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;
365 struct sockaddr __user *addr;
366 int __user *addr_len;
368 unsigned long nofile;
392 /* NOTE: kiocb has the file as the first member, so don't do it here */
400 struct sockaddr __user *addr;
407 struct user_msghdr __user *msg;
413 struct io_buffer *kbuf;
422 struct filename *filename;
423 struct statx __user *buffer;
425 unsigned long nofile;
428 struct io_files_update {
454 struct epoll_event event;
458 struct file *file_out;
459 struct file *file_in;
466 struct io_provide_buf {
475 struct io_async_connect {
476 struct sockaddr_storage address;
479 struct io_async_msghdr {
480 struct iovec fast_iov[UIO_FASTIOV];
482 struct sockaddr __user *uaddr;
484 struct sockaddr_storage addr;
488 struct iovec fast_iov[UIO_FASTIOV];
494 struct io_async_ctx {
496 struct io_async_rw rw;
497 struct io_async_msghdr msg;
498 struct io_async_connect connect;
499 struct io_timeout_data timeout;
504 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
505 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
506 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
507 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
508 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
509 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
516 REQ_F_IOPOLL_COMPLETED_BIT,
517 REQ_F_LINK_TIMEOUT_BIT,
521 REQ_F_TIMEOUT_NOSEQ_BIT,
522 REQ_F_COMP_LOCKED_BIT,
523 REQ_F_NEED_CLEANUP_BIT,
526 REQ_F_BUFFER_SELECTED_BIT,
528 /* not a real bit, just to check we're not overflowing the space */
534 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
535 /* drain existing IO first */
536 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
538 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
539 /* doesn't sever on completion < 0 */
540 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
542 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
543 /* IOSQE_BUFFER_SELECT */
544 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
546 /* already grabbed next link */
547 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
548 /* fail rest of links */
549 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
550 /* on inflight list */
551 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
552 /* read/write uses file position */
553 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
554 /* must not punt to workers */
555 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
556 /* polled IO has completed */
557 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
558 /* has linked timeout */
559 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
560 /* timeout request */
561 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
563 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
564 /* must be punted even for NONBLOCK */
565 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
566 /* no timeout sequence */
567 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
568 /* completion under lock */
569 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
571 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
572 /* in overflow list */
573 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
574 /* already went through poll handler */
575 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
576 /* buffer already selected */
577 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
581 struct io_poll_iocb poll;
582 struct io_wq_work work;
586 * NOTE! Each of the iocb union members has the file pointer
587 * as the first entry in their struct definition. So you can
588 * access the file pointer through any of the sub-structs,
589 * or directly as just 'ki_filp' in this struct.
595 struct io_poll_iocb poll;
596 struct io_accept accept;
598 struct io_cancel cancel;
599 struct io_timeout timeout;
600 struct io_connect connect;
601 struct io_sr_msg sr_msg;
603 struct io_close close;
604 struct io_files_update files_update;
605 struct io_fadvise fadvise;
606 struct io_madvise madvise;
607 struct io_epoll epoll;
608 struct io_splice splice;
609 struct io_provide_buf pbuf;
612 struct io_async_ctx *io;
614 bool needs_fixed_file;
617 struct io_ring_ctx *ctx;
618 struct list_head list;
621 struct task_struct *task;
627 struct list_head link_list;
629 struct list_head inflight_entry;
631 struct percpu_ref *fixed_file_refs;
635 * Only commands that never go async can use the below fields,
636 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
637 * async armed poll handlers for regular commands. The latter
638 * restore the work, if needed.
641 struct callback_head task_work;
642 struct hlist_node hash_node;
643 struct async_poll *apoll;
645 struct io_wq_work work;
649 #define IO_PLUG_THRESHOLD 2
650 #define IO_IOPOLL_BATCH 8
652 struct io_submit_state {
653 struct blk_plug plug;
656 * io_kiocb alloc cache
658 void *reqs[IO_IOPOLL_BATCH];
659 unsigned int free_reqs;
662 * File reference cache
666 unsigned int has_refs;
667 unsigned int used_refs;
668 unsigned int ios_left;
672 /* needs req->io allocated for deferral/async */
673 unsigned async_ctx : 1;
674 /* needs current->mm setup, does mm access */
675 unsigned needs_mm : 1;
676 /* needs req->file assigned */
677 unsigned needs_file : 1;
678 /* needs req->file assigned IFF fd is >= 0 */
679 unsigned fd_non_neg : 1;
680 /* hash wq insertion if file is a regular file */
681 unsigned hash_reg_file : 1;
682 /* unbound wq insertion if file is a non-regular file */
683 unsigned unbound_nonreg_file : 1;
684 /* opcode is not supported by this kernel */
685 unsigned not_supported : 1;
686 /* needs file table */
687 unsigned file_table : 1;
689 unsigned needs_fs : 1;
690 /* set if opcode supports polled "wait" */
692 unsigned pollout : 1;
693 /* op supports buffer selection */
694 unsigned buffer_select : 1;
697 static const struct io_op_def io_op_defs[] = {
698 [IORING_OP_NOP] = {},
699 [IORING_OP_READV] = {
703 .unbound_nonreg_file = 1,
707 [IORING_OP_WRITEV] = {
712 .unbound_nonreg_file = 1,
715 [IORING_OP_FSYNC] = {
718 [IORING_OP_READ_FIXED] = {
720 .unbound_nonreg_file = 1,
723 [IORING_OP_WRITE_FIXED] = {
726 .unbound_nonreg_file = 1,
729 [IORING_OP_POLL_ADD] = {
731 .unbound_nonreg_file = 1,
733 [IORING_OP_POLL_REMOVE] = {},
734 [IORING_OP_SYNC_FILE_RANGE] = {
737 [IORING_OP_SENDMSG] = {
741 .unbound_nonreg_file = 1,
745 [IORING_OP_RECVMSG] = {
749 .unbound_nonreg_file = 1,
754 [IORING_OP_TIMEOUT] = {
758 [IORING_OP_TIMEOUT_REMOVE] = {},
759 [IORING_OP_ACCEPT] = {
762 .unbound_nonreg_file = 1,
766 [IORING_OP_ASYNC_CANCEL] = {},
767 [IORING_OP_LINK_TIMEOUT] = {
771 [IORING_OP_CONNECT] = {
775 .unbound_nonreg_file = 1,
778 [IORING_OP_FALLOCATE] = {
781 [IORING_OP_OPENAT] = {
787 [IORING_OP_CLOSE] = {
791 [IORING_OP_FILES_UPDATE] = {
795 [IORING_OP_STATX] = {
804 .unbound_nonreg_file = 1,
808 [IORING_OP_WRITE] = {
811 .unbound_nonreg_file = 1,
814 [IORING_OP_FADVISE] = {
817 [IORING_OP_MADVISE] = {
823 .unbound_nonreg_file = 1,
829 .unbound_nonreg_file = 1,
833 [IORING_OP_OPENAT2] = {
839 [IORING_OP_EPOLL_CTL] = {
840 .unbound_nonreg_file = 1,
843 [IORING_OP_SPLICE] = {
846 .unbound_nonreg_file = 1,
848 [IORING_OP_PROVIDE_BUFFERS] = {},
849 [IORING_OP_REMOVE_BUFFERS] = {},
852 static void io_wq_submit_work(struct io_wq_work **workptr);
853 static void io_cqring_fill_event(struct io_kiocb *req, long res);
854 static void io_put_req(struct io_kiocb *req);
855 static void __io_double_put_req(struct io_kiocb *req);
856 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
857 static void io_queue_linked_timeout(struct io_kiocb *req);
858 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
859 struct io_uring_files_update *ip,
861 static int io_grab_files(struct io_kiocb *req);
862 static void io_cleanup_req(struct io_kiocb *req);
863 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
864 int fd, struct file **out_file, bool fixed);
865 static void __io_queue_sqe(struct io_kiocb *req,
866 const struct io_uring_sqe *sqe);
868 static struct kmem_cache *req_cachep;
870 static const struct file_operations io_uring_fops;
872 struct sock *io_uring_get_socket(struct file *file)
874 #if defined(CONFIG_UNIX)
875 if (file->f_op == &io_uring_fops) {
876 struct io_ring_ctx *ctx = file->private_data;
878 return ctx->ring_sock->sk;
883 EXPORT_SYMBOL(io_uring_get_socket);
885 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
887 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
889 complete(&ctx->completions[0]);
892 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
894 struct io_ring_ctx *ctx;
897 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
901 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
902 if (!ctx->fallback_req)
905 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
906 if (!ctx->completions)
910 * Use 5 bits less than the max cq entries, that should give us around
911 * 32 entries per hash list if totally full and uniformly spread.
913 hash_bits = ilog2(p->cq_entries);
917 ctx->cancel_hash_bits = hash_bits;
918 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
920 if (!ctx->cancel_hash)
922 __hash_init(ctx->cancel_hash, 1U << hash_bits);
924 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
925 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
928 ctx->flags = p->flags;
929 init_waitqueue_head(&ctx->cq_wait);
930 INIT_LIST_HEAD(&ctx->cq_overflow_list);
931 init_completion(&ctx->completions[0]);
932 init_completion(&ctx->completions[1]);
933 idr_init(&ctx->io_buffer_idr);
934 idr_init(&ctx->personality_idr);
935 mutex_init(&ctx->uring_lock);
936 init_waitqueue_head(&ctx->wait);
937 spin_lock_init(&ctx->completion_lock);
938 INIT_LIST_HEAD(&ctx->poll_list);
939 INIT_LIST_HEAD(&ctx->defer_list);
940 INIT_LIST_HEAD(&ctx->timeout_list);
941 init_waitqueue_head(&ctx->inflight_wait);
942 spin_lock_init(&ctx->inflight_lock);
943 INIT_LIST_HEAD(&ctx->inflight_list);
946 if (ctx->fallback_req)
947 kmem_cache_free(req_cachep, ctx->fallback_req);
948 kfree(ctx->completions);
949 kfree(ctx->cancel_hash);
954 static inline bool __req_need_defer(struct io_kiocb *req)
956 struct io_ring_ctx *ctx = req->ctx;
958 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
959 + atomic_read(&ctx->cached_cq_overflow);
962 static inline bool req_need_defer(struct io_kiocb *req)
964 if (unlikely(req->flags & REQ_F_IO_DRAIN))
965 return __req_need_defer(req);
970 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
972 struct io_kiocb *req;
974 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
975 if (req && !req_need_defer(req)) {
976 list_del_init(&req->list);
983 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
985 struct io_kiocb *req;
987 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
989 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
991 if (!__req_need_defer(req)) {
992 list_del_init(&req->list);
1000 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1002 struct io_rings *rings = ctx->rings;
1004 /* order cqe stores with ring update */
1005 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1007 if (wq_has_sleeper(&ctx->cq_wait)) {
1008 wake_up_interruptible(&ctx->cq_wait);
1009 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1013 static inline void io_req_work_grab_env(struct io_kiocb *req,
1014 const struct io_op_def *def)
1016 if (!req->work.mm && def->needs_mm) {
1017 mmgrab(current->mm);
1018 req->work.mm = current->mm;
1020 if (!req->work.creds)
1021 req->work.creds = get_current_cred();
1022 if (!req->work.fs && def->needs_fs) {
1023 spin_lock(¤t->fs->lock);
1024 if (!current->fs->in_exec) {
1025 req->work.fs = current->fs;
1026 req->work.fs->users++;
1028 req->work.flags |= IO_WQ_WORK_CANCEL;
1030 spin_unlock(¤t->fs->lock);
1032 if (!req->work.task_pid)
1033 req->work.task_pid = task_pid_vnr(current);
1036 static inline void io_req_work_drop_env(struct io_kiocb *req)
1039 mmdrop(req->work.mm);
1040 req->work.mm = NULL;
1042 if (req->work.creds) {
1043 put_cred(req->work.creds);
1044 req->work.creds = NULL;
1047 struct fs_struct *fs = req->work.fs;
1049 spin_lock(&req->work.fs->lock);
1052 spin_unlock(&req->work.fs->lock);
1058 static inline void io_prep_async_work(struct io_kiocb *req,
1059 struct io_kiocb **link)
1061 const struct io_op_def *def = &io_op_defs[req->opcode];
1063 if (req->flags & REQ_F_ISREG) {
1064 if (def->hash_reg_file)
1065 io_wq_hash_work(&req->work, file_inode(req->file));
1067 if (def->unbound_nonreg_file)
1068 req->work.flags |= IO_WQ_WORK_UNBOUND;
1071 io_req_work_grab_env(req, def);
1073 *link = io_prep_linked_timeout(req);
1076 static inline void io_queue_async_work(struct io_kiocb *req)
1078 struct io_ring_ctx *ctx = req->ctx;
1079 struct io_kiocb *link;
1081 io_prep_async_work(req, &link);
1083 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1084 &req->work, req->flags);
1085 io_wq_enqueue(ctx->io_wq, &req->work);
1088 io_queue_linked_timeout(link);
1091 static void io_kill_timeout(struct io_kiocb *req)
1095 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1097 atomic_inc(&req->ctx->cq_timeouts);
1098 list_del_init(&req->list);
1099 req->flags |= REQ_F_COMP_LOCKED;
1100 io_cqring_fill_event(req, 0);
1105 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1107 struct io_kiocb *req, *tmp;
1109 spin_lock_irq(&ctx->completion_lock);
1110 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1111 io_kill_timeout(req);
1112 spin_unlock_irq(&ctx->completion_lock);
1115 static void io_commit_cqring(struct io_ring_ctx *ctx)
1117 struct io_kiocb *req;
1119 while ((req = io_get_timeout_req(ctx)) != NULL)
1120 io_kill_timeout(req);
1122 __io_commit_cqring(ctx);
1124 while ((req = io_get_deferred_req(ctx)) != NULL)
1125 io_queue_async_work(req);
1128 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1130 struct io_rings *rings = ctx->rings;
1133 tail = ctx->cached_cq_tail;
1135 * writes to the cq entry need to come after reading head; the
1136 * control dependency is enough as we're using WRITE_ONCE to
1139 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1142 ctx->cached_cq_tail++;
1143 return &rings->cqes[tail & ctx->cq_mask];
1146 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1150 if (!ctx->eventfd_async)
1152 return io_wq_current_is_worker();
1155 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1157 if (waitqueue_active(&ctx->wait))
1158 wake_up(&ctx->wait);
1159 if (waitqueue_active(&ctx->sqo_wait))
1160 wake_up(&ctx->sqo_wait);
1161 if (io_should_trigger_evfd(ctx))
1162 eventfd_signal(ctx->cq_ev_fd, 1);
1165 /* Returns true if there are no backlogged entries after the flush */
1166 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1168 struct io_rings *rings = ctx->rings;
1169 struct io_uring_cqe *cqe;
1170 struct io_kiocb *req;
1171 unsigned long flags;
1175 if (list_empty_careful(&ctx->cq_overflow_list))
1177 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1178 rings->cq_ring_entries))
1182 spin_lock_irqsave(&ctx->completion_lock, flags);
1184 /* if force is set, the ring is going away. always drop after that */
1186 ctx->cq_overflow_flushed = 1;
1189 while (!list_empty(&ctx->cq_overflow_list)) {
1190 cqe = io_get_cqring(ctx);
1194 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1196 list_move(&req->list, &list);
1197 req->flags &= ~REQ_F_OVERFLOW;
1199 WRITE_ONCE(cqe->user_data, req->user_data);
1200 WRITE_ONCE(cqe->res, req->result);
1201 WRITE_ONCE(cqe->flags, req->cflags);
1203 WRITE_ONCE(ctx->rings->cq_overflow,
1204 atomic_inc_return(&ctx->cached_cq_overflow));
1208 io_commit_cqring(ctx);
1210 clear_bit(0, &ctx->sq_check_overflow);
1211 clear_bit(0, &ctx->cq_check_overflow);
1213 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1214 io_cqring_ev_posted(ctx);
1216 while (!list_empty(&list)) {
1217 req = list_first_entry(&list, struct io_kiocb, list);
1218 list_del(&req->list);
1225 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1227 struct io_ring_ctx *ctx = req->ctx;
1228 struct io_uring_cqe *cqe;
1230 trace_io_uring_complete(ctx, req->user_data, res);
1233 * If we can't get a cq entry, userspace overflowed the
1234 * submission (by quite a lot). Increment the overflow count in
1237 cqe = io_get_cqring(ctx);
1239 WRITE_ONCE(cqe->user_data, req->user_data);
1240 WRITE_ONCE(cqe->res, res);
1241 WRITE_ONCE(cqe->flags, cflags);
1242 } else if (ctx->cq_overflow_flushed) {
1243 WRITE_ONCE(ctx->rings->cq_overflow,
1244 atomic_inc_return(&ctx->cached_cq_overflow));
1246 if (list_empty(&ctx->cq_overflow_list)) {
1247 set_bit(0, &ctx->sq_check_overflow);
1248 set_bit(0, &ctx->cq_check_overflow);
1250 req->flags |= REQ_F_OVERFLOW;
1251 refcount_inc(&req->refs);
1253 req->cflags = cflags;
1254 list_add_tail(&req->list, &ctx->cq_overflow_list);
1258 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1260 __io_cqring_fill_event(req, res, 0);
1263 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1265 struct io_ring_ctx *ctx = req->ctx;
1266 unsigned long flags;
1268 spin_lock_irqsave(&ctx->completion_lock, flags);
1269 __io_cqring_fill_event(req, res, cflags);
1270 io_commit_cqring(ctx);
1271 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1273 io_cqring_ev_posted(ctx);
1276 static void io_cqring_add_event(struct io_kiocb *req, long res)
1278 __io_cqring_add_event(req, res, 0);
1281 static inline bool io_is_fallback_req(struct io_kiocb *req)
1283 return req == (struct io_kiocb *)
1284 ((unsigned long) req->ctx->fallback_req & ~1UL);
1287 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1289 struct io_kiocb *req;
1291 req = ctx->fallback_req;
1292 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1298 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1299 struct io_submit_state *state)
1301 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1302 struct io_kiocb *req;
1305 req = kmem_cache_alloc(req_cachep, gfp);
1308 } else if (!state->free_reqs) {
1312 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1313 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1316 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1317 * retry single alloc to be on the safe side.
