.arg = READ_ONCE(sqe->addr);
4890 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4892 struct io_ring_ctx *ctx = req->ctx;
4893 struct io_uring_files_update up;
4899 up.offset = req->files_update.offset;
4900 up.fds = req->files_update.arg;
4902 mutex_lock(&ctx->uring_lock);
4903 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4904 mutex_unlock(&ctx->uring_lock);
4907 req_set_fail_links(req);
4908 io_cqring_add_event(req, ret);
4913 static int io_req_defer_prep(struct io_kiocb *req,
4914 const struct io_uring_sqe *sqe)
4921 if (io_op_defs[req->opcode].file_table) {
4922 ret = io_grab_files(req);
4927 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4929 switch (req->opcode) {
4932 case IORING_OP_READV:
4933 case IORING_OP_READ_FIXED:
4934 case IORING_OP_READ:
4935 ret = io_read_prep(req, sqe, true);
4937 case IORING_OP_WRITEV:
4938 case IORING_OP_WRITE_FIXED:
4939 case IORING_OP_WRITE:
4940 ret = io_write_prep(req, sqe, true);
4942 case IORING_OP_POLL_ADD:
4943 ret = io_poll_add_prep(req, sqe);
4945 case IORING_OP_POLL_REMOVE:
4946 ret = io_poll_remove_prep(req, sqe);
4948 case IORING_OP_FSYNC:
4949 ret = io_prep_fsync(req, sqe);
4951 case IORING_OP_SYNC_FILE_RANGE:
4952 ret = io_prep_sfr(req, sqe);
4954 case IORING_OP_SENDMSG:
4955 case IORING_OP_SEND:
4956 ret = io_sendmsg_prep(req, sqe);
4958 case IORING_OP_RECVMSG:
4959 case IORING_OP_RECV:
4960 ret = io_recvmsg_prep(req, sqe);
4962 case IORING_OP_CONNECT:
4963 ret = io_connect_prep(req, sqe);
4965 case IORING_OP_TIMEOUT:
4966 ret = io_timeout_prep(req, sqe, false);
4968 case IORING_OP_TIMEOUT_REMOVE:
4969 ret = io_timeout_remove_prep(req, sqe);
4971 case IORING_OP_ASYNC_CANCEL:
4972 ret = io_async_cancel_prep(req, sqe);
4974 case IORING_OP_LINK_TIMEOUT:
4975 ret = io_timeout_prep(req, sqe, true);
4977 case IORING_OP_ACCEPT:
4978 ret = io_accept_prep(req, sqe);
4980 case IORING_OP_FALLOCATE:
4981 ret = io_fallocate_prep(req, sqe);
4983 case IORING_OP_OPENAT:
4984 ret = io_openat_prep(req, sqe);
4986 case IORING_OP_CLOSE:
4987 ret = io_close_prep(req, sqe);
4989 case IORING_OP_FILES_UPDATE:
4990 ret = io_files_update_prep(req, sqe);
4992 case IORING_OP_STATX:
4993 ret = io_statx_prep(req, sqe);
4995 case IORING_OP_FADVISE:
4996 ret = io_fadvise_prep(req, sqe);
4998 case IORING_OP_MADVISE:
4999 ret = io_madvise_prep(req, sqe);
5001 case IORING_OP_OPENAT2:
5002 ret = io_openat2_prep(req, sqe);
5004 case IORING_OP_EPOLL_CTL:
5005 ret = io_epoll_ctl_prep(req, sqe);
5007 case IORING_OP_SPLICE:
5008 ret = io_splice_prep(req, sqe);
5010 case IORING_OP_PROVIDE_BUFFERS:
5011 ret = io_provide_buffers_prep(req, sqe);
5013 case IORING_OP_REMOVE_BUFFERS:
5014 ret = io_remove_buffers_prep(req, sqe);
5017 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5026 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5028 struct io_ring_ctx *ctx = req->ctx;
5031 /* Still need defer if there is pending req in defer list. */
5032 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
5035 if (!req->io && io_alloc_async_ctx(req))
5038 ret = io_req_defer_prep(req, sqe);
5042 spin_lock_irq(&ctx->completion_lock);
5043 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5044 spin_unlock_irq(&ctx->completion_lock);
5048 trace_io_uring_defer(ctx, req, req->user_data);
5049 list_add_tail(&req->list, &ctx->defer_list);
5050 spin_unlock_irq(&ctx->completion_lock);
5051 return -EIOCBQUEUED;
5054 static void io_cleanup_req(struct io_kiocb *req)
5056 struct io_async_ctx *io = req->io;
5058 switch (req->opcode) {
5059 case IORING_OP_READV:
5060 case IORING_OP_READ_FIXED:
5061 case IORING_OP_READ:
5062 if (req->flags & REQ_F_BUFFER_SELECTED)
5063 kfree((void *)(unsigned long)req->rw.addr);
5065 case IORING_OP_WRITEV:
5066 case IORING_OP_WRITE_FIXED:
5067 case IORING_OP_WRITE:
5068 if (io->rw.iov != io->rw.fast_iov)
5071 case IORING_OP_RECVMSG:
5072 if (req->flags & REQ_F_BUFFER_SELECTED)
5073 kfree(req->sr_msg.kbuf);
5075 case IORING_OP_SENDMSG:
5076 if (io->msg.iov != io->msg.fast_iov)
5079 case IORING_OP_RECV:
5080 if (req->flags & REQ_F_BUFFER_SELECTED)
5081 kfree(req->sr_msg.kbuf);
5083 case IORING_OP_OPENAT:
5084 case IORING_OP_OPENAT2:
5085 case IORING_OP_STATX:
5086 putname(req->open.filename);
5088 case IORING_OP_SPLICE:
5089 io_put_file(req, req->splice.file_in,
5090 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5094 req->flags &= ~REQ_F_NEED_CLEANUP;
5097 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5098 bool force_nonblock)
5100 struct io_ring_ctx *ctx = req->ctx;
5103 switch (req->opcode) {
5107 case IORING_OP_READV:
5108 case IORING_OP_READ_FIXED:
5109 case IORING_OP_READ:
5111 ret = io_read_prep(req, sqe, force_nonblock);
5115 ret = io_read(req, force_nonblock);
5117 case IORING_OP_WRITEV:
5118 case IORING_OP_WRITE_FIXED:
5119 case IORING_OP_WRITE:
5121 ret = io_write_prep(req, sqe, force_nonblock);
5125 ret = io_write(req, force_nonblock);
5127 case IORING_OP_FSYNC:
5129 ret = io_prep_fsync(req, sqe);
5133 ret = io_fsync(req, force_nonblock);
5135 case IORING_OP_POLL_ADD:
5137 ret = io_poll_add_prep(req, sqe);
5141 ret = io_poll_add(req);
5143 case IORING_OP_POLL_REMOVE:
5145 ret = io_poll_remove_prep(req, sqe);
5149 ret = io_poll_remove(req);
5151 case IORING_OP_SYNC_FILE_RANGE:
5153 ret = io_prep_sfr(req, sqe);
5157 ret = io_sync_file_range(req, force_nonblock);
5159 case IORING_OP_SENDMSG:
5160 case IORING_OP_SEND:
5162 ret = io_sendmsg_prep(req, sqe);
5166 if (req->opcode == IORING_OP_SENDMSG)
5167 ret = io_sendmsg(req, force_nonblock);
5169 ret = io_send(req, force_nonblock);
5171 case IORING_OP_RECVMSG:
5172 case IORING_OP_RECV:
5174 ret = io_recvmsg_prep(req, sqe);
5178 if (req->opcode == IORING_OP_RECVMSG)
5179 ret = io_recvmsg(req, force_nonblock);
5181 ret = io_recv(req, force_nonblock);
5183 case IORING_OP_TIMEOUT:
5185 ret = io_timeout_prep(req, sqe, false);
5189 ret = io_timeout(req);
5191 case IORING_OP_TIMEOUT_REMOVE:
5193 ret = io_timeout_remove_prep(req, sqe);
5197 ret = io_timeout_remove(req);
5199 case IORING_OP_ACCEPT:
5201 ret = io_accept_prep(req, sqe);
5205 ret = io_accept(req, force_nonblock);
5207 case IORING_OP_CONNECT:
5209 ret = io_connect_prep(req, sqe);
5213 ret = io_connect(req, force_nonblock);
5215 case IORING_OP_ASYNC_CANCEL:
5217 ret = io_async_cancel_prep(req, sqe);
5221 ret = io_async_cancel(req);
5223 case IORING_OP_FALLOCATE:
5225 ret = io_fallocate_prep(req, sqe);
5229 ret = io_fallocate(req, force_nonblock);
5231 case IORING_OP_OPENAT:
5233 ret = io_openat_prep(req, sqe);
5237 ret = io_openat(req, force_nonblock);
5239 case IORING_OP_CLOSE:
5241 ret = io_close_prep(req, sqe);
5245 ret = io_close(req, force_nonblock);
5247 case IORING_OP_FILES_UPDATE:
5249 ret = io_files_update_prep(req, sqe);
5253 ret = io_files_update(req, force_nonblock);
5255 case IORING_OP_STATX:
5257 ret = io_statx_prep(req, sqe);
5261 ret = io_statx(req, force_nonblock);
5263 case IORING_OP_FADVISE:
5265 ret = io_fadvise_prep(req, sqe);
5269 ret = io_fadvise(req, force_nonblock);
5271 case IORING_OP_MADVISE:
5273 ret = io_madvise_prep(req, sqe);
5277 ret = io_madvise(req, force_nonblock);
