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)
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)
2053 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2054 bool force_nonblock)
2056 struct io_ring_ctx *ctx = req->ctx;
2057 struct kiocb *kiocb = &req->rw.kiocb;
2061 if (S_ISREG(file_inode(req->file)->i_mode))
2062 req->flags |= REQ_F_ISREG;
2064 kiocb->ki_pos = READ_ONCE(sqe->off);
2065 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2066 req->flags |= REQ_F_CUR_POS;
2067 kiocb->ki_pos = req->file->f_pos;
2069 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2070 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2071 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2075 ioprio = READ_ONCE(sqe->ioprio);
2077 ret = ioprio_check_cap(ioprio);
2081 kiocb->ki_ioprio = ioprio;
2083 kiocb->ki_ioprio = get_current_ioprio();
2085 /* don't allow async punt if RWF_NOWAIT was requested */
2086 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2087 (req->file->f_flags & O_NONBLOCK))
2088 req->flags |= REQ_F_NOWAIT;
2091 kiocb->ki_flags |= IOCB_NOWAIT;
2093 if (ctx->flags & IORING_SETUP_IOPOLL) {
2094 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2095 !kiocb->ki_filp->f_op->iopoll)
2098 kiocb->ki_flags |= IOCB_HIPRI;
2099 kiocb->ki_complete = io_complete_rw_iopoll;
2102 if (kiocb->ki_flags & IOCB_HIPRI)
2104 kiocb->ki_complete = io_complete_rw;
2107 req->rw.addr = READ_ONCE(sqe->addr);
2108 req->rw.len = READ_ONCE(sqe->len);
2109 /* we own ->private, reuse it for the buffer index / buffer ID */
2110 req->rw.kiocb.private = (void *) (unsigned long)
2111 READ_ONCE(sqe->buf_index);
2115 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2121 case -ERESTARTNOINTR:
2122 case -ERESTARTNOHAND:
2123 case -ERESTART_RESTARTBLOCK:
2125 * We can't just restart the syscall, since previously
2126 * submitted sqes may already be in progress. Just fail this
2132 kiocb->ki_complete(kiocb, ret, 0);
2136 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2138 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2140 if (req->flags & REQ_F_CUR_POS)
2141 req->file->f_pos = kiocb->ki_pos;
2142 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2143 io_complete_rw(kiocb, ret, 0);
2145 io_rw_done(kiocb, ret);
2148 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2149 struct iov_iter *iter)
2151 struct io_ring_ctx *ctx = req->ctx;
2152 size_t len = req->rw.len;
2153 struct io_mapped_ubuf *imu;
2154 unsigned index, buf_index;
2158 /* attempt to use fixed buffers without having provided iovecs */
2159 if (unlikely(!ctx->user_bufs))
2162 buf_index = (unsigned long) req->rw.kiocb.private;
2163 if (unlikely(buf_index >= ctx->nr_user_bufs))
2166 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2167 imu = &ctx->user_bufs[index];
2168 buf_addr = req->rw.addr;
2171 if (buf_addr + len < buf_addr)
2173 /* not inside the mapped region */
2174 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2178 * May not be a start of buffer, set size appropriately
2179 * and advance us to the beginning.
2181 offset = buf_addr - imu->ubuf;
2182 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2186 * Don't use iov_iter_advance() here, as it's really slow for
2187 * using the latter parts of a big fixed buffer - it iterates
2188 * over each segment manually. We can cheat a bit here, because
2191 * 1) it's a BVEC iter, we set it up
2192 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2193 * first and last bvec
2195 * So just find our index, and adjust the iterator afterwards.
2196 * If the offset is within the first bvec (or the whole first
2197 * bvec, just use iov_iter_advance(). This makes it easier
2198 * since we can just skip the first segment, which may not
2199 * be PAGE_SIZE aligned.
2201 const struct bio_vec *bvec = imu->bvec;
2203 if (offset <= bvec->bv_len) {
2204 iov_iter_advance(iter, offset);
2206 unsigned long seg_skip;
2208 /* skip first vec */
2209 offset -= bvec->bv_len;
2210 seg_skip = 1 + (offset >> PAGE_SHIFT);
2212 iter->bvec = bvec + seg_skip;
2213 iter->nr_segs -= seg_skip;
2214 iter->count -= bvec->bv_len + offset;
2215 iter->iov_offset = offset & ~PAGE_MASK;
2222 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2225 mutex_unlock(&ctx->uring_lock);
2228 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2231 * "Normal" inline submissions always hold the uring_lock, since we
2232 * grab it from the system call. Same is true for the SQPOLL offload.
2233 * The only exception is when we've detached the request and issue it
2234 * from an async worker thread, grab the lock for that case.
2237 mutex_lock(&ctx->uring_lock);
2240 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2241 int bgid, struct io_buffer *kbuf,
2244 struct io_buffer *head;
2246 if (req->flags & REQ_F_BUFFER_SELECTED)
2249 io_ring_submit_lock(req->ctx, needs_lock);
2251 lockdep_assert_held(&req->ctx->uring_lock);
2253 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2255 if (!list_empty(&head->list)) {
2256 kbuf = list_last_entry(&head->list, struct io_buffer,
2258 list_del(&kbuf->list);
2261 idr_remove(&req->ctx->io_buffer_idr, bgid);
2263 if (*len > kbuf->len)
2266 kbuf = ERR_PTR(-ENOBUFS);
2269 io_ring_submit_unlock(req->ctx, needs_lock);
2274 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2277 struct io_buffer *kbuf;
2280 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2281 bgid = (int) (unsigned long) req->rw.kiocb.private;
2282 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2285 req->rw.addr = (u64) (unsigned long) kbuf;
2286 req->flags |= REQ_F_BUFFER_SELECTED;
2287 return u64_to_user_ptr(kbuf->addr);
2290 #ifdef CONFIG_COMPAT
2291 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2294 struct compat_iovec __user *uiov;
2295 compat_ssize_t clen;
2299 uiov = u64_to_user_ptr(req->rw.addr);
2300 if (!access_ok(uiov, sizeof(*uiov)))
2302 if (__get_user(clen, &uiov->iov_len))
2308 buf = io_rw_buffer_select(req, &len, needs_lock);
2310 return PTR_ERR(buf);
2311 iov[0].iov_base = buf;
2312 iov[0].iov_len = (compat_size_t) len;
2317 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2320 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2324 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2327 len = iov[0].iov_len;
2330 buf = io_rw_buffer_select(req, &len, needs_lock);
2332 return PTR_ERR(buf);
2333 iov[0].iov_base = buf;
2334 iov[0].iov_len = len;
2338 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2341 if (req->flags & REQ_F_BUFFER_SELECTED)
2345 else if (req->rw.len > 1)
2348 #ifdef CONFIG_COMPAT
2349 if (req->ctx->compat)
2350 return io_compat_import(req, iov, needs_lock);
2353 return __io_iov_buffer_select(req, iov, needs_lock);
2356 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2357 struct iovec **iovec, struct iov_iter *iter,
2360 void __user *buf = u64_to_user_ptr(req->rw.addr);
2361 size_t sqe_len = req->rw.len;
2365 opcode = req->opcode;
2366 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2368 return io_import_fixed(req, rw, iter);
2371 /* buffer index only valid with fixed read/write, or buffer select */
2372 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2375 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2376 if (req->flags & REQ_F_BUFFER_SELECT) {
2377 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2380 return PTR_ERR(buf);
2382 req->rw.len = sqe_len;
2385 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2387 return ret < 0 ? ret : sqe_len;
2391 struct io_async_rw *iorw = &req->io->rw;
2394 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2395 if (iorw->iov == iorw->fast_iov)
2400 if (req->flags & REQ_F_BUFFER_SELECT) {
2401 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2403 ret = (*iovec)->iov_len;
2404 iov_iter_init(iter, rw, *iovec, 1, ret);
2410 #ifdef CONFIG_COMPAT
2411 if (req->ctx->compat)
2412 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2416 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2420 * For files that don't have ->read_iter() and ->write_iter(), handle them
2421 * by looping over ->read() or ->write() manually.
2423 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2424 struct iov_iter *iter)
2429 * Don't support polled IO through this interface, and we can't
2430 * support non-blocking either. For the latter, this just causes
2431 * the kiocb to be handled from an async context.
2433 if (kiocb->ki_flags & IOCB_HIPRI)
2435 if (kiocb->ki_flags & IOCB_NOWAIT)
2438 while (iov_iter_count(iter)) {
2442 if (!iov_iter_is_bvec(iter)) {
2443 iovec = iov_iter_iovec(iter);
2445 /* fixed buffers import bvec */
2446 iovec.iov_base = kmap(iter->bvec->bv_page)
2448 iovec.iov_len = min(iter->count,
2449 iter->bvec->bv_len - iter->iov_offset);
2453 nr = file->f_op->read(file, iovec.iov_base,
2454 iovec.iov_len, &kiocb->ki_pos);
2456 nr = file->f_op->write(file, iovec.iov_base,
2457 iovec.iov_len, &kiocb->ki_pos);
2460 if (iov_iter_is_bvec(iter))
2461 kunmap(iter->bvec->bv_page);
2469 if (nr != iovec.iov_len)
2471 iov_iter_advance(iter, nr);
2477 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2478 struct iovec *iovec, struct iovec *fast_iov,
2479 struct iov_iter *iter)
2481 req->io->rw.nr_segs = iter->nr_segs;
2482 req->io->rw.size = io_size;
2483 req->io->rw.iov = iovec;
2484 if (!req->io->rw.iov) {
2485 req->io->rw.iov = req->io->rw.fast_iov;
2486 if (req->io->rw.iov != fast_iov)
2487 memcpy(req->io->rw.iov, fast_iov,
2488 sizeof(struct iovec) * iter->nr_segs);
2490 req->flags |= REQ_F_NEED_CLEANUP;
2494 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2496 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2497 return req->io == NULL;
2500 static int io_alloc_async_ctx(struct io_kiocb *req)
2502 if (!io_op_defs[req->opcode].async_ctx)
2505 return __io_alloc_async_ctx(req);
2508 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2509 struct iovec *iovec, struct iovec *fast_iov,
2510 struct iov_iter *iter)
2512 if (!io_op_defs[req->opcode].async_ctx)
2515 if (__io_alloc_async_ctx(req))
2518 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2523 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2524 bool force_nonblock)
2526 struct io_async_ctx *io;
2527 struct iov_iter iter;
2530 ret = io_prep_rw(req, sqe, force_nonblock);
2534 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2537 /* either don't need iovec imported or already have it */
2538 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2542 io->rw.iov = io->rw.fast_iov;
2544 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2549 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2553 static int io_read(struct io_kiocb *req, bool force_nonblock)
2555 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2556 struct kiocb *kiocb = &req->rw.kiocb;
2557 struct iov_iter iter;
2559 ssize_t io_size, ret;
2561 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2565 /* Ensure we clear previously set non-block flag */
2566 if (!force_nonblock)
2567 kiocb->ki_flags &= ~IOCB_NOWAIT;
2571 if (req->flags & REQ_F_LINK_HEAD)
2572 req->result = io_size;
2575 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2576 * we know to async punt it even if it was opened O_NONBLOCK
2578 if (force_nonblock && !io_file_supports_async(req->file))
2581 iov_count = iov_iter_count(&iter);
2582 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2586 if (req->file->f_op->read_iter)
2587 ret2 = call_read_iter(req->file, kiocb, &iter);
2589 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2591 /* Catch -EAGAIN return for forced non-blocking submission */
2592 if (!force_nonblock || ret2 != -EAGAIN) {
2593 kiocb_done(kiocb, ret2);
2596 ret = io_setup_async_rw(req, io_size, iovec,
2597 inline_vecs, &iter);
2600 /* any defer here is final, must blocking retry */
2601 if (!(req->flags & REQ_F_NOWAIT))
2602 req->flags |= REQ_F_MUST_PUNT;
2608 req->flags &= ~REQ_F_NEED_CLEANUP;
2612 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2613 bool force_nonblock)
2615 struct io_async_ctx *io;
2616 struct iov_iter iter;
2619 ret = io_prep_rw(req, sqe, force_nonblock);
2623 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2626 req->fsize = rlimit(RLIMIT_FSIZE);
2628 /* either don't need iovec imported or already have it */
2629 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2633 io->rw.iov = io->rw.fast_iov;
2635 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2640 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2644 static int io_write(struct io_kiocb *req, bool force_nonblock)
2646 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2647 struct kiocb *kiocb = &req->rw.kiocb;
2648 struct iov_iter iter;
2650 ssize_t ret, io_size;
2652 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2656 /* Ensure we clear previously set non-block flag */
2657 if (!force_nonblock)
2658 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2662 if (req->flags & REQ_F_LINK_HEAD)
2663 req->result = io_size;
2666 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2667 * we know to async punt it even if it was opened O_NONBLOCK
2669 if (force_nonblock && !io_file_supports_async(req->file))
2672 /* file path doesn't support NOWAIT for non-direct_IO */
2673 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2674 (req->flags & REQ_F_ISREG))
2677 iov_count = iov_iter_count(&iter);
2678 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2683 * Open-code file_start_write here to grab freeze protection,
2684 * which will be released by another thread in
2685 * io_complete_rw(). Fool lockdep by telling it the lock got
2686 * released so that it doesn't complain about the held lock when
2687 * we return to userspace.
2689 if (req->flags & REQ_F_ISREG) {
2690 __sb_start_write(file_inode(req->file)->i_sb,
2691 SB_FREEZE_WRITE, true);
2692 __sb_writers_release(file_inode(req->file)->i_sb,
2695 kiocb->ki_flags |= IOCB_WRITE;
2697 if (!force_nonblock)
2698 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2700 if (req->file->f_op->write_iter)
2701 ret2 = call_write_iter(req->file, kiocb, &iter);
2703 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2705 if (!force_nonblock)
2706 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2709 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2710 * retry them without IOCB_NOWAIT.
2712 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2714 if (!force_nonblock || ret2 != -EAGAIN) {
2715 kiocb_done(kiocb, ret2);
2718 ret = io_setup_async_rw(req, io_size, iovec,
2719 inline_vecs, &iter);
2722 /* any defer here is final, must blocking retry */
2723 req->flags |= REQ_F_MUST_PUNT;
2728 req->flags &= ~REQ_F_NEED_CLEANUP;
2733 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2735 struct io_splice* sp = &req->splice;
2736 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2739 if (req->flags & REQ_F_NEED_CLEANUP)
2743 sp->off_in = READ_ONCE(sqe->splice_off_in);
2744 sp->off_out = READ_ONCE(sqe->off);
2745 sp->len = READ_ONCE(sqe->len);
2746 sp->flags = READ_ONCE(sqe->splice_flags);
2748 if (unlikely(sp->flags & ~valid_flags))
2751 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2752 (sp->flags & SPLICE_F_FD_IN_FIXED));
2755 req->flags |= REQ_F_NEED_CLEANUP;
2757 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2758 req->work.flags |= IO_WQ_WORK_UNBOUND;
2763 static bool io_splice_punt(struct file *file)
2765 if (get_pipe_info(file))
2767 if (!io_file_supports_async(file))
2769 return !(file->f_flags & O_NONBLOCK);
2772 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2774 struct io_splice *sp = &req->splice;
2775 struct file *in = sp->file_in;
2776 struct file *out = sp->file_out;
2777 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2778 loff_t *poff_in, *poff_out;
2781 if (force_nonblock) {
2782 if (io_splice_punt(in) || io_splice_punt(out))
2784 flags |= SPLICE_F_NONBLOCK;
2787 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2788 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2789 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2790 if (force_nonblock && ret == -EAGAIN)
2793 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2794 req->flags &= ~REQ_F_NEED_CLEANUP;
2796 io_cqring_add_event(req, ret);
2798 req_set_fail_links(req);
2804 * IORING_OP_NOP just posts a completion event, nothing else.