1319 if (unlikely(ret <= 0)) {
1320 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1321 if (!state->reqs[0])
1325 state->free_reqs = ret - 1;
1326 req = state->reqs[ret - 1];
1329 req = state->reqs[state->free_reqs];
1334 return io_get_fallback_req(ctx);
1337 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1341 percpu_ref_put(req->fixed_file_refs);
1346 static void __io_req_aux_free(struct io_kiocb *req)
1348 if (req->flags & REQ_F_NEED_CLEANUP)
1349 io_cleanup_req(req);
1353 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1355 put_task_struct(req->task);
1357 io_req_work_drop_env(req);
1360 static void __io_free_req(struct io_kiocb *req)
1362 __io_req_aux_free(req);
1364 if (req->flags & REQ_F_INFLIGHT) {
1365 struct io_ring_ctx *ctx = req->ctx;
1366 unsigned long flags;
1368 spin_lock_irqsave(&ctx->inflight_lock, flags);
1369 list_del(&req->inflight_entry);
1370 if (waitqueue_active(&ctx->inflight_wait))
1371 wake_up(&ctx->inflight_wait);
1372 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1375 percpu_ref_put(&req->ctx->refs);
1376 if (likely(!io_is_fallback_req(req)))
1377 kmem_cache_free(req_cachep, req);
1379 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1383 void *reqs[IO_IOPOLL_BATCH];
1388 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1392 if (rb->need_iter) {
1393 int i, inflight = 0;
1394 unsigned long flags;
1396 for (i = 0; i < rb->to_free; i++) {
1397 struct io_kiocb *req = rb->reqs[i];
1399 if (req->flags & REQ_F_FIXED_FILE) {
1401 percpu_ref_put(req->fixed_file_refs);
1403 if (req->flags & REQ_F_INFLIGHT)
1405 __io_req_aux_free(req);
1410 spin_lock_irqsave(&ctx->inflight_lock, flags);
1411 for (i = 0; i < rb->to_free; i++) {
1412 struct io_kiocb *req = rb->reqs[i];
1414 if (req->flags & REQ_F_INFLIGHT) {
1415 list_del(&req->inflight_entry);
1420 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1422 if (waitqueue_active(&ctx->inflight_wait))
1423 wake_up(&ctx->inflight_wait);
1426 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1427 percpu_ref_put_many(&ctx->refs, rb->to_free);
1428 rb->to_free = rb->need_iter = 0;
1431 static bool io_link_cancel_timeout(struct io_kiocb *req)
1433 struct io_ring_ctx *ctx = req->ctx;
1436 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1438 io_cqring_fill_event(req, -ECANCELED);
1439 io_commit_cqring(ctx);
1440 req->flags &= ~REQ_F_LINK;
1448 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1450 struct io_ring_ctx *ctx = req->ctx;
1451 bool wake_ev = false;
1453 /* Already got next link */
1454 if (req->flags & REQ_F_LINK_NEXT)
1458 * The list should never be empty when we are called here. But could
1459 * potentially happen if the chain is messed up, check to be on the
1462 while (!list_empty(&req->link_list)) {
1463 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1464 struct io_kiocb, link_list);
1466 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1467 (nxt->flags & REQ_F_TIMEOUT))) {
1468 list_del_init(&nxt->link_list);
1469 wake_ev |= io_link_cancel_timeout(nxt);
1470 req->flags &= ~REQ_F_LINK_TIMEOUT;
1474 list_del_init(&req->link_list);
1475 if (!list_empty(&nxt->link_list))
1476 nxt->flags |= REQ_F_LINK;
1481 req->flags |= REQ_F_LINK_NEXT;
1483 io_cqring_ev_posted(ctx);
1487 * Called if REQ_F_LINK is set, and we fail the head request
1489 static void io_fail_links(struct io_kiocb *req)
1491 struct io_ring_ctx *ctx = req->ctx;
1492 unsigned long flags;
1494 spin_lock_irqsave(&ctx->completion_lock, flags);
1496 while (!list_empty(&req->link_list)) {
1497 struct io_kiocb *link = list_first_entry(&req->link_list,
1498 struct io_kiocb, link_list);
1500 list_del_init(&link->link_list);
1501 trace_io_uring_fail_link(req, link);
1503 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1504 link->opcode == IORING_OP_LINK_TIMEOUT) {
1505 io_link_cancel_timeout(link);
1507 io_cqring_fill_event(link, -ECANCELED);
1508 __io_double_put_req(link);
1510 req->flags &= ~REQ_F_LINK_TIMEOUT;
1513 io_commit_cqring(ctx);
1514 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1515 io_cqring_ev_posted(ctx);
1518 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1520 if (likely(!(req->flags & REQ_F_LINK)))
1524 * If LINK is set, we have dependent requests in this chain. If we
1525 * didn't fail this request, queue the first one up, moving any other
1526 * dependencies to the next request. In case of failure, fail the rest
1529 if (req->flags & REQ_F_FAIL_LINK) {
1531 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1532 REQ_F_LINK_TIMEOUT) {
1533 struct io_ring_ctx *ctx = req->ctx;
1534 unsigned long flags;
1537 * If this is a timeout link, we could be racing with the
1538 * timeout timer. Grab the completion lock for this case to
1539 * protect against that.
1541 spin_lock_irqsave(&ctx->completion_lock, flags);
1542 io_req_link_next(req, nxt);
1543 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1545 io_req_link_next(req, nxt);
1549 static void io_free_req(struct io_kiocb *req)
1551 struct io_kiocb *nxt = NULL;
1553 io_req_find_next(req, &nxt);
1557 io_queue_async_work(nxt);
1560 static void io_link_work_cb(struct io_wq_work **workptr)
1562 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1563 struct io_kiocb *link;
1565 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1566 io_queue_linked_timeout(link);
1567 io_wq_submit_work(workptr);
1570 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1572 struct io_kiocb *link;
1573 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1575 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1576 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1578 *workptr = &nxt->work;
1579 link = io_prep_linked_timeout(nxt);
1581 nxt->work.func = io_link_work_cb;
1585 * Drop reference to request, return next in chain (if there is one) if this
1586 * was the last reference to this request.
1588 __attribute__((nonnull))
1589 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1591 if (refcount_dec_and_test(&req->refs)) {
1592 io_req_find_next(req, nxtptr);
1597 static void io_put_req(struct io_kiocb *req)
1599 if (refcount_dec_and_test(&req->refs))
1603 static void io_steal_work(struct io_kiocb *req,
1604 struct io_wq_work **workptr)
1607 * It's in an io-wq worker, so there always should be at least
1608 * one reference, which will be dropped in io_put_work() just
1609 * after the current handler returns.
1611 * It also means, that if the counter dropped to 1, then there is
1612 * no asynchronous users left, so it's safe to steal the next work.
1614 if (refcount_read(&req->refs) == 1) {
1615 struct io_kiocb *nxt = NULL;
1617 io_req_find_next(req, &nxt);
1619 io_wq_assign_next(workptr, nxt);
1624 * Must only be used if we don't need to care about links, usually from
1625 * within the completion handling itself.
1627 static void __io_double_put_req(struct io_kiocb *req)
1629 /* drop both submit and complete references */
1630 if (refcount_sub_and_test(2, &req->refs))
1634 static void io_double_put_req(struct io_kiocb *req)
1636 /* drop both submit and complete references */
1637 if (refcount_sub_and_test(2, &req->refs))
1641 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1643 struct io_rings *rings = ctx->rings;
1645 if (test_bit(0, &ctx->cq_check_overflow)) {
1647 * noflush == true is from the waitqueue handler, just ensure
1648 * we wake up the task, and the next invocation will flush the
1649 * entries. We cannot safely to it from here.
1651 if (noflush && !list_empty(&ctx->cq_overflow_list))
1654 io_cqring_overflow_flush(ctx, false);
1657 /* See comment at the top of this file */
1659 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1662 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1664 struct io_rings *rings = ctx->rings;
1666 /* make sure SQ entry isn't read before tail */
1667 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1670 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1672 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1675 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1678 rb->reqs[rb->to_free++] = req;
1679 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1680 io_free_req_many(req->ctx, rb);
1684 static int io_put_kbuf(struct io_kiocb *req)
1686 struct io_buffer *kbuf;
1689 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1690 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1691 cflags |= IORING_CQE_F_BUFFER;
1698 * Find and free completed poll iocbs
1700 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1701 struct list_head *done)
1703 struct req_batch rb;
1704 struct io_kiocb *req;
1706 rb.to_free = rb.need_iter = 0;
1707 while (!list_empty(done)) {
1710 req = list_first_entry(done, struct io_kiocb, list);
1711 list_del(&req->list);
1713 if (req->flags & REQ_F_BUFFER_SELECTED)
1714 cflags = io_put_kbuf(req);
1716 __io_cqring_fill_event(req, req->result, cflags);
1719 if (refcount_dec_and_test(&req->refs) &&
1720 !io_req_multi_free(&rb, req))
1724 io_commit_cqring(ctx);
1725 if (ctx->flags & IORING_SETUP_SQPOLL)
1726 io_cqring_ev_posted(ctx);
1727 io_free_req_many(ctx, &rb);
1730 static void io_iopoll_queue(struct list_head *again)
1732 struct io_kiocb *req;
1735 req = list_first_entry(again, struct io_kiocb, list);
1736 list_del(&req->list);
1737 refcount_inc(&req->refs);
1738 io_queue_async_work(req);
1739 } while (!list_empty(again));
1742 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1745 struct io_kiocb *req, *tmp;
1752 * Only spin for completions if we don't have multiple devices hanging
1753 * off our complete list, and we're under the requested amount.
1755 spin = !ctx->poll_multi_file && *nr_events < min;
1758 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1759 struct kiocb *kiocb = &req->rw.kiocb;
1762 * Move completed and retryable entries to our local lists.
1763 * If we find a request that requires polling, break out
1764 * and complete those lists first, if we have entries there.
1766 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1767 list_move_tail(&req->list, &done);
1770 if (!list_empty(&done))
1773 if (req->result == -EAGAIN) {
1774 list_move_tail(&req->list, &again);
1777 if (!list_empty(&again))
1780 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1789 if (!list_empty(&done))
1790 io_iopoll_complete(ctx, nr_events, &done);
1792 if (!list_empty(&again))
1793 io_iopoll_queue(&again);
1799 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1800 * non-spinning poll check - we'll still enter the driver poll loop, but only
1801 * as a non-spinning completion check.
1803 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1806 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1809 ret = io_do_iopoll(ctx, nr_events, min);
1812 if (!min || *nr_events >= min)
1820 * We can't just wait for polled events to come to us, we have to actively
1821 * find and complete them.
1823 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1825 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1828 mutex_lock(&ctx->uring_lock);
1829 while (!list_empty(&ctx->poll_list)) {
1830 unsigned int nr_events = 0;
1832 io_iopoll_getevents(ctx, &nr_events, 1);
1835 * Ensure we allow local-to-the-cpu processing to take place,
1836 * in this case we need to ensure that we reap all events.
1840 mutex_unlock(&ctx->uring_lock);
1843 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1846 int iters = 0, ret = 0;
1849 * We disallow the app entering submit/complete with polling, but we
1850 * still need to lock the ring to prevent racing with polled issue
1851 * that got punted to a workqueue.
1853 mutex_lock(&ctx->uring_lock);
1858 * Don't enter poll loop if we already have events pending.
1859 * If we do, we can potentially be spinning for commands that
1860 * already triggered a CQE (eg in error).
1862 if (io_cqring_events(ctx, false))
1866 * If a submit got punted to a workqueue, we can have the
1867 * application entering polling for a command before it gets
1868 * issued. That app will hold the uring_lock for the duration
1869 * of the poll right here, so we need to take a breather every
1870 * now and then to ensure that the issue has a chance to add
1871 * the poll to the issued list. Otherwise we can spin here
1872 * forever, while the workqueue is stuck trying to acquire the
1875 if (!(++iters & 7)) {
1876 mutex_unlock(&ctx->uring_lock);
1877 mutex_lock(&ctx->uring_lock);
1880 if (*nr_events < min)
1881 tmin = min - *nr_events;
1883 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1887 } while (min && !*nr_events && !need_resched());
1889 mutex_unlock(&ctx->uring_lock);
1893 static void kiocb_end_write(struct io_kiocb *req)
1896 * Tell lockdep we inherited freeze protection from submission
1899 if (req->flags & REQ_F_ISREG) {
1900 struct inode *inode = file_inode(req->file);
1902 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1904 file_end_write(req->file);
1907 static inline void req_set_fail_links(struct io_kiocb *req)
1909 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1910 req->flags |= REQ_F_FAIL_LINK;
1913 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1915 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1918 if (kiocb->ki_flags & IOCB_WRITE)
1919 kiocb_end_write(req);
1921 if (res != req->result)
1922 req_set_fail_links(req);
1923 if (req->flags & REQ_F_BUFFER_SELECTED)
1924 cflags = io_put_kbuf(req);
1925 __io_cqring_add_event(req, res, cflags);
1928 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1930 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1932 io_complete_rw_common(kiocb, res);
1936 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1938 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1940 if (kiocb->ki_flags & IOCB_WRITE)
1941 kiocb_end_write(req);
1943 if (res != req->result)
1944 req_set_fail_links(req);
1947 req->flags |= REQ_F_IOPOLL_COMPLETED;
1951 * After the iocb has been issued, it's safe to be found on the poll list.
1952 * Adding the kiocb to the list AFTER submission ensures that we don't
1953 * find it from a io_iopoll_getevents() thread before the issuer is done
1954 * accessing the kiocb cookie.
1956 static void io_iopoll_req_issued(struct io_kiocb *req)
1958 struct io_ring_ctx *ctx = req->ctx;
1961 * Track whether we have multiple files in our lists. This will impact
1962 * how we do polling eventually, not spinning if we're on potentially
1963 * different devices.
1965 if (list_empty(&ctx->poll_list)) {
1966 ctx->poll_multi_file = false;
1967 } else if (!ctx->poll_multi_file) {
1968 struct io_kiocb *list_req;
1970 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1972 if (list_req->file != req->file)
1973 ctx->poll_multi_file = true;
1977 * For fast devices, IO may have already completed. If it has, add
1978 * it to the front so we find it first.
1980 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1981 list_add(&req->list, &ctx->poll_list);
1983 list_add_tail(&req->list, &ctx->poll_list);
1985 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1986 wq_has_sleeper(&ctx->sqo_wait))
1987 wake_up(&ctx->sqo_wait);
1990 static void io_file_put(struct io_submit_state *state)
1993 int diff = state->has_refs - state->used_refs;
1996 fput_many(state->file, diff);
2002 * Get as many references to a file as we have IOs left in this submission,
2003 * assuming most submissions are for one file, or at least that each file
2004 * has more than one submission.
2006 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2012 if (state->fd == fd) {
2019 state->file = fget_many(fd, state->ios_left);
2024 state->has_refs = state->ios_left;
2025 state->used_refs = 1;
2031 * If we tracked the file through the SCM inflight mechanism, we could support
2032 * any file. For now, just ensure that anything potentially problematic is done
2035 static bool io_file_supports_async(struct file *file)
2037 umode_t mode = file_inode(file)->i_mode;
2039 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2041 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2047 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2048 bool force_nonblock)
2050 struct io_ring_ctx *ctx = req->ctx;
2051 struct kiocb *kiocb = &req->rw.kiocb;
2055 if (S_ISREG(file_inode(req->file)->i_mode))
2056 req->flags |= REQ_F_ISREG;
2058 kiocb->ki_pos = READ_ONCE(sqe->off);
2059 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2060 req->flags |= REQ_F_CUR_POS;
2061 kiocb->ki_pos = req->file->f_pos;
2063 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2064 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2065 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2069 ioprio = READ_ONCE(sqe->ioprio);
2071 ret = ioprio_check_cap(ioprio);
2075 kiocb->ki_ioprio = ioprio;
2077 kiocb->ki_ioprio = get_current_ioprio();
2079 /* don't allow async punt if RWF_NOWAIT was requested */
2080 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2081 (req->file->f_flags & O_NONBLOCK))
2082 req->flags |= REQ_F_NOWAIT;
2085 kiocb->ki_flags |= IOCB_NOWAIT;
2087 if (ctx->flags & IORING_SETUP_IOPOLL) {
2088 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2089 !kiocb->ki_filp->f_op->iopoll)
2092 kiocb->ki_flags |= IOCB_HIPRI;
2093 kiocb->ki_complete = io_complete_rw_iopoll;
2096 if (kiocb->ki_flags & IOCB_HIPRI)
2098 kiocb->ki_complete = io_complete_rw;
2101 req->rw.addr = READ_ONCE(sqe->addr);
2102 req->rw.len = READ_ONCE(sqe->len);
2103 /* we own ->private, reuse it for the buffer index / buffer ID */
2104 req->rw.kiocb.private = (void *) (unsigned long)
2105 READ_ONCE(sqe->buf_index);
2109 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2115 case -ERESTARTNOINTR:
2116 case -ERESTARTNOHAND:
2117 case -ERESTART_RESTARTBLOCK:
2119 * We can't just restart the syscall, since previously
2120 * submitted sqes may already be in progress. Just fail this
2126 kiocb->ki_complete(kiocb, ret, 0);
2130 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2132 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2134 if (req->flags & REQ_F_CUR_POS)
2135 req->file->f_pos = kiocb->ki_pos;
2136 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2137 io_complete_rw(kiocb, ret, 0);
2139 io_rw_done(kiocb, ret);
2142 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2143 struct iov_iter *iter)
2145 struct io_ring_ctx *ctx = req->ctx;
2146 size_t len = req->rw.len;
2147 struct io_mapped_ubuf *imu;
2148 unsigned index, buf_index;
2152 /* attempt to use fixed buffers without having provided iovecs */
2153 if (unlikely(!ctx->user_bufs))
2156 buf_index = (unsigned long) req->rw.kiocb.private;
2157 if (unlikely(buf_index >= ctx->nr_user_bufs))
2160 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2161 imu = &ctx->user_bufs[index];
2162 buf_addr = req->rw.addr;
2165 if (buf_addr + len < buf_addr)
2167 /* not inside the mapped region */
2168 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2172 * May not be a start of buffer, set size appropriately
2173 * and advance us to the beginning.
2175 offset = buf_addr - imu->ubuf;
2176 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2180 * Don't use iov_iter_advance() here, as it's really slow for
2181 * using the latter parts of a big fixed buffer - it iterates
2182 * over each segment manually. We can cheat a bit here, because
2185 * 1) it's a BVEC iter, we set it up
2186 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2187 * first and last bvec
2189 * So just find our index, and adjust the iterator afterwards.
2190 * If the offset is within the first bvec (or the whole first
2191 * bvec, just use iov_iter_advance(). This makes it easier
2192 * since we can just skip the first segment, which may not
2193 * be PAGE_SIZE aligned.
2195 const struct bio_vec *bvec = imu->bvec;
2197 if (offset <= bvec->bv_len) {
2198 iov_iter_advance(iter, offset);
2200 unsigned long seg_skip;
2202 /* skip first vec */
2203 offset -= bvec->bv_len;
2204 seg_skip = 1 + (offset >> PAGE_SHIFT);
2206 iter->bvec = bvec + seg_skip;
2207 iter->nr_segs -= seg_skip;
2208 iter->count -= bvec->bv_len + offset;
2209 iter->iov_offset = offset & ~PAGE_MASK;
2216 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2219 mutex_unlock(&ctx->uring_lock);
2222 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2225 * "Normal" inline submissions always hold the uring_lock, since we
2226 * grab it from the system call. Same is true for the SQPOLL offload.
2227 * The only exception is when we've detached the request and issue it
2228 * from an async worker thread, grab the lock for that case.