5279 case IORING_OP_OPENAT2:
5281 ret = io_openat2_prep(req, sqe);
5285 ret = io_openat2(req, force_nonblock);
5287 case IORING_OP_EPOLL_CTL:
5289 ret = io_epoll_ctl_prep(req, sqe);
5293 ret = io_epoll_ctl(req, force_nonblock);
5295 case IORING_OP_SPLICE:
5297 ret = io_splice_prep(req, sqe);
5301 ret = io_splice(req, force_nonblock);
5303 case IORING_OP_PROVIDE_BUFFERS:
5305 ret = io_provide_buffers_prep(req, sqe);
5309 ret = io_provide_buffers(req, force_nonblock);
5311 case IORING_OP_REMOVE_BUFFERS:
5313 ret = io_remove_buffers_prep(req, sqe);
5317 ret = io_remove_buffers(req, force_nonblock);
5327 if (ctx->flags & IORING_SETUP_IOPOLL) {
5328 const bool in_async = io_wq_current_is_worker();
5330 if (req->result == -EAGAIN)
5333 /* workqueue context doesn't hold uring_lock, grab it now */
5335 mutex_lock(&ctx->uring_lock);
5337 io_iopoll_req_issued(req);
5340 mutex_unlock(&ctx->uring_lock);
5346 static void io_wq_submit_work(struct io_wq_work **workptr)
5348 struct io_wq_work *work = *workptr;
5349 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5352 /* if NO_CANCEL is set, we must still run the work */
5353 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5354 IO_WQ_WORK_CANCEL) {
5360 ret = io_issue_sqe(req, NULL, false);
5362 * We can get EAGAIN for polled IO even though we're
5363 * forcing a sync submission from here, since we can't
5364 * wait for request slots on the block side.
5373 req_set_fail_links(req);
5374 io_cqring_add_event(req, ret);
5378 io_steal_work(req, workptr);
5381 static int io_req_needs_file(struct io_kiocb *req, int fd)
5383 if (!io_op_defs[req->opcode].needs_file)
5385 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5390 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5393 struct fixed_file_table *table;
5395 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5396 return table->files[index & IORING_FILE_TABLE_MASK];;
5399 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5400 int fd, struct file **out_file, bool fixed)
5402 struct io_ring_ctx *ctx = req->ctx;
5406 if (unlikely(!ctx->file_data ||
5407 (unsigned) fd >= ctx->nr_user_files))
5409 fd = array_index_nospec(fd, ctx->nr_user_files);
5410 file = io_file_from_index(ctx, fd);
5413 req->fixed_file_refs = ctx->file_data->cur_refs;
5414 percpu_ref_get(req->fixed_file_refs);
5416 trace_io_uring_file_get(ctx, fd);
5417 file = __io_file_get(state, fd);
5418 if (unlikely(!file))
5426 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5427 int fd, unsigned int flags)
5431 if (!io_req_needs_file(req, fd))
5434 fixed = (flags & IOSQE_FIXED_FILE);
5435 if (unlikely(!fixed && req->needs_fixed_file))
5438 return io_file_get(state, req, fd, &req->file, fixed);
5441 static int io_grab_files(struct io_kiocb *req)
5444 struct io_ring_ctx *ctx = req->ctx;
5446 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5448 if (!ctx->ring_file)
5452 spin_lock_irq(&ctx->inflight_lock);
5454 * We use the f_ops->flush() handler to ensure that we can flush
5455 * out work accessing these files if the fd is closed. Check if
5456 * the fd has changed since we started down this path, and disallow
5457 * this operation if it has.
5459 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5460 list_add(&req->inflight_entry, &ctx->inflight_list);
5461 req->flags |= REQ_F_INFLIGHT;
5462 req->work.files = current->files;
5465 spin_unlock_irq(&ctx->inflight_lock);
5471 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5473 struct io_timeout_data *data = container_of(timer,
5474 struct io_timeout_data, timer);
5475 struct io_kiocb *req = data->req;
5476 struct io_ring_ctx *ctx = req->ctx;
5477 struct io_kiocb *prev = NULL;
5478 unsigned long flags;
5480 spin_lock_irqsave(&ctx->completion_lock, flags);
5483 * We don't expect the list to be empty, that will only happen if we
5484 * race with the completion of the linked work.
5486 if (!list_empty(&req->link_list)) {
5487 prev = list_entry(req->link_list.prev, struct io_kiocb,
5489 if (refcount_inc_not_zero(&prev->refs)) {
5490 list_del_init(&req->link_list);
5491 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5496 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5499 req_set_fail_links(prev);
5500 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5503 io_cqring_add_event(req, -ETIME);
5506 return HRTIMER_NORESTART;
5509 static void io_queue_linked_timeout(struct io_kiocb *req)
5511 struct io_ring_ctx *ctx = req->ctx;
5514 * If the list is now empty, then our linked request finished before
5515 * we got a chance to setup the timer
5517 spin_lock_irq(&ctx->completion_lock);
5518 if (!list_empty(&req->link_list)) {
5519 struct io_timeout_data *data = &req->io->timeout;
5521 data->timer.function = io_link_timeout_fn;
5522 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5525 spin_unlock_irq(&ctx->completion_lock);
5527 /* drop submission reference */
5531 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5533 struct io_kiocb *nxt;
5535 if (!(req->flags & REQ_F_LINK_HEAD))
5537 /* for polled retry, if flag is set, we already went through here */
5538 if (req->flags & REQ_F_POLLED)
5541 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5543 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5546 req->flags |= REQ_F_LINK_TIMEOUT;
5550 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5552 struct io_kiocb *linked_timeout;
5553 struct io_kiocb *nxt;
5554 const struct cred *old_creds = NULL;
5558 linked_timeout = io_prep_linked_timeout(req);
5560 if (req->work.creds && req->work.creds != current_cred()) {
5562 revert_creds(old_creds);
5563 if (old_creds == req->work.creds)
5564 old_creds = NULL; /* restored original creds */
5566 old_creds = override_creds(req->work.creds);
5569 ret = io_issue_sqe(req, sqe, true);
5572 * We async punt it if the file wasn't marked NOWAIT, or if the file
5573 * doesn't support non-blocking read/write attempts
5575 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5576 (req->flags & REQ_F_MUST_PUNT))) {
5577 if (io_arm_poll_handler(req)) {
5579 io_queue_linked_timeout(linked_timeout);
5583 if (io_op_defs[req->opcode].file_table) {
5584 ret = io_grab_files(req);
5590 * Queued up for async execution, worker will release
5591 * submit reference when the iocb is actually submitted.
5593 io_queue_async_work(req);
5599 /* drop submission reference */
5600 io_put_req_find_next(req, &nxt);
5602 if (linked_timeout) {
5604 io_queue_linked_timeout(linked_timeout);
5606 io_put_req(linked_timeout);
5609 /* and drop final reference, if we failed */
5611 io_cqring_add_event(req, ret);
5612 req_set_fail_links(req);
5618 if (req->flags & REQ_F_FORCE_ASYNC)
5624 revert_creds(old_creds);
5627 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5631 ret = io_req_defer(req, sqe);
5633 if (ret != -EIOCBQUEUED) {
5635 io_cqring_add_event(req, ret);
5636 req_set_fail_links(req);
5637 io_double_put_req(req);
5639 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5640 ret = io_req_defer_prep(req, sqe);
5641 if (unlikely(ret < 0))
5644 * Never try inline submit of IOSQE_ASYNC is set, go straight
5645 * to async execution.