2806 static int io_nop(struct io_kiocb *req)
2808 struct io_ring_ctx *ctx = req->ctx;
2810 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2813 io_cqring_add_event(req, 0);
2818 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2820 struct io_ring_ctx *ctx = req->ctx;
2825 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2827 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2830 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2831 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2834 req->sync.off = READ_ONCE(sqe->off);
2835 req->sync.len = READ_ONCE(sqe->len);
2839 static bool io_req_cancelled(struct io_kiocb *req)
2841 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2842 req_set_fail_links(req);
2843 io_cqring_add_event(req, -ECANCELED);
2851 static void __io_fsync(struct io_kiocb *req)
2853 loff_t end = req->sync.off + req->sync.len;
2856 ret = vfs_fsync_range(req->file, req->sync.off,
2857 end > 0 ? end : LLONG_MAX,
2858 req->sync.flags & IORING_FSYNC_DATASYNC);
2860 req_set_fail_links(req);
2861 io_cqring_add_event(req, ret);
2865 static void io_fsync_finish(struct io_wq_work **workptr)
2867 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2869 if (io_req_cancelled(req))
2872 io_steal_work(req, workptr);
2875 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2877 /* fsync always requires a blocking context */
2878 if (force_nonblock) {
2879 req->work.func = io_fsync_finish;
2886 static void __io_fallocate(struct io_kiocb *req)
2890 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2891 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2893 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2895 req_set_fail_links(req);
2896 io_cqring_add_event(req, ret);
2900 static void io_fallocate_finish(struct io_wq_work **workptr)
2902 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2904 if (io_req_cancelled(req))
2906 __io_fallocate(req);
2907 io_steal_work(req, workptr);
2910 static int io_fallocate_prep(struct io_kiocb *req,
2911 const struct io_uring_sqe *sqe)
2913 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2916 req->sync.off = READ_ONCE(sqe->off);
2917 req->sync.len = READ_ONCE(sqe->addr);
2918 req->sync.mode = READ_ONCE(sqe->len);
2919 req->fsize = rlimit(RLIMIT_FSIZE);
2923 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2925 /* fallocate always requiring blocking context */
2926 if (force_nonblock) {
2927 req->work.func = io_fallocate_finish;
2931 __io_fallocate(req);
2935 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2937 const char __user *fname;
2940 if (sqe->ioprio || sqe->buf_index)
2942 if (req->flags & REQ_F_FIXED_FILE)
2944 if (req->flags & REQ_F_NEED_CLEANUP)
2947 req->open.dfd = READ_ONCE(sqe->fd);
2948 req->open.how.mode = READ_ONCE(sqe->len);
2949 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2950 req->open.how.flags = READ_ONCE(sqe->open_flags);
2951 if (force_o_largefile())
2952 req->open.how.flags |= O_LARGEFILE;
2954 req->open.filename = getname(fname);
2955 if (IS_ERR(req->open.filename)) {
2956 ret = PTR_ERR(req->open.filename);
2957 req->open.filename = NULL;
2961 req->open.nofile = rlimit(RLIMIT_NOFILE);
2962 req->flags |= REQ_F_NEED_CLEANUP;
2966 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2968 struct open_how __user *how;
2969 const char __user *fname;
2973 if (sqe->ioprio || sqe->buf_index)
2975 if (req->flags & REQ_F_FIXED_FILE)
2977 if (req->flags & REQ_F_NEED_CLEANUP)
2980 req->open.dfd = READ_ONCE(sqe->fd);
2981 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2982 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2983 len = READ_ONCE(sqe->len);
2985 if (len < OPEN_HOW_SIZE_VER0)
2988 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2993 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2994 req->open.how.flags |= O_LARGEFILE;
2996 req->open.filename = getname(fname);
2997 if (IS_ERR(req->open.filename)) {
2998 ret = PTR_ERR(req->open.filename);
2999 req->open.filename = NULL;
3003 req->open.nofile = rlimit(RLIMIT_NOFILE);
3004 req->flags |= REQ_F_NEED_CLEANUP;
3008 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3010 struct open_flags op;
3017 ret = build_open_flags(&req->open.how, &op);
3021 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3025 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3028 ret = PTR_ERR(file);
3030 fsnotify_open(file);
3031 fd_install(ret, file);
3034 putname(req->open.filename);
3035 req->flags &= ~REQ_F_NEED_CLEANUP;
3037 req_set_fail_links(req);
3038 io_cqring_add_event(req, ret);
3043 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3045 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3046 return io_openat2(req, force_nonblock);
3049 static int io_remove_buffers_prep(struct io_kiocb *req,
3050 const struct io_uring_sqe *sqe)
3052 struct io_provide_buf *p = &req->pbuf;
3055 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3058 tmp = READ_ONCE(sqe->fd);
3059 if (!tmp || tmp > USHRT_MAX)
3062 memset(p, 0, sizeof(*p));
3064 p->bgid = READ_ONCE(sqe->buf_group);
3068 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3069 int bgid, unsigned nbufs)
3073 /* shouldn't happen */
3077 /* the head kbuf is the list itself */
3078 while (!list_empty(&buf->list)) {
3079 struct io_buffer *nxt;
3081 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3082 list_del(&nxt->list);
3089 idr_remove(&ctx->io_buffer_idr, bgid);
3094 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3096 struct io_provide_buf *p = &req->pbuf;
3097 struct io_ring_ctx *ctx = req->ctx;
3098 struct io_buffer *head;
3101 io_ring_submit_lock(ctx, !force_nonblock);
3103 lockdep_assert_held(&ctx->uring_lock);
3106 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3108 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3110 io_ring_submit_lock(ctx, !force_nonblock);
3112 req_set_fail_links(req);
3113 io_cqring_add_event(req, ret);
3118 static int io_provide_buffers_prep(struct io_kiocb *req,
3119 const struct io_uring_sqe *sqe)
3121 struct io_provide_buf *p = &req->pbuf;
3124 if (sqe->ioprio || sqe->rw_flags)
3127 tmp = READ_ONCE(sqe->fd);
3128 if (!tmp || tmp > USHRT_MAX)
3131 p->addr = READ_ONCE(sqe->addr);
3132 p->len = READ_ONCE(sqe->len);
3134 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3137 p->bgid = READ_ONCE(sqe->buf_group);
3138 tmp = READ_ONCE(sqe->off);
3139 if (tmp > USHRT_MAX)
3145 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3147 struct io_buffer *buf;
3148 u64 addr = pbuf->addr;
3149 int i, bid = pbuf->bid;
3151 for (i = 0; i < pbuf->nbufs; i++) {
3152 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3157 buf->len = pbuf->len;
3162 INIT_LIST_HEAD(&buf->list);
3165 list_add_tail(&buf->list, &(*head)->list);
3169 return i ? i : -ENOMEM;
3172 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3174 struct io_provide_buf *p = &req->pbuf;
3175 struct io_ring_ctx *ctx = req->ctx;
3176 struct io_buffer *head, *list;
3179 io_ring_submit_lock(ctx, !force_nonblock);
3181 lockdep_assert_held(&ctx->uring_lock);
3183 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3185 ret = io_add_buffers(p, &head);
3190 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3193 __io_remove_buffers(ctx, head, p->bgid, -1U);
3198 io_ring_submit_unlock(ctx, !force_nonblock);
3200 req_set_fail_links(req);
3201 io_cqring_add_event(req, ret);
3206 static int io_epoll_ctl_prep(struct io_kiocb *req,
3207 const struct io_uring_sqe *sqe)
3209 #if defined(CONFIG_EPOLL)
3210 if (sqe->ioprio || sqe->buf_index)
3213 req->epoll.epfd = READ_ONCE(sqe->fd);
3214 req->epoll.op = READ_ONCE(sqe->len);
3215 req->epoll.fd = READ_ONCE(sqe->off);
3217 if (ep_op_has_event(req->epoll.op)) {
3218 struct epoll_event __user *ev;
3220 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3221 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3231 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3233 #if defined(CONFIG_EPOLL)
3234 struct io_epoll *ie = &req->epoll;
3237 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3238 if (force_nonblock && ret == -EAGAIN)
3242 req_set_fail_links(req);
3243 io_cqring_add_event(req, ret);
3251 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3253 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3254 if (sqe->ioprio || sqe->buf_index || sqe->off)
3257 req->madvise.addr = READ_ONCE(sqe->addr);
3258 req->madvise.len = READ_ONCE(sqe->len);
3259 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3266 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3268 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3269 struct io_madvise *ma = &req->madvise;
3275 ret = do_madvise(ma->addr, ma->len, ma->advice);
3277 req_set_fail_links(req);
3278 io_cqring_add_event(req, ret);
3286 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3288 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3291 req->fadvise.offset = READ_ONCE(sqe->off);
3292 req->fadvise.len = READ_ONCE(sqe->len);
3293 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3297 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3299 struct io_fadvise *fa = &req->fadvise;
3302 if (force_nonblock) {
3303 switch (fa->advice) {
3304 case POSIX_FADV_NORMAL:
3305 case POSIX_FADV_RANDOM:
3306 case POSIX_FADV_SEQUENTIAL:
3313 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3315 req_set_fail_links(req);
3316 io_cqring_add_event(req, ret);
3321 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3323 const char __user *fname;
3324 unsigned lookup_flags;
3327 if (sqe->ioprio || sqe->buf_index)
3329 if (req->flags & REQ_F_FIXED_FILE)
3331 if (req->flags & REQ_F_NEED_CLEANUP)
3334 req->open.dfd = READ_ONCE(sqe->fd);
3335 req->open.mask = READ_ONCE(sqe->len);
3336 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3337 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3338 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3340 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3343 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3344 if (IS_ERR(req->open.filename)) {
3345 ret = PTR_ERR(req->open.filename);
3346 req->open.filename = NULL;
3350 req->flags |= REQ_F_NEED_CLEANUP;
3354 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3356 struct io_open *ctx = &req->open;
3357 unsigned lookup_flags;
3362 if (force_nonblock) {
3363 /* only need file table for an actual valid fd */
3364 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3365 req->flags |= REQ_F_NO_FILE_TABLE;
3369 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3373 /* filename_lookup() drops it, keep a reference */
3374 ctx->filename->refcnt++;
3376 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3381 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3383 if (retry_estale(ret, lookup_flags)) {
3384 lookup_flags |= LOOKUP_REVAL;
3388 ret = cp_statx(&stat, ctx->buffer);
3390 putname(ctx->filename);
3391 req->flags &= ~REQ_F_NEED_CLEANUP;
3393 req_set_fail_links(req);
3394 io_cqring_add_event(req, ret);
3399 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3402 * If we queue this for async, it must not be cancellable. That would
3403 * leave the 'file' in an undeterminate state.
3405 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3407 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3408 sqe->rw_flags || sqe->buf_index)
3410 if (req->flags & REQ_F_FIXED_FILE)
3413 req->close.fd = READ_ONCE(sqe->fd);
3414 if (req->file->f_op == &io_uring_fops ||
3415 req->close.fd == req->ctx->ring_fd)
3421 /* only called when __close_fd_get_file() is done */
3422 static void __io_close_finish(struct io_kiocb *req)
3426 ret = filp_close(req->close.put_file, req->work.files);
3428 req_set_fail_links(req);
3429 io_cqring_add_event(req, ret);
3430 fput(req->close.put_file);
3434 static void io_close_finish(struct io_wq_work **workptr)
3436 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3438 /* not cancellable, don't do io_req_cancelled() */
3439 __io_close_finish(req);
3440 io_steal_work(req, workptr);
3443 static int io_close(struct io_kiocb *req, bool force_nonblock)
3447 req->close.put_file = NULL;
3448 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3452 /* if the file has a flush method, be safe and punt to async */
3453 if (req->close.put_file->f_op->flush && force_nonblock) {
3454 /* submission ref will be dropped, take it for async */
3455 refcount_inc(&req->refs);
3457 req->work.func = io_close_finish;
3459 * Do manual async queue here to avoid grabbing files - we don't
3460 * need the files, and it'll cause io_close_finish() to close
3461 * the file again and cause a double CQE entry for this request
3463 io_queue_async_work(req);
3468 * No ->flush(), safely close from here and just punt the
3469 * fput() to async context.
