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/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
102 u32 head ____cacheline_aligned_in_smp;
103 u32 tail ____cacheline_aligned_in_smp;
107 * This data is shared with the application through the mmap at offsets
108 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
110 * The offsets to the member fields are published through struct
111 * io_sqring_offsets when calling io_uring_setup.
115 * Head and tail offsets into the ring; the offsets need to be
116 * masked to get valid indices.
118 * The kernel controls head of the sq ring and the tail of the cq ring,
119 * and the application controls tail of the sq ring and the head of the
122 struct io_uring sq, cq;
124 * Bitmasks to apply to head and tail offsets (constant, equals
127 u32 sq_ring_mask, cq_ring_mask;
128 /* Ring sizes (constant, power of 2) */
129 u32 sq_ring_entries, cq_ring_entries;
131 * Number of invalid entries dropped by the kernel due to
132 * invalid index stored in array
134 * Written by the kernel, shouldn't be modified by the
135 * application (i.e. get number of "new events" by comparing to
138 * After a new SQ head value was read by the application this
139 * counter includes all submissions that were dropped reaching
140 * the new SQ head (and possibly more).
146 * Written by the kernel, shouldn't be modified by the
149 * The application needs a full memory barrier before checking
150 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
156 * Written by the application, shouldn't be modified by the
161 * Number of completion events lost because the queue was full;
162 * this should be avoided by the application by making sure
163 * there are not more requests pending than there is space in
164 * the completion queue.
166 * Written by the kernel, shouldn't be modified by the
167 * application (i.e. get number of "new events" by comparing to
170 * As completion events come in out of order this counter is not
171 * ordered with any other data.
175 * Ring buffer of completion events.
177 * The kernel writes completion events fresh every time they are
178 * produced, so the application is allowed to modify pending
181 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
184 struct io_mapped_ubuf {
187 struct bio_vec *bvec;
188 unsigned int nr_bvecs;
191 struct fixed_file_table {
195 struct fixed_file_ref_node {
196 struct percpu_ref refs;
197 struct list_head node;
198 struct list_head file_list;
199 struct fixed_file_data *file_data;
200 struct llist_node llist;
203 struct fixed_file_data {
204 struct fixed_file_table *table;
205 struct io_ring_ctx *ctx;
207 struct percpu_ref *cur_refs;
208 struct percpu_ref refs;
209 struct completion done;
210 struct list_head ref_list;
215 struct list_head list;
223 struct percpu_ref refs;
224 } ____cacheline_aligned_in_smp;
228 unsigned int compat: 1;
229 unsigned int account_mem: 1;
230 unsigned int cq_overflow_flushed: 1;
231 unsigned int drain_next: 1;
232 unsigned int eventfd_async: 1;
235 * Ring buffer of indices into array of io_uring_sqe, which is
236 * mmapped by the application using the IORING_OFF_SQES offset.
238 * This indirection could e.g. be used to assign fixed
239 * io_uring_sqe entries to operations and only submit them to
240 * the queue when needed.
242 * The kernel modifies neither the indices array nor the entries
246 unsigned cached_sq_head;
249 unsigned sq_thread_idle;
250 unsigned cached_sq_dropped;
251 atomic_t cached_cq_overflow;
252 unsigned long sq_check_overflow;
254 struct list_head defer_list;
255 struct list_head timeout_list;
256 struct list_head cq_overflow_list;
258 wait_queue_head_t inflight_wait;
259 struct io_uring_sqe *sq_sqes;
260 } ____cacheline_aligned_in_smp;
262 struct io_rings *rings;
266 struct task_struct *sqo_thread; /* if using sq thread polling */
267 struct mm_struct *sqo_mm;
268 wait_queue_head_t sqo_wait;
271 * If used, fixed file set. Writers must ensure that ->refs is dead,
272 * readers must ensure that ->refs is alive as long as the file* is
273 * used. Only updated through io_uring_register(2).
275 struct fixed_file_data *file_data;
276 unsigned nr_user_files;
278 struct file *ring_file;
280 /* if used, fixed mapped user buffers */
281 unsigned nr_user_bufs;
282 struct io_mapped_ubuf *user_bufs;
284 struct user_struct *user;
286 const struct cred *creds;
288 struct completion ref_comp;
289 struct completion sq_thread_comp;
291 /* if all else fails... */
292 struct io_kiocb *fallback_req;
294 #if defined(CONFIG_UNIX)
295 struct socket *ring_sock;
298 struct idr io_buffer_idr;
300 struct idr personality_idr;
303 unsigned cached_cq_tail;
306 atomic_t cq_timeouts;
307 unsigned long cq_check_overflow;
308 struct wait_queue_head cq_wait;
309 struct fasync_struct *cq_fasync;
310 struct eventfd_ctx *cq_ev_fd;
311 } ____cacheline_aligned_in_smp;
314 struct mutex uring_lock;
315 wait_queue_head_t wait;
316 } ____cacheline_aligned_in_smp;
319 spinlock_t completion_lock;
322 * ->poll_list is protected by the ctx->uring_lock for
323 * io_uring instances that don't use IORING_SETUP_SQPOLL.