5647 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5648 io_queue_async_work(req);
5650 __io_queue_sqe(req, sqe);
5654 static inline void io_queue_link_head(struct io_kiocb *req)
5656 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5657 io_cqring_add_event(req, -ECANCELED);
5658 io_double_put_req(req);
5660 io_queue_sqe(req, NULL);
5663 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5664 struct io_submit_state *state, struct io_kiocb **link)
5666 struct io_ring_ctx *ctx = req->ctx;
5670 * If we already have a head request, queue this one for async
5671 * submittal once the head completes. If we don't have a head but
5672 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5673 * submitted sync once the chain is complete. If none of those
5674 * conditions are true (normal request), then just queue it.
5677 struct io_kiocb *head = *link;
5680 * Taking sequential execution of a link, draining both sides
5681 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5682 * requests in the link. So, it drains the head and the
5683 * next after the link request. The last one is done via
5684 * drain_next flag to persist the effect across calls.
5686 if (req->flags & REQ_F_IO_DRAIN) {
5687 head->flags |= REQ_F_IO_DRAIN;
5688 ctx->drain_next = 1;
5690 if (io_alloc_async_ctx(req))
5693 ret = io_req_defer_prep(req, sqe);
5695 /* fail even hard links since we don't submit */
5696 head->flags |= REQ_F_FAIL_LINK;
5699 trace_io_uring_link(ctx, req, head);
5700 list_add_tail(&req->link_list, &head->link_list);
5702 /* last request of a link, enqueue the link */
5703 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5704 io_queue_link_head(head);
5708 if (unlikely(ctx->drain_next)) {
5709 req->flags |= REQ_F_IO_DRAIN;
5710 ctx->drain_next = 0;
5712 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5713 req->flags |= REQ_F_LINK_HEAD;
5714 INIT_LIST_HEAD(&req->link_list);
5716 if (io_alloc_async_ctx(req))
5719 ret = io_req_defer_prep(req, sqe);
5721 req->flags |= REQ_F_FAIL_LINK;
5724 io_queue_sqe(req, sqe);
5732 * Batched submission is done, ensure local IO is flushed out.
5734 static void io_submit_state_end(struct io_submit_state *state)
5736 blk_finish_plug(&state->plug);
5738 if (state->free_reqs)
5739 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5743 * Start submission side cache.
5745 static void io_submit_state_start(struct io_submit_state *state,
5746 unsigned int max_ios)
5748 blk_start_plug(&state->plug);
5749 state->free_reqs = 0;
5751 state->ios_left = max_ios;
5754 static void io_commit_sqring(struct io_ring_ctx *ctx)
5756 struct io_rings *rings = ctx->rings;
5759 * Ensure any loads from the SQEs are done at this point,
5760 * since once we write the new head, the application could
5761 * write new data to them.
5763 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5767 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5768 * that is mapped by userspace. This means that care needs to be taken to
5769 * ensure that reads are stable, as we cannot rely on userspace always
5770 * being a good citizen. If members of the sqe are validated and then later
5771 * used, it's important that those reads are done through READ_ONCE() to
5772 * prevent a re-load down the line.
5774 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5776 u32 *sq_array = ctx->sq_array;
5780 * The cached sq head (or cq tail) serves two purposes:
5782 * 1) allows us to batch the cost of updating the user visible
5784 * 2) allows the kernel side to track the head on its own, even
5785 * though the application is the one updating it.
5787 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5788 if (likely(head < ctx->sq_entries))
5789 return &ctx->sq_sqes[head];
5791 /* drop invalid entries */
5792 ctx->cached_sq_dropped++;
5793 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5797 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5799 ctx->cached_sq_head++;
5802 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5803 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5804 IOSQE_BUFFER_SELECT)
5806 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5807 const struct io_uring_sqe *sqe,
5808 struct io_submit_state *state, bool async)
5810 unsigned int sqe_flags;
5814 * All io need record the previous position, if LINK vs DARIN,
5815 * it can be used to mark the position of the first IO in the
5818 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5819 req->opcode = READ_ONCE(sqe->opcode);
5820 req->user_data = READ_ONCE(sqe->user_data);
5825 /* one is dropped after submission, the other at completion */
5826 refcount_set(&req->refs, 2);
5829 req->needs_fixed_file = async;
5830 INIT_IO_WORK(&req->work, io_wq_submit_work);
5832 if (unlikely(req->opcode >= IORING_OP_LAST))
5835 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5836 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5838 use_mm(ctx->sqo_mm);
5841 sqe_flags = READ_ONCE(sqe->flags);
5842 /* enforce forwards compatibility on users */
5843 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5846 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5847 !io_op_defs[req->opcode].buffer_select)
5850 id = READ_ONCE(sqe->personality);
5852 req->work.creds = idr_find(&ctx->personality_idr, id);
5853 if (unlikely(!req->work.creds))
5855 get_cred(req->work.creds);
5858 /* same numerical values with corresponding REQ_F_*, safe to copy */
5859 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5860 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5861 IOSQE_BUFFER_SELECT | IOSQE_IO_LINK);
5863 fd = READ_ONCE(sqe->fd);
5864 return io_req_set_file(state, req, fd, sqe_flags);
5867 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5868 struct file *ring_file, int ring_fd, bool async)
5870 struct io_submit_state state, *statep = NULL;
5871 struct io_kiocb *link = NULL;
5872 int i, submitted = 0;
5874 /* if we have a backlog and couldn't flush it all, return BUSY */
5875 if (test_bit(0, &ctx->sq_check_overflow)) {
5876 if (!list_empty(&ctx->cq_overflow_list) &&
5877 !io_cqring_overflow_flush(ctx, false))
5881 /* make sure SQ entry isn't read before tail */
5882 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5884 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5887 if (nr > IO_PLUG_THRESHOLD) {
5888 io_submit_state_start(&state, nr);
5892 ctx->ring_fd = ring_fd;
5893 ctx->ring_file = ring_file;
5895 for (i = 0; i < nr; i++) {
5896 const struct io_uring_sqe *sqe;
5897 struct io_kiocb *req;
5900 sqe = io_get_sqe(ctx);
5901 if (unlikely(!sqe)) {
5902 io_consume_sqe(ctx);
5905 req = io_alloc_req(ctx, statep);
5906 if (unlikely(!req)) {
5908 submitted = -EAGAIN;
5912 err = io_init_req(ctx, req, sqe, statep, async);
5913 io_consume_sqe(ctx);
5914 /* will complete beyond this point, count as submitted */
5917 if (unlikely(err)) {
5919 io_cqring_add_event(req, err);
5920 io_double_put_req(req);
5924 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5926 err = io_submit_sqe(req, sqe, statep, &link);
5931 if (unlikely(submitted != nr)) {
5932 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5934 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5937 io_queue_link_head(link);
5939 io_submit_state_end(&state);
5941 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5942 io_commit_sqring(ctx);
5947 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5949 struct mm_struct *mm = current->mm;
5957 static int io_sq_thread(void *data)
5959 struct io_ring_ctx *ctx = data;
5960 const struct cred *old_cred;
5961 mm_segment_t old_fs;
5963 unsigned long timeout;
5966 complete(&ctx->completions[1]);
5970 old_cred = override_creds(ctx->creds);
5972 timeout = jiffies + ctx->sq_thread_idle;
5973 while (!kthread_should_park()) {
5974 unsigned int to_submit;
5976 if (!list_empty(&ctx->poll_list)) {
5977 unsigned nr_events = 0;
5979 mutex_lock(&ctx->uring_lock);
5980 if (!list_empty(&ctx->poll_list))
5981 io_iopoll_getevents(ctx, &nr_events, 0);
5983 timeout = jiffies + ctx->sq_thread_idle;
5984 mutex_unlock(&ctx->uring_lock);
5987 to_submit = io_sqring_entries(ctx);
5990 * If submit got -EBUSY, flag us as needing the application
5991 * to enter the kernel to reap and flush events.
5993 if (!to_submit || ret == -EBUSY) {
5995 * Drop cur_mm before scheduling, we can't hold it for
5996 * long periods (or over schedule()). Do this before
5997 * adding ourselves to the waitqueue, as the unuse/drop
6000 io_sq_thread_drop_mm(ctx);
6003 * We're polling. If we're within the defined idle
6004 * period, then let us spin without work before going
6005 * to sleep. The exception is if we got EBUSY doing
6006 * more IO, we should wait for the application to
6007 * reap events and wake us up.