3471 __io_close_finish(req);
3475 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3477 struct io_ring_ctx *ctx = req->ctx;
3482 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3484 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3487 req->sync.off = READ_ONCE(sqe->off);
3488 req->sync.len = READ_ONCE(sqe->len);
3489 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3493 static void __io_sync_file_range(struct io_kiocb *req)
3497 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3500 req_set_fail_links(req);
3501 io_cqring_add_event(req, ret);
3506 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3508 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3510 if (io_req_cancelled(req))
3512 __io_sync_file_range(req);
3513 io_put_req(req); /* put submission ref */
3516 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3518 /* sync_file_range always requires a blocking context */
3519 if (force_nonblock) {
3520 req->work.func = io_sync_file_range_finish;
3524 __io_sync_file_range(req);
3528 #if defined(CONFIG_NET)
3529 static int io_setup_async_msg(struct io_kiocb *req,
3530 struct io_async_msghdr *kmsg)
3534 if (io_alloc_async_ctx(req)) {
3535 if (kmsg->iov != kmsg->fast_iov)
3539 req->flags |= REQ_F_NEED_CLEANUP;
3540 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3544 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3546 struct io_sr_msg *sr = &req->sr_msg;
3547 struct io_async_ctx *io = req->io;
3550 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3551 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3552 sr->len = READ_ONCE(sqe->len);
3554 #ifdef CONFIG_COMPAT
3555 if (req->ctx->compat)
3556 sr->msg_flags |= MSG_CMSG_COMPAT;
3559 if (!io || req->opcode == IORING_OP_SEND)
3561 /* iovec is already imported */
3562 if (req->flags & REQ_F_NEED_CLEANUP)
3565 io->msg.iov = io->msg.fast_iov;
3566 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3569 req->flags |= REQ_F_NEED_CLEANUP;
3573 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3575 struct io_async_msghdr *kmsg = NULL;
3576 struct socket *sock;
3579 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3582 sock = sock_from_file(req->file, &ret);
3584 struct io_async_ctx io;
3588 kmsg = &req->io->msg;
3589 kmsg->msg.msg_name = &req->io->msg.addr;
3590 /* if iov is set, it's allocated already */
3592 kmsg->iov = kmsg->fast_iov;
3593 kmsg->msg.msg_iter.iov = kmsg->iov;
3595 struct io_sr_msg *sr = &req->sr_msg;
3598 kmsg->msg.msg_name = &io.msg.addr;
3600 io.msg.iov = io.msg.fast_iov;
3601 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3602 sr->msg_flags, &io.msg.iov);
3607 flags = req->sr_msg.msg_flags;
3608 if (flags & MSG_DONTWAIT)
3609 req->flags |= REQ_F_NOWAIT;
3610 else if (force_nonblock)
3611 flags |= MSG_DONTWAIT;
3613 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3614 if (force_nonblock && ret == -EAGAIN)
3615 return io_setup_async_msg(req, kmsg);
3616 if (ret == -ERESTARTSYS)
3620 if (kmsg && kmsg->iov != kmsg->fast_iov)
3622 req->flags &= ~REQ_F_NEED_CLEANUP;
3623 io_cqring_add_event(req, ret);
3625 req_set_fail_links(req);
3630 static int io_send(struct io_kiocb *req, bool force_nonblock)
3632 struct socket *sock;
3635 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3638 sock = sock_from_file(req->file, &ret);
3640 struct io_sr_msg *sr = &req->sr_msg;
3645 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3650 msg.msg_name = NULL;
3651 msg.msg_control = NULL;
3652 msg.msg_controllen = 0;
3653 msg.msg_namelen = 0;
3655 flags = req->sr_msg.msg_flags;
3656 if (flags & MSG_DONTWAIT)
3657 req->flags |= REQ_F_NOWAIT;
3658 else if (force_nonblock)
3659 flags |= MSG_DONTWAIT;
3661 msg.msg_flags = flags;
3662 ret = sock_sendmsg(sock, &msg);
3663 if (force_nonblock && ret == -EAGAIN)
3665 if (ret == -ERESTARTSYS)
3669 io_cqring_add_event(req, ret);
3671 req_set_fail_links(req);
3676 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3678 struct io_sr_msg *sr = &req->sr_msg;
3679 struct iovec __user *uiov;
3683 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3688 if (req->flags & REQ_F_BUFFER_SELECT) {
3691 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3693 sr->len = io->msg.iov[0].iov_len;
3694 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3698 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3699 &io->msg.iov, &io->msg.msg.msg_iter);
3707 #ifdef CONFIG_COMPAT
3708 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3709 struct io_async_ctx *io)
3711 struct compat_msghdr __user *msg_compat;
3712 struct io_sr_msg *sr = &req->sr_msg;
3713 struct compat_iovec __user *uiov;
3718 msg_compat = (struct compat_msghdr __user *) sr->msg;
3719 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3724 uiov = compat_ptr(ptr);
3725 if (req->flags & REQ_F_BUFFER_SELECT) {
3726 compat_ssize_t clen;
3730 if (!access_ok(uiov, sizeof(*uiov)))
3732 if (__get_user(clen, &uiov->iov_len))
3736 sr->len = io->msg.iov[0].iov_len;
3739 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3741 &io->msg.msg.msg_iter);
3750 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3752 io->msg.iov = io->msg.fast_iov;
3754 #ifdef CONFIG_COMPAT
3755 if (req->ctx->compat)
3756 return __io_compat_recvmsg_copy_hdr(req, io);
3759 return __io_recvmsg_copy_hdr(req, io);
3762 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3763 int *cflags, bool needs_lock)
3765 struct io_sr_msg *sr = &req->sr_msg;
3766 struct io_buffer *kbuf;
3768 if (!(req->flags & REQ_F_BUFFER_SELECT))
3771 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3776 req->flags |= REQ_F_BUFFER_SELECTED;
3778 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3779 *cflags |= IORING_CQE_F_BUFFER;
3783 static int io_recvmsg_prep(struct io_kiocb *req,
3784 const struct io_uring_sqe *sqe)
3786 struct io_sr_msg *sr = &req->sr_msg;
3787 struct io_async_ctx *io = req->io;
3790 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3791 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3792 sr->len = READ_ONCE(sqe->len);
3793 sr->bgid = READ_ONCE(sqe->buf_group);
3795 #ifdef CONFIG_COMPAT
3796 if (req->ctx->compat)
3797 sr->msg_flags |= MSG_CMSG_COMPAT;
3800 if (!io || req->opcode == IORING_OP_RECV)
3802 /* iovec is already imported */
3803 if (req->flags & REQ_F_NEED_CLEANUP)
3806 ret = io_recvmsg_copy_hdr(req, io);
3808 req->flags |= REQ_F_NEED_CLEANUP;
3812 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3814 struct io_async_msghdr *kmsg = NULL;
3815 struct socket *sock;
3816 int ret, cflags = 0;
3818 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3821 sock = sock_from_file(req->file, &ret);
3823 struct io_buffer *kbuf;
3824 struct io_async_ctx io;
3828 kmsg = &req->io->msg;
3829 kmsg->msg.msg_name = &req->io->msg.addr;
3830 /* if iov is set, it's allocated already */
3832 kmsg->iov = kmsg->fast_iov;
3833 kmsg->msg.msg_iter.iov = kmsg->iov;
3836 kmsg->msg.msg_name = &io.msg.addr;
3838 ret = io_recvmsg_copy_hdr(req, &io);
3843 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3845 return PTR_ERR(kbuf);
3847 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3848 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3849 1, req->sr_msg.len);
3852 flags = req->sr_msg.msg_flags;
3853 if (flags & MSG_DONTWAIT)
3854 req->flags |= REQ_F_NOWAIT;
3855 else if (force_nonblock)
3856 flags |= MSG_DONTWAIT;
3858 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3859 kmsg->uaddr, flags);
3860 if (force_nonblock && ret == -EAGAIN)
3861 return io_setup_async_msg(req, kmsg);
3862 if (ret == -ERESTARTSYS)
3866 if (kmsg && kmsg->iov != kmsg->fast_iov)
3868 req->flags &= ~REQ_F_NEED_CLEANUP;
3869 __io_cqring_add_event(req, ret, cflags);
3871 req_set_fail_links(req);
3876 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3878 struct io_buffer *kbuf = NULL;
3879 struct socket *sock;
3880 int ret, cflags = 0;
3882 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3885 sock = sock_from_file(req->file, &ret);
3887 struct io_sr_msg *sr = &req->sr_msg;
3888 void __user *buf = sr->buf;
3893 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3895 return PTR_ERR(kbuf);
3897 buf = u64_to_user_ptr(kbuf->addr);
3899 ret = import_single_range(READ, buf, sr->len, &iov,
3906 req->flags |= REQ_F_NEED_CLEANUP;
3907 msg.msg_name = NULL;
3908 msg.msg_control = NULL;
3909 msg.msg_controllen = 0;
3910 msg.msg_namelen = 0;
3911 msg.msg_iocb = NULL;
3914 flags = req->sr_msg.msg_flags;
3915 if (flags & MSG_DONTWAIT)
3916 req->flags |= REQ_F_NOWAIT;
3917 else if (force_nonblock)
3918 flags |= MSG_DONTWAIT;
3920 ret = sock_recvmsg(sock, &msg, flags);
3921 if (force_nonblock && ret == -EAGAIN)
3923 if (ret == -ERESTARTSYS)
3928 req->flags &= ~REQ_F_NEED_CLEANUP;
3929 __io_cqring_add_event(req, ret, cflags);
3931 req_set_fail_links(req);
3936 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3938 struct io_accept *accept = &req->accept;
3940 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3942 if (sqe->ioprio || sqe->len || sqe->buf_index)
3945 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3946 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3947 accept->flags = READ_ONCE(sqe->accept_flags);
3948 accept->nofile = rlimit(RLIMIT_NOFILE);
3952 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3954 struct io_accept *accept = &req->accept;
3955 unsigned file_flags;
3958 file_flags = force_nonblock ? O_NONBLOCK : 0;
3959 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3960 accept->addr_len, accept->flags,
3962 if (ret == -EAGAIN && force_nonblock)
3964 if (ret == -ERESTARTSYS)
3967 req_set_fail_links(req);
3968 io_cqring_add_event(req, ret);
3973 static void io_accept_finish(struct io_wq_work **workptr)
3975 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3977 if (io_req_cancelled(req))
3979 __io_accept(req, false);
3980 io_steal_work(req, workptr);
3983 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3987 ret = __io_accept(req, force_nonblock);
3988 if (ret == -EAGAIN && force_nonblock) {
3989 req->work.func = io_accept_finish;
3995 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3997 struct io_connect *conn = &req->connect;
3998 struct io_async_ctx *io = req->io;
4000 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4002 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4005 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4006 conn->addr_len = READ_ONCE(sqe->addr2);
4011 return move_addr_to_kernel(conn->addr, conn->addr_len,
4012 &io->connect.address);
4015 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4017 struct io_async_ctx __io, *io;
4018 unsigned file_flags;
4024 ret = move_addr_to_kernel(req->connect.addr,
4025 req->connect.addr_len,
4026 &__io.connect.address);
4032 file_flags = force_nonblock ? O_NONBLOCK : 0;
4034 ret = __sys_connect_file(req->file, &io->connect.address,
4035 req->connect.addr_len, file_flags);
4036 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4039 if (io_alloc_async_ctx(req)) {
4043 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4046 if (ret == -ERESTARTSYS)
4050 req_set_fail_links(req);
4051 io_cqring_add_event(req, ret);
4055 #else /* !CONFIG_NET */
4056 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4061 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4066 static int io_send(struct io_kiocb *req, bool force_nonblock)
4071 static int io_recvmsg_prep(struct io_kiocb *req,
4072 const struct io_uring_sqe *sqe)
4077 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4082 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4087 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4092 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4097 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4102 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4106 #endif /* CONFIG_NET */
4108 struct io_poll_table {
4109 struct poll_table_struct pt;
4110 struct io_kiocb *req;
4114 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4115 struct wait_queue_head *head)
4117 if (unlikely(poll->head)) {
4118 pt->error = -EINVAL;
4124 add_wait_queue(head, &poll->wait);
4127 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4128 struct poll_table_struct *p)
4130 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4132 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4135 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4136 __poll_t mask, task_work_func_t func)
4138 struct task_struct *tsk;
4141 /* for instances that support it check for an event match first: */
4142 if (mask && !(mask & poll->events))
4145 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4147 list_del_init(&poll->wait.entry);
4151 init_task_work(&req->task_work, func);
4153 * If this fails, then the task is exiting. Punt to one of the io-wq
4154 * threads to ensure the work gets run, we can't always rely on exit
4155 * cancelation taking care of this.
4157 ret = task_work_add(tsk, &req->task_work, true);
4158 if (unlikely(ret)) {
4159 tsk = io_wq_get_task(req->ctx->io_wq);
4160 task_work_add(tsk, &req->task_work, true);
4162 wake_up_process(tsk);
4166 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4167 __acquires(&req->ctx->completion_lock)
4169 struct io_ring_ctx *ctx = req->ctx;
4171 if (!req->result && !READ_ONCE(poll->canceled)) {
4172 struct poll_table_struct pt = { ._key = poll->events };
4174 req->result = vfs_poll(req->file, &pt) & poll->events;
4177 spin_lock_irq(&ctx->completion_lock);
4178 if (!req->result && !READ_ONCE(poll->canceled)) {
4179 add_wait_queue(poll->head, &poll->wait);
4186 static void io_async_task_func(struct callback_head *cb)
4188 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4189 struct async_poll *apoll = req->apoll;
4190 struct io_ring_ctx *ctx = req->ctx;
4193 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4195 if (io_poll_rewait(req, &apoll->poll)) {
4196 spin_unlock_irq(&ctx->completion_lock);
4200 if (hash_hashed(&req->hash_node))
4201 hash_del(&req->hash_node);
4203 canceled = READ_ONCE(apoll->poll.canceled);
4205 io_cqring_fill_event(req, -ECANCELED);
4206 io_commit_cqring(ctx);
4209 spin_unlock_irq(&ctx->completion_lock);
4211 /* restore ->work in case we need to retry again */
4212 memcpy(&req->work, &apoll->work, sizeof(req->work));
4216 io_cqring_ev_posted(ctx);
4217 req_set_fail_links(req);
4218 io_double_put_req(req);
4222 __set_current_state(TASK_RUNNING);
4223 mutex_lock(&ctx->uring_lock);
4224 __io_queue_sqe(req, NULL);
4225 mutex_unlock(&ctx->uring_lock);
4230 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4233 struct io_kiocb *req = wait->private;
4234 struct io_poll_iocb *poll = &req->apoll->poll;
4236 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4239 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4242 static void io_poll_req_insert(struct io_kiocb *req)
4244 struct io_ring_ctx *ctx = req->ctx;
4245 struct hlist_head *list;
4247 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4248 hlist_add_head(&req->hash_node, list);
4251 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4252 struct io_poll_iocb *poll,
4253 struct io_poll_table *ipt, __poll_t mask,
4254 wait_queue_func_t wake_func)
4255 __acquires(&ctx->completion_lock)
4257 struct io_ring_ctx *ctx = req->ctx;
4258 bool cancel = false;
4260 poll->file = req->file;
4262 poll->done = poll->canceled = false;
4263 poll->events = mask;
4265 ipt->pt._key = mask;
4267 ipt->error = -EINVAL;
4269 INIT_LIST_HEAD(&poll->wait.entry);
4270 init_waitqueue_func_entry(&poll->wait, wake_func);
4271 poll->wait.private = req;
4273 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4275 spin_lock_irq(&ctx->completion_lock);
4276 if (likely(poll->head)) {
4277 spin_lock(&poll->head->lock);
4278 if (unlikely(list_empty(&poll->wait.entry))) {
4284 if (mask || ipt->error)
4285 list_del_init(&poll->wait.entry);
4287 WRITE_ONCE(poll->canceled, true);
4288 else if (!poll->done) /* actually waiting for an event */
4289 io_poll_req_insert(req);
4290 spin_unlock(&poll->head->lock);
4296 static bool io_arm_poll_handler(struct io_kiocb *req)
4298 const struct io_op_def *def = &io_op_defs[req->opcode];
4299 struct io_ring_ctx *ctx = req->ctx;
4300 struct async_poll *apoll;
4301 struct io_poll_table ipt;
4304 if (!req->file || !file_can_poll(req->file))
4306 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4308 if (!def->pollin && !def->pollout)
4311 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4312 if (unlikely(!apoll))
4315 req->flags |= REQ_F_POLLED;
4316 memcpy(&apoll->work, &req->work, sizeof(req->work));
4318 get_task_struct(current);
4319 req->task = current;
4321 INIT_HLIST_NODE(&req->hash_node);
4325 mask |= POLLIN | POLLRDNORM;
4327 mask |= POLLOUT | POLLWRNORM;
4328 mask |= POLLERR | POLLPRI;
4330 ipt.pt._qproc = io_async_queue_proc;
4332 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4336 apoll->poll.done = true;
4337 spin_unlock_irq(&ctx->completion_lock);
4338 memcpy(&req->work, &apoll->work, sizeof(req->work));
4342 spin_unlock_irq(&ctx->completion_lock);
4343 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4344 apoll->poll.events);
4348 static bool __io_poll_remove_one(struct io_kiocb *req,
4349 struct io_poll_iocb *poll)
4351 bool do_complete = false;
4353 spin_lock(&poll->head->lock);
4354 WRITE_ONCE(poll->canceled, true);
4355 if (!list_empty(&poll->wait.entry)) {
4356 list_del_init(&poll->wait.entry);
4359 spin_unlock(&poll->head->lock);
4363 static bool io_poll_remove_one(struct io_kiocb *req)
4365 struct async_poll *apoll = NULL;
4368 if (req->opcode == IORING_OP_POLL_ADD) {
4369 do_complete = __io_poll_remove_one(req, &req->poll);
4372 /* non-poll requests have submit ref still */
4373 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4378 hash_del(&req->hash_node);
4380 if (do_complete && apoll) {
4382 * restore ->work because we need to call io_req_work_drop_env.