324 * For SQPOLL, only the single threaded io_sq_thread() will
325 * manipulate the list, hence no extra locking is needed there.
327 struct list_head poll_list;
328 struct hlist_head *cancel_hash;
329 unsigned cancel_hash_bits;
330 bool poll_multi_file;
332 spinlock_t inflight_lock;
333 struct list_head inflight_list;
334 } ____cacheline_aligned_in_smp;
336 struct delayed_work file_put_work;
337 struct llist_head file_put_llist;
339 struct work_struct exit_work;
343 * First field must be the file pointer in all the
344 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
346 struct io_poll_iocb {
349 struct wait_queue_head *head;
355 struct wait_queue_entry wait;
360 struct file *put_file;
364 struct io_timeout_data {
365 struct io_kiocb *req;
366 struct hrtimer timer;
367 struct timespec64 ts;
368 enum hrtimer_mode mode;
373 struct sockaddr __user *addr;
374 int __user *addr_len;
376 unsigned long nofile;
401 /* NOTE: kiocb has the file as the first member, so don't do it here */
409 struct sockaddr __user *addr;
416 struct user_msghdr __user *msg;
422 struct io_buffer *kbuf;
428 struct filename *filename;
430 unsigned long nofile;
433 struct io_files_update {
459 struct epoll_event event;
463 struct file *file_out;
464 struct file *file_in;
471 struct io_provide_buf {
485 const char __user *filename;
486 struct statx __user *buffer;
489 struct io_async_connect {
490 struct sockaddr_storage address;
493 struct io_async_msghdr {
494 struct iovec fast_iov[UIO_FASTIOV];
496 struct sockaddr __user *uaddr;
498 struct sockaddr_storage addr;
502 struct iovec fast_iov[UIO_FASTIOV];
508 struct io_async_ctx {
510 struct io_async_rw rw;
511 struct io_async_msghdr msg;
512 struct io_async_connect connect;
513 struct io_timeout_data timeout;
518 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
519 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
520 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
521 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
522 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
523 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
531 REQ_F_LINK_TIMEOUT_BIT,
535 REQ_F_TIMEOUT_NOSEQ_BIT,
536 REQ_F_COMP_LOCKED_BIT,
537 REQ_F_NEED_CLEANUP_BIT,
540 REQ_F_BUFFER_SELECTED_BIT,
541 REQ_F_NO_FILE_TABLE_BIT,
542 REQ_F_QUEUE_TIMEOUT_BIT,
543 REQ_F_WORK_INITIALIZED_BIT,
545 /* not a real bit, just to check we're not overflowing the space */
551 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
552 /* drain existing IO first */
553 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
555 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
556 /* doesn't sever on completion < 0 */
557 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
559 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
560 /* IOSQE_BUFFER_SELECT */
561 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
564 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
565 /* already grabbed next link */
566 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
567 /* fail rest of links */
568 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
569 /* on inflight list */
570 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
571 /* read/write uses file position */
572 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
573 /* must not punt to workers */
574 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
575 /* has linked timeout */
576 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
577 /* timeout request */
578 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
580 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
581 /* must be punted even for NONBLOCK */
582 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
583 /* no timeout sequence */
584 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
585 /* completion under lock */
586 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
588 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
589 /* in overflow list */
590 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
591 /* already went through poll handler */
592 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
593 /* buffer already selected */
594 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
595 /* doesn't need file table for this request */
596 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
597 /* needs to queue linked timeout */
598 REQ_F_QUEUE_TIMEOUT = BIT(REQ_F_QUEUE_TIMEOUT_BIT),
599 /* io_wq_work is initialized */
600 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
604 struct io_poll_iocb poll;
605 struct io_wq_work work;
609 * NOTE! Each of the iocb union members has the file pointer
610 * as the first entry in their struct definition. So you can
611 * access the file pointer through any of the sub-structs,
612 * or directly as just 'ki_filp' in this struct.