6009 if (!list_empty(&ctx->poll_list) ||
6010 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6011 !percpu_ref_is_dying(&ctx->refs))) {
6012 if (current->task_works)
6018 prepare_to_wait(&ctx->sqo_wait, &wait,
6019 TASK_INTERRUPTIBLE);
6022 * While doing polled IO, before going to sleep, we need
6023 * to check if there are new reqs added to poll_list, it
6024 * is because reqs may have been punted to io worker and
6025 * will be added to poll_list later, hence check the
6028 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6029 !list_empty_careful(&ctx->poll_list)) {
6030 finish_wait(&ctx->sqo_wait, &wait);
6034 /* Tell userspace we may need a wakeup call */
6035 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6036 /* make sure to read SQ tail after writing flags */
6039 to_submit = io_sqring_entries(ctx);
6040 if (!to_submit || ret == -EBUSY) {
6041 if (kthread_should_park()) {
6042 finish_wait(&ctx->sqo_wait, &wait);
6045 if (current->task_works) {
6047 finish_wait(&ctx->sqo_wait, &wait);
6050 if (signal_pending(current))
6051 flush_signals(current);
6053 finish_wait(&ctx->sqo_wait, &wait);
6055 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6058 finish_wait(&ctx->sqo_wait, &wait);
6060 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6063 mutex_lock(&ctx->uring_lock);
6064 ret = io_submit_sqes(ctx, to_submit, NULL, -1, true);
6065 mutex_unlock(&ctx->uring_lock);
6066 timeout = jiffies + ctx->sq_thread_idle;
6069 if (current->task_works)
6073 io_sq_thread_drop_mm(ctx);
6074 revert_creds(old_cred);
6081 struct io_wait_queue {
6082 struct wait_queue_entry wq;
6083 struct io_ring_ctx *ctx;
6085 unsigned nr_timeouts;
6088 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6090 struct io_ring_ctx *ctx = iowq->ctx;
6093 * Wake up if we have enough events, or if a timeout occurred since we
6094 * started waiting. For timeouts, we always want to return to userspace,
6095 * regardless of event count.
6097 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6098 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6101 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6102 int wake_flags, void *key)
6104 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6107 /* use noflush == true, as we can't safely rely on locking context */
6108 if (!io_should_wake(iowq, true))
6111 return autoremove_wake_function(curr, mode, wake_flags, key);
6115 * Wait until events become available, if we don't already have some. The
6116 * application must reap them itself, as they reside on the shared cq ring.
6118 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6119 const sigset_t __user *sig, size_t sigsz)
6121 struct io_wait_queue iowq = {
6124 .func = io_wake_function,
6125 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6128 .to_wait = min_events,
6130 struct io_rings *rings = ctx->rings;
6134 if (io_cqring_events(ctx, false) >= min_events)
6136 if (!current->task_works)
6142 #ifdef CONFIG_COMPAT
6143 if (in_compat_syscall())
6144 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6148 ret = set_user_sigmask(sig, sigsz);
6154 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6155 trace_io_uring_cqring_wait(ctx, min_events);
6157 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6158 TASK_INTERRUPTIBLE);
6159 if (current->task_works)
6161 if (io_should_wake(&iowq, false))
6164 if (signal_pending(current)) {
6169 finish_wait(&ctx->wait, &iowq.wq);
6171 restore_saved_sigmask_unless(ret == -EINTR);
6173 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6176 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6178 #if defined(CONFIG_UNIX)
6179 if (ctx->ring_sock) {
6180 struct sock *sock = ctx->ring_sock->sk;
6181 struct sk_buff *skb;
6183 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6189 for (i = 0; i < ctx->nr_user_files; i++) {
6192 file = io_file_from_index(ctx, i);
6199 static void io_file_ref_kill(struct percpu_ref *ref)
6201 struct fixed_file_data *data;
6203 data = container_of(ref, struct fixed_file_data, refs);
6204 complete(&data->done);
6207 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6209 struct fixed_file_data *data = ctx->file_data;
6210 struct fixed_file_ref_node *ref_node = NULL;
6211 unsigned nr_tables, i;
6212 unsigned long flags;
6217 spin_lock_irqsave(&data->lock, flags);
6218 if (!list_empty(&data->ref_list))
6219 ref_node = list_first_entry(&data->ref_list,
6220 struct fixed_file_ref_node, node);
6221 spin_unlock_irqrestore(&data->lock, flags);
6223 percpu_ref_kill(&ref_node->refs);
6225 percpu_ref_kill(&data->refs);
6227 /* wait for all refs nodes to complete */
6228 wait_for_completion(&data->done);
6230 __io_sqe_files_unregister(ctx);
6231 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6232 for (i = 0; i < nr_tables; i++)
6233 kfree(data->table[i].files);
6235 percpu_ref_exit(&data->refs);
6237 ctx->file_data = NULL;
6238 ctx->nr_user_files = 0;
6242 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6244 if (ctx->sqo_thread) {
6245 wait_for_completion(&ctx->completions[1]);
6247 * The park is a bit of a work-around, without it we get
6248 * warning spews on shutdown with SQPOLL set and affinity
6249 * set to a single CPU.
6251 kthread_park(ctx->sqo_thread);
6252 kthread_stop(ctx->sqo_thread);
6253 ctx->sqo_thread = NULL;
6257 static void io_finish_async(struct io_ring_ctx *ctx)
6259 io_sq_thread_stop(ctx);
6262 io_wq_destroy(ctx->io_wq);
6267 #if defined(CONFIG_UNIX)
6269 * Ensure the UNIX gc is aware of our file set, so we are certain that
6270 * the io_uring can be safely unregistered on process exit, even if we have
6271 * loops in the file referencing.
6273 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6275 struct sock *sk = ctx->ring_sock->sk;
6276 struct scm_fp_list *fpl;
6277 struct sk_buff *skb;
6280 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6284 skb = alloc_skb(0, GFP_KERNEL);
6293 fpl->user = get_uid(ctx->user);
6294 for (i = 0; i < nr; i++) {
6295 struct file *file = io_file_from_index(ctx, i + offset);
6299 fpl->fp[nr_files] = get_file(file);
6300 unix_inflight(fpl->user, fpl->fp[nr_files]);
6305 fpl->max = SCM_MAX_FD;
6306 fpl->count = nr_files;
6307 UNIXCB(skb).fp = fpl;
6308 skb->destructor = unix_destruct_scm;
6309 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6310 skb_queue_head(&sk->sk_receive_queue, skb);
6312 for (i = 0; i < nr_files; i++)
6323 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6324 * causes regular reference counting to break down. We rely on the UNIX
6325 * garbage collection to take care of this problem for us.
6327 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6329 unsigned left, total;
6333 left = ctx->nr_user_files;
6335 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6337 ret = __io_sqe_files_scm(ctx, this_files, total);
6341 total += this_files;
6347 while (total < ctx->nr_user_files) {
6348 struct file *file = io_file_from_index(ctx, total);
6358 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6364 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6369 for (i = 0; i < nr_tables; i++) {
6370 struct fixed_file_table *table = &ctx->file_data->table[i];
6371 unsigned this_files;
6373 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6374 table->files = kcalloc(this_files, sizeof(struct file *),
6378 nr_files -= this_files;
6384 for (i = 0; i < nr_tables; i++) {
6385 struct fixed_file_table *table = &ctx->file_data->table[i];
6386 kfree(table->files);
6391 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6393 #if defined(CONFIG_UNIX)
6394 struct sock *sock = ctx->ring_sock->sk;
6395 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6396 struct sk_buff *skb;
6399 __skb_queue_head_init(&list);
6402 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6403 * remove this entry and rearrange the file array.