4384 memcpy(&req->work, &apoll->work, sizeof(req->work));
4389 io_cqring_fill_event(req, -ECANCELED);
4390 io_commit_cqring(req->ctx);
4391 req->flags |= REQ_F_COMP_LOCKED;
4398 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4400 struct hlist_node *tmp;
4401 struct io_kiocb *req;
4404 spin_lock_irq(&ctx->completion_lock);
4405 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4406 struct hlist_head *list;
4408 list = &ctx->cancel_hash[i];
4409 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4410 posted += io_poll_remove_one(req);
4412 spin_unlock_irq(&ctx->completion_lock);
4415 io_cqring_ev_posted(ctx);
4418 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4420 struct hlist_head *list;
4421 struct io_kiocb *req;
4423 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4424 hlist_for_each_entry(req, list, hash_node) {
4425 if (sqe_addr != req->user_data)
4427 if (io_poll_remove_one(req))
4435 static int io_poll_remove_prep(struct io_kiocb *req,
4436 const struct io_uring_sqe *sqe)
4438 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4440 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4444 req->poll.addr = READ_ONCE(sqe->addr);
4449 * Find a running poll command that matches one specified in sqe->addr,
4450 * and remove it if found.
4452 static int io_poll_remove(struct io_kiocb *req)
4454 struct io_ring_ctx *ctx = req->ctx;
4458 addr = req->poll.addr;
4459 spin_lock_irq(&ctx->completion_lock);
4460 ret = io_poll_cancel(ctx, addr);
4461 spin_unlock_irq(&ctx->completion_lock);
4463 io_cqring_add_event(req, ret);
4465 req_set_fail_links(req);
4470 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4472 struct io_ring_ctx *ctx = req->ctx;
4474 req->poll.done = true;
4475 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4476 io_commit_cqring(ctx);
4479 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4481 struct io_ring_ctx *ctx = req->ctx;
4482 struct io_poll_iocb *poll = &req->poll;
4484 if (io_poll_rewait(req, poll)) {
4485 spin_unlock_irq(&ctx->completion_lock);
4489 hash_del(&req->hash_node);
4490 io_poll_complete(req, req->result, 0);
4491 req->flags |= REQ_F_COMP_LOCKED;
4492 io_put_req_find_next(req, nxt);
4493 spin_unlock_irq(&ctx->completion_lock);
4495 io_cqring_ev_posted(ctx);
4498 static void io_poll_task_func(struct callback_head *cb)
4500 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4501 struct io_kiocb *nxt = NULL;
4503 io_poll_task_handler(req, &nxt);
4505 struct io_ring_ctx *ctx = nxt->ctx;
4507 mutex_lock(&ctx->uring_lock);
4508 __io_queue_sqe(nxt, NULL);
4509 mutex_unlock(&ctx->uring_lock);
4513 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4516 struct io_kiocb *req = wait->private;
4517 struct io_poll_iocb *poll = &req->poll;
4519 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4522 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4523 struct poll_table_struct *p)
4525 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4527 __io_queue_proc(&pt->req->poll, pt, head);
4530 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4532 struct io_poll_iocb *poll = &req->poll;
4535 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4537 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4542 events = READ_ONCE(sqe->poll_events);
4543 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4545 get_task_struct(current);
4546 req->task = current;
4550 static int io_poll_add(struct io_kiocb *req)
4552 struct io_poll_iocb *poll = &req->poll;
4553 struct io_ring_ctx *ctx = req->ctx;
4554 struct io_poll_table ipt;
4557 INIT_HLIST_NODE(&req->hash_node);
4558 INIT_LIST_HEAD(&req->list);
4559 ipt.pt._qproc = io_poll_queue_proc;
4561 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4564 if (mask) { /* no async, we'd stolen it */
4566 io_poll_complete(req, mask, 0);
4568 spin_unlock_irq(&ctx->completion_lock);
4571 io_cqring_ev_posted(ctx);
4577 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4579 struct io_timeout_data *data = container_of(timer,
4580 struct io_timeout_data, timer);
4581 struct io_kiocb *req = data->req;
4582 struct io_ring_ctx *ctx = req->ctx;
4583 unsigned long flags;
4585 atomic_inc(&ctx->cq_timeouts);
4587 spin_lock_irqsave(&ctx->completion_lock, flags);
4589 * We could be racing with timeout deletion. If the list is empty,
4590 * then timeout lookup already found it and will be handling it.
4592 if (!list_empty(&req->list)) {
4593 struct io_kiocb *prev;
4596 * Adjust the reqs sequence before the current one because it
4597 * will consume a slot in the cq_ring and the cq_tail
4598 * pointer will be increased, otherwise other timeout reqs may
4599 * return in advance without waiting for enough wait_nr.
4602 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4604 list_del_init(&req->list);
4607 io_cqring_fill_event(req, -ETIME);
4608 io_commit_cqring(ctx);
4609 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4611 io_cqring_ev_posted(ctx);
4612 req_set_fail_links(req);
4614 return HRTIMER_NORESTART;
4617 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4619 struct io_kiocb *req;
4622 list_for_each_entry(req, &ctx->timeout_list, list) {
4623 if (user_data == req->user_data) {
4624 list_del_init(&req->list);
4633 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4637 req_set_fail_links(req);
4638 io_cqring_fill_event(req, -ECANCELED);
4643 static int io_timeout_remove_prep(struct io_kiocb *req,
4644 const struct io_uring_sqe *sqe)
4646 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4648 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4651 req->timeout.addr = READ_ONCE(sqe->addr);
4652 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4653 if (req->timeout.flags)
4660 * Remove or update an existing timeout command
4662 static int io_timeout_remove(struct io_kiocb *req)
4664 struct io_ring_ctx *ctx = req->ctx;
4667 spin_lock_irq(&ctx->completion_lock);
4668 ret = io_timeout_cancel(ctx, req->timeout.addr);
4670 io_cqring_fill_event(req, ret);
4671 io_commit_cqring(ctx);
4672 spin_unlock_irq(&ctx->completion_lock);
4673 io_cqring_ev_posted(ctx);
4675 req_set_fail_links(req);
4680 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4681 bool is_timeout_link)
4683 struct io_timeout_data *data;
4686 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4688 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4690 if (sqe->off && is_timeout_link)
4692 flags = READ_ONCE(sqe->timeout_flags);
4693 if (flags & ~IORING_TIMEOUT_ABS)
4696 req->timeout.count = READ_ONCE(sqe->off);
4698 if (!req->io && io_alloc_async_ctx(req))
4701 data = &req->io->timeout;
4703 req->flags |= REQ_F_TIMEOUT;
4705 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4708 if (flags & IORING_TIMEOUT_ABS)
4709 data->mode = HRTIMER_MODE_ABS;
4711 data->mode = HRTIMER_MODE_REL;
4713 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4717 static int io_timeout(struct io_kiocb *req)
4719 struct io_ring_ctx *ctx = req->ctx;
4720 struct io_timeout_data *data;
4721 struct list_head *entry;
4723 u32 count = req->timeout.count;
4724 u32 seq = req->sequence;
4726 data = &req->io->timeout;
4729 * sqe->off holds how many events that need to occur for this
4730 * timeout event to be satisfied. If it isn't set, then this is
4731 * a pure timeout request, sequence isn't used.
4734 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4735 spin_lock_irq(&ctx->completion_lock);
4736 entry = ctx->timeout_list.prev;
4740 req->sequence = seq + count;
4743 * Insertion sort, ensuring the first entry in the list is always
4744 * the one we need first.
4746 spin_lock_irq(&ctx->completion_lock);
4747 list_for_each_prev(entry, &ctx->timeout_list) {
4748 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4750 long long tmp, tmp_nxt;
4751 u32 nxt_offset = nxt->timeout.count;
4753 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4757 * Since seq + count can overflow, use type long
4760 tmp = (long long)seq + count;
4761 nxt_seq = nxt->sequence - nxt_offset;
4762 tmp_nxt = (long long)nxt_seq + nxt_offset;
4765 * cached_sq_head may overflow, and it will never overflow twice
4766 * once there is some timeout req still be valid.
4775 * Sequence of reqs after the insert one and itself should
4776 * be adjusted because each timeout req consumes a slot.
4781 req->sequence -= span;
4783 list_add(&req->list, entry);
4784 data->timer.function = io_timeout_fn;
4785 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4786 spin_unlock_irq(&ctx->completion_lock);
4790 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4792 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4794 return req->user_data == (unsigned long) data;
4797 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4799 enum io_wq_cancel cancel_ret;
4802 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4803 switch (cancel_ret) {
4804 case IO_WQ_CANCEL_OK:
4807 case IO_WQ_CANCEL_RUNNING:
4810 case IO_WQ_CANCEL_NOTFOUND:
4818 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4819 struct io_kiocb *req, __u64 sqe_addr,
4822 unsigned long flags;
4825 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4826 if (ret != -ENOENT) {
4827 spin_lock_irqsave(&ctx->completion_lock, flags);
4831 spin_lock_irqsave(&ctx->completion_lock, flags);
4832 ret = io_timeout_cancel(ctx, sqe_addr);
4835 ret = io_poll_cancel(ctx, sqe_addr);
4839 io_cqring_fill_event(req, ret);
4840 io_commit_cqring(ctx);
4841 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4842 io_cqring_ev_posted(ctx);
4845 req_set_fail_links(req);
4849 static int io_async_cancel_prep(struct io_kiocb *req,
4850 const struct io_uring_sqe *sqe)
4852 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4854 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4858 req->cancel.addr = READ_ONCE(sqe->addr);
4862 static int io_async_cancel(struct io_kiocb *req)
4864 struct io_ring_ctx *ctx = req->ctx;
4866 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4870 static int io_files_update_prep(struct io_kiocb *req,
4871 const struct io_uring_sqe *sqe)
4873 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4876 req->files_update.offset = READ_ONCE(sqe->off);
4877 req->files_update.nr_args = READ_ONCE(sqe->len);
4878 if (!req->files_update.nr_args)
4880 req->files_update.arg = READ_ONCE(sqe->addr);
4884 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4886 struct io_ring_ctx *ctx = req->ctx;
4887 struct io_uring_files_update up;
4893 up.offset = req->files_update.offset;
4894 up.fds = req->files_update.arg;
4896 mutex_lock(&ctx->uring_lock);
4897 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4898 mutex_unlock(&ctx->uring_lock);
4901 req_set_fail_links(req);
4902 io_cqring_add_event(req, ret);
4907 static int io_req_defer_prep(struct io_kiocb *req,
4908 const struct io_uring_sqe *sqe)
4915 if (io_op_defs[req->opcode].file_table) {
4916 ret = io_grab_files(req);
4921 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4923 switch (req->opcode) {
4926 case IORING_OP_READV:
4927 case IORING_OP_READ_FIXED:
4928 case IORING_OP_READ:
4929 ret = io_read_prep(req, sqe, true);
4931 case IORING_OP_WRITEV:
4932 case IORING_OP_WRITE_FIXED:
4933 case IORING_OP_WRITE:
4934 ret = io_write_prep(req, sqe, true);
4936 case IORING_OP_POLL_ADD:
4937 ret = io_poll_add_prep(req, sqe);
4939 case IORING_OP_POLL_REMOVE:
4940 ret = io_poll_remove_prep(req, sqe);
4942 case IORING_OP_FSYNC:
4943 ret = io_prep_fsync(req, sqe);
4945 case IORING_OP_SYNC_FILE_RANGE:
4946 ret = io_prep_sfr(req, sqe);
4948 case IORING_OP_SENDMSG:
4949 case IORING_OP_SEND:
4950 ret = io_sendmsg_prep(req, sqe);
4952 case IORING_OP_RECVMSG:
4953 case IORING_OP_RECV:
4954 ret = io_recvmsg_prep(req, sqe);
4956 case IORING_OP_CONNECT:
4957 ret = io_connect_prep(req, sqe);
4959 case IORING_OP_TIMEOUT:
4960 ret = io_timeout_prep(req, sqe, false);
4962 case IORING_OP_TIMEOUT_REMOVE:
4963 ret = io_timeout_remove_prep(req, sqe);
4965 case IORING_OP_ASYNC_CANCEL:
4966 ret = io_async_cancel_prep(req, sqe);
4968 case IORING_OP_LINK_TIMEOUT:
4969 ret = io_timeout_prep(req, sqe, true);
4971 case IORING_OP_ACCEPT:
4972 ret = io_accept_prep(req, sqe);
4974 case IORING_OP_FALLOCATE:
4975 ret = io_fallocate_prep(req, sqe);
4977 case IORING_OP_OPENAT:
4978 ret = io_openat_prep(req, sqe);
4980 case IORING_OP_CLOSE:
4981 ret = io_close_prep(req, sqe);
4983 case IORING_OP_FILES_UPDATE:
4984 ret = io_files_update_prep(req, sqe);
4986 case IORING_OP_STATX:
4987 ret = io_statx_prep(req, sqe);
4989 case IORING_OP_FADVISE:
4990 ret = io_fadvise_prep(req, sqe);
4992 case IORING_OP_MADVISE:
4993 ret = io_madvise_prep(req, sqe);
4995 case IORING_OP_OPENAT2:
4996 ret = io_openat2_prep(req, sqe);
4998 case IORING_OP_EPOLL_CTL:
4999 ret = io_epoll_ctl_prep(req, sqe);
5001 case IORING_OP_SPLICE:
5002 ret = io_splice_prep(req, sqe);
5004 case IORING_OP_PROVIDE_BUFFERS:
5005 ret = io_provide_buffers_prep(req, sqe);
5007 case IORING_OP_REMOVE_BUFFERS:
5008 ret = io_remove_buffers_prep(req, sqe);
5011 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5020 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5022 struct io_ring_ctx *ctx = req->ctx;
5025 /* Still need defer if there is pending req in defer list. */
5026 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
5029 if (!req->io && io_alloc_async_ctx(req))
5032 ret = io_req_defer_prep(req, sqe);
5036 spin_lock_irq(&ctx->completion_lock);
5037 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5038 spin_unlock_irq(&ctx->completion_lock);
5042 trace_io_uring_defer(ctx, req, req->user_data);
5043 list_add_tail(&req->list, &ctx->defer_list);
5044 spin_unlock_irq(&ctx->completion_lock);
5045 return -EIOCBQUEUED;
5048 static void io_cleanup_req(struct io_kiocb *req)
5050 struct io_async_ctx *io = req->io;
5052 switch (req->opcode) {
5053 case IORING_OP_READV:
5054 case IORING_OP_READ_FIXED:
5055 case IORING_OP_READ:
5056 if (req->flags & REQ_F_BUFFER_SELECTED)
5057 kfree((void *)(unsigned long)req->rw.addr);
5059 case IORING_OP_WRITEV:
5060 case IORING_OP_WRITE_FIXED:
5061 case IORING_OP_WRITE:
5062 if (io->rw.iov != io->rw.fast_iov)
5065 case IORING_OP_RECVMSG:
5066 if (req->flags & REQ_F_BUFFER_SELECTED)
5067 kfree(req->sr_msg.kbuf);
5069 case IORING_OP_SENDMSG:
5070 if (io->msg.iov != io->msg.fast_iov)
5073 case IORING_OP_RECV:
5074 if (req->flags & REQ_F_BUFFER_SELECTED)
5075 kfree(req->sr_msg.kbuf);
5077 case IORING_OP_OPENAT:
5078 case IORING_OP_OPENAT2:
5079 case IORING_OP_STATX:
5080 putname(req->open.filename);
5082 case IORING_OP_SPLICE:
5083 io_put_file(req, req->splice.file_in,
5084 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5088 req->flags &= ~REQ_F_NEED_CLEANUP;
5091 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5092 bool force_nonblock)
5094 struct io_ring_ctx *ctx = req->ctx;
5097 switch (req->opcode) {
5101 case IORING_OP_READV:
5102 case IORING_OP_READ_FIXED:
5103 case IORING_OP_READ:
5105 ret = io_read_prep(req, sqe, force_nonblock);
5109 ret = io_read(req, force_nonblock);
5111 case IORING_OP_WRITEV:
5112 case IORING_OP_WRITE_FIXED:
5113 case IORING_OP_WRITE:
5115 ret = io_write_prep(req, sqe, force_nonblock);
5119 ret = io_write(req, force_nonblock);
5121 case IORING_OP_FSYNC:
5123 ret = io_prep_fsync(req, sqe);
5127 ret = io_fsync(req, force_nonblock);
5129 case IORING_OP_POLL_ADD:
5131 ret = io_poll_add_prep(req, sqe);
5135 ret = io_poll_add(req);
5137 case IORING_OP_POLL_REMOVE:
5139 ret = io_poll_remove_prep(req, sqe);
5143 ret = io_poll_remove(req);
5145 case IORING_OP_SYNC_FILE_RANGE:
5147 ret = io_prep_sfr(req, sqe);
5151 ret = io_sync_file_range(req, force_nonblock);
5153 case IORING_OP_SENDMSG:
5154 case IORING_OP_SEND:
5156 ret = io_sendmsg_prep(req, sqe);
5160 if (req->opcode == IORING_OP_SENDMSG)
5161 ret = io_sendmsg(req, force_nonblock);
5163 ret = io_send(req, force_nonblock);
5165 case IORING_OP_RECVMSG:
5166 case IORING_OP_RECV:
5168 ret = io_recvmsg_prep(req, sqe);
5172 if (req->opcode == IORING_OP_RECVMSG)
5173 ret = io_recvmsg(req, force_nonblock);
5175 ret = io_recv(req, force_nonblock);
5177 case IORING_OP_TIMEOUT:
5179 ret = io_timeout_prep(req, sqe, false);
5183 ret = io_timeout(req);
5185 case IORING_OP_TIMEOUT_REMOVE:
5187 ret = io_timeout_remove_prep(req, sqe);
5191 ret = io_timeout_remove(req);
5193 case IORING_OP_ACCEPT:
5195 ret = io_accept_prep(req, sqe);
5199 ret = io_accept(req, force_nonblock);
5201 case IORING_OP_CONNECT:
5203 ret = io_connect_prep(req, sqe);
5207 ret = io_connect(req, force_nonblock);
5209 case IORING_OP_ASYNC_CANCEL:
5211 ret = io_async_cancel_prep(req, sqe);
5215 ret = io_async_cancel(req);
5217 case IORING_OP_FALLOCATE:
5219 ret = io_fallocate_prep(req, sqe);
5223 ret = io_fallocate(req, force_nonblock);
5225 case IORING_OP_OPENAT:
5227 ret = io_openat_prep(req, sqe);
5231 ret = io_openat(req, force_nonblock);
5233 case IORING_OP_CLOSE:
5235 ret = io_close_prep(req, sqe);
5239 ret = io_close(req, force_nonblock);
5241 case IORING_OP_FILES_UPDATE:
5243 ret = io_files_update_prep(req, sqe);
5247 ret = io_files_update(req, force_nonblock);
5249 case IORING_OP_STATX:
5251 ret = io_statx_prep(req, sqe);
5255 ret = io_statx(req, force_nonblock);
5257 case IORING_OP_FADVISE:
5259 ret = io_fadvise_prep(req, sqe);
5263 ret = io_fadvise(req, force_nonblock);
5265 case IORING_OP_MADVISE:
5267 ret = io_madvise_prep(req, sqe);
5271 ret = io_madvise(req, force_nonblock);
5273 case IORING_OP_OPENAT2:
5275 ret = io_openat2_prep(req, sqe);
5279 ret = io_openat2(req, force_nonblock);
5281 case IORING_OP_EPOLL_CTL:
5283 ret = io_epoll_ctl_prep(req, sqe);
5287 ret = io_epoll_ctl(req, force_nonblock);
5289 case IORING_OP_SPLICE:
5291 ret = io_splice_prep(req, sqe);
5295 ret = io_splice(req, force_nonblock);
5297 case IORING_OP_PROVIDE_BUFFERS:
5299 ret = io_provide_buffers_prep(req, sqe);
5303 ret = io_provide_buffers(req, force_nonblock);
5305 case IORING_OP_REMOVE_BUFFERS:
5307 ret = io_remove_buffers_prep(req, sqe);
5311 ret = io_remove_buffers(req, force_nonblock);
5321 if (ctx->flags & IORING_SETUP_IOPOLL) {
5322 const bool in_async = io_wq_current_is_worker();
5324 if (req->result == -EAGAIN)
5327 /* workqueue context doesn't hold uring_lock, grab it now */
5329 mutex_lock(&ctx->uring_lock);
5331 io_iopoll_req_issued(req);
5334 mutex_unlock(&ctx->uring_lock);
5340 static void io_wq_submit_work(struct io_wq_work **workptr)
5342 struct io_wq_work *work = *workptr;
5343 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5346 /* if NO_CANCEL is set, we must still run the work */
5347 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5348 IO_WQ_WORK_CANCEL) {
5354 ret = io_issue_sqe(req, NULL, false);
5356 * We can get EAGAIN for polled IO even though we're
5357 * forcing a sync submission from here, since we can't
5358 * wait for request slots on the block side.
5367 req_set_fail_links(req);
5368 io_cqring_add_event(req, ret);
5372 io_steal_work(req, workptr);
5375 static int io_req_needs_file(struct io_kiocb *req, int fd)
5377 if (!io_op_defs[req->opcode].needs_file)
5379 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5384 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5387 struct fixed_file_table *table;
5389 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5390 return table->files[index & IORING_FILE_TABLE_MASK];;
5393 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5394 int fd, struct file **out_file, bool fixed)
5396 struct io_ring_ctx *ctx = req->ctx;
5400 if (unlikely(!ctx->file_data ||
5401 (unsigned) fd >= ctx->nr_user_files))
5403 fd = array_index_nospec(fd, ctx->nr_user_files);
5404 file = io_file_from_index(ctx, fd);
5407 req->fixed_file_refs = ctx->file_data->cur_refs;
5408 percpu_ref_get(req->fixed_file_refs);
5410 trace_io_uring_file_get(ctx, fd);
5411 file = __io_file_get(state, fd);
5412 if (unlikely(!file))
5420 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5421 int fd, unsigned int flags)
5425 if (!io_req_needs_file(req, fd))
5428 fixed = (flags & IOSQE_FIXED_FILE);
5429 if (unlikely(!fixed && req->needs_fixed_file))
5432 return io_file_get(state, req, fd, &req->file, fixed);
5435 static int io_grab_files(struct io_kiocb *req)
5438 struct io_ring_ctx *ctx = req->ctx;
5440 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5442 if (!ctx->ring_file)
5446 spin_lock_irq(&ctx->inflight_lock);
5448 * We use the f_ops->flush() handler to ensure that we can flush
5449 * out work accessing these files if the fd is closed. Check if
5450 * the fd has changed since we started down this path, and disallow
5451 * this operation if it has.
5453 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5454 list_add(&req->inflight_entry, &ctx->inflight_list);
5455 req->flags |= REQ_F_INFLIGHT;
5456 req->work.files = current->files;
5459 spin_unlock_irq(&ctx->inflight_lock);
5465 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5467 struct io_timeout_data *data = container_of(timer,
5468 struct io_timeout_data, timer);
5469 struct io_kiocb *req = data->req;
5470 struct io_ring_ctx *ctx = req->ctx;
5471 struct io_kiocb *prev = NULL;
5472 unsigned long flags;
5474 spin_lock_irqsave(&ctx->completion_lock, flags);
5477 * We don't expect the list to be empty, that will only happen if we
5478 * race with the completion of the linked work.
5480 if (!list_empty(&req->link_list)) {
5481 prev = list_entry(req->link_list.prev, struct io_kiocb,
5483 if (refcount_inc_not_zero(&prev->refs)) {
5484 list_del_init(&req->link_list);
5485 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5490 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5493 req_set_fail_links(prev);
5494 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5497 io_cqring_add_event(req, -ETIME);
5500 return HRTIMER_NORESTART;
5503 static void io_queue_linked_timeout(struct io_kiocb *req)
5505 struct io_ring_ctx *ctx = req->ctx;
5508 * If the list is now empty, then our linked request finished before
5509 * we got a chance to setup the timer
5511 spin_lock_irq(&ctx->completion_lock);
5512 if (!list_empty(&req->link_list)) {
5513 struct io_timeout_data *data = &req->io->timeout;
5515 data->timer.function = io_link_timeout_fn;
5516 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5519 spin_unlock_irq(&ctx->completion_lock);
5521 /* drop submission reference */
5525 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5527 struct io_kiocb *nxt;
5529 if (!(req->flags & REQ_F_LINK_HEAD))
5531 /* for polled retry, if flag is set, we already went through here */
5532 if (req->flags & REQ_F_POLLED)
5535 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5537 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5540 req->flags |= REQ_F_LINK_TIMEOUT;
5544 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5546 struct io_kiocb *linked_timeout;
5547 struct io_kiocb *nxt;
5548 const struct cred *old_creds = NULL;
5552 linked_timeout = io_prep_linked_timeout(req);
5554 if (req->work.creds && req->work.creds != current_cred()) {
5556 revert_creds(old_creds);
5557 if (old_creds == req->work.creds)
5558 old_creds = NULL; /* restored original creds */
5560 old_creds = override_creds(req->work.creds);
5563 ret = io_issue_sqe(req, sqe, true);
5566 * We async punt it if the file wasn't marked NOWAIT, or if the file
5567 * doesn't support non-blocking read/write attempts
5569 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5570 (req->flags & REQ_F_MUST_PUNT))) {
5571 if (io_arm_poll_handler(req)) {
5573 io_queue_linked_timeout(linked_timeout);
5577 if (io_op_defs[req->opcode].file_table) {
5578 ret = io_grab_files(req);
5584 * Queued up for async execution, worker will release
5585 * submit reference when the iocb is actually submitted.
5587 io_queue_async_work(req);
5593 /* drop submission reference */
5594 io_put_req_find_next(req, &nxt);
5596 if (linked_timeout) {
5598 io_queue_linked_timeout(linked_timeout);
5600 io_put_req(linked_timeout);
5603 /* and drop final reference, if we failed */
5605 io_cqring_add_event(req, ret);
5606 req_set_fail_links(req);
5612 if (req->flags & REQ_F_FORCE_ASYNC)
5618 revert_creds(old_creds);
5621 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5625 ret = io_req_defer(req, sqe);
5627 if (ret != -EIOCBQUEUED) {
5629 io_cqring_add_event(req, ret);
5630 req_set_fail_links(req);
5631 io_double_put_req(req);
5633 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5634 ret = io_req_defer_prep(req, sqe);
5635 if (unlikely(ret < 0))
5638 * Never try inline submit of IOSQE_ASYNC is set, go straight
5639 * to async execution.
5641 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5642 io_queue_async_work(req);
5644 __io_queue_sqe(req, sqe);
5648 static inline void io_queue_link_head(struct io_kiocb *req)
5650 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5651 io_cqring_add_event(req, -ECANCELED);
5652 io_double_put_req(req);
5654 io_queue_sqe(req, NULL);
5657 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5658 struct io_submit_state *state, struct io_kiocb **link)
5660 struct io_ring_ctx *ctx = req->ctx;
5664 * If we already have a head request, queue this one for async
5665 * submittal once the head completes. If we don't have a head but
5666 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5667 * submitted sync once the chain is complete. If none of those
5668 * conditions are true (normal request), then just queue it.
5671 struct io_kiocb *head = *link;
5674 * Taking sequential execution of a link, draining both sides
5675 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5676 * requests in the link. So, it drains the head and the
5677 * next after the link request. The last one is done via
5678 * drain_next flag to persist the effect across calls.
5680 if (req->flags & REQ_F_IO_DRAIN) {
5681 head->flags |= REQ_F_IO_DRAIN;
5682 ctx->drain_next = 1;
5684 if (io_alloc_async_ctx(req))
5687 ret = io_req_defer_prep(req, sqe);
5689 /* fail even hard links since we don't submit */
5690 head->flags |= REQ_F_FAIL_LINK;
5693 trace_io_uring_link(ctx, req, head);
5694 list_add_tail(&req->link_list, &head->link_list);
5696 /* last request of a link, enqueue the link */
5697 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5698 io_queue_link_head(head);
5702 if (unlikely(ctx->drain_next)) {
5703 req->flags |= REQ_F_IO_DRAIN;
5704 ctx->drain_next = 0;
5706 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5707 req->flags |= REQ_F_LINK_HEAD;
5708 INIT_LIST_HEAD(&req->link_list);
5710 if (io_alloc_async_ctx(req))
5713 ret = io_req_defer_prep(req, sqe);
5715 req->flags |= REQ_F_FAIL_LINK;
5718 io_queue_sqe(req, sqe);
5726 * Batched submission is done, ensure local IO is flushed out.
5728 static void io_submit_state_end(struct io_submit_state *state)
5730 blk_finish_plug(&state->plug);
5732 if (state->free_reqs)
5733 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5737 * Start submission side cache.
5739 static void io_submit_state_start(struct io_submit_state *state,
5740 unsigned int max_ios)
5742 blk_start_plug(&state->plug);
5743 state->free_reqs = 0;
5745 state->ios_left = max_ios;
5748 static void io_commit_sqring(struct io_ring_ctx *ctx)
5750 struct io_rings *rings = ctx->rings;
5753 * Ensure any loads from the SQEs are done at this point,
5754 * since once we write the new head, the application could
5755 * write new data to them.
5757 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5761 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5762 * that is mapped by userspace. This means that care needs to be taken to
5763 * ensure that reads are stable, as we cannot rely on userspace always
5764 * being a good citizen. If members of the sqe are validated and then later
5765 * used, it's important that those reads are done through READ_ONCE() to
5766 * prevent a re-load down the line.
5768 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5770 u32 *sq_array = ctx->sq_array;
5774 * The cached sq head (or cq tail) serves two purposes:
5776 * 1) allows us to batch the cost of updating the user visible
5778 * 2) allows the kernel side to track the head on its own, even
5779 * though the application is the one updating it.