618 struct io_poll_iocb poll;
619 struct io_accept accept;
621 struct io_cancel cancel;
622 struct io_timeout timeout;
623 struct io_connect connect;
624 struct io_sr_msg sr_msg;
626 struct io_close close;
627 struct io_files_update files_update;
628 struct io_fadvise fadvise;
629 struct io_madvise madvise;
630 struct io_epoll epoll;
631 struct io_splice splice;
632 struct io_provide_buf pbuf;
633 struct io_statx statx;
636 struct io_async_ctx *io;
639 /* polled IO has completed */
644 struct io_ring_ctx *ctx;
645 struct list_head list;
648 struct task_struct *task;
654 struct list_head link_list;
656 struct list_head inflight_entry;
658 struct percpu_ref *fixed_file_refs;
662 * Only commands that never go async can use the below fields,
663 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
664 * async armed poll handlers for regular commands. The latter
665 * restore the work, if needed.
668 struct callback_head task_work;
669 struct hlist_node hash_node;
670 struct async_poll *apoll;
672 struct io_wq_work work;
676 #define IO_PLUG_THRESHOLD 2
677 #define IO_IOPOLL_BATCH 8
679 struct io_submit_state {
680 struct blk_plug plug;
683 * io_kiocb alloc cache
685 void *reqs[IO_IOPOLL_BATCH];
686 unsigned int free_reqs;
689 * File reference cache
693 unsigned int has_refs;
694 unsigned int used_refs;
695 unsigned int ios_left;
699 /* needs req->io allocated for deferral/async */
700 unsigned async_ctx : 1;
701 /* needs current->mm setup, does mm access */
702 unsigned needs_mm : 1;
703 /* needs req->file assigned */
704 unsigned needs_file : 1;
705 /* don't fail if file grab fails */
706 unsigned needs_file_no_error : 1;
707 /* hash wq insertion if file is a regular file */
708 unsigned hash_reg_file : 1;
709 /* unbound wq insertion if file is a non-regular file */
710 unsigned unbound_nonreg_file : 1;
711 /* opcode is not supported by this kernel */
712 unsigned not_supported : 1;
713 /* needs file table */
714 unsigned file_table : 1;
716 unsigned needs_fs : 1;
717 /* set if opcode supports polled "wait" */
719 unsigned pollout : 1;
720 /* op supports buffer selection */
721 unsigned buffer_select : 1;
724 static const struct io_op_def io_op_defs[] = {
725 [IORING_OP_NOP] = {},
726 [IORING_OP_READV] = {
730 .unbound_nonreg_file = 1,
734 [IORING_OP_WRITEV] = {
739 .unbound_nonreg_file = 1,
742 [IORING_OP_FSYNC] = {
745 [IORING_OP_READ_FIXED] = {
747 .unbound_nonreg_file = 1,
750 [IORING_OP_WRITE_FIXED] = {
753 .unbound_nonreg_file = 1,
756 [IORING_OP_POLL_ADD] = {
758 .unbound_nonreg_file = 1,
760 [IORING_OP_POLL_REMOVE] = {},
761 [IORING_OP_SYNC_FILE_RANGE] = {
764 [IORING_OP_SENDMSG] = {
768 .unbound_nonreg_file = 1,
772 [IORING_OP_RECVMSG] = {
776 .unbound_nonreg_file = 1,
781 [IORING_OP_TIMEOUT] = {
785 [IORING_OP_TIMEOUT_REMOVE] = {},
786 [IORING_OP_ACCEPT] = {
789 .unbound_nonreg_file = 1,
793 [IORING_OP_ASYNC_CANCEL] = {},
794 [IORING_OP_LINK_TIMEOUT] = {
798 [IORING_OP_CONNECT] = {
802 .unbound_nonreg_file = 1,
805 [IORING_OP_FALLOCATE] = {
808 [IORING_OP_OPENAT] = {
812 [IORING_OP_CLOSE] = {
814 .needs_file_no_error = 1,
817 [IORING_OP_FILES_UPDATE] = {
821 [IORING_OP_STATX] = {
829 .unbound_nonreg_file = 1,
833 [IORING_OP_WRITE] = {
836 .unbound_nonreg_file = 1,
839 [IORING_OP_FADVISE] = {
842 [IORING_OP_MADVISE] = {
848 .unbound_nonreg_file = 1,
854 .unbound_nonreg_file = 1,
858 [IORING_OP_OPENAT2] = {
862 [IORING_OP_EPOLL_CTL] = {
863 .unbound_nonreg_file = 1,
866 [IORING_OP_SPLICE] = {
869 .unbound_nonreg_file = 1,
871 [IORING_OP_PROVIDE_BUFFERS] = {},
872 [IORING_OP_REMOVE_BUFFERS] = {},
876 .unbound_nonreg_file = 1,
880 static void io_wq_submit_work(struct io_wq_work **workptr);
881 static void io_cqring_fill_event(struct io_kiocb *req, long res);
882 static void io_put_req(struct io_kiocb *req);
883 static void __io_double_put_req(struct io_kiocb *req);
884 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
885 static void io_queue_linked_timeout(struct io_kiocb *req);