6405 skb = skb_dequeue(head);
6407 struct scm_fp_list *fp;
6409 fp = UNIXCB(skb).fp;
6410 for (i = 0; i < fp->count; i++) {
6413 if (fp->fp[i] != file)
6416 unix_notinflight(fp->user, fp->fp[i]);
6417 left = fp->count - 1 - i;
6419 memmove(&fp->fp[i], &fp->fp[i + 1],
6420 left * sizeof(struct file *));
6427 __skb_queue_tail(&list, skb);
6437 __skb_queue_tail(&list, skb);
6439 skb = skb_dequeue(head);
6442 if (skb_peek(&list)) {
6443 spin_lock_irq(&head->lock);
6444 while ((skb = __skb_dequeue(&list)) != NULL)
6445 __skb_queue_tail(head, skb);
6446 spin_unlock_irq(&head->lock);
6453 struct io_file_put {
6454 struct list_head list;
6458 static void io_file_put_work(struct work_struct *work)
6460 struct fixed_file_ref_node *ref_node;
6461 struct fixed_file_data *file_data;
6462 struct io_ring_ctx *ctx;
6463 struct io_file_put *pfile, *tmp;
6464 unsigned long flags;
6466 ref_node = container_of(work, struct fixed_file_ref_node, work);
6467 file_data = ref_node->file_data;
6468 ctx = file_data->ctx;
6470 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6471 list_del_init(&pfile->list);
6472 io_ring_file_put(ctx, pfile->file);
6476 spin_lock_irqsave(&file_data->lock, flags);
6477 list_del_init(&ref_node->node);
6478 spin_unlock_irqrestore(&file_data->lock, flags);
6480 percpu_ref_exit(&ref_node->refs);
6482 percpu_ref_put(&file_data->refs);
6485 static void io_file_data_ref_zero(struct percpu_ref *ref)
6487 struct fixed_file_ref_node *ref_node;
6489 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6491 queue_work(system_wq, &ref_node->work);
6494 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6495 struct io_ring_ctx *ctx)
6497 struct fixed_file_ref_node *ref_node;
6499 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6501 return ERR_PTR(-ENOMEM);
6503 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6506 return ERR_PTR(-ENOMEM);
6508 INIT_LIST_HEAD(&ref_node->node);
6509 INIT_LIST_HEAD(&ref_node->file_list);
6510 INIT_WORK(&ref_node->work, io_file_put_work);
6511 ref_node->file_data = ctx->file_data;
6516 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6518 percpu_ref_exit(&ref_node->refs);
6522 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6525 __s32 __user *fds = (__s32 __user *) arg;
6530 struct fixed_file_ref_node *ref_node;
6531 unsigned long flags;
6537 if (nr_args > IORING_MAX_FIXED_FILES)
6540 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6541 if (!ctx->file_data)
6543 ctx->file_data->ctx = ctx;
6544 init_completion(&ctx->file_data->done);
6545 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6546 spin_lock_init(&ctx->file_data->lock);
6548 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6549 ctx->file_data->table = kcalloc(nr_tables,
6550 sizeof(struct fixed_file_table),
6552 if (!ctx->file_data->table) {
6553 kfree(ctx->file_data);
6554 ctx->file_data = NULL;
6558 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6559 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6560 kfree(ctx->file_data->table);
6561 kfree(ctx->file_data);
6562 ctx->file_data = NULL;
6566 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6567 percpu_ref_exit(&ctx->file_data->refs);
6568 kfree(ctx->file_data->table);
6569 kfree(ctx->file_data);
6570 ctx->file_data = NULL;
6574 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6575 struct fixed_file_table *table;
6579 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6581 /* allow sparse sets */
6587 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6588 index = i & IORING_FILE_TABLE_MASK;
6596 * Don't allow io_uring instances to be registered. If UNIX
6597 * isn't enabled, then this causes a reference cycle and this
6598 * instance can never get freed. If UNIX is enabled we'll
6599 * handle it just fine, but there's still no point in allowing
6600 * a ring fd as it doesn't support regular read/write anyway.
6602 if (file->f_op == &io_uring_fops) {
6607 table->files[index] = file;
6611 for (i = 0; i < ctx->nr_user_files; i++) {
6612 file = io_file_from_index(ctx, i);
6616 for (i = 0; i < nr_tables; i++)
6617 kfree(ctx->file_data->table[i].files);
6619 kfree(ctx->file_data->table);
6620 kfree(ctx->file_data);
6621 ctx->file_data = NULL;
6622 ctx->nr_user_files = 0;
6626 ret = io_sqe_files_scm(ctx);
6628 io_sqe_files_unregister(ctx);
6632 ref_node = alloc_fixed_file_ref_node(ctx);
6633 if (IS_ERR(ref_node)) {
6634 io_sqe_files_unregister(ctx);
6635 return PTR_ERR(ref_node);
6638 ctx->file_data->cur_refs = &ref_node->refs;
6639 spin_lock_irqsave(&ctx->file_data->lock, flags);
6640 list_add(&ref_node->node, &ctx->file_data->ref_list);
6641 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6642 percpu_ref_get(&ctx->file_data->refs);
6646 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6649 #if defined(CONFIG_UNIX)
6650 struct sock *sock = ctx->ring_sock->sk;
6651 struct sk_buff_head *head = &sock->sk_receive_queue;
6652 struct sk_buff *skb;
6655 * See if we can merge this file into an existing skb SCM_RIGHTS
6656 * file set. If there's no room, fall back to allocating a new skb
6657 * and filling it in.
6659 spin_lock_irq(&head->lock);
6660 skb = skb_peek(head);
6662 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6664 if (fpl->count < SCM_MAX_FD) {
6665 __skb_unlink(skb, head);
6666 spin_unlock_irq(&head->lock);
6667 fpl->fp[fpl->count] = get_file(file);
6668 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6670 spin_lock_irq(&head->lock);
6671 __skb_queue_head(head, skb);
6676 spin_unlock_irq(&head->lock);
6683 return __io_sqe_files_scm(ctx, 1, index);
6689 static int io_queue_file_removal(struct fixed_file_data *data,
6692 struct io_file_put *pfile;
6693 struct percpu_ref *refs = data->cur_refs;
6694 struct fixed_file_ref_node *ref_node;
6696 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6700 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6702 list_add(&pfile->list, &ref_node->file_list);
6707 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6708 struct io_uring_files_update *up,
6711 struct fixed_file_data *data = ctx->file_data;
6712 struct fixed_file_ref_node *ref_node;
6717 unsigned long flags;
6718 bool needs_switch = false;
6720 if (check_add_overflow(up->offset, nr_args, &done))
6722 if (done > ctx->nr_user_files)
6725 ref_node = alloc_fixed_file_ref_node(ctx);
6726 if (IS_ERR(ref_node))
6727 return PTR_ERR(ref_node);
6730 fds = u64_to_user_ptr(up->fds);
6732 struct fixed_file_table *table;
6736 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6740 i = array_index_nospec(up->offset, ctx->nr_user_files);
6741 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6742 index = i & IORING_FILE_TABLE_MASK;
6743 if (table->files[index]) {
6744 file = io_file_from_index(ctx, index);
6745 err = io_queue_file_removal(data, file);
6748 table->files[index] = NULL;
6749 needs_switch = true;
6758 * Don't allow io_uring instances to be registered. If
6759 * UNIX isn't enabled, then this causes a reference
6760 * cycle and this instance can never get freed. If UNIX
6761 * is enabled we'll handle it just fine, but there's
6762 * still no point in allowing a ring fd as it doesn't
6763 * support regular read/write anyway.