5781 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5782 if (likely(head < ctx->sq_entries))
5783 return &ctx->sq_sqes[head];
5785 /* drop invalid entries */
5786 ctx->cached_sq_dropped++;
5787 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5791 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5793 ctx->cached_sq_head++;
5796 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5797 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5798 IOSQE_BUFFER_SELECT)
5800 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5801 const struct io_uring_sqe *sqe,
5802 struct io_submit_state *state, bool async)
5804 unsigned int sqe_flags;
5808 * All io need record the previous position, if LINK vs DARIN,
5809 * it can be used to mark the position of the first IO in the
5812 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5813 req->opcode = READ_ONCE(sqe->opcode);
5814 req->user_data = READ_ONCE(sqe->user_data);
5819 /* one is dropped after submission, the other at completion */
5820 refcount_set(&req->refs, 2);
5823 req->needs_fixed_file = async;
5824 INIT_IO_WORK(&req->work, io_wq_submit_work);
5826 if (unlikely(req->opcode >= IORING_OP_LAST))
5829 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5830 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5832 use_mm(ctx->sqo_mm);
5835 sqe_flags = READ_ONCE(sqe->flags);
5836 /* enforce forwards compatibility on users */
5837 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5840 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5841 !io_op_defs[req->opcode].buffer_select)
5844 id = READ_ONCE(sqe->personality);
5846 req->work.creds = idr_find(&ctx->personality_idr, id);
5847 if (unlikely(!req->work.creds))
5849 get_cred(req->work.creds);
5852 /* same numerical values with corresponding REQ_F_*, safe to copy */
5853 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5854 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5855 IOSQE_BUFFER_SELECT | IOSQE_IO_LINK);
5857 fd = READ_ONCE(sqe->fd);
5858 return io_req_set_file(state, req, fd, sqe_flags);
5861 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5862 struct file *ring_file, int ring_fd, bool async)
5864 struct io_submit_state state, *statep = NULL;
5865 struct io_kiocb *link = NULL;
5866 int i, submitted = 0;
5868 /* if we have a backlog and couldn't flush it all, return BUSY */
5869 if (test_bit(0, &ctx->sq_check_overflow)) {
5870 if (!list_empty(&ctx->cq_overflow_list) &&
5871 !io_cqring_overflow_flush(ctx, false))
5875 /* make sure SQ entry isn't read before tail */
5876 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5878 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5881 if (nr > IO_PLUG_THRESHOLD) {
5882 io_submit_state_start(&state, nr);
5886 ctx->ring_fd = ring_fd;
5887 ctx->ring_file = ring_file;
5889 for (i = 0; i < nr; i++) {
5890 const struct io_uring_sqe *sqe;
5891 struct io_kiocb *req;
5894 sqe = io_get_sqe(ctx);
5895 if (unlikely(!sqe)) {
5896 io_consume_sqe(ctx);
5899 req = io_alloc_req(ctx, statep);
5900 if (unlikely(!req)) {
5902 submitted = -EAGAIN;
5906 err = io_init_req(ctx, req, sqe, statep, async);
5907 io_consume_sqe(ctx);
5908 /* will complete beyond this point, count as submitted */
5911 if (unlikely(err)) {
5913 io_cqring_add_event(req, err);
5914 io_double_put_req(req);
5918 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5920 err = io_submit_sqe(req, sqe, statep, &link);
5925 if (unlikely(submitted != nr)) {
5926 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5928 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5931 io_queue_link_head(link);
5933 io_submit_state_end(&state);
5935 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5936 io_commit_sqring(ctx);
5941 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5943 struct mm_struct *mm = current->mm;
5951 static int io_sq_thread(void *data)
5953 struct io_ring_ctx *ctx = data;
5954 const struct cred *old_cred;
5955 mm_segment_t old_fs;
5957 unsigned long timeout;
5960 complete(&ctx->completions[1]);
5964 old_cred = override_creds(ctx->creds);
5966 timeout = jiffies + ctx->sq_thread_idle;
5967 while (!kthread_should_park()) {
5968 unsigned int to_submit;
5970 if (!list_empty(&ctx->poll_list)) {
5971 unsigned nr_events = 0;
5973 mutex_lock(&ctx->uring_lock);
5974 if (!list_empty(&ctx->poll_list))
5975 io_iopoll_getevents(ctx, &nr_events, 0);
5977 timeout = jiffies + ctx->sq_thread_idle;
5978 mutex_unlock(&ctx->uring_lock);
5981 to_submit = io_sqring_entries(ctx);
5984 * If submit got -EBUSY, flag us as needing the application
5985 * to enter the kernel to reap and flush events.
5987 if (!to_submit || ret == -EBUSY) {
5989 * Drop cur_mm before scheduling, we can't hold it for
5990 * long periods (or over schedule()). Do this before
5991 * adding ourselves to the waitqueue, as the unuse/drop
5994 io_sq_thread_drop_mm(ctx);
5997 * We're polling. If we're within the defined idle
5998 * period, then let us spin without work before going
5999 * to sleep. The exception is if we got EBUSY doing
6000 * more IO, we should wait for the application to
6001 * reap events and wake us up.
6003 if (!list_empty(&ctx->poll_list) ||
6004 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6005 !percpu_ref_is_dying(&ctx->refs))) {
6006 if (current->task_works)
6012 prepare_to_wait(&ctx->sqo_wait, &wait,
6013 TASK_INTERRUPTIBLE);
6016 * While doing polled IO, before going to sleep, we need
6017 * to check if there are new reqs added to poll_list, it
6018 * is because reqs may have been punted to io worker and
6019 * will be added to poll_list later, hence check the
6022 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6023 !list_empty_careful(&ctx->poll_list)) {
6024 finish_wait(&ctx->sqo_wait, &wait);
6028 /* Tell userspace we may need a wakeup call */
6029 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6030 /* make sure to read SQ tail after writing flags */
6033 to_submit = io_sqring_entries(ctx);
6034 if (!to_submit || ret == -EBUSY) {
6035 if (kthread_should_park()) {
6036 finish_wait(&ctx->sqo_wait, &wait);
6039 if (current->task_works) {
6041 finish_wait(&ctx->sqo_wait, &wait);
6044 if (signal_pending(current))
6045 flush_signals(current);
6047 finish_wait(&ctx->sqo_wait, &wait);
6049 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6052 finish_wait(&ctx->sqo_wait, &wait);
6054 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6057 mutex_lock(&ctx->uring_lock);
6058 ret = io_submit_sqes(ctx, to_submit, NULL, -1, true);
6059 mutex_unlock(&ctx->uring_lock);
6060 timeout = jiffies + ctx->sq_thread_idle;
6063 if (current->task_works)
6067 io_sq_thread_drop_mm(ctx);
6068 revert_creds(old_cred);
6075 struct io_wait_queue {
6076 struct wait_queue_entry wq;
6077 struct io_ring_ctx *ctx;
6079 unsigned nr_timeouts;
6082 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6084 struct io_ring_ctx *ctx = iowq->ctx;
6087 * Wake up if we have enough events, or if a timeout occurred since we
6088 * started waiting. For timeouts, we always want to return to userspace,
6089 * regardless of event count.
6091 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6092 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6095 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6096 int wake_flags, void *key)
6098 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6101 /* use noflush == true, as we can't safely rely on locking context */
6102 if (!io_should_wake(iowq, true))
6105 return autoremove_wake_function(curr, mode, wake_flags, key);
6109 * Wait until events become available, if we don't already have some. The
6110 * application must reap them itself, as they reside on the shared cq ring.
6112 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6113 const sigset_t __user *sig, size_t sigsz)
6115 struct io_wait_queue iowq = {
6118 .func = io_wake_function,
6119 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6122 .to_wait = min_events,
6124 struct io_rings *rings = ctx->rings;
6128 if (io_cqring_events(ctx, false) >= min_events)
6130 if (!current->task_works)
6136 #ifdef CONFIG_COMPAT
6137 if (in_compat_syscall())
6138 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6142 ret = set_user_sigmask(sig, sigsz);
6148 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6149 trace_io_uring_cqring_wait(ctx, min_events);
6151 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6152 TASK_INTERRUPTIBLE);
6153 if (current->task_works)
6155 if (io_should_wake(&iowq, false))
6158 if (signal_pending(current)) {
6163 finish_wait(&ctx->wait, &iowq.wq);
6165 restore_saved_sigmask_unless(ret == -EINTR);
6167 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6170 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6172 #if defined(CONFIG_UNIX)
6173 if (ctx->ring_sock) {
6174 struct sock *sock = ctx->ring_sock->sk;
6175 struct sk_buff *skb;
6177 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6183 for (i = 0; i < ctx->nr_user_files; i++) {
6186 file = io_file_from_index(ctx, i);
6193 static void io_file_ref_kill(struct percpu_ref *ref)
6195 struct fixed_file_data *data;
6197 data = container_of(ref, struct fixed_file_data, refs);
6198 complete(&data->done);
6201 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6203 struct fixed_file_data *data = ctx->file_data;
6204 struct fixed_file_ref_node *ref_node = NULL;
6205 unsigned nr_tables, i;
6206 unsigned long flags;
6211 spin_lock_irqsave(&data->lock, flags);
6212 if (!list_empty(&data->ref_list))
6213 ref_node = list_first_entry(&data->ref_list,
6214 struct fixed_file_ref_node, node);
6215 spin_unlock_irqrestore(&data->lock, flags);
6217 percpu_ref_kill(&ref_node->refs);
6219 percpu_ref_kill(&data->refs);
6221 /* wait for all refs nodes to complete */
6222 wait_for_completion(&data->done);
6224 __io_sqe_files_unregister(ctx);
6225 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6226 for (i = 0; i < nr_tables; i++)
6227 kfree(data->table[i].files);
6229 percpu_ref_exit(&data->refs);
6231 ctx->file_data = NULL;
6232 ctx->nr_user_files = 0;
6236 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6238 if (ctx->sqo_thread) {
6239 wait_for_completion(&ctx->completions[1]);
6241 * The park is a bit of a work-around, without it we get
6242 * warning spews on shutdown with SQPOLL set and affinity
6243 * set to a single CPU.
6245 kthread_park(ctx->sqo_thread);
6246 kthread_stop(ctx->sqo_thread);
6247 ctx->sqo_thread = NULL;
6251 static void io_finish_async(struct io_ring_ctx *ctx)
6253 io_sq_thread_stop(ctx);
6256 io_wq_destroy(ctx->io_wq);
6261 #if defined(CONFIG_UNIX)
6263 * Ensure the UNIX gc is aware of our file set, so we are certain that
6264 * the io_uring can be safely unregistered on process exit, even if we have
6265 * loops in the file referencing.
6267 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6269 struct sock *sk = ctx->ring_sock->sk;
6270 struct scm_fp_list *fpl;
6271 struct sk_buff *skb;
6274 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6278 skb = alloc_skb(0, GFP_KERNEL);
6287 fpl->user = get_uid(ctx->user);
6288 for (i = 0; i < nr; i++) {
6289 struct file *file = io_file_from_index(ctx, i + offset);
6293 fpl->fp[nr_files] = get_file(file);
6294 unix_inflight(fpl->user, fpl->fp[nr_files]);
6299 fpl->max = SCM_MAX_FD;
6300 fpl->count = nr_files;
6301 UNIXCB(skb).fp = fpl;
6302 skb->destructor = unix_destruct_scm;
6303 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6304 skb_queue_head(&sk->sk_receive_queue, skb);
6306 for (i = 0; i < nr_files; i++)
6317 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6318 * causes regular reference counting to break down. We rely on the UNIX
6319 * garbage collection to take care of this problem for us.
6321 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6323 unsigned left, total;
6327 left = ctx->nr_user_files;
6329 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6331 ret = __io_sqe_files_scm(ctx, this_files, total);
6335 total += this_files;
6341 while (total < ctx->nr_user_files) {
6342 struct file *file = io_file_from_index(ctx, total);
6352 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6358 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6363 for (i = 0; i < nr_tables; i++) {
6364 struct fixed_file_table *table = &ctx->file_data->table[i];
6365 unsigned this_files;
6367 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6368 table->files = kcalloc(this_files, sizeof(struct file *),
6372 nr_files -= this_files;
6378 for (i = 0; i < nr_tables; i++) {
6379 struct fixed_file_table *table = &ctx->file_data->table[i];
6380 kfree(table->files);
6385 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6387 #if defined(CONFIG_UNIX)
6388 struct sock *sock = ctx->ring_sock->sk;
6389 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6390 struct sk_buff *skb;
6393 __skb_queue_head_init(&list);
6396 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6397 * remove this entry and rearrange the file array.
6399 skb = skb_dequeue(head);
6401 struct scm_fp_list *fp;
6403 fp = UNIXCB(skb).fp;
6404 for (i = 0; i < fp->count; i++) {
6407 if (fp->fp[i] != file)
6410 unix_notinflight(fp->user, fp->fp[i]);
6411 left = fp->count - 1 - i;
6413 memmove(&fp->fp[i], &fp->fp[i + 1],
6414 left * sizeof(struct file *));
6421 __skb_queue_tail(&list, skb);
6431 __skb_queue_tail(&list, skb);
6433 skb = skb_dequeue(head);
6436 if (skb_peek(&list)) {
6437 spin_lock_irq(&head->lock);
6438 while ((skb = __skb_dequeue(&list)) != NULL)
6439 __skb_queue_tail(head, skb);
6440 spin_unlock_irq(&head->lock);
6447 struct io_file_put {
6448 struct list_head list;
6452 static void io_file_put_work(struct work_struct *work)
6454 struct fixed_file_ref_node *ref_node;
6455 struct fixed_file_data *file_data;
6456 struct io_ring_ctx *ctx;
6457 struct io_file_put *pfile, *tmp;
6458 unsigned long flags;
6460 ref_node = container_of(work, struct fixed_file_ref_node, work);
6461 file_data = ref_node->file_data;
6462 ctx = file_data->ctx;
6464 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6465 list_del_init(&pfile->list);
6466 io_ring_file_put(ctx, pfile->file);
6470 spin_lock_irqsave(&file_data->lock, flags);
6471 list_del_init(&ref_node->node);
6472 spin_unlock_irqrestore(&file_data->lock, flags);
6474 percpu_ref_exit(&ref_node->refs);
6476 percpu_ref_put(&file_data->refs);
6479 static void io_file_data_ref_zero(struct percpu_ref *ref)
6481 struct fixed_file_ref_node *ref_node;
6483 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6485 queue_work(system_wq, &ref_node->work);
6488 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6489 struct io_ring_ctx *ctx)
6491 struct fixed_file_ref_node *ref_node;
6493 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6495 return ERR_PTR(-ENOMEM);
6497 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6500 return ERR_PTR(-ENOMEM);
6502 INIT_LIST_HEAD(&ref_node->node);
6503 INIT_LIST_HEAD(&ref_node->file_list);
6504 INIT_WORK(&ref_node->work, io_file_put_work);
6505 ref_node->file_data = ctx->file_data;
6510 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6512 percpu_ref_exit(&ref_node->refs);
6516 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6519 __s32 __user *fds = (__s32 __user *) arg;
6524 struct fixed_file_ref_node *ref_node;
6525 unsigned long flags;
6531 if (nr_args > IORING_MAX_FIXED_FILES)
6534 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6535 if (!ctx->file_data)
6537 ctx->file_data->ctx = ctx;
6538 init_completion(&ctx->file_data->done);
6539 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6540 spin_lock_init(&ctx->file_data->lock);
6542 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6543 ctx->file_data->table = kcalloc(nr_tables,
6544 sizeof(struct fixed_file_table),
6546 if (!ctx->file_data->table) {
6547 kfree(ctx->file_data);
6548 ctx->file_data = NULL;
6552 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6553 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6554 kfree(ctx->file_data->table);
6555 kfree(ctx->file_data);
6556 ctx->file_data = NULL;
6560 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6561 percpu_ref_exit(&ctx->file_data->refs);
6562 kfree(ctx->file_data->table);
6563 kfree(ctx->file_data);
6564 ctx->file_data = NULL;
6568 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6569 struct fixed_file_table *table;
6573 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6575 /* allow sparse sets */
6581 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6582 index = i & IORING_FILE_TABLE_MASK;
6590 * Don't allow io_uring instances to be registered. If UNIX
6591 * isn't enabled, then this causes a reference cycle and this
6592 * instance can never get freed. If UNIX is enabled we'll
6593 * handle it just fine, but there's still no point in allowing
6594 * a ring fd as it doesn't support regular read/write anyway.