2231 mutex_lock(&ctx->uring_lock);
2234 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2235 int bgid, struct io_buffer *kbuf,
2238 struct io_buffer *head;
2240 if (req->flags & REQ_F_BUFFER_SELECTED)
2243 io_ring_submit_lock(req->ctx, needs_lock);
2245 lockdep_assert_held(&req->ctx->uring_lock);
2247 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2249 if (!list_empty(&head->list)) {
2250 kbuf = list_last_entry(&head->list, struct io_buffer,
2252 list_del(&kbuf->list);
2255 idr_remove(&req->ctx->io_buffer_idr, bgid);
2257 if (*len > kbuf->len)
2260 kbuf = ERR_PTR(-ENOBUFS);
2263 io_ring_submit_unlock(req->ctx, needs_lock);
2268 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2271 struct io_buffer *kbuf;
2274 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2275 bgid = (int) (unsigned long) req->rw.kiocb.private;
2276 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2279 req->rw.addr = (u64) (unsigned long) kbuf;
2280 req->flags |= REQ_F_BUFFER_SELECTED;
2281 return u64_to_user_ptr(kbuf->addr);
2284 #ifdef CONFIG_COMPAT
2285 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2288 struct compat_iovec __user *uiov;
2289 compat_ssize_t clen;
2293 uiov = u64_to_user_ptr(req->rw.addr);
2294 if (!access_ok(uiov, sizeof(*uiov)))
2296 if (__get_user(clen, &uiov->iov_len))
2302 buf = io_rw_buffer_select(req, &len, needs_lock);
2304 return PTR_ERR(buf);
2305 iov[0].iov_base = buf;
2306 iov[0].iov_len = (compat_size_t) len;
2311 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2314 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2318 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2321 len = iov[0].iov_len;
2324 buf = io_rw_buffer_select(req, &len, needs_lock);
2326 return PTR_ERR(buf);
2327 iov[0].iov_base = buf;
2328 iov[0].iov_len = len;
2332 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2335 if (req->flags & REQ_F_BUFFER_SELECTED)
2339 else if (req->rw.len > 1)
2342 #ifdef CONFIG_COMPAT
2343 if (req->ctx->compat)
2344 return io_compat_import(req, iov, needs_lock);
2347 return __io_iov_buffer_select(req, iov, needs_lock);
2350 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2351 struct iovec **iovec, struct iov_iter *iter,
2354 void __user *buf = u64_to_user_ptr(req->rw.addr);
2355 size_t sqe_len = req->rw.len;
2359 opcode = req->opcode;
2360 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2362 return io_import_fixed(req, rw, iter);
2365 /* buffer index only valid with fixed read/write, or buffer select */
2366 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2369 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2370 if (req->flags & REQ_F_BUFFER_SELECT) {
2371 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2374 return PTR_ERR(buf);
2376 req->rw.len = sqe_len;
2379 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2381 return ret < 0 ? ret : sqe_len;
2385 struct io_async_rw *iorw = &req->io->rw;
2388 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2389 if (iorw->iov == iorw->fast_iov)
2394 if (req->flags & REQ_F_BUFFER_SELECT) {
2395 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2397 ret = (*iovec)->iov_len;
2398 iov_iter_init(iter, rw, *iovec, 1, ret);
2404 #ifdef CONFIG_COMPAT
2405 if (req->ctx->compat)
2406 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2410 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2414 * For files that don't have ->read_iter() and ->write_iter(), handle them
2415 * by looping over ->read() or ->write() manually.
2417 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2418 struct iov_iter *iter)
2423 * Don't support polled IO through this interface, and we can't
2424 * support non-blocking either. For the latter, this just causes
2425 * the kiocb to be handled from an async context.
2427 if (kiocb->ki_flags & IOCB_HIPRI)
2429 if (kiocb->ki_flags & IOCB_NOWAIT)
2432 while (iov_iter_count(iter)) {
2436 if (!iov_iter_is_bvec(iter)) {
2437 iovec = iov_iter_iovec(iter);
2439 /* fixed buffers import bvec */
2440 iovec.iov_base = kmap(iter->bvec->bv_page)
2442 iovec.iov_len = min(iter->count,
2443 iter->bvec->bv_len - iter->iov_offset);
2447 nr = file->f_op->read(file, iovec.iov_base,
2448 iovec.iov_len, &kiocb->ki_pos);
2450 nr = file->f_op->write(file, iovec.iov_base,
2451 iovec.iov_len, &kiocb->ki_pos);
2454 if (iov_iter_is_bvec(iter))
2455 kunmap(iter->bvec->bv_page);
2463 if (nr != iovec.iov_len)
2465 iov_iter_advance(iter, nr);
2471 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2472 struct iovec *iovec, struct iovec *fast_iov,
2473 struct iov_iter *iter)
2475 req->io->rw.nr_segs = iter->nr_segs;
2476 req->io->rw.size = io_size;
2477 req->io->rw.iov = iovec;
2478 if (!req->io->rw.iov) {
2479 req->io->rw.iov = req->io->rw.fast_iov;
2480 if (req->io->rw.iov != fast_iov)
2481 memcpy(req->io->rw.iov, fast_iov,
2482 sizeof(struct iovec) * iter->nr_segs);
2484 req->flags |= REQ_F_NEED_CLEANUP;
2488 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2490 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2491 return req->io == NULL;
2494 static int io_alloc_async_ctx(struct io_kiocb *req)
2496 if (!io_op_defs[req->opcode].async_ctx)
2499 return __io_alloc_async_ctx(req);
2502 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2503 struct iovec *iovec, struct iovec *fast_iov,
2504 struct iov_iter *iter)
2506 if (!io_op_defs[req->opcode].async_ctx)
2509 if (__io_alloc_async_ctx(req))
2512 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2517 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2518 bool force_nonblock)
2520 struct io_async_ctx *io;
2521 struct iov_iter iter;
2524 ret = io_prep_rw(req, sqe, force_nonblock);
2528 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2531 /* either don't need iovec imported or already have it */
2532 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2536 io->rw.iov = io->rw.fast_iov;
2538 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2543 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2547 static int io_read(struct io_kiocb *req, bool force_nonblock)
2549 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2550 struct kiocb *kiocb = &req->rw.kiocb;
2551 struct iov_iter iter;
2553 ssize_t io_size, ret;
2555 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2559 /* Ensure we clear previously set non-block flag */
2560 if (!force_nonblock)
2561 kiocb->ki_flags &= ~IOCB_NOWAIT;
2565 if (req->flags & REQ_F_LINK)
2566 req->result = io_size;
2569 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2570 * we know to async punt it even if it was opened O_NONBLOCK
2572 if (force_nonblock && !io_file_supports_async(req->file))
2575 iov_count = iov_iter_count(&iter);
2576 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2580 if (req->file->f_op->read_iter)
2581 ret2 = call_read_iter(req->file, kiocb, &iter);
2583 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2585 /* Catch -EAGAIN return for forced non-blocking submission */
2586 if (!force_nonblock || ret2 != -EAGAIN) {
2587 kiocb_done(kiocb, ret2);
2590 ret = io_setup_async_rw(req, io_size, iovec,
2591 inline_vecs, &iter);
2594 /* any defer here is final, must blocking retry */
2595 if (!(req->flags & REQ_F_NOWAIT))
2596 req->flags |= REQ_F_MUST_PUNT;
2602 req->flags &= ~REQ_F_NEED_CLEANUP;
2606 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2607 bool force_nonblock)
2609 struct io_async_ctx *io;
2610 struct iov_iter iter;
2613 ret = io_prep_rw(req, sqe, force_nonblock);
2617 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2620 req->fsize = rlimit(RLIMIT_FSIZE);
2622 /* either don't need iovec imported or already have it */
2623 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2627 io->rw.iov = io->rw.fast_iov;
2629 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2634 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2638 static int io_write(struct io_kiocb *req, bool force_nonblock)
2640 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2641 struct kiocb *kiocb = &req->rw.kiocb;
2642 struct iov_iter iter;
2644 ssize_t ret, io_size;
2646 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2650 /* Ensure we clear previously set non-block flag */
2651 if (!force_nonblock)
2652 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2656 if (req->flags & REQ_F_LINK)
2657 req->result = io_size;
2660 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2661 * we know to async punt it even if it was opened O_NONBLOCK
2663 if (force_nonblock && !io_file_supports_async(req->file))
2666 /* file path doesn't support NOWAIT for non-direct_IO */
2667 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2668 (req->flags & REQ_F_ISREG))
2671 iov_count = iov_iter_count(&iter);
2672 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2677 * Open-code file_start_write here to grab freeze protection,
2678 * which will be released by another thread in
2679 * io_complete_rw(). Fool lockdep by telling it the lock got
2680 * released so that it doesn't complain about the held lock when
2681 * we return to userspace.
2683 if (req->flags & REQ_F_ISREG) {
2684 __sb_start_write(file_inode(req->file)->i_sb,
2685 SB_FREEZE_WRITE, true);
2686 __sb_writers_release(file_inode(req->file)->i_sb,
2689 kiocb->ki_flags |= IOCB_WRITE;
2691 if (!force_nonblock)
2692 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2694 if (req->file->f_op->write_iter)
2695 ret2 = call_write_iter(req->file, kiocb, &iter);
2697 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2699 if (!force_nonblock)
2700 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2703 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2704 * retry them without IOCB_NOWAIT.
2706 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2708 if (!force_nonblock || ret2 != -EAGAIN) {
2709 kiocb_done(kiocb, ret2);
2712 ret = io_setup_async_rw(req, io_size, iovec,
2713 inline_vecs, &iter);
2716 /* any defer here is final, must blocking retry */
2717 req->flags |= REQ_F_MUST_PUNT;
2722 req->flags &= ~REQ_F_NEED_CLEANUP;
2727 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2729 struct io_splice* sp = &req->splice;
2730 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2733 if (req->flags & REQ_F_NEED_CLEANUP)
2737 sp->off_in = READ_ONCE(sqe->splice_off_in);
2738 sp->off_out = READ_ONCE(sqe->off);
2739 sp->len = READ_ONCE(sqe->len);
2740 sp->flags = READ_ONCE(sqe->splice_flags);
2742 if (unlikely(sp->flags & ~valid_flags))
2745 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2746 (sp->flags & SPLICE_F_FD_IN_FIXED));
2749 req->flags |= REQ_F_NEED_CLEANUP;
2751 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2752 req->work.flags |= IO_WQ_WORK_UNBOUND;
2757 static bool io_splice_punt(struct file *file)
2759 if (get_pipe_info(file))
2761 if (!io_file_supports_async(file))
2763 return !(file->f_mode & O_NONBLOCK);
2766 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2768 struct io_splice *sp = &req->splice;
2769 struct file *in = sp->file_in;
2770 struct file *out = sp->file_out;
2771 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2772 loff_t *poff_in, *poff_out;
2775 if (force_nonblock) {
2776 if (io_splice_punt(in) || io_splice_punt(out))
2778 flags |= SPLICE_F_NONBLOCK;
2781 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2782 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2783 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2784 if (force_nonblock && ret == -EAGAIN)
2787 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2788 req->flags &= ~REQ_F_NEED_CLEANUP;
2790 io_cqring_add_event(req, ret);
2792 req_set_fail_links(req);
2798 * IORING_OP_NOP just posts a completion event, nothing else.
2800 static int io_nop(struct io_kiocb *req)
2802 struct io_ring_ctx *ctx = req->ctx;
2804 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2807 io_cqring_add_event(req, 0);
2812 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2814 struct io_ring_ctx *ctx = req->ctx;
2819 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2821 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2824 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2825 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2828 req->sync.off = READ_ONCE(sqe->off);
2829 req->sync.len = READ_ONCE(sqe->len);
2833 static bool io_req_cancelled(struct io_kiocb *req)
2835 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2836 req_set_fail_links(req);
2837 io_cqring_add_event(req, -ECANCELED);
2845 static void __io_fsync(struct io_kiocb *req)
2847 loff_t end = req->sync.off + req->sync.len;
2850 ret = vfs_fsync_range(req->file, req->sync.off,
2851 end > 0 ? end : LLONG_MAX,
2852 req->sync.flags & IORING_FSYNC_DATASYNC);
2854 req_set_fail_links(req);
2855 io_cqring_add_event(req, ret);
2859 static void io_fsync_finish(struct io_wq_work **workptr)
2861 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2863 if (io_req_cancelled(req))
2866 io_steal_work(req, workptr);
2869 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2871 /* fsync always requires a blocking context */
2872 if (force_nonblock) {
2873 req->work.func = io_fsync_finish;
2880 static void __io_fallocate(struct io_kiocb *req)
2884 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2885 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2887 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2889 req_set_fail_links(req);
2890 io_cqring_add_event(req, ret);
2894 static void io_fallocate_finish(struct io_wq_work **workptr)
2896 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2898 if (io_req_cancelled(req))
2900 __io_fallocate(req);
2901 io_steal_work(req, workptr);
2904 static int io_fallocate_prep(struct io_kiocb *req,
2905 const struct io_uring_sqe *sqe)
2907 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2910 req->sync.off = READ_ONCE(sqe->off);
2911 req->sync.len = READ_ONCE(sqe->addr);
2912 req->sync.mode = READ_ONCE(sqe->len);
2913 req->fsize = rlimit(RLIMIT_FSIZE);
2917 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2919 /* fallocate always requiring blocking context */
2920 if (force_nonblock) {
2921 req->work.func = io_fallocate_finish;
2925 __io_fallocate(req);
2929 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2931 const char __user *fname;
2934 if (sqe->ioprio || sqe->buf_index)
2936 if (req->flags & REQ_F_FIXED_FILE)
2938 if (req->flags & REQ_F_NEED_CLEANUP)
2941 req->open.dfd = READ_ONCE(sqe->fd);
2942 req->open.how.mode = READ_ONCE(sqe->len);
2943 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2944 req->open.how.flags = READ_ONCE(sqe->open_flags);
2945 if (force_o_largefile())
2946 req->open.how.flags |= O_LARGEFILE;
2948 req->open.filename = getname(fname);
2949 if (IS_ERR(req->open.filename)) {
2950 ret = PTR_ERR(req->open.filename);
2951 req->open.filename = NULL;
2955 req->open.nofile = rlimit(RLIMIT_NOFILE);
2956 req->flags |= REQ_F_NEED_CLEANUP;
2960 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2962 struct open_how __user *how;
2963 const char __user *fname;
2967 if (sqe->ioprio || sqe->buf_index)
2969 if (req->flags & REQ_F_FIXED_FILE)
2971 if (req->flags & REQ_F_NEED_CLEANUP)
2974 req->open.dfd = READ_ONCE(sqe->fd);
2975 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2976 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2977 len = READ_ONCE(sqe->len);
2979 if (len < OPEN_HOW_SIZE_VER0)
2982 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2987 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2988 req->open.how.flags |= O_LARGEFILE;
2990 req->open.filename = getname(fname);
2991 if (IS_ERR(req->open.filename)) {
2992 ret = PTR_ERR(req->open.filename);
2993 req->open.filename = NULL;
2997 req->open.nofile = rlimit(RLIMIT_NOFILE);
2998 req->flags |= REQ_F_NEED_CLEANUP;
3002 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3004 struct open_flags op;
3011 ret = build_open_flags(&req->open.how, &op);
3015 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3019 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3022 ret = PTR_ERR(file);
3024 fsnotify_open(file);
3025 fd_install(ret, file);
3028 putname(req->open.filename);
3029 req->flags &= ~REQ_F_NEED_CLEANUP;
3031 req_set_fail_links(req);
3032 io_cqring_add_event(req, ret);
3037 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3039 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3040 return io_openat2(req, force_nonblock);
3043 static int io_remove_buffers_prep(struct io_kiocb *req,
3044 const struct io_uring_sqe *sqe)
3046 struct io_provide_buf *p = &req->pbuf;
3049 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3052 tmp = READ_ONCE(sqe->fd);
3053 if (!tmp || tmp > USHRT_MAX)
3056 memset(p, 0, sizeof(*p));
3058 p->bgid = READ_ONCE(sqe->buf_group);
3062 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3063 int bgid, unsigned nbufs)
3067 /* shouldn't happen */
3071 /* the head kbuf is the list itself */
3072 while (!list_empty(&buf->list)) {
3073 struct io_buffer *nxt;
3075 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3076 list_del(&nxt->list);
3083 idr_remove(&ctx->io_buffer_idr, bgid);
3088 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3090 struct io_provide_buf *p = &req->pbuf;
3091 struct io_ring_ctx *ctx = req->ctx;
3092 struct io_buffer *head;
3095 io_ring_submit_lock(ctx, !force_nonblock);
3097 lockdep_assert_held(&ctx->uring_lock);
3100 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3102 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3104 io_ring_submit_lock(ctx, !force_nonblock);
3106 req_set_fail_links(req);
3107 io_cqring_add_event(req, ret);
3112 static int io_provide_buffers_prep(struct io_kiocb *req,
3113 const struct io_uring_sqe *sqe)
3115 struct io_provide_buf *p = &req->pbuf;
3118 if (sqe->ioprio || sqe->rw_flags)
3121 tmp = READ_ONCE(sqe->fd);
3122 if (!tmp || tmp > USHRT_MAX)
3125 p->addr = READ_ONCE(sqe->addr);
3126 p->len = READ_ONCE(sqe->len);
3128 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3131 p->bgid = READ_ONCE(sqe->buf_group);
3132 tmp = READ_ONCE(sqe->off);
3133 if (tmp > USHRT_MAX)
3139 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3141 struct io_buffer *buf;
3142 u64 addr = pbuf->addr;
3143 int i, bid = pbuf->bid;
3145 for (i = 0; i < pbuf->nbufs; i++) {
3146 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3151 buf->len = pbuf->len;
3156 INIT_LIST_HEAD(&buf->list);
3159 list_add_tail(&buf->list, &(*head)->list);
3163 return i ? i : -ENOMEM;
3166 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3168 struct io_provide_buf *p = &req->pbuf;
3169 struct io_ring_ctx *ctx = req->ctx;
3170 struct io_buffer *head, *list;
3173 io_ring_submit_lock(ctx, !force_nonblock);
3175 lockdep_assert_held(&ctx->uring_lock);
3177 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3179 ret = io_add_buffers(p, &head);
3184 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3187 __io_remove_buffers(ctx, head, p->bgid, -1U);
3192 io_ring_submit_unlock(ctx, !force_nonblock);
3194 req_set_fail_links(req);
3195 io_cqring_add_event(req, ret);
3200 static int io_epoll_ctl_prep(struct io_kiocb *req,
3201 const struct io_uring_sqe *sqe)
3203 #if defined(CONFIG_EPOLL)
3204 if (sqe->ioprio || sqe->buf_index)
3207 req->epoll.epfd = READ_ONCE(sqe->fd);
3208 req->epoll.op = READ_ONCE(sqe->len);
3209 req->epoll.fd = READ_ONCE(sqe->off);
3211 if (ep_op_has_event(req->epoll.op)) {
3212 struct epoll_event __user *ev;
3214 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3215 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3225 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3227 #if defined(CONFIG_EPOLL)
3228 struct io_epoll *ie = &req->epoll;
3231 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3232 if (force_nonblock && ret == -EAGAIN)
3236 req_set_fail_links(req);
3237 io_cqring_add_event(req, ret);
3245 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3247 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3248 if (sqe->ioprio || sqe->buf_index || sqe->off)
3251 req->madvise.addr = READ_ONCE(sqe->addr);
3252 req->madvise.len = READ_ONCE(sqe->len);
3253 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3260 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3262 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3263 struct io_madvise *ma = &req->madvise;
3269 ret = do_madvise(ma->addr, ma->len, ma->advice);
3271 req_set_fail_links(req);
3272 io_cqring_add_event(req, ret);
3280 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3282 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3285 req->fadvise.offset = READ_ONCE(sqe->off);
3286 req->fadvise.len = READ_ONCE(sqe->len);
3287 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3291 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3293 struct io_fadvise *fa = &req->fadvise;
3296 if (force_nonblock) {
3297 switch (fa->advice) {
3298 case POSIX_FADV_NORMAL:
3299 case POSIX_FADV_RANDOM:
3300 case POSIX_FADV_SEQUENTIAL:
3307 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3309 req_set_fail_links(req);
3310 io_cqring_add_event(req, ret);
3315 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3317 const char __user *fname;
3318 unsigned lookup_flags;
3321 if (sqe->ioprio || sqe->buf_index)
3323 if (req->flags & REQ_F_FIXED_FILE)
3325 if (req->flags & REQ_F_NEED_CLEANUP)
3328 req->open.dfd = READ_ONCE(sqe->fd);
3329 req->open.mask = READ_ONCE(sqe->len);
3330 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3331 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3332 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3334 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3337 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3338 if (IS_ERR(req->open.filename)) {
3339 ret = PTR_ERR(req->open.filename);
3340 req->open.filename = NULL;
3344 req->flags |= REQ_F_NEED_CLEANUP;
3348 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3350 struct io_open *ctx = &req->open;
3351 unsigned lookup_flags;
3359 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3363 /* filename_lookup() drops it, keep a reference */
3364 ctx->filename->refcnt++;
3366 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3371 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3373 if (retry_estale(ret, lookup_flags)) {
3374 lookup_flags |= LOOKUP_REVAL;
3378 ret = cp_statx(&stat, ctx->buffer);
3380 putname(ctx->filename);
3381 req->flags &= ~REQ_F_NEED_CLEANUP;
3383 req_set_fail_links(req);
3384 io_cqring_add_event(req, ret);
3389 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3392 * If we queue this for async, it must not be cancellable. That would
3393 * leave the 'file' in an undeterminate state.