6765 if (file->f_op == &io_uring_fops) {
6770 table->files[index] = file;
6771 err = io_sqe_file_register(ctx, file, i);
6781 percpu_ref_kill(data->cur_refs);
6782 spin_lock_irqsave(&data->lock, flags);
6783 list_add(&ref_node->node, &data->ref_list);
6784 data->cur_refs = &ref_node->refs;
6785 spin_unlock_irqrestore(&data->lock, flags);
6786 percpu_ref_get(&ctx->file_data->refs);
6788 destroy_fixed_file_ref_node(ref_node);
6790 return done ? done : err;
6793 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6796 struct io_uring_files_update up;
6798 if (!ctx->file_data)
6802 if (copy_from_user(&up, arg, sizeof(up)))
6807 return __io_sqe_files_update(ctx, &up, nr_args);
6810 static void io_free_work(struct io_wq_work *work)
6812 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6814 /* Consider that io_steal_work() relies on this ref */
6818 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6819 struct io_uring_params *p)
6821 struct io_wq_data data;
6823 struct io_ring_ctx *ctx_attach;
6824 unsigned int concurrency;
6827 data.user = ctx->user;
6828 data.free_work = io_free_work;
6830 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6831 /* Do QD, or 4 * CPUS, whatever is smallest */
6832 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6834 ctx->io_wq = io_wq_create(concurrency, &data);
6835 if (IS_ERR(ctx->io_wq)) {
6836 ret = PTR_ERR(ctx->io_wq);
6842 f = fdget(p->wq_fd);
6846 if (f.file->f_op != &io_uring_fops) {
6851 ctx_attach = f.file->private_data;
6852 /* @io_wq is protected by holding the fd */
6853 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6858 ctx->io_wq = ctx_attach->io_wq;
6864 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6865 struct io_uring_params *p)
6869 init_waitqueue_head(&ctx->sqo_wait);
6870 mmgrab(current->mm);
6871 ctx->sqo_mm = current->mm;
6873 if (ctx->flags & IORING_SETUP_SQPOLL) {
6875 if (!capable(CAP_SYS_ADMIN))
6878 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6879 if (!ctx->sq_thread_idle)
6880 ctx->sq_thread_idle = HZ;
6882 if (p->flags & IORING_SETUP_SQ_AFF) {
6883 int cpu = p->sq_thread_cpu;
6886 if (cpu >= nr_cpu_ids)
6888 if (!cpu_online(cpu))
6891 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6895 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6898 if (IS_ERR(ctx->sqo_thread)) {
6899 ret = PTR_ERR(ctx->sqo_thread);
6900 ctx->sqo_thread = NULL;
6903 wake_up_process(ctx->sqo_thread);
6904 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6905 /* Can't have SQ_AFF without SQPOLL */
6910 ret = io_init_wq_offload(ctx, p);
6916 io_finish_async(ctx);
6917 mmdrop(ctx->sqo_mm);
6922 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6924 atomic_long_sub(nr_pages, &user->locked_vm);
6927 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6929 unsigned long page_limit, cur_pages, new_pages;
6931 /* Don't allow more pages than we can safely lock */
6932 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6935 cur_pages = atomic_long_read(&user->locked_vm);
6936 new_pages = cur_pages + nr_pages;
6937 if (new_pages > page_limit)
6939 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6940 new_pages) != cur_pages);
6945 static void io_mem_free(void *ptr)
6952 page = virt_to_head_page(ptr);
6953 if (put_page_testzero(page))
6954 free_compound_page(page);
6957 static void *io_mem_alloc(size_t size)
6959 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6962 return (void *) __get_free_pages(gfp_flags, get_order(size));
6965 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6968 struct io_rings *rings;
6969 size_t off, sq_array_size;
6971 off = struct_size(rings, cqes, cq_entries);
6972 if (off == SIZE_MAX)
6976 off = ALIGN(off, SMP_CACHE_BYTES);
6981 sq_array_size = array_size(sizeof(u32), sq_entries);
6982 if (sq_array_size == SIZE_MAX)
6985 if (check_add_overflow(off, sq_array_size, &off))
6994 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6998 pages = (size_t)1 << get_order(
6999 rings_size(sq_entries, cq_entries, NULL));
7000 pages += (size_t)1 << get_order(
7001 array_size(sizeof(struct io_uring_sqe), sq_entries));
7006 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7010 if (!ctx->user_bufs)
7013 for (i = 0; i < ctx->nr_user_bufs; i++) {
7014 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7016 for (j = 0; j < imu->nr_bvecs; j++)
7017 unpin_user_page(imu->bvec[j].bv_page);
7019 if (ctx->account_mem)
7020 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7025 kfree(ctx->user_bufs);
7026 ctx->user_bufs = NULL;
7027 ctx->nr_user_bufs = 0;
7031 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7032 void __user *arg, unsigned index)
7034 struct iovec __user *src;
7036 #ifdef CONFIG_COMPAT
7038 struct compat_iovec __user *ciovs;
7039 struct compat_iovec ciov;
7041 ciovs = (struct compat_iovec __user *) arg;
7042 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7045 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7046 dst->iov_len = ciov.iov_len;
7050 src = (struct iovec __user *) arg;
7051 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7056 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7059 struct vm_area_struct **vmas = NULL;
7060 struct page **pages = NULL;
7061 int i, j, got_pages = 0;
7066 if (!nr_args || nr_args > UIO_MAXIOV)
7069 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7071 if (!ctx->user_bufs)
7074 for (i = 0; i < nr_args; i++) {
7075 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7076 unsigned long off, start, end, ubuf;
7081 ret = io_copy_iov(ctx, &iov, arg, i);
7086 * Don't impose further limits on the size and buffer
7087 * constraints here, we'll -EINVAL later when IO is
7088 * submitted if they are wrong.
7091 if (!iov.iov_base || !iov.iov_len)
7094 /* arbitrary limit, but we need something */
7095 if (iov.iov_len > SZ_1G)
7098 ubuf = (unsigned long) iov.iov_base;
7099 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7100 start = ubuf >> PAGE_SHIFT;
7101 nr_pages = end - start;
7103 if (ctx->account_mem) {
7104 ret = io_account_mem(ctx->user, nr_pages);
7110 if (!pages || nr_pages > got_pages) {
7113 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7115 vmas = kvmalloc_array(nr_pages,
7116 sizeof(struct vm_area_struct *),
7118 if (!pages || !vmas) {
7120 if (ctx->account_mem)
7121 io_unaccount_mem(ctx->user, nr_pages);
7124 got_pages = nr_pages;
7127 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7131 if (ctx->account_mem)
7132 io_unaccount_mem(ctx->user, nr_pages);
7137 down_read(¤t->mm->mmap_sem);
7138 pret = pin_user_pages(ubuf, nr_pages,
7139 FOLL_WRITE | FOLL_LONGTERM,
7141 if (pret == nr_pages) {
7142 /* don't support file backed memory */
7143 for (j = 0; j < nr_pages; j++) {
7144 struct vm_area_struct *vma = vmas[j];
7147 !is_file_hugepages(vma->vm_file)) {
7153 ret = pret < 0 ? pret : -EFAULT;
7155 up_read(¤t->mm->mmap_sem);
7158 * if we did partial map, or found file backed vmas,
7159 * release any pages we did get
7162 unpin_user_pages(pages, pret);
7163 if (ctx->account_mem)
7164 io_unaccount_mem(ctx->user, nr_pages);
7169 off = ubuf & ~PAGE_MASK;
7171 for (j = 0; j < nr_pages; j++) {
7174 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7175 imu->bvec[j].bv_page = pages[j];
7176 imu->bvec[j].bv_len = vec_len;
7177 imu->bvec[j].bv_offset = off;
7181 /* store original address for later verification */
7183 imu->len = iov.iov_len;
7184 imu->nr_bvecs = nr_pages;
7186 ctx->nr_user_bufs++;
7194 io_sqe_buffer_unregister(ctx);
7198 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7200 __s32 __user *fds = arg;
7206 if (copy_from_user(&fd, fds, sizeof(*fds)))
7209 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7210 if (IS_ERR(ctx->cq_ev_fd)) {
7211 int ret = PTR_ERR(ctx->cq_ev_fd);
7212 ctx->cq_ev_fd = NULL;
7219 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7221 if (ctx->cq_ev_fd) {
7222 eventfd_ctx_put(ctx->cq_ev_fd);
7223 ctx->cq_ev_fd = NULL;
7230 static int __io_destroy_buffers(int id, void *p, void *data)
7232 struct io_ring_ctx *ctx = data;
7233 struct io_buffer *buf = p;
7235 __io_remove_buffers(ctx, buf, id, -1U);
7239 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7241 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7242 idr_destroy(&ctx->io_buffer_idr);
7245 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7247 io_finish_async(ctx);
7249 mmdrop(ctx->sqo_mm);
7251 io_iopoll_reap_events(ctx);
7252 io_sqe_buffer_unregister(ctx);
7253 io_sqe_files_unregister(ctx);
7254 io_eventfd_unregister(ctx);
7255 io_destroy_buffers(ctx);
7256 idr_destroy(&ctx->personality_idr);
7258 #if defined(CONFIG_UNIX)
7259 if (ctx->ring_sock) {
7260 ctx->ring_sock->file = NULL; /* so that iput() is called */
7261 sock_release(ctx->ring_sock);
7265 io_mem_free(ctx->rings);
7266 io_mem_free(ctx->sq_sqes);
7268 percpu_ref_exit(&ctx->refs);
7269 if (ctx->account_mem)
7270 io_unaccount_mem(ctx->user,
7271 ring_pages(ctx->sq_entries, ctx->cq_entries));
7272 free_uid(ctx->user);
7273 put_cred(ctx->creds);
7274 kfree(ctx->completions);
7275 kfree(ctx->cancel_hash);
7276 kmem_cache_free(req_cachep, ctx->fallback_req);
7280 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7282 struct io_ring_ctx *ctx = file->private_data;
7285 poll_wait(file, &ctx->cq_wait, wait);
7287 * synchronizes with barrier from wq_has_sleeper call in
7291 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7292 ctx->rings->sq_ring_entries)
7293 mask |= EPOLLOUT | EPOLLWRNORM;
7294 if (io_cqring_events(ctx, false))
7295 mask |= EPOLLIN | EPOLLRDNORM;
7300 static int io_uring_fasync(int fd, struct file *file, int on)
7302 struct io_ring_ctx *ctx = file->private_data;
7304 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7307 static int io_remove_personalities(int id, void *p, void *data)
7309 struct io_ring_ctx *ctx = data;
7310 const struct cred *cred;
7312 cred = idr_remove(&ctx->personality_idr, id);
7318 static void io_ring_exit_work(struct work_struct *work)
7320 struct io_ring_ctx *ctx;
7322 ctx = container_of(work, struct io_ring_ctx, exit_work);
7324 io_cqring_overflow_flush(ctx, true);
7326 wait_for_completion(&ctx->completions[0]);
7327 io_ring_ctx_free(ctx);
7330 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7332 mutex_lock(&ctx->uring_lock);
7333 percpu_ref_kill(&ctx->refs);
7334 mutex_unlock(&ctx->uring_lock);
7337 * Wait for sq thread to idle, if we have one. It won't spin on new
7338 * work after we've killed the ctx ref above. This is important to do
7339 * before we cancel existing commands, as the thread could otherwise
7340 * be queueing new work post that. If that's work we need to cancel,
7341 * it could cause shutdown to hang.