6596 if (file->f_op == &io_uring_fops) {
6601 table->files[index] = file;
6605 for (i = 0; i < ctx->nr_user_files; i++) {
6606 file = io_file_from_index(ctx, i);
6610 for (i = 0; i < nr_tables; i++)
6611 kfree(ctx->file_data->table[i].files);
6613 kfree(ctx->file_data->table);
6614 kfree(ctx->file_data);
6615 ctx->file_data = NULL;
6616 ctx->nr_user_files = 0;
6620 ret = io_sqe_files_scm(ctx);
6622 io_sqe_files_unregister(ctx);
6626 ref_node = alloc_fixed_file_ref_node(ctx);
6627 if (IS_ERR(ref_node)) {
6628 io_sqe_files_unregister(ctx);
6629 return PTR_ERR(ref_node);
6632 ctx->file_data->cur_refs = &ref_node->refs;
6633 spin_lock_irqsave(&ctx->file_data->lock, flags);
6634 list_add(&ref_node->node, &ctx->file_data->ref_list);
6635 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6636 percpu_ref_get(&ctx->file_data->refs);
6640 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6643 #if defined(CONFIG_UNIX)
6644 struct sock *sock = ctx->ring_sock->sk;
6645 struct sk_buff_head *head = &sock->sk_receive_queue;
6646 struct sk_buff *skb;
6649 * See if we can merge this file into an existing skb SCM_RIGHTS
6650 * file set. If there's no room, fall back to allocating a new skb
6651 * and filling it in.
6653 spin_lock_irq(&head->lock);
6654 skb = skb_peek(head);
6656 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6658 if (fpl->count < SCM_MAX_FD) {
6659 __skb_unlink(skb, head);
6660 spin_unlock_irq(&head->lock);
6661 fpl->fp[fpl->count] = get_file(file);
6662 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6664 spin_lock_irq(&head->lock);
6665 __skb_queue_head(head, skb);
6670 spin_unlock_irq(&head->lock);
6677 return __io_sqe_files_scm(ctx, 1, index);
6683 static int io_queue_file_removal(struct fixed_file_data *data,
6686 struct io_file_put *pfile;
6687 struct percpu_ref *refs = data->cur_refs;
6688 struct fixed_file_ref_node *ref_node;
6690 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6694 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6696 list_add(&pfile->list, &ref_node->file_list);
6701 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6702 struct io_uring_files_update *up,
6705 struct fixed_file_data *data = ctx->file_data;
6706 struct fixed_file_ref_node *ref_node;
6711 unsigned long flags;
6712 bool needs_switch = false;
6714 if (check_add_overflow(up->offset, nr_args, &done))
6716 if (done > ctx->nr_user_files)
6719 ref_node = alloc_fixed_file_ref_node(ctx);
6720 if (IS_ERR(ref_node))
6721 return PTR_ERR(ref_node);
6724 fds = u64_to_user_ptr(up->fds);
6726 struct fixed_file_table *table;
6730 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6734 i = array_index_nospec(up->offset, ctx->nr_user_files);
6735 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6736 index = i & IORING_FILE_TABLE_MASK;
6737 if (table->files[index]) {
6738 file = io_file_from_index(ctx, index);
6739 err = io_queue_file_removal(data, file);
6742 table->files[index] = NULL;
6743 needs_switch = true;
6752 * Don't allow io_uring instances to be registered. If
6753 * UNIX isn't enabled, then this causes a reference
6754 * cycle and this instance can never get freed. If UNIX
6755 * is enabled we'll handle it just fine, but there's
6756 * still no point in allowing a ring fd as it doesn't
6757 * support regular read/write anyway.
6759 if (file->f_op == &io_uring_fops) {
6764 table->files[index] = file;
6765 err = io_sqe_file_register(ctx, file, i);
6775 percpu_ref_kill(data->cur_refs);
6776 spin_lock_irqsave(&data->lock, flags);
6777 list_add(&ref_node->node, &data->ref_list);
6778 data->cur_refs = &ref_node->refs;
6779 spin_unlock_irqrestore(&data->lock, flags);
6780 percpu_ref_get(&ctx->file_data->refs);
6782 destroy_fixed_file_ref_node(ref_node);
6784 return done ? done : err;
6787 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6790 struct io_uring_files_update up;
6792 if (!ctx->file_data)
6796 if (copy_from_user(&up, arg, sizeof(up)))
6801 return __io_sqe_files_update(ctx, &up, nr_args);
6804 static void io_free_work(struct io_wq_work *work)
6806 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6808 /* Consider that io_steal_work() relies on this ref */
6812 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6813 struct io_uring_params *p)
6815 struct io_wq_data data;
6817 struct io_ring_ctx *ctx_attach;
6818 unsigned int concurrency;
6821 data.user = ctx->user;
6822 data.free_work = io_free_work;
6824 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6825 /* Do QD, or 4 * CPUS, whatever is smallest */
6826 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6828 ctx->io_wq = io_wq_create(concurrency, &data);
6829 if (IS_ERR(ctx->io_wq)) {
6830 ret = PTR_ERR(ctx->io_wq);
6836 f = fdget(p->wq_fd);
6840 if (f.file->f_op != &io_uring_fops) {
6845 ctx_attach = f.file->private_data;
6846 /* @io_wq is protected by holding the fd */
6847 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6852 ctx->io_wq = ctx_attach->io_wq;
6858 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6859 struct io_uring_params *p)
6863 init_waitqueue_head(&ctx->sqo_wait);
6864 mmgrab(current->mm);
6865 ctx->sqo_mm = current->mm;
6867 if (ctx->flags & IORING_SETUP_SQPOLL) {
6869 if (!capable(CAP_SYS_ADMIN))
6872 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6873 if (!ctx->sq_thread_idle)
6874 ctx->sq_thread_idle = HZ;
6876 if (p->flags & IORING_SETUP_SQ_AFF) {
6877 int cpu = p->sq_thread_cpu;
6880 if (cpu >= nr_cpu_ids)
6882 if (!cpu_online(cpu))
6885 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6889 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6892 if (IS_ERR(ctx->sqo_thread)) {
6893 ret = PTR_ERR(ctx->sqo_thread);
6894 ctx->sqo_thread = NULL;
6897 wake_up_process(ctx->sqo_thread);
6898 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6899 /* Can't have SQ_AFF without SQPOLL */
6904 ret = io_init_wq_offload(ctx, p);
6910 io_finish_async(ctx);
6911 mmdrop(ctx->sqo_mm);
6916 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6918 atomic_long_sub(nr_pages, &user->locked_vm);
6921 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6923 unsigned long page_limit, cur_pages, new_pages;
6925 /* Don't allow more pages than we can safely lock */
6926 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6929 cur_pages = atomic_long_read(&user->locked_vm);
6930 new_pages = cur_pages + nr_pages;
6931 if (new_pages > page_limit)
6933 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6934 new_pages) != cur_pages);
6939 static void io_mem_free(void *ptr)
6946 page = virt_to_head_page(ptr);
6947 if (put_page_testzero(page))
6948 free_compound_page(page);
6951 static void *io_mem_alloc(size_t size)
6953 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6956 return (void *) __get_free_pages(gfp_flags, get_order(size));
6959 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6962 struct io_rings *rings;
6963 size_t off, sq_array_size;
6965 off = struct_size(rings, cqes, cq_entries);
6966 if (off == SIZE_MAX)
6970 off = ALIGN(off, SMP_CACHE_BYTES);
6975 sq_array_size = array_size(sizeof(u32), sq_entries);
6976 if (sq_array_size == SIZE_MAX)
6979 if (check_add_overflow(off, sq_array_size, &off))
6988 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6992 pages = (size_t)1 << get_order(
6993 rings_size(sq_entries, cq_entries, NULL));
6994 pages += (size_t)1 << get_order(
6995 array_size(sizeof(struct io_uring_sqe), sq_entries));
7000 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7004 if (!ctx->user_bufs)
7007 for (i = 0; i < ctx->nr_user_bufs; i++) {
7008 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7010 for (j = 0; j < imu->nr_bvecs; j++)
7011 unpin_user_page(imu->bvec[j].bv_page);
7013 if (ctx->account_mem)
7014 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7019 kfree(ctx->user_bufs);
7020 ctx->user_bufs = NULL;
7021 ctx->nr_user_bufs = 0;
7025 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7026 void __user *arg, unsigned index)
7028 struct iovec __user *src;
7030 #ifdef CONFIG_COMPAT
7032 struct compat_iovec __user *ciovs;
7033 struct compat_iovec ciov;
7035 ciovs = (struct compat_iovec __user *) arg;
7036 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7039 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7040 dst->iov_len = ciov.iov_len;
7044 src = (struct iovec __user *) arg;
7045 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7050 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7053 struct vm_area_struct **vmas = NULL;
7054 struct page **pages = NULL;
7055 int i, j, got_pages = 0;
7060 if (!nr_args || nr_args > UIO_MAXIOV)
7063 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7065 if (!ctx->user_bufs)
7068 for (i = 0; i < nr_args; i++) {
7069 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7070 unsigned long off, start, end, ubuf;
7075 ret = io_copy_iov(ctx, &iov, arg, i);
7080 * Don't impose further limits on the size and buffer
7081 * constraints here, we'll -EINVAL later when IO is
7082 * submitted if they are wrong.
7085 if (!iov.iov_base || !iov.iov_len)
7088 /* arbitrary limit, but we need something */
7089 if (iov.iov_len > SZ_1G)
7092 ubuf = (unsigned long) iov.iov_base;
7093 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7094 start = ubuf >> PAGE_SHIFT;
7095 nr_pages = end - start;
7097 if (ctx->account_mem) {
7098 ret = io_account_mem(ctx->user, nr_pages);
7104 if (!pages || nr_pages > got_pages) {
7107 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7109 vmas = kvmalloc_array(nr_pages,
7110 sizeof(struct vm_area_struct *),
7112 if (!pages || !vmas) {
7114 if (ctx->account_mem)
7115 io_unaccount_mem(ctx->user, nr_pages);
7118 got_pages = nr_pages;
7121 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7125 if (ctx->account_mem)
7126 io_unaccount_mem(ctx->user, nr_pages);
7131 down_read(¤t->mm->mmap_sem);
7132 pret = pin_user_pages(ubuf, nr_pages,
7133 FOLL_WRITE | FOLL_LONGTERM,
7135 if (pret == nr_pages) {
7136 /* don't support file backed memory */
7137 for (j = 0; j < nr_pages; j++) {
7138 struct vm_area_struct *vma = vmas[j];
7141 !is_file_hugepages(vma->vm_file)) {
7147 ret = pret < 0 ? pret : -EFAULT;
7149 up_read(¤t->mm->mmap_sem);
7152 * if we did partial map, or found file backed vmas,
7153 * release any pages we did get
7156 unpin_user_pages(pages, pret);
7157 if (ctx->account_mem)
7158 io_unaccount_mem(ctx->user, nr_pages);
7163 off = ubuf & ~PAGE_MASK;
7165 for (j = 0; j < nr_pages; j++) {
7168 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7169 imu->bvec[j].bv_page = pages[j];
7170 imu->bvec[j].bv_len = vec_len;
7171 imu->bvec[j].bv_offset = off;
7175 /* store original address for later verification */
7177 imu->len = iov.iov_len;
7178 imu->nr_bvecs = nr_pages;
7180 ctx->nr_user_bufs++;
7188 io_sqe_buffer_unregister(ctx);
7192 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7194 __s32 __user *fds = arg;
7200 if (copy_from_user(&fd, fds, sizeof(*fds)))
7203 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7204 if (IS_ERR(ctx->cq_ev_fd)) {
7205 int ret = PTR_ERR(ctx->cq_ev_fd);
7206 ctx->cq_ev_fd = NULL;
7213 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7215 if (ctx->cq_ev_fd) {
7216 eventfd_ctx_put(ctx->cq_ev_fd);
7217 ctx->cq_ev_fd = NULL;
7224 static int __io_destroy_buffers(int id, void *p, void *data)
7226 struct io_ring_ctx *ctx = data;
7227 struct io_buffer *buf = p;
7229 __io_remove_buffers(ctx, buf, id, -1U);
7233 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7235 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7236 idr_destroy(&ctx->io_buffer_idr);
7239 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7241 io_finish_async(ctx);
7243 mmdrop(ctx->sqo_mm);
7245 io_iopoll_reap_events(ctx);
7246 io_sqe_buffer_unregister(ctx);
7247 io_sqe_files_unregister(ctx);
7248 io_eventfd_unregister(ctx);
7249 io_destroy_buffers(ctx);
7250 idr_destroy(&ctx->personality_idr);
7252 #if defined(CONFIG_UNIX)
7253 if (ctx->ring_sock) {
7254 ctx->ring_sock->file = NULL; /* so that iput() is called */
7255 sock_release(ctx->ring_sock);
7259 io_mem_free(ctx->rings);
7260 io_mem_free(ctx->sq_sqes);
7262 percpu_ref_exit(&ctx->refs);
7263 if (ctx->account_mem)
7264 io_unaccount_mem(ctx->user,
7265 ring_pages(ctx->sq_entries, ctx->cq_entries));
7266 free_uid(ctx->user);
7267 put_cred(ctx->creds);
7268 kfree(ctx->completions);
7269 kfree(ctx->cancel_hash);
7270 kmem_cache_free(req_cachep, ctx->fallback_req);
7274 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7276 struct io_ring_ctx *ctx = file->private_data;
7279 poll_wait(file, &ctx->cq_wait, wait);
7281 * synchronizes with barrier from wq_has_sleeper call in
7285 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7286 ctx->rings->sq_ring_entries)
7287 mask |= EPOLLOUT | EPOLLWRNORM;
7288 if (io_cqring_events(ctx, false))
7289 mask |= EPOLLIN | EPOLLRDNORM;
7294 static int io_uring_fasync(int fd, struct file *file, int on)
7296 struct io_ring_ctx *ctx = file->private_data;
7298 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7301 static int io_remove_personalities(int id, void *p, void *data)
7303 struct io_ring_ctx *ctx = data;
7304 const struct cred *cred;
7306 cred = idr_remove(&ctx->personality_idr, id);
7312 static void io_ring_exit_work(struct work_struct *work)
7314 struct io_ring_ctx *ctx;
7316 ctx = container_of(work, struct io_ring_ctx, exit_work);
7318 io_cqring_overflow_flush(ctx, true);
7320 wait_for_completion(&ctx->completions[0]);
7321 io_ring_ctx_free(ctx);
7324 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7326 mutex_lock(&ctx->uring_lock);
7327 percpu_ref_kill(&ctx->refs);
7328 mutex_unlock(&ctx->uring_lock);
7331 * Wait for sq thread to idle, if we have one. It won't spin on new
7332 * work after we've killed the ctx ref above. This is important to do
7333 * before we cancel existing commands, as the thread could otherwise
7334 * be queueing new work post that. If that's work we need to cancel,
7335 * it could cause shutdown to hang.
7337 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7340 io_kill_timeouts(ctx);
7341 io_poll_remove_all(ctx);
7344 io_wq_cancel_all(ctx->io_wq);
7346 io_iopoll_reap_events(ctx);
7347 /* if we failed setting up the ctx, we might not have any rings */
7349 io_cqring_overflow_flush(ctx, true);
7350 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7351 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7352 queue_work(system_wq, &ctx->exit_work);
7355 static int io_uring_release(struct inode *inode, struct file *file)
7357 struct io_ring_ctx *ctx = file->private_data;
7359 file->private_data = NULL;
7360 io_ring_ctx_wait_and_kill(ctx);
7364 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7365 struct files_struct *files)
7367 struct io_kiocb *req;
7370 while (!list_empty_careful(&ctx->inflight_list)) {
7371 struct io_kiocb *cancel_req = NULL;
7373 spin_lock_irq(&ctx->inflight_lock);
7374 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7375 if (req->work.files != files)
7377 /* req is being completed, ignore */
7378 if (!refcount_inc_not_zero(&req->refs))
7384 prepare_to_wait(&ctx->inflight_wait, &wait,
7385 TASK_UNINTERRUPTIBLE);
7386 spin_unlock_irq(&ctx->inflight_lock);
7388 /* We need to keep going until we don't find a matching req */
7392 if (cancel_req->flags & REQ_F_OVERFLOW) {
7393 spin_lock_irq(&ctx->completion_lock);
7394 list_del(&cancel_req->list);
7395 cancel_req->flags &= ~REQ_F_OVERFLOW;
7396 if (list_empty(&ctx->cq_overflow_list)) {
7397 clear_bit(0, &ctx->sq_check_overflow);
7398 clear_bit(0, &ctx->cq_check_overflow);
7400 spin_unlock_irq(&ctx->completion_lock);
7402 WRITE_ONCE(ctx->rings->cq_overflow,
7403 atomic_inc_return(&ctx->cached_cq_overflow));
7406 * Put inflight ref and overflow ref. If that's
7407 * all we had, then we're done with this request.