3395 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3397 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3398 sqe->rw_flags || sqe->buf_index)
3400 if (req->flags & REQ_F_FIXED_FILE)
3403 req->close.fd = READ_ONCE(sqe->fd);
3404 if (req->file->f_op == &io_uring_fops ||
3405 req->close.fd == req->ctx->ring_fd)
3411 /* only called when __close_fd_get_file() is done */
3412 static void __io_close_finish(struct io_kiocb *req)
3416 ret = filp_close(req->close.put_file, req->work.files);
3418 req_set_fail_links(req);
3419 io_cqring_add_event(req, ret);
3420 fput(req->close.put_file);
3424 static void io_close_finish(struct io_wq_work **workptr)
3426 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3428 /* not cancellable, don't do io_req_cancelled() */
3429 __io_close_finish(req);
3430 io_steal_work(req, workptr);
3433 static int io_close(struct io_kiocb *req, bool force_nonblock)
3437 req->close.put_file = NULL;
3438 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3442 /* if the file has a flush method, be safe and punt to async */
3443 if (req->close.put_file->f_op->flush && force_nonblock) {
3444 /* submission ref will be dropped, take it for async */
3445 refcount_inc(&req->refs);
3447 req->work.func = io_close_finish;
3449 * Do manual async queue here to avoid grabbing files - we don't
3450 * need the files, and it'll cause io_close_finish() to close
3451 * the file again and cause a double CQE entry for this request
3453 io_queue_async_work(req);
3458 * No ->flush(), safely close from here and just punt the
3459 * fput() to async context.
3461 __io_close_finish(req);
3465 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3467 struct io_ring_ctx *ctx = req->ctx;
3472 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3474 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3477 req->sync.off = READ_ONCE(sqe->off);
3478 req->sync.len = READ_ONCE(sqe->len);
3479 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3483 static void __io_sync_file_range(struct io_kiocb *req)
3487 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3490 req_set_fail_links(req);
3491 io_cqring_add_event(req, ret);
3496 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3498 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3500 if (io_req_cancelled(req))
3502 __io_sync_file_range(req);
3503 io_put_req(req); /* put submission ref */
3506 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3508 /* sync_file_range always requires a blocking context */
3509 if (force_nonblock) {
3510 req->work.func = io_sync_file_range_finish;
3514 __io_sync_file_range(req);
3518 #if defined(CONFIG_NET)
3519 static int io_setup_async_msg(struct io_kiocb *req,
3520 struct io_async_msghdr *kmsg)
3524 if (io_alloc_async_ctx(req)) {
3525 if (kmsg->iov != kmsg->fast_iov)
3529 req->flags |= REQ_F_NEED_CLEANUP;
3530 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3534 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3536 struct io_sr_msg *sr = &req->sr_msg;
3537 struct io_async_ctx *io = req->io;
3540 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3541 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3542 sr->len = READ_ONCE(sqe->len);
3544 #ifdef CONFIG_COMPAT
3545 if (req->ctx->compat)
3546 sr->msg_flags |= MSG_CMSG_COMPAT;
3549 if (!io || req->opcode == IORING_OP_SEND)
3551 /* iovec is already imported */
3552 if (req->flags & REQ_F_NEED_CLEANUP)
3555 io->msg.iov = io->msg.fast_iov;
3556 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3559 req->flags |= REQ_F_NEED_CLEANUP;
3563 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3565 struct io_async_msghdr *kmsg = NULL;
3566 struct socket *sock;
3569 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3572 sock = sock_from_file(req->file, &ret);
3574 struct io_async_ctx io;
3578 kmsg = &req->io->msg;
3579 kmsg->msg.msg_name = &req->io->msg.addr;
3580 /* if iov is set, it's allocated already */
3582 kmsg->iov = kmsg->fast_iov;
3583 kmsg->msg.msg_iter.iov = kmsg->iov;
3585 struct io_sr_msg *sr = &req->sr_msg;
3588 kmsg->msg.msg_name = &io.msg.addr;
3590 io.msg.iov = io.msg.fast_iov;
3591 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3592 sr->msg_flags, &io.msg.iov);
3597 flags = req->sr_msg.msg_flags;
3598 if (flags & MSG_DONTWAIT)
3599 req->flags |= REQ_F_NOWAIT;
3600 else if (force_nonblock)
3601 flags |= MSG_DONTWAIT;
3603 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3604 if (force_nonblock && ret == -EAGAIN)
3605 return io_setup_async_msg(req, kmsg);
3606 if (ret == -ERESTARTSYS)
3610 if (kmsg && kmsg->iov != kmsg->fast_iov)
3612 req->flags &= ~REQ_F_NEED_CLEANUP;
3613 io_cqring_add_event(req, ret);
3615 req_set_fail_links(req);
3620 static int io_send(struct io_kiocb *req, bool force_nonblock)
3622 struct socket *sock;
3625 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3628 sock = sock_from_file(req->file, &ret);
3630 struct io_sr_msg *sr = &req->sr_msg;
3635 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3640 msg.msg_name = NULL;
3641 msg.msg_control = NULL;
3642 msg.msg_controllen = 0;
3643 msg.msg_namelen = 0;
3645 flags = req->sr_msg.msg_flags;
3646 if (flags & MSG_DONTWAIT)
3647 req->flags |= REQ_F_NOWAIT;
3648 else if (force_nonblock)
3649 flags |= MSG_DONTWAIT;
3651 msg.msg_flags = flags;
3652 ret = sock_sendmsg(sock, &msg);
3653 if (force_nonblock && ret == -EAGAIN)
3655 if (ret == -ERESTARTSYS)
3659 io_cqring_add_event(req, ret);
3661 req_set_fail_links(req);
3666 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3668 struct io_sr_msg *sr = &req->sr_msg;
3669 struct iovec __user *uiov;
3673 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3678 if (req->flags & REQ_F_BUFFER_SELECT) {
3681 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3683 sr->len = io->msg.iov[0].iov_len;
3684 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3688 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3689 &io->msg.iov, &io->msg.msg.msg_iter);
3697 #ifdef CONFIG_COMPAT
3698 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3699 struct io_async_ctx *io)
3701 struct compat_msghdr __user *msg_compat;
3702 struct io_sr_msg *sr = &req->sr_msg;
3703 struct compat_iovec __user *uiov;
3708 msg_compat = (struct compat_msghdr __user *) sr->msg;
3709 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3714 uiov = compat_ptr(ptr);
3715 if (req->flags & REQ_F_BUFFER_SELECT) {
3716 compat_ssize_t clen;
3720 if (!access_ok(uiov, sizeof(*uiov)))
3722 if (__get_user(clen, &uiov->iov_len))
3726 sr->len = io->msg.iov[0].iov_len;
3729 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3731 &io->msg.msg.msg_iter);
3740 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3742 io->msg.iov = io->msg.fast_iov;
3744 #ifdef CONFIG_COMPAT
3745 if (req->ctx->compat)
3746 return __io_compat_recvmsg_copy_hdr(req, io);
3749 return __io_recvmsg_copy_hdr(req, io);
3752 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3753 int *cflags, bool needs_lock)
3755 struct io_sr_msg *sr = &req->sr_msg;
3756 struct io_buffer *kbuf;
3758 if (!(req->flags & REQ_F_BUFFER_SELECT))
3761 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3766 req->flags |= REQ_F_BUFFER_SELECTED;
3768 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3769 *cflags |= IORING_CQE_F_BUFFER;
3773 static int io_recvmsg_prep(struct io_kiocb *req,
3774 const struct io_uring_sqe *sqe)
3776 struct io_sr_msg *sr = &req->sr_msg;
3777 struct io_async_ctx *io = req->io;
3780 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3781 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3782 sr->len = READ_ONCE(sqe->len);
3783 sr->bgid = READ_ONCE(sqe->buf_group);
3785 #ifdef CONFIG_COMPAT
3786 if (req->ctx->compat)
3787 sr->msg_flags |= MSG_CMSG_COMPAT;
3790 if (!io || req->opcode == IORING_OP_RECV)
3792 /* iovec is already imported */
3793 if (req->flags & REQ_F_NEED_CLEANUP)
3796 ret = io_recvmsg_copy_hdr(req, io);
3798 req->flags |= REQ_F_NEED_CLEANUP;
3802 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3804 struct io_async_msghdr *kmsg = NULL;
3805 struct socket *sock;
3806 int ret, cflags = 0;
3808 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3811 sock = sock_from_file(req->file, &ret);
3813 struct io_buffer *kbuf;
3814 struct io_async_ctx io;
3818 kmsg = &req->io->msg;
3819 kmsg->msg.msg_name = &req->io->msg.addr;
3820 /* if iov is set, it's allocated already */
3822 kmsg->iov = kmsg->fast_iov;
3823 kmsg->msg.msg_iter.iov = kmsg->iov;
3826 kmsg->msg.msg_name = &io.msg.addr;
3828 ret = io_recvmsg_copy_hdr(req, &io);
3833 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3835 return PTR_ERR(kbuf);
3837 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3838 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3839 1, req->sr_msg.len);
3842 flags = req->sr_msg.msg_flags;
3843 if (flags & MSG_DONTWAIT)
3844 req->flags |= REQ_F_NOWAIT;
3845 else if (force_nonblock)
3846 flags |= MSG_DONTWAIT;
3848 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3849 kmsg->uaddr, flags);
3850 if (force_nonblock && ret == -EAGAIN)
3851 return io_setup_async_msg(req, kmsg);
3852 if (ret == -ERESTARTSYS)
3856 if (kmsg && kmsg->iov != kmsg->fast_iov)
3858 req->flags &= ~REQ_F_NEED_CLEANUP;
3859 __io_cqring_add_event(req, ret, cflags);
3861 req_set_fail_links(req);
3866 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3868 struct io_buffer *kbuf = NULL;
3869 struct socket *sock;
3870 int ret, cflags = 0;
3872 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3875 sock = sock_from_file(req->file, &ret);
3877 struct io_sr_msg *sr = &req->sr_msg;
3878 void __user *buf = sr->buf;
3883 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3885 return PTR_ERR(kbuf);
3887 buf = u64_to_user_ptr(kbuf->addr);
3889 ret = import_single_range(READ, buf, sr->len, &iov,
3896 req->flags |= REQ_F_NEED_CLEANUP;
3897 msg.msg_name = NULL;
3898 msg.msg_control = NULL;
3899 msg.msg_controllen = 0;
3900 msg.msg_namelen = 0;
3901 msg.msg_iocb = NULL;
3904 flags = req->sr_msg.msg_flags;
3905 if (flags & MSG_DONTWAIT)
3906 req->flags |= REQ_F_NOWAIT;
3907 else if (force_nonblock)
3908 flags |= MSG_DONTWAIT;
3910 ret = sock_recvmsg(sock, &msg, flags);
3911 if (force_nonblock && ret == -EAGAIN)
3913 if (ret == -ERESTARTSYS)
3918 req->flags &= ~REQ_F_NEED_CLEANUP;
3919 __io_cqring_add_event(req, ret, cflags);
3921 req_set_fail_links(req);
3926 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3928 struct io_accept *accept = &req->accept;
3930 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3932 if (sqe->ioprio || sqe->len || sqe->buf_index)
3935 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3936 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3937 accept->flags = READ_ONCE(sqe->accept_flags);
3938 accept->nofile = rlimit(RLIMIT_NOFILE);
3942 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3944 struct io_accept *accept = &req->accept;
3945 unsigned file_flags;
3948 file_flags = force_nonblock ? O_NONBLOCK : 0;
3949 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3950 accept->addr_len, accept->flags,
3952 if (ret == -EAGAIN && force_nonblock)
3954 if (ret == -ERESTARTSYS)
3957 req_set_fail_links(req);
3958 io_cqring_add_event(req, ret);
3963 static void io_accept_finish(struct io_wq_work **workptr)
3965 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3967 if (io_req_cancelled(req))
3969 __io_accept(req, false);
3970 io_steal_work(req, workptr);
3973 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3977 ret = __io_accept(req, force_nonblock);
3978 if (ret == -EAGAIN && force_nonblock) {
3979 req->work.func = io_accept_finish;
3985 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3987 struct io_connect *conn = &req->connect;
3988 struct io_async_ctx *io = req->io;
3990 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3992 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3995 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3996 conn->addr_len = READ_ONCE(sqe->addr2);
4001 return move_addr_to_kernel(conn->addr, conn->addr_len,
4002 &io->connect.address);
4005 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4007 struct io_async_ctx __io, *io;
4008 unsigned file_flags;
4014 ret = move_addr_to_kernel(req->connect.addr,
4015 req->connect.addr_len,
4016 &__io.connect.address);
4022 file_flags = force_nonblock ? O_NONBLOCK : 0;
4024 ret = __sys_connect_file(req->file, &io->connect.address,
4025 req->connect.addr_len, file_flags);
4026 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4029 if (io_alloc_async_ctx(req)) {
4033 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4036 if (ret == -ERESTARTSYS)
4040 req_set_fail_links(req);
4041 io_cqring_add_event(req, ret);
4045 #else /* !CONFIG_NET */
4046 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4051 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4056 static int io_send(struct io_kiocb *req, bool force_nonblock)
4061 static int io_recvmsg_prep(struct io_kiocb *req,
4062 const struct io_uring_sqe *sqe)
4067 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4072 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4077 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4082 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4087 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4092 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4096 #endif /* CONFIG_NET */
4098 struct io_poll_table {
4099 struct poll_table_struct pt;
4100 struct io_kiocb *req;
4104 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4105 struct wait_queue_head *head)
4107 if (unlikely(poll->head)) {
4108 pt->error = -EINVAL;
4114 add_wait_queue(head, &poll->wait);
4117 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4118 struct poll_table_struct *p)
4120 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4122 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4125 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4126 __poll_t mask, task_work_func_t func)
4128 struct task_struct *tsk;
4131 /* for instances that support it check for an event match first: */
4132 if (mask && !(mask & poll->events))
4135 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4137 list_del_init(&poll->wait.entry);
4141 init_task_work(&req->task_work, func);
4143 * If this fails, then the task is exiting. Punt to one of the io-wq
4144 * threads to ensure the work gets run, we can't always rely on exit
4145 * cancelation taking care of this.
4147 ret = task_work_add(tsk, &req->task_work, true);
4148 if (unlikely(ret)) {
4149 tsk = io_wq_get_task(req->ctx->io_wq);
4150 task_work_add(tsk, &req->task_work, true);
4152 wake_up_process(tsk);
4156 static void io_async_task_func(struct callback_head *cb)
4158 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4159 struct async_poll *apoll = req->apoll;
4160 struct io_ring_ctx *ctx = req->ctx;
4162 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4164 WARN_ON_ONCE(!list_empty(&req->apoll->poll.wait.entry));
4166 if (hash_hashed(&req->hash_node)) {
4167 spin_lock_irq(&ctx->completion_lock);
4168 hash_del(&req->hash_node);
4169 spin_unlock_irq(&ctx->completion_lock);
4172 /* restore ->work in case we need to retry again */
4173 memcpy(&req->work, &apoll->work, sizeof(req->work));
4175 __set_current_state(TASK_RUNNING);
4176 mutex_lock(&ctx->uring_lock);
4177 __io_queue_sqe(req, NULL);
4178 mutex_unlock(&ctx->uring_lock);
4183 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4186 struct io_kiocb *req = wait->private;
4187 struct io_poll_iocb *poll = &req->apoll->poll;
4189 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4192 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4195 static void io_poll_req_insert(struct io_kiocb *req)
4197 struct io_ring_ctx *ctx = req->ctx;
4198 struct hlist_head *list;
4200 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4201 hlist_add_head(&req->hash_node, list);
4204 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4205 struct io_poll_iocb *poll,
4206 struct io_poll_table *ipt, __poll_t mask,
4207 wait_queue_func_t wake_func)
4208 __acquires(&ctx->completion_lock)
4210 struct io_ring_ctx *ctx = req->ctx;
4211 bool cancel = false;
4213 poll->file = req->file;
4215 poll->done = poll->canceled = false;
4216 poll->events = mask;
4218 ipt->pt._key = mask;
4220 ipt->error = -EINVAL;
4222 INIT_LIST_HEAD(&poll->wait.entry);
4223 init_waitqueue_func_entry(&poll->wait, wake_func);
4224 poll->wait.private = req;
4226 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4228 spin_lock_irq(&ctx->completion_lock);
4229 if (likely(poll->head)) {
4230 spin_lock(&poll->head->lock);
4231 if (unlikely(list_empty(&poll->wait.entry))) {
4237 if (mask || ipt->error)
4238 list_del_init(&poll->wait.entry);
4240 WRITE_ONCE(poll->canceled, true);
4241 else if (!poll->done) /* actually waiting for an event */
4242 io_poll_req_insert(req);
4243 spin_unlock(&poll->head->lock);
4249 static bool io_arm_poll_handler(struct io_kiocb *req)
4251 const struct io_op_def *def = &io_op_defs[req->opcode];
4252 struct io_ring_ctx *ctx = req->ctx;
4253 struct async_poll *apoll;
4254 struct io_poll_table ipt;
4257 if (!req->file || !file_can_poll(req->file))
4259 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4261 if (!def->pollin && !def->pollout)
4264 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4265 if (unlikely(!apoll))
4268 req->flags |= REQ_F_POLLED;
4269 memcpy(&apoll->work, &req->work, sizeof(req->work));
4271 get_task_struct(current);
4272 req->task = current;
4274 INIT_HLIST_NODE(&req->hash_node);
4278 mask |= POLLIN | POLLRDNORM;
4280 mask |= POLLOUT | POLLWRNORM;
4281 mask |= POLLERR | POLLPRI;
4283 ipt.pt._qproc = io_async_queue_proc;
4285 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4289 apoll->poll.done = true;
4290 spin_unlock_irq(&ctx->completion_lock);
4291 memcpy(&req->work, &apoll->work, sizeof(req->work));
4295 spin_unlock_irq(&ctx->completion_lock);
4296 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4297 apoll->poll.events);
4301 static bool __io_poll_remove_one(struct io_kiocb *req,
4302 struct io_poll_iocb *poll)
4304 bool do_complete = false;
4306 spin_lock(&poll->head->lock);
4307 WRITE_ONCE(poll->canceled, true);
4308 if (!list_empty(&poll->wait.entry)) {
4309 list_del_init(&poll->wait.entry);
4312 spin_unlock(&poll->head->lock);
4316 static bool io_poll_remove_one(struct io_kiocb *req)
4320 if (req->opcode == IORING_OP_POLL_ADD) {
4321 do_complete = __io_poll_remove_one(req, &req->poll);
4323 /* non-poll requests have submit ref still */
4324 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4329 hash_del(&req->hash_node);
4332 io_cqring_fill_event(req, -ECANCELED);
4333 io_commit_cqring(req->ctx);
4334 req->flags |= REQ_F_COMP_LOCKED;
4341 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4343 struct hlist_node *tmp;
4344 struct io_kiocb *req;
4347 spin_lock_irq(&ctx->completion_lock);
4348 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4349 struct hlist_head *list;
4351 list = &ctx->cancel_hash[i];
4352 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4353 io_poll_remove_one(req);
4355 spin_unlock_irq(&ctx->completion_lock);
4357 io_cqring_ev_posted(ctx);
4360 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4362 struct hlist_head *list;
4363 struct io_kiocb *req;
4365 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4366 hlist_for_each_entry(req, list, hash_node) {
4367 if (sqe_addr != req->user_data)
4369 if (io_poll_remove_one(req))
4377 static int io_poll_remove_prep(struct io_kiocb *req,
4378 const struct io_uring_sqe *sqe)
4380 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4382 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4386 req->poll.addr = READ_ONCE(sqe->addr);
4391 * Find a running poll command that matches one specified in sqe->addr,
4392 * and remove it if found.