7343 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7346 io_kill_timeouts(ctx);
7347 io_poll_remove_all(ctx);
7350 io_wq_cancel_all(ctx->io_wq);
7352 io_iopoll_reap_events(ctx);
7353 /* if we failed setting up the ctx, we might not have any rings */
7355 io_cqring_overflow_flush(ctx, true);
7356 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7357 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7358 queue_work(system_wq, &ctx->exit_work);
7361 static int io_uring_release(struct inode *inode, struct file *file)
7363 struct io_ring_ctx *ctx = file->private_data;
7365 file->private_data = NULL;
7366 io_ring_ctx_wait_and_kill(ctx);
7370 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7371 struct files_struct *files)
7373 struct io_kiocb *req;
7376 while (!list_empty_careful(&ctx->inflight_list)) {
7377 struct io_kiocb *cancel_req = NULL;
7379 spin_lock_irq(&ctx->inflight_lock);
7380 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7381 if (req->work.files != files)
7383 /* req is being completed, ignore */
7384 if (!refcount_inc_not_zero(&req->refs))
7390 prepare_to_wait(&ctx->inflight_wait, &wait,
7391 TASK_UNINTERRUPTIBLE);
7392 spin_unlock_irq(&ctx->inflight_lock);
7394 /* We need to keep going until we don't find a matching req */
7398 if (cancel_req->flags & REQ_F_OVERFLOW) {
7399 spin_lock_irq(&ctx->completion_lock);
7400 list_del(&cancel_req->list);
7401 cancel_req->flags &= ~REQ_F_OVERFLOW;
7402 if (list_empty(&ctx->cq_overflow_list)) {
7403 clear_bit(0, &ctx->sq_check_overflow);
7404 clear_bit(0, &ctx->cq_check_overflow);
7406 spin_unlock_irq(&ctx->completion_lock);
7408 WRITE_ONCE(ctx->rings->cq_overflow,
7409 atomic_inc_return(&ctx->cached_cq_overflow));
7412 * Put inflight ref and overflow ref. If that's
7413 * all we had, then we're done with this request.
7415 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7416 io_put_req(cancel_req);
7421 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7422 io_put_req(cancel_req);
7425 finish_wait(&ctx->inflight_wait, &wait);
7428 static int io_uring_flush(struct file *file, void *data)
7430 struct io_ring_ctx *ctx = file->private_data;
7432 io_uring_cancel_files(ctx, data);
7435 * If the task is going away, cancel work it may have pending
7437 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7438 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7443 static void *io_uring_validate_mmap_request(struct file *file,
7444 loff_t pgoff, size_t sz)
7446 struct io_ring_ctx *ctx = file->private_data;
7447 loff_t offset = pgoff << PAGE_SHIFT;
7452 case IORING_OFF_SQ_RING:
7453 case IORING_OFF_CQ_RING:
7456 case IORING_OFF_SQES:
7460 return ERR_PTR(-EINVAL);
7463 page = virt_to_head_page(ptr);
7464 if (sz > page_size(page))
7465 return ERR_PTR(-EINVAL);
7472 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7474 size_t sz = vma->vm_end - vma->vm_start;
7478 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7480 return PTR_ERR(ptr);
7482 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7483 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7486 #else /* !CONFIG_MMU */
7488 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7490 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7493 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7495 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7498 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7499 unsigned long addr, unsigned long len,
7500 unsigned long pgoff, unsigned long flags)
7504 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7506 return PTR_ERR(ptr);
7508 return (unsigned long) ptr;
7511 #endif /* !CONFIG_MMU */
7513 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7514 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7517 struct io_ring_ctx *ctx;
7522 if (current->task_works)
7525 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7533 if (f.file->f_op != &io_uring_fops)
7537 ctx = f.file->private_data;
7538 if (!percpu_ref_tryget(&ctx->refs))
7542 * For SQ polling, the thread will do all submissions and completions.
7543 * Just return the requested submit count, and wake the thread if
7547 if (ctx->flags & IORING_SETUP_SQPOLL) {
7548 if (!list_empty_careful(&ctx->cq_overflow_list))
7549 io_cqring_overflow_flush(ctx, false);
7550 if (flags & IORING_ENTER_SQ_WAKEUP)
7551 wake_up(&ctx->sqo_wait);
7552 submitted = to_submit;
7553 } else if (to_submit) {
7554 mutex_lock(&ctx->uring_lock);
7555 submitted = io_submit_sqes(ctx, to_submit, f.file, fd, false);
7556 mutex_unlock(&ctx->uring_lock);
7558 if (submitted != to_submit)
7561 if (flags & IORING_ENTER_GETEVENTS) {
7562 unsigned nr_events = 0;
7564 min_complete = min(min_complete, ctx->cq_entries);
7567 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7568 * space applications don't need to do io completion events
7569 * polling again, they can rely on io_sq_thread to do polling
7570 * work, which can reduce cpu usage and uring_lock contention.
7572 if (ctx->flags & IORING_SETUP_IOPOLL &&
7573 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7574 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7576 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7581 percpu_ref_put(&ctx->refs);
7584 return submitted ? submitted : ret;
7587 #ifdef CONFIG_PROC_FS
7588 static int io_uring_show_cred(int id, void *p, void *data)
7590 const struct cred *cred = p;
7591 struct seq_file *m = data;
7592 struct user_namespace *uns = seq_user_ns(m);
7593 struct group_info *gi;
7598 seq_printf(m, "%5d\n", id);
7599 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7600 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7601 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7602 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7603 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7604 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7605 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7606 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7607 seq_puts(m, "\n\tGroups:\t");
7608 gi = cred->group_info;
7609 for (g = 0; g < gi->ngroups; g++) {
7610 seq_put_decimal_ull(m, g ? " " : "",
7611 from_kgid_munged(uns, gi->gid[g]));
7613 seq_puts(m, "\n\tCapEff:\t");
7614 cap = cred->cap_effective;
7615 CAP_FOR_EACH_U32(__capi)
7616 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7621 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7625 mutex_lock(&ctx->uring_lock);
7626 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7627 for (i = 0; i < ctx->nr_user_files; i++) {
7628 struct fixed_file_table *table;
7631 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7632 f = table->files[i & IORING_FILE_TABLE_MASK];
7634 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7636 seq_printf(m, "%5u: <none>\n", i);
7638 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7639 for (i = 0; i < ctx->nr_user_bufs; i++) {
7640 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7642 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7643 (unsigned int) buf->len);
7645 if (!idr_is_empty(&ctx->personality_idr)) {
7646 seq_printf(m, "Personalities:\n");
7647 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7649 seq_printf(m, "PollList:\n");
7650 spin_lock_irq(&ctx->completion_lock);
7651 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7652 struct hlist_head *list = &ctx->cancel_hash[i];
7653 struct io_kiocb *req;
7655 hlist_for_each_entry(req, list, hash_node)
7656 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7657 req->task->task_works != NULL);
7659 spin_unlock_irq(&ctx->completion_lock);
7660 mutex_unlock(&ctx->uring_lock);
7663 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7665 struct io_ring_ctx *ctx = f->private_data;
7667 if (percpu_ref_tryget(&ctx->refs)) {
7668 __io_uring_show_fdinfo(ctx, m);
7669 percpu_ref_put(&ctx->refs);
7674 static const struct file_operations io_uring_fops = {
7675 .release = io_uring_release,
7676 .flush = io_uring_flush,
7677 .mmap = io_uring_mmap,
7679 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7680 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7682 .poll = io_uring_poll,
7683 .fasync = io_uring_fasync,
7684 #ifdef CONFIG_PROC_FS
7685 .show_fdinfo = io_uring_show_fdinfo,
7689 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7690 struct io_uring_params *p)
7692 struct io_rings *rings;
7693 size_t size, sq_array_offset;
7695 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7696 if (size == SIZE_MAX)
7699 rings = io_mem_alloc(size);
7704 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7705 rings->sq_ring_mask = p->sq_entries - 1;
7706 rings->cq_ring_mask = p->cq_entries - 1;
7707 rings->sq_ring_entries = p->sq_entries;
7708 rings->cq_ring_entries = p->cq_entries;
7709 ctx->sq_mask = rings->sq_ring_mask;
7710 ctx->cq_mask = rings->cq_ring_mask;
7711 ctx->sq_entries = rings->sq_ring_entries;
7712 ctx->cq_entries = rings->cq_ring_entries;
7714 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7715 if (size == SIZE_MAX) {
7716 io_mem_free(ctx->rings);
7721 ctx->sq_sqes = io_mem_alloc(size);
7722 if (!ctx->sq_sqes) {
7723 io_mem_free(ctx->rings);
7732 * Allocate an anonymous fd, this is what constitutes the application
7733 * visible backing of an io_uring instance. The application mmaps this
7734 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7735 * we have to tie this fd to a socket for file garbage collection purposes.