7409 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7410 io_put_req(cancel_req);
7415 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7416 io_put_req(cancel_req);
7419 finish_wait(&ctx->inflight_wait, &wait);
7422 static int io_uring_flush(struct file *file, void *data)
7424 struct io_ring_ctx *ctx = file->private_data;
7426 io_uring_cancel_files(ctx, data);
7429 * If the task is going away, cancel work it may have pending
7431 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7432 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7437 static void *io_uring_validate_mmap_request(struct file *file,
7438 loff_t pgoff, size_t sz)
7440 struct io_ring_ctx *ctx = file->private_data;
7441 loff_t offset = pgoff << PAGE_SHIFT;
7446 case IORING_OFF_SQ_RING:
7447 case IORING_OFF_CQ_RING:
7450 case IORING_OFF_SQES:
7454 return ERR_PTR(-EINVAL);
7457 page = virt_to_head_page(ptr);
7458 if (sz > page_size(page))
7459 return ERR_PTR(-EINVAL);
7466 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7468 size_t sz = vma->vm_end - vma->vm_start;
7472 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7474 return PTR_ERR(ptr);
7476 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7477 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7480 #else /* !CONFIG_MMU */
7482 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7484 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7487 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7489 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7492 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7493 unsigned long addr, unsigned long len,
7494 unsigned long pgoff, unsigned long flags)
7498 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7500 return PTR_ERR(ptr);
7502 return (unsigned long) ptr;
7505 #endif /* !CONFIG_MMU */
7507 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7508 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7511 struct io_ring_ctx *ctx;
7516 if (current->task_works)
7519 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7527 if (f.file->f_op != &io_uring_fops)
7531 ctx = f.file->private_data;
7532 if (!percpu_ref_tryget(&ctx->refs))
7536 * For SQ polling, the thread will do all submissions and completions.
7537 * Just return the requested submit count, and wake the thread if
7541 if (ctx->flags & IORING_SETUP_SQPOLL) {
7542 if (!list_empty_careful(&ctx->cq_overflow_list))
7543 io_cqring_overflow_flush(ctx, false);
7544 if (flags & IORING_ENTER_SQ_WAKEUP)
7545 wake_up(&ctx->sqo_wait);
7546 submitted = to_submit;
7547 } else if (to_submit) {
7548 mutex_lock(&ctx->uring_lock);
7549 submitted = io_submit_sqes(ctx, to_submit, f.file, fd, false);
7550 mutex_unlock(&ctx->uring_lock);
7552 if (submitted != to_submit)
7555 if (flags & IORING_ENTER_GETEVENTS) {
7556 unsigned nr_events = 0;
7558 min_complete = min(min_complete, ctx->cq_entries);
7561 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7562 * space applications don't need to do io completion events
7563 * polling again, they can rely on io_sq_thread to do polling
7564 * work, which can reduce cpu usage and uring_lock contention.
7566 if (ctx->flags & IORING_SETUP_IOPOLL &&
7567 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7568 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7570 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7575 percpu_ref_put(&ctx->refs);
7578 return submitted ? submitted : ret;
7581 #ifdef CONFIG_PROC_FS
7582 static int io_uring_show_cred(int id, void *p, void *data)
7584 const struct cred *cred = p;
7585 struct seq_file *m = data;
7586 struct user_namespace *uns = seq_user_ns(m);
7587 struct group_info *gi;
7592 seq_printf(m, "%5d\n", id);
7593 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7594 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7595 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7596 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7597 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7598 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7599 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7600 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7601 seq_puts(m, "\n\tGroups:\t");
7602 gi = cred->group_info;
7603 for (g = 0; g < gi->ngroups; g++) {
7604 seq_put_decimal_ull(m, g ? " " : "",
7605 from_kgid_munged(uns, gi->gid[g]));
7607 seq_puts(m, "\n\tCapEff:\t");
7608 cap = cred->cap_effective;
7609 CAP_FOR_EACH_U32(__capi)
7610 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7615 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7619 mutex_lock(&ctx->uring_lock);
7620 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7621 for (i = 0; i < ctx->nr_user_files; i++) {
7622 struct fixed_file_table *table;
7625 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7626 f = table->files[i & IORING_FILE_TABLE_MASK];
7628 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7630 seq_printf(m, "%5u: <none>\n", i);
7632 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7633 for (i = 0; i < ctx->nr_user_bufs; i++) {
7634 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7636 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7637 (unsigned int) buf->len);
7639 if (!idr_is_empty(&ctx->personality_idr)) {
7640 seq_printf(m, "Personalities:\n");
7641 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7643 seq_printf(m, "PollList:\n");
7644 spin_lock_irq(&ctx->completion_lock);
7645 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7646 struct hlist_head *list = &ctx->cancel_hash[i];
7647 struct io_kiocb *req;
7649 hlist_for_each_entry(req, list, hash_node)
7650 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7651 req->task->task_works != NULL);
7653 spin_unlock_irq(&ctx->completion_lock);
7654 mutex_unlock(&ctx->uring_lock);
7657 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7659 struct io_ring_ctx *ctx = f->private_data;
7661 if (percpu_ref_tryget(&ctx->refs)) {
7662 __io_uring_show_fdinfo(ctx, m);
7663 percpu_ref_put(&ctx->refs);
7668 static const struct file_operations io_uring_fops = {
7669 .release = io_uring_release,
7670 .flush = io_uring_flush,
7671 .mmap = io_uring_mmap,
7673 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7674 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7676 .poll = io_uring_poll,
7677 .fasync = io_uring_fasync,
7678 #ifdef CONFIG_PROC_FS
7679 .show_fdinfo = io_uring_show_fdinfo,
7683 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7684 struct io_uring_params *p)
7686 struct io_rings *rings;
7687 size_t size, sq_array_offset;
7689 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7690 if (size == SIZE_MAX)
7693 rings = io_mem_alloc(size);
7698 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7699 rings->sq_ring_mask = p->sq_entries - 1;
7700 rings->cq_ring_mask = p->cq_entries - 1;
7701 rings->sq_ring_entries = p->sq_entries;
7702 rings->cq_ring_entries = p->cq_entries;
7703 ctx->sq_mask = rings->sq_ring_mask;
7704 ctx->cq_mask = rings->cq_ring_mask;
7705 ctx->sq_entries = rings->sq_ring_entries;
7706 ctx->cq_entries = rings->cq_ring_entries;
7708 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7709 if (size == SIZE_MAX) {
7710 io_mem_free(ctx->rings);
7715 ctx->sq_sqes = io_mem_alloc(size);
7716 if (!ctx->sq_sqes) {
7717 io_mem_free(ctx->rings);
7726 * Allocate an anonymous fd, this is what constitutes the application
7727 * visible backing of an io_uring instance. The application mmaps this
7728 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7729 * we have to tie this fd to a socket for file garbage collection purposes.
7731 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7736 #if defined(CONFIG_UNIX)
7737 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7743 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7747 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7748 O_RDWR | O_CLOEXEC);
7751 ret = PTR_ERR(file);
7755 #if defined(CONFIG_UNIX)
7756 ctx->ring_sock->file = file;
7758 fd_install(ret, file);
7761 #if defined(CONFIG_UNIX)
7762 sock_release(ctx->ring_sock);
7763 ctx->ring_sock = NULL;
7768 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7770 struct user_struct *user = NULL;
7771 struct io_ring_ctx *ctx;
7777 if (entries > IORING_MAX_ENTRIES) {
7778 if (!(p->flags & IORING_SETUP_CLAMP))
7780 entries = IORING_MAX_ENTRIES;
7784 * Use twice as many entries for the CQ ring. It's possible for the
7785 * application to drive a higher depth than the size of the SQ ring,
7786 * since the sqes are only used at submission time. This allows for
7787 * some flexibility in overcommitting a bit. If the application has
7788 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7789 * of CQ ring entries manually.
7791 p->sq_entries = roundup_pow_of_two(entries);
7792 if (p->flags & IORING_SETUP_CQSIZE) {
7794 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7795 * to a power-of-two, if it isn't already. We do NOT impose
7796 * any cq vs sq ring sizing.
7798 if (p->cq_entries < p->sq_entries)
7800 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7801 if (!(p->flags & IORING_SETUP_CLAMP))
7803 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7805 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7807 p->cq_entries = 2 * p->sq_entries;
7810 user = get_uid(current_user());
7811 account_mem = !capable(CAP_IPC_LOCK);
7814 ret = io_account_mem(user,
7815 ring_pages(p->sq_entries, p->cq_entries));
7822 ctx = io_ring_ctx_alloc(p);
7825 io_unaccount_mem(user, ring_pages(p->sq_entries,
7830 ctx->compat = in_compat_syscall();
7831 ctx->account_mem = account_mem;
7833 ctx->creds = get_current_cred();
7835 ret = io_allocate_scq_urings(ctx, p);
7839 ret = io_sq_offload_start(ctx, p);
7843 memset(&p->sq_off, 0, sizeof(p->sq_off));
7844 p->sq_off.head = offsetof(struct io_rings, sq.head);
7845 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7846 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7847 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7848 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7849 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7850 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7852 memset(&p->cq_off, 0, sizeof(p->cq_off));
7853 p->cq_off.head = offsetof(struct io_rings, cq.head);
7854 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7855 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7856 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7857 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7858 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7861 * Install ring fd as the very last thing, so we don't risk someone
7862 * having closed it before we finish setup
7864 ret = io_uring_get_fd(ctx);
7868 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7869 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7870 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7871 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7874 io_ring_ctx_wait_and_kill(ctx);
7879 * Sets up an aio uring context, and returns the fd. Applications asks for a
7880 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7881 * params structure passed in.
7883 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7885 struct io_uring_params p;
7889 if (copy_from_user(&p, params, sizeof(p)))
7891 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7896 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7897 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7898 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7901 ret = io_uring_create(entries, &p);
7905 if (copy_to_user(params, &p, sizeof(p)))
7911 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7912 struct io_uring_params __user *, params)
7914 return io_uring_setup(entries, params);
7917 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7919 struct io_uring_probe *p;
7923 size = struct_size(p, ops, nr_args);
7924 if (size == SIZE_MAX)
7926 p = kzalloc(size, GFP_KERNEL);
7931 if (copy_from_user(p, arg, size))
7934 if (memchr_inv(p, 0, size))
7937 p->last_op = IORING_OP_LAST - 1;
7938 if (nr_args > IORING_OP_LAST)
7939 nr_args = IORING_OP_LAST;
7941 for (i = 0; i < nr_args; i++) {
7943 if (!io_op_defs[i].not_supported)
7944 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7949 if (copy_to_user(arg, p, size))
7956 static int io_register_personality(struct io_ring_ctx *ctx)
7958 const struct cred *creds = get_current_cred();
7961 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7962 USHRT_MAX, GFP_KERNEL);
7968 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7970 const struct cred *old_creds;
7972 old_creds = idr_remove(&ctx->personality_idr, id);
7974 put_cred(old_creds);
7981 static bool io_register_op_must_quiesce(int op)
7984 case IORING_UNREGISTER_FILES:
7985 case IORING_REGISTER_FILES_UPDATE:
7986 case IORING_REGISTER_PROBE:
7987 case IORING_REGISTER_PERSONALITY:
7988 case IORING_UNREGISTER_PERSONALITY:
7995 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7996 void __user *arg, unsigned nr_args)
7997 __releases(ctx->uring_lock)
7998 __acquires(ctx->uring_lock)
8003 * We're inside the ring mutex, if the ref is already dying, then
8004 * someone else killed the ctx or is already going through
8005 * io_uring_register().
8007 if (percpu_ref_is_dying(&ctx->refs))
8010 if (io_register_op_must_quiesce(opcode)) {
8011 percpu_ref_kill(&ctx->refs);
8014 * Drop uring mutex before waiting for references to exit. If
8015 * another thread is currently inside io_uring_enter() it might
8016 * need to grab the uring_lock to make progress. If we hold it
8017 * here across the drain wait, then we can deadlock. It's safe
8018 * to drop the mutex here, since no new references will come in
8019 * after we've killed the percpu ref.
8021 mutex_unlock(&ctx->uring_lock);
8022 ret = wait_for_completion_interruptible(&ctx->completions[0]);
8023 mutex_lock(&ctx->uring_lock);
8025 percpu_ref_resurrect(&ctx->refs);
8032 case IORING_REGISTER_BUFFERS:
8033 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8035 case IORING_UNREGISTER_BUFFERS:
8039 ret = io_sqe_buffer_unregister(ctx);
8041 case IORING_REGISTER_FILES:
8042 ret = io_sqe_files_register(ctx, arg, nr_args);
8044 case IORING_UNREGISTER_FILES:
8048 ret = io_sqe_files_unregister(ctx);
8050 case IORING_REGISTER_FILES_UPDATE:
8051 ret = io_sqe_files_update(ctx, arg, nr_args);
8053 case IORING_REGISTER_EVENTFD:
8054 case IORING_REGISTER_EVENTFD_ASYNC:
8058 ret = io_eventfd_register(ctx, arg);
8061 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8062 ctx->eventfd_async = 1;
8064 ctx->eventfd_async = 0;
8066 case IORING_UNREGISTER_EVENTFD:
8070 ret = io_eventfd_unregister(ctx);
8072 case IORING_REGISTER_PROBE:
8074 if (!arg || nr_args > 256)
8076 ret = io_probe(ctx, arg, nr_args);
8078 case IORING_REGISTER_PERSONALITY:
8082 ret = io_register_personality(ctx);
8084 case IORING_UNREGISTER_PERSONALITY:
8088 ret = io_unregister_personality(ctx, nr_args);
8095 if (io_register_op_must_quiesce(opcode)) {
8096 /* bring the ctx back to life */
8097 percpu_ref_reinit(&ctx->refs);
8099 reinit_completion(&ctx->completions[0]);
8104 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8105 void __user *, arg, unsigned int, nr_args)
8107 struct io_ring_ctx *ctx;
8116 if (f.file->f_op != &io_uring_fops)
8119 ctx = f.file->private_data;
8121 mutex_lock(&ctx->uring_lock);
8122 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8123 mutex_unlock(&ctx->uring_lock);
8124 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8125 ctx->cq_ev_fd != NULL, ret);
8131 static int __init io_uring_init(void)
8133 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8134 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8135 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8138 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8139 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8140 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8141 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8142 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8143 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8144 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8145 BUILD_BUG_SQE_ELEM(8, __u64, off);
8146 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8147 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8148 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8149 BUILD_BUG_SQE_ELEM(24, __u32, len);
8150 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8151 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8152 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8153 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8154 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8155 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8156 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8157 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8158 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8159 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8160 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8161 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8162 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8163 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8164 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8165 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8166 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8167 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8169 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8170 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8171 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8174 __initcall(io_uring_init);