4394 static int io_poll_remove(struct io_kiocb *req)
4396 struct io_ring_ctx *ctx = req->ctx;
4400 addr = req->poll.addr;
4401 spin_lock_irq(&ctx->completion_lock);
4402 ret = io_poll_cancel(ctx, addr);
4403 spin_unlock_irq(&ctx->completion_lock);
4405 io_cqring_add_event(req, ret);
4407 req_set_fail_links(req);
4412 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4414 struct io_ring_ctx *ctx = req->ctx;
4416 req->poll.done = true;
4417 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4418 io_commit_cqring(ctx);
4421 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4423 struct io_ring_ctx *ctx = req->ctx;
4424 struct io_poll_iocb *poll = &req->poll;
4426 if (!req->result && !READ_ONCE(poll->canceled)) {
4427 struct poll_table_struct pt = { ._key = poll->events };
4429 req->result = vfs_poll(req->file, &pt) & poll->events;
4432 spin_lock_irq(&ctx->completion_lock);
4433 if (!req->result && !READ_ONCE(poll->canceled)) {
4434 add_wait_queue(poll->head, &poll->wait);
4435 spin_unlock_irq(&ctx->completion_lock);
4438 hash_del(&req->hash_node);
4439 io_poll_complete(req, req->result, 0);
4440 req->flags |= REQ_F_COMP_LOCKED;
4441 io_put_req_find_next(req, nxt);
4442 spin_unlock_irq(&ctx->completion_lock);
4444 io_cqring_ev_posted(ctx);
4447 static void io_poll_task_func(struct callback_head *cb)
4449 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4450 struct io_kiocb *nxt = NULL;
4452 io_poll_task_handler(req, &nxt);
4454 struct io_ring_ctx *ctx = nxt->ctx;
4456 mutex_lock(&ctx->uring_lock);
4457 __io_queue_sqe(nxt, NULL);
4458 mutex_unlock(&ctx->uring_lock);
4462 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4465 struct io_kiocb *req = wait->private;
4466 struct io_poll_iocb *poll = &req->poll;
4468 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4471 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4472 struct poll_table_struct *p)
4474 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4476 __io_queue_proc(&pt->req->poll, pt, head);
4479 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4481 struct io_poll_iocb *poll = &req->poll;
4484 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4486 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4491 events = READ_ONCE(sqe->poll_events);
4492 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4494 get_task_struct(current);
4495 req->task = current;
4499 static int io_poll_add(struct io_kiocb *req)
4501 struct io_poll_iocb *poll = &req->poll;
4502 struct io_ring_ctx *ctx = req->ctx;
4503 struct io_poll_table ipt;
4506 INIT_HLIST_NODE(&req->hash_node);
4507 INIT_LIST_HEAD(&req->list);
4508 ipt.pt._qproc = io_poll_queue_proc;
4510 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4513 if (mask) { /* no async, we'd stolen it */
4515 io_poll_complete(req, mask, 0);
4517 spin_unlock_irq(&ctx->completion_lock);
4520 io_cqring_ev_posted(ctx);
4526 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4528 struct io_timeout_data *data = container_of(timer,
4529 struct io_timeout_data, timer);
4530 struct io_kiocb *req = data->req;
4531 struct io_ring_ctx *ctx = req->ctx;
4532 unsigned long flags;
4534 atomic_inc(&ctx->cq_timeouts);
4536 spin_lock_irqsave(&ctx->completion_lock, flags);
4538 * We could be racing with timeout deletion. If the list is empty,
4539 * then timeout lookup already found it and will be handling it.
4541 if (!list_empty(&req->list)) {
4542 struct io_kiocb *prev;
4545 * Adjust the reqs sequence before the current one because it
4546 * will consume a slot in the cq_ring and the cq_tail
4547 * pointer will be increased, otherwise other timeout reqs may
4548 * return in advance without waiting for enough wait_nr.
4551 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4553 list_del_init(&req->list);
4556 io_cqring_fill_event(req, -ETIME);
4557 io_commit_cqring(ctx);
4558 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4560 io_cqring_ev_posted(ctx);
4561 req_set_fail_links(req);
4563 return HRTIMER_NORESTART;
4566 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4568 struct io_kiocb *req;
4571 list_for_each_entry(req, &ctx->timeout_list, list) {
4572 if (user_data == req->user_data) {
4573 list_del_init(&req->list);
4582 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4586 req_set_fail_links(req);
4587 io_cqring_fill_event(req, -ECANCELED);
4592 static int io_timeout_remove_prep(struct io_kiocb *req,
4593 const struct io_uring_sqe *sqe)
4595 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4597 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4600 req->timeout.addr = READ_ONCE(sqe->addr);
4601 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4602 if (req->timeout.flags)
4609 * Remove or update an existing timeout command
4611 static int io_timeout_remove(struct io_kiocb *req)
4613 struct io_ring_ctx *ctx = req->ctx;
4616 spin_lock_irq(&ctx->completion_lock);
4617 ret = io_timeout_cancel(ctx, req->timeout.addr);
4619 io_cqring_fill_event(req, ret);
4620 io_commit_cqring(ctx);
4621 spin_unlock_irq(&ctx->completion_lock);
4622 io_cqring_ev_posted(ctx);
4624 req_set_fail_links(req);
4629 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4630 bool is_timeout_link)
4632 struct io_timeout_data *data;
4635 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4637 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4639 if (sqe->off && is_timeout_link)
4641 flags = READ_ONCE(sqe->timeout_flags);
4642 if (flags & ~IORING_TIMEOUT_ABS)
4645 req->timeout.count = READ_ONCE(sqe->off);
4647 if (!req->io && io_alloc_async_ctx(req))
4650 data = &req->io->timeout;
4652 req->flags |= REQ_F_TIMEOUT;
4654 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4657 if (flags & IORING_TIMEOUT_ABS)
4658 data->mode = HRTIMER_MODE_ABS;
4660 data->mode = HRTIMER_MODE_REL;
4662 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4666 static int io_timeout(struct io_kiocb *req)
4669 struct io_ring_ctx *ctx = req->ctx;
4670 struct io_timeout_data *data;
4671 struct list_head *entry;
4674 data = &req->io->timeout;
4677 * sqe->off holds how many events that need to occur for this
4678 * timeout event to be satisfied. If it isn't set, then this is
4679 * a pure timeout request, sequence isn't used.
4681 count = req->timeout.count;
4683 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4684 spin_lock_irq(&ctx->completion_lock);
4685 entry = ctx->timeout_list.prev;
4689 req->sequence = ctx->cached_sq_head + count - 1;
4690 data->seq_offset = count;
4693 * Insertion sort, ensuring the first entry in the list is always
4694 * the one we need first.
4696 spin_lock_irq(&ctx->completion_lock);
4697 list_for_each_prev(entry, &ctx->timeout_list) {
4698 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4699 unsigned nxt_sq_head;
4700 long long tmp, tmp_nxt;
4701 u32 nxt_offset = nxt->io->timeout.seq_offset;
4703 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4707 * Since cached_sq_head + count - 1 can overflow, use type long
4710 tmp = (long long)ctx->cached_sq_head + count - 1;
4711 nxt_sq_head = nxt->sequence - nxt_offset + 1;
4712 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
4715 * cached_sq_head may overflow, and it will never overflow twice
4716 * once there is some timeout req still be valid.
4718 if (ctx->cached_sq_head < nxt_sq_head)
4725 * Sequence of reqs after the insert one and itself should
4726 * be adjusted because each timeout req consumes a slot.
4731 req->sequence -= span;
4733 list_add(&req->list, entry);
4734 data->timer.function = io_timeout_fn;
4735 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4736 spin_unlock_irq(&ctx->completion_lock);
4740 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4742 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4744 return req->user_data == (unsigned long) data;
4747 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4749 enum io_wq_cancel cancel_ret;
4752 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4753 switch (cancel_ret) {
4754 case IO_WQ_CANCEL_OK:
4757 case IO_WQ_CANCEL_RUNNING:
4760 case IO_WQ_CANCEL_NOTFOUND:
4768 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4769 struct io_kiocb *req, __u64 sqe_addr,
4772 unsigned long flags;
4775 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4776 if (ret != -ENOENT) {
4777 spin_lock_irqsave(&ctx->completion_lock, flags);
4781 spin_lock_irqsave(&ctx->completion_lock, flags);
4782 ret = io_timeout_cancel(ctx, sqe_addr);
4785 ret = io_poll_cancel(ctx, sqe_addr);
4789 io_cqring_fill_event(req, ret);
4790 io_commit_cqring(ctx);
4791 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4792 io_cqring_ev_posted(ctx);
4795 req_set_fail_links(req);
4799 static int io_async_cancel_prep(struct io_kiocb *req,
4800 const struct io_uring_sqe *sqe)
4802 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4804 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4808 req->cancel.addr = READ_ONCE(sqe->addr);
4812 static int io_async_cancel(struct io_kiocb *req)
4814 struct io_ring_ctx *ctx = req->ctx;
4816 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4820 static int io_files_update_prep(struct io_kiocb *req,
4821 const struct io_uring_sqe *sqe)
4823 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4826 req->files_update.offset = READ_ONCE(sqe->off);
4827 req->files_update.nr_args = READ_ONCE(sqe->len);
4828 if (!req->files_update.nr_args)
4830 req->files_update.arg = READ_ONCE(sqe->addr);
4834 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4836 struct io_ring_ctx *ctx = req->ctx;
4837 struct io_uring_files_update up;
4843 up.offset = req->files_update.offset;
4844 up.fds = req->files_update.arg;
4846 mutex_lock(&ctx->uring_lock);
4847 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4848 mutex_unlock(&ctx->uring_lock);
4851 req_set_fail_links(req);
4852 io_cqring_add_event(req, ret);
4857 static int io_req_defer_prep(struct io_kiocb *req,
4858 const struct io_uring_sqe *sqe)
4865 if (io_op_defs[req->opcode].file_table) {
4866 ret = io_grab_files(req);
4871 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4873 switch (req->opcode) {
4876 case IORING_OP_READV:
4877 case IORING_OP_READ_FIXED:
4878 case IORING_OP_READ:
4879 ret = io_read_prep(req, sqe, true);
4881 case IORING_OP_WRITEV:
4882 case IORING_OP_WRITE_FIXED:
4883 case IORING_OP_WRITE:
4884 ret = io_write_prep(req, sqe, true);
4886 case IORING_OP_POLL_ADD:
4887 ret = io_poll_add_prep(req, sqe);
4889 case IORING_OP_POLL_REMOVE:
4890 ret = io_poll_remove_prep(req, sqe);
4892 case IORING_OP_FSYNC:
4893 ret = io_prep_fsync(req, sqe);
4895 case IORING_OP_SYNC_FILE_RANGE:
4896 ret = io_prep_sfr(req, sqe);
4898 case IORING_OP_SENDMSG:
4899 case IORING_OP_SEND:
4900 ret = io_sendmsg_prep(req, sqe);
4902 case IORING_OP_RECVMSG:
4903 case IORING_OP_RECV:
4904 ret = io_recvmsg_prep(req, sqe);
4906 case IORING_OP_CONNECT:
4907 ret = io_connect_prep(req, sqe);
4909 case IORING_OP_TIMEOUT:
4910 ret = io_timeout_prep(req, sqe, false);
4912 case IORING_OP_TIMEOUT_REMOVE:
4913 ret = io_timeout_remove_prep(req, sqe);
4915 case IORING_OP_ASYNC_CANCEL:
4916 ret = io_async_cancel_prep(req, sqe);
4918 case IORING_OP_LINK_TIMEOUT:
4919 ret = io_timeout_prep(req, sqe, true);
4921 case IORING_OP_ACCEPT:
4922 ret = io_accept_prep(req, sqe);
4924 case IORING_OP_FALLOCATE:
4925 ret = io_fallocate_prep(req, sqe);
4927 case IORING_OP_OPENAT:
4928 ret = io_openat_prep(req, sqe);
4930 case IORING_OP_CLOSE:
4931 ret = io_close_prep(req, sqe);
4933 case IORING_OP_FILES_UPDATE:
4934 ret = io_files_update_prep(req, sqe);
4936 case IORING_OP_STATX:
4937 ret = io_statx_prep(req, sqe);
4939 case IORING_OP_FADVISE:
4940 ret = io_fadvise_prep(req, sqe);
4942 case IORING_OP_MADVISE:
4943 ret = io_madvise_prep(req, sqe);
4945 case IORING_OP_OPENAT2:
4946 ret = io_openat2_prep(req, sqe);
4948 case IORING_OP_EPOLL_CTL:
4949 ret = io_epoll_ctl_prep(req, sqe);
4951 case IORING_OP_SPLICE:
4952 ret = io_splice_prep(req, sqe);
4954 case IORING_OP_PROVIDE_BUFFERS:
4955 ret = io_provide_buffers_prep(req, sqe);
4957 case IORING_OP_REMOVE_BUFFERS:
4958 ret = io_remove_buffers_prep(req, sqe);
4961 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4970 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4972 struct io_ring_ctx *ctx = req->ctx;
4975 /* Still need defer if there is pending req in defer list. */
4976 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4979 if (!req->io && io_alloc_async_ctx(req))
4982 ret = io_req_defer_prep(req, sqe);
4986 spin_lock_irq(&ctx->completion_lock);
4987 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4988 spin_unlock_irq(&ctx->completion_lock);
4992 trace_io_uring_defer(ctx, req, req->user_data);
4993 list_add_tail(&req->list, &ctx->defer_list);
4994 spin_unlock_irq(&ctx->completion_lock);
4995 return -EIOCBQUEUED;
4998 static void io_cleanup_req(struct io_kiocb *req)
5000 struct io_async_ctx *io = req->io;
5002 switch (req->opcode) {
5003 case IORING_OP_READV:
5004 case IORING_OP_READ_FIXED:
5005 case IORING_OP_READ:
5006 if (req->flags & REQ_F_BUFFER_SELECTED)
5007 kfree((void *)(unsigned long)req->rw.addr);
5009 case IORING_OP_WRITEV:
5010 case IORING_OP_WRITE_FIXED:
5011 case IORING_OP_WRITE:
5012 if (io->rw.iov != io->rw.fast_iov)
5015 case IORING_OP_RECVMSG:
5016 if (req->flags & REQ_F_BUFFER_SELECTED)
5017 kfree(req->sr_msg.kbuf);
5019 case IORING_OP_SENDMSG:
5020 if (io->msg.iov != io->msg.fast_iov)
5023 case IORING_OP_RECV:
5024 if (req->flags & REQ_F_BUFFER_SELECTED)
5025 kfree(req->sr_msg.kbuf);
5027 case IORING_OP_OPENAT:
5028 case IORING_OP_OPENAT2:
5029 case IORING_OP_STATX:
5030 putname(req->open.filename);
5032 case IORING_OP_SPLICE:
5033 io_put_file(req, req->splice.file_in,
5034 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5038 req->flags &= ~REQ_F_NEED_CLEANUP;
5041 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5042 bool force_nonblock)
5044 struct io_ring_ctx *ctx = req->ctx;
5047 switch (req->opcode) {
5051 case IORING_OP_READV:
5052 case IORING_OP_READ_FIXED:
5053 case IORING_OP_READ:
5055 ret = io_read_prep(req, sqe, force_nonblock);
5059 ret = io_read(req, force_nonblock);
5061 case IORING_OP_WRITEV:
5062 case IORING_OP_WRITE_FIXED:
5063 case IORING_OP_WRITE:
5065 ret = io_write_prep(req, sqe, force_nonblock);
5069 ret = io_write(req, force_nonblock);
5071 case IORING_OP_FSYNC:
5073 ret = io_prep_fsync(req, sqe);
5077 ret = io_fsync(req, force_nonblock);
5079 case IORING_OP_POLL_ADD:
5081 ret = io_poll_add_prep(req, sqe);
5085 ret = io_poll_add(req);
5087 case IORING_OP_POLL_REMOVE:
5089 ret = io_poll_remove_prep(req, sqe);
5093 ret = io_poll_remove(req);
5095 case IORING_OP_SYNC_FILE_RANGE:
5097 ret = io_prep_sfr(req, sqe);
5101 ret = io_sync_file_range(req, force_nonblock);
5103 case IORING_OP_SENDMSG:
5104 case IORING_OP_SEND:
5106 ret = io_sendmsg_prep(req, sqe);
5110 if (req->opcode == IORING_OP_SENDMSG)
5111 ret = io_sendmsg(req, force_nonblock);
5113 ret = io_send(req, force_nonblock);
5115 case IORING_OP_RECVMSG:
5116 case IORING_OP_RECV:
5118 ret = io_recvmsg_prep(req, sqe);
5122 if (req->opcode == IORING_OP_RECVMSG)
5123 ret = io_recvmsg(req, force_nonblock);
5125 ret = io_recv(req, force_nonblock);
5127 case IORING_OP_TIMEOUT:
5129 ret = io_timeout_prep(req, sqe, false);
5133 ret = io_timeout(req);
5135 case IORING_OP_TIMEOUT_REMOVE:
5137 ret = io_timeout_remove_prep(req, sqe);
5141 ret = io_timeout_remove(req);
5143 case IORING_OP_ACCEPT:
5145 ret = io_accept_prep(req, sqe);
5149 ret = io_accept(req, force_nonblock);
5151 case IORING_OP_CONNECT:
5153 ret = io_connect_prep(req, sqe);
5157 ret = io_connect(req, force_nonblock);
5159 case IORING_OP_ASYNC_CANCEL:
5161 ret = io_async_cancel_prep(req, sqe);
5165 ret = io_async_cancel(req);
5167 case IORING_OP_FALLOCATE:
5169 ret = io_fallocate_prep(req, sqe);
5173 ret = io_fallocate(req, force_nonblock);
5175 case IORING_OP_OPENAT:
5177 ret = io_openat_prep(req, sqe);
5181 ret = io_openat(req, force_nonblock);
5183 case IORING_OP_CLOSE:
5185 ret = io_close_prep(req, sqe);
5189 ret = io_close(req, force_nonblock);
5191 case IORING_OP_FILES_UPDATE:
5193 ret = io_files_update_prep(req, sqe);
5197 ret = io_files_update(req, force_nonblock);
5199 case IORING_OP_STATX:
5201 ret = io_statx_prep(req, sqe);
5205 ret = io_statx(req, force_nonblock);
5207 case IORING_OP_FADVISE:
5209 ret = io_fadvise_prep(req, sqe);
5213 ret = io_fadvise(req, force_nonblock);
5215 case IORING_OP_MADVISE:
5217 ret = io_madvise_prep(req, sqe);
5221 ret = io_madvise(req, force_nonblock);
5223 case IORING_OP_OPENAT2:
5225 ret = io_openat2_prep(req, sqe);
5229 ret = io_openat2(req, force_nonblock);
5231 case IORING_OP_EPOLL_CTL:
5233 ret = io_epoll_ctl_prep(req, sqe);
5237 ret = io_epoll_ctl(req, force_nonblock);
5239 case IORING_OP_SPLICE:
5241 ret = io_splice_prep(req, sqe);
5245 ret = io_splice(req, force_nonblock);
5247 case IORING_OP_PROVIDE_BUFFERS:
5249 ret = io_provide_buffers_prep(req, sqe);
5253 ret = io_provide_buffers(req, force_nonblock);
5255 case IORING_OP_REMOVE_BUFFERS:
5257 ret = io_remove_buffers_prep(req, sqe);
5261 ret = io_remove_buffers(req, force_nonblock);
5271 if (ctx->flags & IORING_SETUP_IOPOLL) {
5272 const bool in_async = io_wq_current_is_worker();
5274 if (req->result == -EAGAIN)
5277 /* workqueue context doesn't hold uring_lock, grab it now */
5279 mutex_lock(&ctx->uring_lock);
5281 io_iopoll_req_issued(req);
5284 mutex_unlock(&ctx->uring_lock);
5290 static void io_wq_submit_work(struct io_wq_work **workptr)
5292 struct io_wq_work *work = *workptr;
5293 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5296 /* if NO_CANCEL is set, we must still run the work */
5297 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5298 IO_WQ_WORK_CANCEL) {
5304 ret = io_issue_sqe(req, NULL, false);
5306 * We can get EAGAIN for polled IO even though we're
5307 * forcing a sync submission from here, since we can't
5308 * wait for request slots on the block side.