7737 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7742 #if defined(CONFIG_UNIX)
7743 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7749 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7753 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7754 O_RDWR | O_CLOEXEC);
7757 ret = PTR_ERR(file);
7761 #if defined(CONFIG_UNIX)
7762 ctx->ring_sock->file = file;
7764 fd_install(ret, file);
7767 #if defined(CONFIG_UNIX)
7768 sock_release(ctx->ring_sock);
7769 ctx->ring_sock = NULL;
7774 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7776 struct user_struct *user = NULL;
7777 struct io_ring_ctx *ctx;
7783 if (entries > IORING_MAX_ENTRIES) {
7784 if (!(p->flags & IORING_SETUP_CLAMP))
7786 entries = IORING_MAX_ENTRIES;
7790 * Use twice as many entries for the CQ ring. It's possible for the
7791 * application to drive a higher depth than the size of the SQ ring,
7792 * since the sqes are only used at submission time. This allows for
7793 * some flexibility in overcommitting a bit. If the application has
7794 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7795 * of CQ ring entries manually.
7797 p->sq_entries = roundup_pow_of_two(entries);
7798 if (p->flags & IORING_SETUP_CQSIZE) {
7800 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7801 * to a power-of-two, if it isn't already. We do NOT impose
7802 * any cq vs sq ring sizing.
7804 if (p->cq_entries < p->sq_entries)
7806 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7807 if (!(p->flags & IORING_SETUP_CLAMP))
7809 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7811 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7813 p->cq_entries = 2 * p->sq_entries;
7816 user = get_uid(current_user());
7817 account_mem = !capable(CAP_IPC_LOCK);
7820 ret = io_account_mem(user,
7821 ring_pages(p->sq_entries, p->cq_entries));
7828 ctx = io_ring_ctx_alloc(p);
7831 io_unaccount_mem(user, ring_pages(p->sq_entries,
7836 ctx->compat = in_compat_syscall();
7837 ctx->account_mem = account_mem;
7839 ctx->creds = get_current_cred();
7841 ret = io_allocate_scq_urings(ctx, p);
7845 ret = io_sq_offload_start(ctx, p);
7849 memset(&p->sq_off, 0, sizeof(p->sq_off));
7850 p->sq_off.head = offsetof(struct io_rings, sq.head);
7851 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7852 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7853 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7854 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7855 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7856 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7858 memset(&p->cq_off, 0, sizeof(p->cq_off));
7859 p->cq_off.head = offsetof(struct io_rings, cq.head);
7860 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7861 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7862 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7863 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7864 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7867 * Install ring fd as the very last thing, so we don't risk someone
7868 * having closed it before we finish setup
7870 ret = io_uring_get_fd(ctx);
7874 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7875 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7876 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7877 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7880 io_ring_ctx_wait_and_kill(ctx);
7885 * Sets up an aio uring context, and returns the fd. Applications asks for a
7886 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7887 * params structure passed in.
7889 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7891 struct io_uring_params p;
7895 if (copy_from_user(&p, params, sizeof(p)))
7897 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7902 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7903 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7904 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7907 ret = io_uring_create(entries, &p);
7911 if (copy_to_user(params, &p, sizeof(p)))
7917 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7918 struct io_uring_params __user *, params)
7920 return io_uring_setup(entries, params);
7923 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7925 struct io_uring_probe *p;
7929 size = struct_size(p, ops, nr_args);
7930 if (size == SIZE_MAX)
7932 p = kzalloc(size, GFP_KERNEL);
7937 if (copy_from_user(p, arg, size))
7940 if (memchr_inv(p, 0, size))
7943 p->last_op = IORING_OP_LAST - 1;
7944 if (nr_args > IORING_OP_LAST)
7945 nr_args = IORING_OP_LAST;
7947 for (i = 0; i < nr_args; i++) {
7949 if (!io_op_defs[i].not_supported)
7950 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7955 if (copy_to_user(arg, p, size))
7962 static int io_register_personality(struct io_ring_ctx *ctx)
7964 const struct cred *creds = get_current_cred();
7967 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7968 USHRT_MAX, GFP_KERNEL);
7974 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7976 const struct cred *old_creds;
7978 old_creds = idr_remove(&ctx->personality_idr, id);
7980 put_cred(old_creds);
7987 static bool io_register_op_must_quiesce(int op)
7990 case IORING_UNREGISTER_FILES:
7991 case IORING_REGISTER_FILES_UPDATE:
7992 case IORING_REGISTER_PROBE:
7993 case IORING_REGISTER_PERSONALITY:
7994 case IORING_UNREGISTER_PERSONALITY:
8001 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8002 void __user *arg, unsigned nr_args)
8003 __releases(ctx->uring_lock)
8004 __acquires(ctx->uring_lock)
8009 * We're inside the ring mutex, if the ref is already dying, then
8010 * someone else killed the ctx or is already going through
8011 * io_uring_register().
8013 if (percpu_ref_is_dying(&ctx->refs))
8016 if (io_register_op_must_quiesce(opcode)) {
8017 percpu_ref_kill(&ctx->refs);
8020 * Drop uring mutex before waiting for references to exit. If
8021 * another thread is currently inside io_uring_enter() it might
8022 * need to grab the uring_lock to make progress. If we hold it
8023 * here across the drain wait, then we can deadlock. It's safe
8024 * to drop the mutex here, since no new references will come in
8025 * after we've killed the percpu ref.
8027 mutex_unlock(&ctx->uring_lock);
8028 ret = wait_for_completion_interruptible(&ctx->completions[0]);
8029 mutex_lock(&ctx->uring_lock);
8031 percpu_ref_resurrect(&ctx->refs);
8038 case IORING_REGISTER_BUFFERS:
8039 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8041 case IORING_UNREGISTER_BUFFERS:
8045 ret = io_sqe_buffer_unregister(ctx);
8047 case IORING_REGISTER_FILES:
8048 ret = io_sqe_files_register(ctx, arg, nr_args);
8050 case IORING_UNREGISTER_FILES:
8054 ret = io_sqe_files_unregister(ctx);
8056 case IORING_REGISTER_FILES_UPDATE:
8057 ret = io_sqe_files_update(ctx, arg, nr_args);
8059 case IORING_REGISTER_EVENTFD:
8060 case IORING_REGISTER_EVENTFD_ASYNC:
8064 ret = io_eventfd_register(ctx, arg);
8067 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8068 ctx->eventfd_async = 1;
8070 ctx->eventfd_async = 0;
8072 case IORING_UNREGISTER_EVENTFD:
8076 ret = io_eventfd_unregister(ctx);
8078 case IORING_REGISTER_PROBE:
8080 if (!arg || nr_args > 256)
8082 ret = io_probe(ctx, arg, nr_args);
8084 case IORING_REGISTER_PERSONALITY:
8088 ret = io_register_personality(ctx);
8090 case IORING_UNREGISTER_PERSONALITY:
8094 ret = io_unregister_personality(ctx, nr_args);
8101 if (io_register_op_must_quiesce(opcode)) {
8102 /* bring the ctx back to life */
8103 percpu_ref_reinit(&ctx->refs);
8105 reinit_completion(&ctx->completions[0]);
8110 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8111 void __user *, arg, unsigned int, nr_args)
8113 struct io_ring_ctx *ctx;
8122 if (f.file->f_op != &io_uring_fops)
8125 ctx = f.file->private_data;
8127 mutex_lock(&ctx->uring_lock);
8128 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8129 mutex_unlock(&ctx->uring_lock);
8130 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8131 ctx->cq_ev_fd != NULL, ret);
8137 static int __init io_uring_init(void)
8139 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8140 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8141 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8144 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8145 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8146 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8147 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8148 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8149 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8150 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8151 BUILD_BUG_SQE_ELEM(8, __u64, off);
8152 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8153 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8154 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8155 BUILD_BUG_SQE_ELEM(24, __u32, len);
8156 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8157 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8158 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8159 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8160 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8161 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8162 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8163 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8164 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8165 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8166 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8167 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8168 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8169 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8170 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8171 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8172 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8173 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8175 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8176 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8177 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8180 __initcall(io_uring_init);