5317 req_set_fail_links(req);
5318 io_cqring_add_event(req, ret);
5322 io_steal_work(req, workptr);
5325 static int io_req_needs_file(struct io_kiocb *req, int fd)
5327 if (!io_op_defs[req->opcode].needs_file)
5329 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5334 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5337 struct fixed_file_table *table;
5339 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5340 return table->files[index & IORING_FILE_TABLE_MASK];;
5343 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5344 int fd, struct file **out_file, bool fixed)
5346 struct io_ring_ctx *ctx = req->ctx;
5350 if (unlikely(!ctx->file_data ||
5351 (unsigned) fd >= ctx->nr_user_files))
5353 fd = array_index_nospec(fd, ctx->nr_user_files);
5354 file = io_file_from_index(ctx, fd);
5357 req->fixed_file_refs = ctx->file_data->cur_refs;
5358 percpu_ref_get(req->fixed_file_refs);
5360 trace_io_uring_file_get(ctx, fd);
5361 file = __io_file_get(state, fd);
5362 if (unlikely(!file))
5370 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5371 int fd, unsigned int flags)
5375 if (!io_req_needs_file(req, fd))
5378 fixed = (flags & IOSQE_FIXED_FILE);
5379 if (unlikely(!fixed && req->needs_fixed_file))
5382 return io_file_get(state, req, fd, &req->file, fixed);
5385 static int io_grab_files(struct io_kiocb *req)
5388 struct io_ring_ctx *ctx = req->ctx;
5390 if (req->work.files)
5392 if (!ctx->ring_file)
5396 spin_lock_irq(&ctx->inflight_lock);
5398 * We use the f_ops->flush() handler to ensure that we can flush
5399 * out work accessing these files if the fd is closed. Check if
5400 * the fd has changed since we started down this path, and disallow
5401 * this operation if it has.
5403 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5404 list_add(&req->inflight_entry, &ctx->inflight_list);
5405 req->flags |= REQ_F_INFLIGHT;
5406 req->work.files = current->files;
5409 spin_unlock_irq(&ctx->inflight_lock);
5415 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5417 struct io_timeout_data *data = container_of(timer,
5418 struct io_timeout_data, timer);
5419 struct io_kiocb *req = data->req;
5420 struct io_ring_ctx *ctx = req->ctx;
5421 struct io_kiocb *prev = NULL;
5422 unsigned long flags;
5424 spin_lock_irqsave(&ctx->completion_lock, flags);
5427 * We don't expect the list to be empty, that will only happen if we
5428 * race with the completion of the linked work.
5430 if (!list_empty(&req->link_list)) {
5431 prev = list_entry(req->link_list.prev, struct io_kiocb,
5433 if (refcount_inc_not_zero(&prev->refs)) {
5434 list_del_init(&req->link_list);
5435 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5440 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5443 req_set_fail_links(prev);
5444 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5447 io_cqring_add_event(req, -ETIME);
5450 return HRTIMER_NORESTART;
5453 static void io_queue_linked_timeout(struct io_kiocb *req)
5455 struct io_ring_ctx *ctx = req->ctx;
5458 * If the list is now empty, then our linked request finished before
5459 * we got a chance to setup the timer
5461 spin_lock_irq(&ctx->completion_lock);
5462 if (!list_empty(&req->link_list)) {
5463 struct io_timeout_data *data = &req->io->timeout;
5465 data->timer.function = io_link_timeout_fn;
5466 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5469 spin_unlock_irq(&ctx->completion_lock);
5471 /* drop submission reference */
5475 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5477 struct io_kiocb *nxt;
5479 if (!(req->flags & REQ_F_LINK))
5481 /* for polled retry, if flag is set, we already went through here */
5482 if (req->flags & REQ_F_POLLED)
5485 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5487 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5490 req->flags |= REQ_F_LINK_TIMEOUT;
5494 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5496 struct io_kiocb *linked_timeout;
5497 struct io_kiocb *nxt;
5498 const struct cred *old_creds = NULL;
5502 linked_timeout = io_prep_linked_timeout(req);
5504 if (req->work.creds && req->work.creds != current_cred()) {
5506 revert_creds(old_creds);
5507 if (old_creds == req->work.creds)
5508 old_creds = NULL; /* restored original creds */
5510 old_creds = override_creds(req->work.creds);
5513 ret = io_issue_sqe(req, sqe, true);
5516 * We async punt it if the file wasn't marked NOWAIT, or if the file
5517 * doesn't support non-blocking read/write attempts
5519 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5520 (req->flags & REQ_F_MUST_PUNT))) {
5521 if (io_arm_poll_handler(req)) {
5523 io_queue_linked_timeout(linked_timeout);
5527 if (io_op_defs[req->opcode].file_table) {
5528 ret = io_grab_files(req);
5534 * Queued up for async execution, worker will release
5535 * submit reference when the iocb is actually submitted.
5537 io_queue_async_work(req);
5543 /* drop submission reference */
5544 io_put_req_find_next(req, &nxt);
5546 if (linked_timeout) {
5548 io_queue_linked_timeout(linked_timeout);
5550 io_put_req(linked_timeout);
5553 /* and drop final reference, if we failed */
5555 io_cqring_add_event(req, ret);
5556 req_set_fail_links(req);
5562 if (req->flags & REQ_F_FORCE_ASYNC)
5568 revert_creds(old_creds);
5571 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5575 ret = io_req_defer(req, sqe);
5577 if (ret != -EIOCBQUEUED) {
5579 io_cqring_add_event(req, ret);
5580 req_set_fail_links(req);
5581 io_double_put_req(req);
5583 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5584 ret = io_req_defer_prep(req, sqe);
5585 if (unlikely(ret < 0))
5588 * Never try inline submit of IOSQE_ASYNC is set, go straight
5589 * to async execution.
5591 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5592 io_queue_async_work(req);
5594 __io_queue_sqe(req, sqe);
5598 static inline void io_queue_link_head(struct io_kiocb *req)
5600 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5601 io_cqring_add_event(req, -ECANCELED);
5602 io_double_put_req(req);
5604 io_queue_sqe(req, NULL);
5607 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5608 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5609 IOSQE_BUFFER_SELECT)
5611 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5612 struct io_submit_state *state, struct io_kiocb **link)
5614 struct io_ring_ctx *ctx = req->ctx;
5615 unsigned int sqe_flags;
5618 sqe_flags = READ_ONCE(sqe->flags);
5620 /* enforce forwards compatibility on users */
5621 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
5626 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5627 !io_op_defs[req->opcode].buffer_select) {
5632 id = READ_ONCE(sqe->personality);
5634 req->work.creds = idr_find(&ctx->personality_idr, id);
5635 if (unlikely(!req->work.creds)) {
5639 get_cred(req->work.creds);
5642 /* same numerical values with corresponding REQ_F_*, safe to copy */
5643 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5644 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5645 IOSQE_BUFFER_SELECT);
5647 fd = READ_ONCE(sqe->fd);
5648 ret = io_req_set_file(state, req, fd, sqe_flags);
5649 if (unlikely(ret)) {
5651 io_cqring_add_event(req, ret);
5652 io_double_put_req(req);
5657 * If we already have a head request, queue this one for async
5658 * submittal once the head completes. If we don't have a head but
5659 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5660 * submitted sync once the chain is complete. If none of those
5661 * conditions are true (normal request), then just queue it.
5664 struct io_kiocb *head = *link;
5667 * Taking sequential execution of a link, draining both sides
5668 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5669 * requests in the link. So, it drains the head and the
5670 * next after the link request. The last one is done via
5671 * drain_next flag to persist the effect across calls.
5673 if (sqe_flags & IOSQE_IO_DRAIN) {
5674 head->flags |= REQ_F_IO_DRAIN;
5675 ctx->drain_next = 1;
5677 if (io_alloc_async_ctx(req)) {
5682 ret = io_req_defer_prep(req, sqe);
5684 /* fail even hard links since we don't submit */
5685 head->flags |= REQ_F_FAIL_LINK;
5688 trace_io_uring_link(ctx, req, head);
5689 list_add_tail(&req->link_list, &head->link_list);
5691 /* last request of a link, enqueue the link */
5692 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
5693 io_queue_link_head(head);
5697 if (unlikely(ctx->drain_next)) {
5698 req->flags |= REQ_F_IO_DRAIN;
5699 req->ctx->drain_next = 0;
5701 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
5702 req->flags |= REQ_F_LINK;
5703 INIT_LIST_HEAD(&req->link_list);
5705 if (io_alloc_async_ctx(req)) {
5709 ret = io_req_defer_prep(req, sqe);
5711 req->flags |= REQ_F_FAIL_LINK;
5714 io_queue_sqe(req, sqe);
5722 * Batched submission is done, ensure local IO is flushed out.
5724 static void io_submit_state_end(struct io_submit_state *state)
5726 blk_finish_plug(&state->plug);
5728 if (state->free_reqs)
5729 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5733 * Start submission side cache.
5735 static void io_submit_state_start(struct io_submit_state *state,
5736 unsigned int max_ios)
5738 blk_start_plug(&state->plug);
5739 state->free_reqs = 0;
5741 state->ios_left = max_ios;
5744 static void io_commit_sqring(struct io_ring_ctx *ctx)
5746 struct io_rings *rings = ctx->rings;
5749 * Ensure any loads from the SQEs are done at this point,
5750 * since once we write the new head, the application could
5751 * write new data to them.
5753 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5757 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5758 * that is mapped by userspace. This means that care needs to be taken to
5759 * ensure that reads are stable, as we cannot rely on userspace always
5760 * being a good citizen. If members of the sqe are validated and then later
5761 * used, it's important that those reads are done through READ_ONCE() to
5762 * prevent a re-load down the line.
5764 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5766 u32 *sq_array = ctx->sq_array;
5770 * The cached sq head (or cq tail) serves two purposes:
5772 * 1) allows us to batch the cost of updating the user visible
5774 * 2) allows the kernel side to track the head on its own, even
5775 * though the application is the one updating it.
5777 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5778 if (likely(head < ctx->sq_entries))
5779 return &ctx->sq_sqes[head];
5781 /* drop invalid entries */
5782 ctx->cached_sq_dropped++;
5783 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5787 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5789 ctx->cached_sq_head++;
5792 static void io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5793 const struct io_uring_sqe *sqe)
5796 * All io need record the previous position, if LINK vs DARIN,
5797 * it can be used to mark the position of the first IO in the
5800 req->sequence = ctx->cached_sq_head;
5801 req->opcode = READ_ONCE(sqe->opcode);
5802 req->user_data = READ_ONCE(sqe->user_data);
5807 /* one is dropped after submission, the other at completion */
5808 refcount_set(&req->refs, 2);
5811 INIT_IO_WORK(&req->work, io_wq_submit_work);
5814 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5815 struct file *ring_file, int ring_fd,
5816 struct mm_struct **mm, bool async)
5818 struct io_submit_state state, *statep = NULL;
5819 struct io_kiocb *link = NULL;
5820 int i, submitted = 0;
5821 bool mm_fault = false;
5823 /* if we have a backlog and couldn't flush it all, return BUSY */
5824 if (test_bit(0, &ctx->sq_check_overflow)) {
5825 if (!list_empty(&ctx->cq_overflow_list) &&
5826 !io_cqring_overflow_flush(ctx, false))
5830 /* make sure SQ entry isn't read before tail */
5831 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5833 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5836 if (nr > IO_PLUG_THRESHOLD) {
5837 io_submit_state_start(&state, nr);
5841 ctx->ring_fd = ring_fd;
5842 ctx->ring_file = ring_file;
5844 for (i = 0; i < nr; i++) {
5845 const struct io_uring_sqe *sqe;
5846 struct io_kiocb *req;
5849 sqe = io_get_sqe(ctx);
5850 if (unlikely(!sqe)) {
5851 io_consume_sqe(ctx);
5854 req = io_alloc_req(ctx, statep);
5855 if (unlikely(!req)) {
5857 submitted = -EAGAIN;
5861 io_init_req(ctx, req, sqe);
5862 io_consume_sqe(ctx);
5863 /* will complete beyond this point, count as submitted */
5866 if (unlikely(req->opcode >= IORING_OP_LAST)) {
5869 io_cqring_add_event(req, err);
5870 io_double_put_req(req);
5874 if (io_op_defs[req->opcode].needs_mm && !*mm) {
5875 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
5876 if (unlikely(mm_fault)) {
5880 use_mm(ctx->sqo_mm);
5884 req->needs_fixed_file = async;
5885 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5887 if (!io_submit_sqe(req, sqe, statep, &link))
5891 if (unlikely(submitted != nr)) {
5892 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5894 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5897 io_queue_link_head(link);
5899 io_submit_state_end(&state);
5901 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5902 io_commit_sqring(ctx);
5907 static int io_sq_thread(void *data)
5909 struct io_ring_ctx *ctx = data;
5910 struct mm_struct *cur_mm = NULL;
5911 const struct cred *old_cred;
5912 mm_segment_t old_fs;
5914 unsigned long timeout;
5917 complete(&ctx->completions[1]);
5921 old_cred = override_creds(ctx->creds);
5923 timeout = jiffies + ctx->sq_thread_idle;
5924 while (!kthread_should_park()) {
5925 unsigned int to_submit;
5927 if (!list_empty(&ctx->poll_list)) {
5928 unsigned nr_events = 0;
5930 mutex_lock(&ctx->uring_lock);
5931 if (!list_empty(&ctx->poll_list))
5932 io_iopoll_getevents(ctx, &nr_events, 0);
5934 timeout = jiffies + ctx->sq_thread_idle;
5935 mutex_unlock(&ctx->uring_lock);
5938 to_submit = io_sqring_entries(ctx);
5941 * If submit got -EBUSY, flag us as needing the application
5942 * to enter the kernel to reap and flush events.
5944 if (!to_submit || ret == -EBUSY) {
5946 * Drop cur_mm before scheduling, we can't hold it for
5947 * long periods (or over schedule()). Do this before
5948 * adding ourselves to the waitqueue, as the unuse/drop
5958 * We're polling. If we're within the defined idle
5959 * period, then let us spin without work before going
5960 * to sleep. The exception is if we got EBUSY doing
5961 * more IO, we should wait for the application to
5962 * reap events and wake us up.
5964 if (!list_empty(&ctx->poll_list) ||
5965 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5966 !percpu_ref_is_dying(&ctx->refs))) {
5967 if (current->task_works)
5973 prepare_to_wait(&ctx->sqo_wait, &wait,
5974 TASK_INTERRUPTIBLE);
5977 * While doing polled IO, before going to sleep, we need
5978 * to check if there are new reqs added to poll_list, it
5979 * is because reqs may have been punted to io worker and
5980 * will be added to poll_list later, hence check the
5983 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
5984 !list_empty_careful(&ctx->poll_list)) {
5985 finish_wait(&ctx->sqo_wait, &wait);
5989 /* Tell userspace we may need a wakeup call */
5990 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5991 /* make sure to read SQ tail after writing flags */
5994 to_submit = io_sqring_entries(ctx);
5995 if (!to_submit || ret == -EBUSY) {
5996 if (kthread_should_park()) {
5997 finish_wait(&ctx->sqo_wait, &wait);
6000 if (current->task_works) {
6002 finish_wait(&ctx->sqo_wait, &wait);
6005 if (signal_pending(current))
6006 flush_signals(current);
6008 finish_wait(&ctx->sqo_wait, &wait);
6010 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6013 finish_wait(&ctx->sqo_wait, &wait);
6015 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6018 mutex_lock(&ctx->uring_lock);
6019 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
6020 mutex_unlock(&ctx->uring_lock);
6021 timeout = jiffies + ctx->sq_thread_idle;
6024 if (current->task_works)
6032 revert_creds(old_cred);
6039 struct io_wait_queue {
6040 struct wait_queue_entry wq;
6041 struct io_ring_ctx *ctx;
6043 unsigned nr_timeouts;
6046 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6048 struct io_ring_ctx *ctx = iowq->ctx;
6051 * Wake up if we have enough events, or if a timeout occurred since we
6052 * started waiting. For timeouts, we always want to return to userspace,
6053 * regardless of event count.
6055 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6056 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6059 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6060 int wake_flags, void *key)
6062 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6065 /* use noflush == true, as we can't safely rely on locking context */
6066 if (!io_should_wake(iowq, true))
6069 return autoremove_wake_function(curr, mode, wake_flags, key);
6073 * Wait until events become available, if we don't already have some. The
6074 * application must reap them itself, as they reside on the shared cq ring.
6076 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6077 const sigset_t __user *sig, size_t sigsz)
6079 struct io_wait_queue iowq = {
6082 .func = io_wake_function,
6083 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6086 .to_wait = min_events,
6088 struct io_rings *rings = ctx->rings;
6092 if (io_cqring_events(ctx, false) >= min_events)
6094 if (!current->task_works)
6100 #ifdef CONFIG_COMPAT
6101 if (in_compat_syscall())
6102 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6106 ret = set_user_sigmask(sig, sigsz);
6112 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6113 trace_io_uring_cqring_wait(ctx, min_events);
6115 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6116 TASK_INTERRUPTIBLE);
6117 if (current->task_works)
6119 if (io_should_wake(&iowq, false))
6122 if (signal_pending(current)) {
6127 finish_wait(&ctx->wait, &iowq.wq);
6129 restore_saved_sigmask_unless(ret == -EINTR);
6131 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6134 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6136 #if defined(CONFIG_UNIX)
6137 if (ctx->ring_sock) {
6138 struct sock *sock = ctx->ring_sock->sk;
6139 struct sk_buff *skb;
6141 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6147 for (i = 0; i < ctx->nr_user_files; i++) {
6150 file = io_file_from_index(ctx, i);
6157 static void io_file_ref_kill(struct percpu_ref *ref)
6159 struct fixed_file_data *data;
6161 data = container_of(ref, struct fixed_file_data, refs);
6162 complete(&data->done);
6165 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6167 struct fixed_file_data *data = ctx->file_data;
6168 struct fixed_file_ref_node *ref_node = NULL;
6169 unsigned nr_tables, i;
6170 unsigned long flags;
6175 spin_lock_irqsave(&data->lock, flags);
6176 if (!list_empty(&data->ref_list))
6177 ref_node = list_first_entry(&data->ref_list,
6178 struct fixed_file_ref_node, node);
6179 spin_unlock_irqrestore(&data->lock, flags);
6181 percpu_ref_kill(&ref_node->refs);
6183 percpu_ref_kill(&data->refs);
6185 /* wait for all refs nodes to complete */
6186 wait_for_completion(&data->done);
6188 __io_sqe_files_unregister(ctx);
6189 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6190 for (i = 0; i < nr_tables; i++)
6191 kfree(data->table[i].files);
6193 percpu_ref_exit(&data->refs);
6195 ctx->file_data = NULL;
6196 ctx->nr_user_files = 0;
6200 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6202 if (ctx->sqo_thread) {
6203 wait_for_completion(&ctx->completions[1]);
6205 * The park is a bit of a work-around, without it we get
6206 * warning spews on shutdown with SQPOLL set and affinity
6207 * set to a single CPU.
6209 kthread_park(ctx->sqo_thread);
6210 kthread_stop(ctx->sqo_thread);
6211 ctx->sqo_thread = NULL;
6215 static void io_finish_async(struct io_ring_ctx *ctx)
6217 io_sq_thread_stop(ctx);
6220 io_wq_destroy(ctx->io_wq);
6225 #if defined(CONFIG_UNIX)
6227 * Ensure the UNIX gc is aware of our file set, so we are certain that
6228 * the io_uring can be safely unregistered on process exit, even if we have
6229 * loops in the file referencing.
6231 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6233 struct sock *sk = ctx->ring_sock->sk;
6234 struct scm_fp_list *fpl;
6235 struct sk_buff *skb;
6238 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6242 skb = alloc_skb(0, GFP_KERNEL);
6251 fpl->user = get_uid(ctx->user);
6252 for (i = 0; i < nr; i++) {
6253 struct file *file = io_file_from_index(ctx, i + offset);
6257 fpl->fp[nr_files] = get_file(file);
6258 unix_inflight(fpl->user, fpl->fp[nr_files]);
6263 fpl->max = SCM_MAX_FD;
6264 fpl->count = nr_files;
6265 UNIXCB(skb).fp = fpl;
6266 skb->destructor = unix_destruct_scm;
6267 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6268 skb_queue_head(&sk->sk_receive_queue, skb);
6270 for (i = 0; i < nr_files; i++)
6281 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6282 * causes regular reference counting to break down. We rely on the UNIX
6283 * garbage collection to take care of this problem for us.
6285 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6287 unsigned left, total;
6291 left = ctx->nr_user_files;
6293 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6295 ret = __io_sqe_files_scm(ctx, this_files, total);
6299 total += this_files;
6305 while (total < ctx->nr_user_files) {
6306 struct file *file = io_file_from_index(ctx, total);
6316 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6322 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6327 for (i = 0; i < nr_tables; i++) {
6328 struct fixed_file_table *table = &ctx->file_data->table[i];
6329 unsigned this_files;
6331 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6332 table->files = kcalloc(this_files, sizeof(struct file *),
6336 nr_files -= this_files;
6342 for (i = 0; i < nr_tables; i++) {
6343 struct fixed_file_table *table = &ctx->file_data->table[i];
6344 kfree(table->files);
6349 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6351 #if defined(CONFIG_UNIX)
6352 struct sock *sock = ctx->ring_sock->sk;
6353 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6354 struct sk_buff *skb;
6357 __skb_queue_head_init(&list);
6360 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6361 * remove this entry and rearrange the file array.
6363 skb = skb_dequeue(head);
6365 struct scm_fp_list *fp;
6367 fp = UNIXCB(skb).fp;
6368 for (i = 0; i < fp->count; i++) {
6371 if (fp->fp[i] != file)
6374 unix_notinflight(fp->user, fp->fp[i]);
6375 left = fp->count - 1 - i;
6377 memmove(&fp->fp[i], &fp->fp[i + 1],
6378 left * sizeof(struct file *));
6385 __skb_queue_tail(&list, skb);
6395 __skb_queue_tail(&list, skb);
6397 skb = skb_dequeue(head);
6400 if (skb_peek(&list)) {
6401 spin_lock_irq(&head->lock);
6402 while ((skb = __skb_dequeue(&list)) != NULL)
6403 __skb_queue_tail(head, skb);
6404 spin_unlock_irq(&head->lock);
6411 struct io_file_put {
6412 struct list_head list;
6416 static void io_file_put_work(struct work_struct *work)
6418 struct fixed_file_ref_node *ref_node;
6419 struct fixed_file_data *file_data;
6420 struct io_ring_ctx *ctx;
6421 struct io_file_put *pfile, *tmp;
6422 unsigned long flags;
6424 ref_node = container_of(work, struct fixed_file_ref_node, work);
6425 file_data = ref_node->file_data;
6426 ctx = file_data->ctx;
6428 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6429 list_del_init(&pfile->list);
6430 io_ring_file_put(ctx, pfile->file);
6434 spin_lock_irqsave(&file_data->lock, flags);
6435 list_del_init(&ref_node->node);
6436 spin_unlock_irqrestore(&file_data->lock, flags);
6438 percpu_ref_exit(&ref_node->refs);
6440 percpu_ref_put(&file_data->refs);
6443 static void io_file_data_ref_zero(struct percpu_ref *ref)
6445 struct fixed_file_ref_node *ref_node;
6447 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6449 queue_work(system_wq, &ref_node->work);
6452 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6453 struct io_ring_ctx *ctx)
6455 struct fixed_file_ref_node *ref_node;
6457 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6459 return ERR_PTR(-ENOMEM);
6461 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6464 return ERR_PTR(-ENOMEM);
6466 INIT_LIST_HEAD(&ref_node->node);
6467 INIT_LIST_HEAD(&ref_node->file_list);
6468 INIT_WORK(&ref_node->work, io_file_put_work);
6469 ref_node->file_data = ctx->file_data;
6474 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6476 percpu_ref_exit(&ref_node->refs);
6480 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6483 __s32 __user *fds = (__s32 __user *) arg;
6488 struct fixed_file_ref_node *ref_node;
6489 unsigned long flags;
6495 if (nr_args > IORING_MAX_FIXED_FILES)
6498 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6499 if (!ctx->file_data)
6501 ctx->file_data->ctx = ctx;
6502 init_completion(&ctx->file_data->done);
6503 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6504 spin_lock_init(&ctx->file_data->lock);
6506 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6507 ctx->file_data->table = kcalloc(nr_tables,
6508 sizeof(struct fixed_file_table),
6510 if (!ctx->file_data->table) {
6511 kfree(ctx->file_data);
6512 ctx->file_data = NULL;
6516 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6517 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6518 kfree(ctx->file_data->table);
6519 kfree(ctx->file_data);
6520 ctx->file_data = NULL;
6524 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6525 percpu_ref_exit(&ctx->file_data->refs);
6526 kfree(ctx->file_data->table);
6527 kfree(ctx->file_data);
6528 ctx->file_data = NULL;
6532 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6533 struct fixed_file_table *table;
6537 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6539 /* allow sparse sets */
6545 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6546 index = i & IORING_FILE_TABLE_MASK;
6554 * Don't allow io_uring instances to be registered. If UNIX
6555 * isn't enabled, then this causes a reference cycle and this
6556 * instance can never get freed. If UNIX is enabled we'll
6557 * handle it just fine, but there's still no point in allowing
6558 * a ring fd as it doesn't support regular read/write anyway.
6560 if (file->f_op == &io_uring_fops) {
6565 table->files[index] = file;
6569 for (i = 0; i < ctx->nr_user_files; i++) {
6570 file = io_file_from_index(ctx, i);
6574 for (i = 0; i < nr_tables; i++)
6575 kfree(ctx->file_data->table[i].files);
6577 kfree(ctx->file_data->table);
6578 kfree(ctx->file_data);
6579 ctx->file_data = NULL;
6580 ctx->nr_user_files = 0;
6584 ret = io_sqe_files_scm(ctx);
6586 io_sqe_files_unregister(ctx);
6590 ref_node = alloc_fixed_file_ref_node(ctx);
6591 if (IS_ERR(ref_node)) {
6592 io_sqe_files_unregister(ctx);
6593 return PTR_ERR(ref_node);
6596 ctx->file_data->cur_refs = &ref_node->refs;
6597 spin_lock_irqsave(&ctx->file_data->lock, flags);
6598 list_add(&ref_node->node, &ctx->file_data->ref_list);
6599 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6600 percpu_ref_get(&ctx->file_data->refs);
6604 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6607 #if defined(CONFIG_UNIX)
6608 struct sock *sock = ctx->ring_sock->sk;
6609 struct sk_buff_head *head = &sock->sk_receive_queue;
6610 struct sk_buff *skb;
6613 * See if we can merge this file into an existing skb SCM_RIGHTS
6614 * file set. If there's no room, fall back to allocating a new skb
6615 * and filling it in.
6617 spin_lock_irq(&head->lock);
6618 skb = skb_peek(head);
6620 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6622 if (fpl->count < SCM_MAX_FD) {
6623 __skb_unlink(skb, head);
6624 spin_unlock_irq(&head->lock);
6625 fpl->fp[fpl->count] = get_file(file);
6626 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6628 spin_lock_irq(&head->lock);
6629 __skb_queue_head(head, skb);
6634 spin_unlock_irq(&head->lock);
6641 return __io_sqe_files_scm(ctx, 1, index);
6647 static int io_queue_file_removal(struct fixed_file_data *data,
6650 struct io_file_put *pfile;
6651 struct percpu_ref *refs = data->cur_refs;
6652 struct fixed_file_ref_node *ref_node;
6654 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6658 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6660 list_add(&pfile->list, &ref_node->file_list);
6665 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6666 struct io_uring_files_update *up,
6669 struct fixed_file_data *data = ctx->file_data;
6670 struct fixed_file_ref_node *ref_node;
6675 unsigned long flags;
6676 bool needs_switch = false;
6678 if (check_add_overflow(up->offset, nr_args, &done))
6680 if (done > ctx->nr_user_files)
6683 ref_node = alloc_fixed_file_ref_node(ctx);
6684 if (IS_ERR(ref_node))
6685 return PTR_ERR(ref_node);
6688 fds = u64_to_user_ptr(up->fds);
6690 struct fixed_file_table *table;
6694 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6698 i = array_index_nospec(up->offset, ctx->nr_user_files);
6699 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6700 index = i & IORING_FILE_TABLE_MASK;
6701 if (table->files[index]) {
6702 file = io_file_from_index(ctx, index);
6703 err = io_queue_file_removal(data, file);
6706 table->files[index] = NULL;
6707 needs_switch = true;
6716 * Don't allow io_uring instances to be registered. If
6717 * UNIX isn't enabled, then this causes a reference
6718 * cycle and this instance can never get freed. If UNIX
6719 * is enabled we'll handle it just fine, but there's
6720 * still no point in allowing a ring fd as it doesn't
6721 * support regular read/write anyway.
6723 if (file->f_op == &io_uring_fops) {
6728 table->files[index] = file;
6729 err = io_sqe_file_register(ctx, file, i);
6739 percpu_ref_kill(data->cur_refs);
6740 spin_lock_irqsave(&data->lock, flags);
6741 list_add(&ref_node->node, &data->ref_list);
6742 data->cur_refs = &ref_node->refs;
6743 spin_unlock_irqrestore(&data->lock, flags);
6744 percpu_ref_get(&ctx->file_data->refs);
6746 destroy_fixed_file_ref_node(ref_node);
6748 return done ? done : err;
6751 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6754 struct io_uring_files_update up;
6756 if (!ctx->file_data)
6760 if (copy_from_user(&up, arg, sizeof(up)))
6765 return __io_sqe_files_update(ctx, &up, nr_args);
6768 static void io_free_work(struct io_wq_work *work)
6770 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6772 /* Consider that io_steal_work() relies on this ref */
6776 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6777 struct io_uring_params *p)
6779 struct io_wq_data data;
6781 struct io_ring_ctx *ctx_attach;
6782 unsigned int concurrency;
6785 data.user = ctx->user;
6786 data.free_work = io_free_work;
6788 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6789 /* Do QD, or 4 * CPUS, whatever is smallest */
6790 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6792 ctx->io_wq = io_wq_create(concurrency, &data);
6793 if (IS_ERR(ctx->io_wq)) {
6794 ret = PTR_ERR(ctx->io_wq);
6800 f = fdget(p->wq_fd);
6804 if (f.file->f_op != &io_uring_fops) {
6809 ctx_attach = f.file->private_data;
6810 /* @io_wq is protected by holding the fd */
6811 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6816 ctx->io_wq = ctx_attach->io_wq;
6822 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6823 struct io_uring_params *p)
6827 init_waitqueue_head(&ctx->sqo_wait);
6828 mmgrab(current->mm);
6829 ctx->sqo_mm = current->mm;
6831 if (ctx->flags & IORING_SETUP_SQPOLL) {
6833 if (!capable(CAP_SYS_ADMIN))
6836 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6837 if (!ctx->sq_thread_idle)
6838 ctx->sq_thread_idle = HZ;
6840 if (p->flags & IORING_SETUP_SQ_AFF) {
6841 int cpu = p->sq_thread_cpu;
6844 if (cpu >= nr_cpu_ids)
6846 if (!cpu_online(cpu))
6849 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6853 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6856 if (IS_ERR(ctx->sqo_thread)) {
6857 ret = PTR_ERR(ctx->sqo_thread);
6858 ctx->sqo_thread = NULL;
6861 wake_up_process(ctx->sqo_thread);
6862 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6863 /* Can't have SQ_AFF without SQPOLL */
6868 ret = io_init_wq_offload(ctx, p);
6874 io_finish_async(ctx);
6875 mmdrop(ctx->sqo_mm);
6880 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6882 atomic_long_sub(nr_pages, &user->locked_vm);
6885 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6887 unsigned long page_limit, cur_pages, new_pages;
6889 /* Don't allow more pages than we can safely lock */
6890 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6893 cur_pages = atomic_long_read(&user->locked_vm);
6894 new_pages = cur_pages + nr_pages;
6895 if (new_pages > page_limit)
6897 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6898 new_pages) != cur_pages);
6903 static void io_mem_free(void *ptr)
6910 page = virt_to_head_page(ptr);
6911 if (put_page_testzero(page))
6912 free_compound_page(page);
6915 static void *io_mem_alloc(size_t size)
6917 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6920 return (void *) __get_free_pages(gfp_flags, get_order(size));
6923 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6926 struct io_rings *rings;
6927 size_t off, sq_array_size;
6929 off = struct_size(rings, cqes, cq_entries);
6930 if (off == SIZE_MAX)
6934 off = ALIGN(off, SMP_CACHE_BYTES);
6939 sq_array_size = array_size(sizeof(u32), sq_entries);
6940 if (sq_array_size == SIZE_MAX)
6943 if (check_add_overflow(off, sq_array_size, &off))
6952 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6956 pages = (size_t)1 << get_order(
6957 rings_size(sq_entries, cq_entries, NULL));
6958 pages += (size_t)1 << get_order(
6959 array_size(sizeof(struct io_uring_sqe), sq_entries));
6964 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6968 if (!ctx->user_bufs)
6971 for (i = 0; i < ctx->nr_user_bufs; i++) {
6972 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6974 for (j = 0; j < imu->nr_bvecs; j++)
6975 unpin_user_page(imu->bvec[j].bv_page);
6977 if (ctx->account_mem)
6978 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6983 kfree(ctx->user_bufs);
6984 ctx->user_bufs = NULL;
6985 ctx->nr_user_bufs = 0;
6989 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6990 void __user *arg, unsigned index)
6992 struct iovec __user *src;
6994 #ifdef CONFIG_COMPAT
6996 struct compat_iovec __user *ciovs;
6997 struct compat_iovec ciov;
6999 ciovs = (struct compat_iovec __user *) arg;
7000 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7003 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7004 dst->iov_len = ciov.iov_len;
7008 src = (struct iovec __user *) arg;
7009 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7014 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7017 struct vm_area_struct **vmas = NULL;
7018 struct page **pages = NULL;
7019 int i, j, got_pages = 0;
7024 if (!nr_args || nr_args > UIO_MAXIOV)
7027 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7029 if (!ctx->user_bufs)
7032 for (i = 0; i < nr_args; i++) {
7033 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7034 unsigned long off, start, end, ubuf;
7039 ret = io_copy_iov(ctx, &iov, arg, i);
7044 * Don't impose further limits on the size and buffer
7045 * constraints here, we'll -EINVAL later when IO is
7046 * submitted if they are wrong.
7049 if (!iov.iov_base || !iov.iov_len)
7052 /* arbitrary limit, but we need something */
7053 if (iov.iov_len > SZ_1G)
7056 ubuf = (unsigned long) iov.iov_base;
7057 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7058 start = ubuf >> PAGE_SHIFT;
7059 nr_pages = end - start;
7061 if (ctx->account_mem) {
7062 ret = io_account_mem(ctx->user, nr_pages);
7068 if (!pages || nr_pages > got_pages) {
7071 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7073 vmas = kvmalloc_array(nr_pages,
7074 sizeof(struct vm_area_struct *),
7076 if (!pages || !vmas) {
7078 if (ctx->account_mem)
7079 io_unaccount_mem(ctx->user, nr_pages);
7082 got_pages = nr_pages;
7085 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7089 if (ctx->account_mem)
7090 io_unaccount_mem(ctx->user, nr_pages);
7095 down_read(¤t->mm->mmap_sem);
7096 pret = pin_user_pages(ubuf, nr_pages,
7097 FOLL_WRITE | FOLL_LONGTERM,
7099 if (pret == nr_pages) {
7100 /* don't support file backed memory */
7101 for (j = 0; j < nr_pages; j++) {
7102 struct vm_area_struct *vma = vmas[j];
7105 !is_file_hugepages(vma->vm_file)) {
7111 ret = pret < 0 ? pret : -EFAULT;
7113 up_read(¤t->mm->mmap_sem);
7116 * if we did partial map, or found file backed vmas,
7117 * release any pages we did get
7120 unpin_user_pages(pages, pret);
7121 if (ctx->account_mem)
7122 io_unaccount_mem(ctx->user, nr_pages);
7127 off = ubuf & ~PAGE_MASK;
7129 for (j = 0; j < nr_pages; j++) {
7132 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7133 imu->bvec[j].bv_page = pages[j];
7134 imu->bvec[j].bv_len = vec_len;
7135 imu->bvec[j].bv_offset = off;
7139 /* store original address for later verification */
7141 imu->len = iov.iov_len;
7142 imu->nr_bvecs = nr_pages;
7144 ctx->nr_user_bufs++;
7152 io_sqe_buffer_unregister(ctx);
7156 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7158 __s32 __user *fds = arg;
7164 if (copy_from_user(&fd, fds, sizeof(*fds)))
7167 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7168 if (IS_ERR(ctx->cq_ev_fd)) {
7169 int ret = PTR_ERR(ctx->cq_ev_fd);
7170 ctx->cq_ev_fd = NULL;
7177 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7179 if (ctx->cq_ev_fd) {
7180 eventfd_ctx_put(ctx->cq_ev_fd);
7181 ctx->cq_ev_fd = NULL;
7188 static int __io_destroy_buffers(int id, void *p, void *data)
7190 struct io_ring_ctx *ctx = data;
7191 struct io_buffer *buf = p;
7193 __io_remove_buffers(ctx, buf, id, -1U);
7197 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7199 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7200 idr_destroy(&ctx->io_buffer_idr);
7203 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7205 io_finish_async(ctx);
7207 mmdrop(ctx->sqo_mm);
7209 io_iopoll_reap_events(ctx);
7210 io_sqe_buffer_unregister(ctx);
7211 io_sqe_files_unregister(ctx);
7212 io_eventfd_unregister(ctx);
7213 io_destroy_buffers(ctx);
7214 idr_destroy(&ctx->personality_idr);
7216 #if defined(CONFIG_UNIX)
7217 if (ctx->ring_sock) {
7218 ctx->ring_sock->file = NULL; /* so that iput() is called */
7219 sock_release(ctx->ring_sock);
7223 io_mem_free(ctx->rings);
7224 io_mem_free(ctx->sq_sqes);
7226 percpu_ref_exit(&ctx->refs);
7227 if (ctx->account_mem)
7228 io_unaccount_mem(ctx->user,
7229 ring_pages(ctx->sq_entries, ctx->cq_entries));
7230 free_uid(ctx->user);
7231 put_cred(ctx->creds);
7232 kfree(ctx->completions);
7233 kfree(ctx->cancel_hash);
7234 kmem_cache_free(req_cachep, ctx->fallback_req);
7238 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7240 struct io_ring_ctx *ctx = file->private_data;
7243 poll_wait(file, &ctx->cq_wait, wait);
7245 * synchronizes with barrier from wq_has_sleeper call in
7249 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7250 ctx->rings->sq_ring_entries)
7251 mask |= EPOLLOUT | EPOLLWRNORM;
7252 if (io_cqring_events(ctx, false))
7253 mask |= EPOLLIN | EPOLLRDNORM;
7258 static int io_uring_fasync(int fd, struct file *file, int on)
7260 struct io_ring_ctx *ctx = file->private_data;
7262 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7265 static int io_remove_personalities(int id, void *p, void *data)
7267 struct io_ring_ctx *ctx = data;
7268 const struct cred *cred;
7270 cred = idr_remove(&ctx->personality_idr, id);
7276 static void io_ring_exit_work(struct work_struct *work)
7278 struct io_ring_ctx *ctx;
7280 ctx = container_of(work, struct io_ring_ctx, exit_work);
7282 io_cqring_overflow_flush(ctx, true);
7284 wait_for_completion(&ctx->completions[0]);
7285 io_ring_ctx_free(ctx);
7288 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7290 mutex_lock(&ctx->uring_lock);
7291 percpu_ref_kill(&ctx->refs);
7292 mutex_unlock(&ctx->uring_lock);
7295 * Wait for sq thread to idle, if we have one. It won't spin on new
7296 * work after we've killed the ctx ref above. This is important to do
7297 * before we cancel existing commands, as the thread could otherwise
7298 * be queueing new work post that. If that's work we need to cancel,
7299 * it could cause shutdown to hang.
7301 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7304 io_kill_timeouts(ctx);
7305 io_poll_remove_all(ctx);
7308 io_wq_cancel_all(ctx->io_wq);
7310 io_iopoll_reap_events(ctx);
7311 /* if we failed setting up the ctx, we might not have any rings */
7313 io_cqring_overflow_flush(ctx, true);
7314 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7315 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7316 queue_work(system_wq, &ctx->exit_work);
7319 static int io_uring_release(struct inode *inode, struct file *file)
7321 struct io_ring_ctx *ctx = file->private_data;
7323 file->private_data = NULL;
7324 io_ring_ctx_wait_and_kill(ctx);
7328 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7329 struct files_struct *files)
7331 struct io_kiocb *req;
7334 while (!list_empty_careful(&ctx->inflight_list)) {
7335 struct io_kiocb *cancel_req = NULL;
7337 spin_lock_irq(&ctx->inflight_lock);
7338 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7339 if (req->work.files != files)
7341 /* req is being completed, ignore */
7342 if (!refcount_inc_not_zero(&req->refs))
7348 prepare_to_wait(&ctx->inflight_wait, &wait,
7349 TASK_UNINTERRUPTIBLE);
7350 spin_unlock_irq(&ctx->inflight_lock);
7352 /* We need to keep going until we don't find a matching req */
7356 if (cancel_req->flags & REQ_F_OVERFLOW) {
7357 spin_lock_irq(&ctx->completion_lock);
7358 list_del(&cancel_req->list);
7359 cancel_req->flags &= ~REQ_F_OVERFLOW;
7360 if (list_empty(&ctx->cq_overflow_list)) {
7361 clear_bit(0, &ctx->sq_check_overflow);
7362 clear_bit(0, &ctx->cq_check_overflow);
7364 spin_unlock_irq(&ctx->completion_lock);
7366 WRITE_ONCE(ctx->rings->cq_overflow,
7367 atomic_inc_return(&ctx->cached_cq_overflow));
7370 * Put inflight ref and overflow ref. If that's
7371 * all we had, then we're done with this request.
7373 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7374 io_put_req(cancel_req);
7379 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7380 io_put_req(cancel_req);
7383 finish_wait(&ctx->inflight_wait, &wait);
7386 static int io_uring_flush(struct file *file, void *data)
7388 struct io_ring_ctx *ctx = file->private_data;
7390 io_uring_cancel_files(ctx, data);
7393 * If the task is going away, cancel work it may have pending
7395 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7396 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7401 static void *io_uring_validate_mmap_request(struct file *file,
7402 loff_t pgoff, size_t sz)
7404 struct io_ring_ctx *ctx = file->private_data;
7405 loff_t offset = pgoff << PAGE_SHIFT;
7410 case IORING_OFF_SQ_RING:
7411 case IORING_OFF_CQ_RING:
7414 case IORING_OFF_SQES:
7418 return ERR_PTR(-EINVAL);
7421 page = virt_to_head_page(ptr);
7422 if (sz > page_size(page))
7423 return ERR_PTR(-EINVAL);
7430 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7432 size_t sz = vma->vm_end - vma->vm_start;
7436 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7438 return PTR_ERR(ptr);
7440 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7441 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7444 #else /* !CONFIG_MMU */
7446 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7448 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7451 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7453 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7456 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7457 unsigned long addr, unsigned long len,
7458 unsigned long pgoff, unsigned long flags)
7462 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7464 return PTR_ERR(ptr);
7466 return (unsigned long) ptr;
7469 #endif /* !CONFIG_MMU */
7471 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7472 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7475 struct io_ring_ctx *ctx;
7480 if (current->task_works)
7483 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7491 if (f.file->f_op != &io_uring_fops)
7495 ctx = f.file->private_data;
7496 if (!percpu_ref_tryget(&ctx->refs))
7500 * For SQ polling, the thread will do all submissions and completions.
7501 * Just return the requested submit count, and wake the thread if
7505 if (ctx->flags & IORING_SETUP_SQPOLL) {
7506 if (!list_empty_careful(&ctx->cq_overflow_list))
7507 io_cqring_overflow_flush(ctx, false);
7508 if (flags & IORING_ENTER_SQ_WAKEUP)
7509 wake_up(&ctx->sqo_wait);
7510 submitted = to_submit;
7511 } else if (to_submit) {
7512 struct mm_struct *cur_mm;
7514 mutex_lock(&ctx->uring_lock);
7515 /* already have mm, so io_submit_sqes() won't try to grab it */
7516 cur_mm = ctx->sqo_mm;
7517 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
7519 mutex_unlock(&ctx->uring_lock);
7521 if (submitted != to_submit)
7524 if (flags & IORING_ENTER_GETEVENTS) {
7525 unsigned nr_events = 0;
7527 min_complete = min(min_complete, ctx->cq_entries);
7530 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7531 * space applications don't need to do io completion events
7532 * polling again, they can rely on io_sq_thread to do polling
7533 * work, which can reduce cpu usage and uring_lock contention.
7535 if (ctx->flags & IORING_SETUP_IOPOLL &&
7536 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7537 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7539 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7544 percpu_ref_put(&ctx->refs);
7547 return submitted ? submitted : ret;
7550 #ifdef CONFIG_PROC_FS
7551 static int io_uring_show_cred(int id, void *p, void *data)
7553 const struct cred *cred = p;
7554 struct seq_file *m = data;
7555 struct user_namespace *uns = seq_user_ns(m);
7556 struct group_info *gi;
7561 seq_printf(m, "%5d\n", id);
7562 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7563 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7564 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7565 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7566 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7567 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7568 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7569 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7570 seq_puts(m, "\n\tGroups:\t");
7571 gi = cred->group_info;
7572 for (g = 0; g < gi->ngroups; g++) {
7573 seq_put_decimal_ull(m, g ? " " : "",
7574 from_kgid_munged(uns, gi->gid[g]));
7576 seq_puts(m, "\n\tCapEff:\t");
7577 cap = cred->cap_effective;
7578 CAP_FOR_EACH_U32(__capi)
7579 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7584 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7588 mutex_lock(&ctx->uring_lock);
7589 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7590 for (i = 0; i < ctx->nr_user_files; i++) {
7591 struct fixed_file_table *table;
7594 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7595 f = table->files[i & IORING_FILE_TABLE_MASK];
7597 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7599 seq_printf(m, "%5u: <none>\n", i);
7601 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7602 for (i = 0; i < ctx->nr_user_bufs; i++) {
7603 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7605 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7606 (unsigned int) buf->len);
7608 if (!idr_is_empty(&ctx->personality_idr)) {
7609 seq_printf(m, "Personalities:\n");
7610 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7612 seq_printf(m, "PollList:\n");
7613 spin_lock_irq(&ctx->completion_lock);
7614 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7615 struct hlist_head *list = &ctx->cancel_hash[i];
7616 struct io_kiocb *req;
7618 hlist_for_each_entry(req, list, hash_node)
7619 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7620 req->task->task_works != NULL);
7622 spin_unlock_irq(&ctx->completion_lock);
7623 mutex_unlock(&ctx->uring_lock);
7626 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7628 struct io_ring_ctx *ctx = f->private_data;
7630 if (percpu_ref_tryget(&ctx->refs)) {
7631 __io_uring_show_fdinfo(ctx, m);
7632 percpu_ref_put(&ctx->refs);
7637 static const struct file_operations io_uring_fops = {
7638 .release = io_uring_release,
7639 .flush = io_uring_flush,
7640 .mmap = io_uring_mmap,
7642 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7643 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7645 .poll = io_uring_poll,
7646 .fasync = io_uring_fasync,
7647 #ifdef CONFIG_PROC_FS
7648 .show_fdinfo = io_uring_show_fdinfo,
7652 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7653 struct io_uring_params *p)
7655 struct io_rings *rings;
7656 size_t size, sq_array_offset;
7658 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7659 if (size == SIZE_MAX)
7662 rings = io_mem_alloc(size);
7667 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7668 rings->sq_ring_mask = p->sq_entries - 1;
7669 rings->cq_ring_mask = p->cq_entries - 1;
7670 rings->sq_ring_entries = p->sq_entries;
7671 rings->cq_ring_entries = p->cq_entries;
7672 ctx->sq_mask = rings->sq_ring_mask;
7673 ctx->cq_mask = rings->cq_ring_mask;
7674 ctx->sq_entries = rings->sq_ring_entries;
7675 ctx->cq_entries = rings->cq_ring_entries;
7677 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7678 if (size == SIZE_MAX) {
7679 io_mem_free(ctx->rings);
7684 ctx->sq_sqes = io_mem_alloc(size);
7685 if (!ctx->sq_sqes) {
7686 io_mem_free(ctx->rings);
7695 * Allocate an anonymous fd, this is what constitutes the application
7696 * visible backing of an io_uring instance. The application mmaps this
7697 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7698 * we have to tie this fd to a socket for file garbage collection purposes.
7700 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7705 #if defined(CONFIG_UNIX)
7706 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7712 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7716 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7717 O_RDWR | O_CLOEXEC);
7720 ret = PTR_ERR(file);
7724 #if defined(CONFIG_UNIX)
7725 ctx->ring_sock->file = file;
7727 fd_install(ret, file);
7730 #if defined(CONFIG_UNIX)
7731 sock_release(ctx->ring_sock);
7732 ctx->ring_sock = NULL;
7737 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7739 struct user_struct *user = NULL;
7740 struct io_ring_ctx *ctx;
7746 if (entries > IORING_MAX_ENTRIES) {
7747 if (!(p->flags & IORING_SETUP_CLAMP))
7749 entries = IORING_MAX_ENTRIES;
7753 * Use twice as many entries for the CQ ring. It's possible for the
7754 * application to drive a higher depth than the size of the SQ ring,
7755 * since the sqes are only used at submission time. This allows for
7756 * some flexibility in overcommitting a bit. If the application has
7757 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7758 * of CQ ring entries manually.
7760 p->sq_entries = roundup_pow_of_two(entries);
7761 if (p->flags & IORING_SETUP_CQSIZE) {
7763 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7764 * to a power-of-two, if it isn't already. We do NOT impose
7765 * any cq vs sq ring sizing.
7767 if (p->cq_entries < p->sq_entries)
7769 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7770 if (!(p->flags & IORING_SETUP_CLAMP))
7772 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7774 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7776 p->cq_entries = 2 * p->sq_entries;
7779 user = get_uid(current_user());
7780 account_mem = !capable(CAP_IPC_LOCK);
7783 ret = io_account_mem(user,
7784 ring_pages(p->sq_entries, p->cq_entries));
7791 ctx = io_ring_ctx_alloc(p);
7794 io_unaccount_mem(user, ring_pages(p->sq_entries,
7799 ctx->compat = in_compat_syscall();
7800 ctx->account_mem = account_mem;
7802 ctx->creds = get_current_cred();
7804 ret = io_allocate_scq_urings(ctx, p);
7808 ret = io_sq_offload_start(ctx, p);
7812 memset(&p->sq_off, 0, sizeof(p->sq_off));
7813 p->sq_off.head = offsetof(struct io_rings, sq.head);
7814 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7815 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7816 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7817 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7818 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7819 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7821 memset(&p->cq_off, 0, sizeof(p->cq_off));
7822 p->cq_off.head = offsetof(struct io_rings, cq.head);
7823 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7824 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7825 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7826 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7827 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7830 * Install ring fd as the very last thing, so we don't risk someone
7831 * having closed it before we finish setup
7833 ret = io_uring_get_fd(ctx);
7837 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7838 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7839 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7840 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7843 io_ring_ctx_wait_and_kill(ctx);
7848 * Sets up an aio uring context, and returns the fd. Applications asks for a
7849 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7850 * params structure passed in.
7852 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7854 struct io_uring_params p;
7858 if (copy_from_user(&p, params, sizeof(p)))
7860 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7865 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7866 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7867 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7870 ret = io_uring_create(entries, &p);
7874 if (copy_to_user(params, &p, sizeof(p)))
7880 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7881 struct io_uring_params __user *, params)
7883 return io_uring_setup(entries, params);
7886 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7888 struct io_uring_probe *p;
7892 size = struct_size(p, ops, nr_args);
7893 if (size == SIZE_MAX)
7895 p = kzalloc(size, GFP_KERNEL);
7900 if (copy_from_user(p, arg, size))
7903 if (memchr_inv(p, 0, size))
7906 p->last_op = IORING_OP_LAST - 1;
7907 if (nr_args > IORING_OP_LAST)
7908 nr_args = IORING_OP_LAST;
7910 for (i = 0; i < nr_args; i++) {
7912 if (!io_op_defs[i].not_supported)
7913 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7918 if (copy_to_user(arg, p, size))
7925 static int io_register_personality(struct io_ring_ctx *ctx)
7927 const struct cred *creds = get_current_cred();
7930 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7931 USHRT_MAX, GFP_KERNEL);
7937 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7939 const struct cred *old_creds;
7941 old_creds = idr_remove(&ctx->personality_idr, id);
7943 put_cred(old_creds);
7950 static bool io_register_op_must_quiesce(int op)
7953 case IORING_UNREGISTER_FILES:
7954 case IORING_REGISTER_FILES_UPDATE:
7955 case IORING_REGISTER_PROBE:
7956 case IORING_REGISTER_PERSONALITY:
7957 case IORING_UNREGISTER_PERSONALITY:
7964 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7965 void __user *arg, unsigned nr_args)
7966 __releases(ctx->uring_lock)
7967 __acquires(ctx->uring_lock)
7972 * We're inside the ring mutex, if the ref is already dying, then
7973 * someone else killed the ctx or is already going through
7974 * io_uring_register().
7976 if (percpu_ref_is_dying(&ctx->refs))
7979 if (io_register_op_must_quiesce(opcode)) {
7980 percpu_ref_kill(&ctx->refs);
7983 * Drop uring mutex before waiting for references to exit. If
7984 * another thread is currently inside io_uring_enter() it might
7985 * need to grab the uring_lock to make progress. If we hold it
7986 * here across the drain wait, then we can deadlock. It's safe
7987 * to drop the mutex here, since no new references will come in
7988 * after we've killed the percpu ref.
7990 mutex_unlock(&ctx->uring_lock);
7991 ret = wait_for_completion_interruptible(&ctx->completions[0]);
7992 mutex_lock(&ctx->uring_lock);
7994 percpu_ref_resurrect(&ctx->refs);
8001 case IORING_REGISTER_BUFFERS:
8002 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8004 case IORING_UNREGISTER_BUFFERS:
8008 ret = io_sqe_buffer_unregister(ctx);
8010 case IORING_REGISTER_FILES:
8011 ret = io_sqe_files_register(ctx, arg, nr_args);
8013 case IORING_UNREGISTER_FILES:
8017 ret = io_sqe_files_unregister(ctx);
8019 case IORING_REGISTER_FILES_UPDATE:
8020 ret = io_sqe_files_update(ctx, arg, nr_args);
8022 case IORING_REGISTER_EVENTFD:
8023 case IORING_REGISTER_EVENTFD_ASYNC:
8027 ret = io_eventfd_register(ctx, arg);
8030 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8031 ctx->eventfd_async = 1;
8033 ctx->eventfd_async = 0;
8035 case IORING_UNREGISTER_EVENTFD:
8039 ret = io_eventfd_unregister(ctx);
8041 case IORING_REGISTER_PROBE:
8043 if (!arg || nr_args > 256)
8045 ret = io_probe(ctx, arg, nr_args);
8047 case IORING_REGISTER_PERSONALITY:
8051 ret = io_register_personality(ctx);
8053 case IORING_UNREGISTER_PERSONALITY:
8057 ret = io_unregister_personality(ctx, nr_args);
8064 if (io_register_op_must_quiesce(opcode)) {
8065 /* bring the ctx back to life */
8066 percpu_ref_reinit(&ctx->refs);
8068 reinit_completion(&ctx->completions[0]);
8073 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8074 void __user *, arg, unsigned int, nr_args)
8076 struct io_ring_ctx *ctx;
8085 if (f.file->f_op != &io_uring_fops)
8088 ctx = f.file->private_data;
8090 mutex_lock(&ctx->uring_lock);
8091 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8092 mutex_unlock(&ctx->uring_lock);
8093 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8094 ctx->cq_ev_fd != NULL, ret);
8100 static int __init io_uring_init(void)
8102 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8103 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8104 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8107 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8108 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8109 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8110 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8111 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8112 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8113 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8114 BUILD_BUG_SQE_ELEM(8, __u64, off);
8115 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8116 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8117 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8118 BUILD_BUG_SQE_ELEM(24, __u32, len);
8119 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8120 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8121 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8122 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8123 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8124 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8125 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8126 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8127 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8128 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8129 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8130 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8131 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8132 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8133 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8134 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8135 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8136 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8138 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8139 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8140 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8143 __initcall(io_uring_init);