886 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
887 struct io_uring_files_update *ip,
889 static int io_grab_files(struct io_kiocb *req);
890 static void io_cleanup_req(struct io_kiocb *req);
891 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
892 int fd, struct file **out_file, bool fixed);
893 static void __io_queue_sqe(struct io_kiocb *req,
894 const struct io_uring_sqe *sqe);
896 static struct kmem_cache *req_cachep;
898 static const struct file_operations io_uring_fops;
900 struct sock *io_uring_get_socket(struct file *file)
902 #if defined(CONFIG_UNIX)
903 if (file->f_op == &io_uring_fops) {
904 struct io_ring_ctx *ctx = file->private_data;
906 return ctx->ring_sock->sk;
911 EXPORT_SYMBOL(io_uring_get_socket);
913 static void io_file_put_work(struct work_struct *work);
916 * Note: must call io_req_init_async() for the first time you
917 * touch any members of io_wq_work.
919 static inline void io_req_init_async(struct io_kiocb *req)
921 if (req->flags & REQ_F_WORK_INITIALIZED)
924 memset(&req->work, 0, sizeof(req->work));
925 req->flags |= REQ_F_WORK_INITIALIZED;
928 static inline bool io_async_submit(struct io_ring_ctx *ctx)
930 return ctx->flags & IORING_SETUP_SQPOLL;
933 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
935 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
937 complete(&ctx->ref_comp);
940 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
942 struct io_ring_ctx *ctx;
945 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
949 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
950 if (!ctx->fallback_req)
954 * Use 5 bits less than the max cq entries, that should give us around
955 * 32 entries per hash list if totally full and uniformly spread.
957 hash_bits = ilog2(p->cq_entries);
961 ctx->cancel_hash_bits = hash_bits;
962 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
964 if (!ctx->cancel_hash)
966 __hash_init(ctx->cancel_hash, 1U << hash_bits);
968 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
969 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
972 ctx->flags = p->flags;
973 init_waitqueue_head(&ctx->sqo_wait);
974 init_waitqueue_head(&ctx->cq_wait);
975 INIT_LIST_HEAD(&ctx->cq_overflow_list);
976 init_completion(&ctx->ref_comp);
977 init_completion(&ctx->sq_thread_comp);
978 idr_init(&ctx->io_buffer_idr);
979 idr_init(&ctx->personality_idr);
980 mutex_init(&ctx->uring_lock);
981 init_waitqueue_head(&ctx->wait);
982 spin_lock_init(&ctx->completion_lock);
983 INIT_LIST_HEAD(&ctx->poll_list);
984 INIT_LIST_HEAD(&ctx->defer_list);
985 INIT_LIST_HEAD(&ctx->timeout_list);
986 init_waitqueue_head(&ctx->inflight_wait);
987 spin_lock_init(&ctx->inflight_lock);
988 INIT_LIST_HEAD(&ctx->inflight_list);
989 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
990 init_llist_head(&ctx->file_put_llist);
993 if (ctx->fallback_req)
994 kmem_cache_free(req_cachep, ctx->fallback_req);
995 kfree(ctx->cancel_hash);
1000 static inline bool __req_need_defer(struct io_kiocb *req)
1002 struct io_ring_ctx *ctx = req->ctx;
1004 return req->sequence != ctx->cached_cq_tail
1005 + atomic_read(&ctx->cached_cq_overflow);
1008 static inline bool req_need_defer(struct io_kiocb *req)
1010 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1011 return __req_need_defer(req);
1016 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1018 struct io_rings *rings = ctx->rings;
1020 /* order cqe stores with ring update */
1021 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1023 if (wq_has_sleeper(&ctx->cq_wait)) {
1024 wake_up_interruptible(&ctx->cq_wait);
1025 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1029 static inline void io_req_work_grab_env(struct io_kiocb *req,
1030 const struct io_op_def *def)
1032 if (!req->work.mm && def->needs_mm) {
1033 mmgrab(current->mm);
1034 req->work.mm = current->mm;
1036 if (!req->work.creds)
1037 req->work.creds = get_current_cred();
1038 if (!req->work.fs && def->needs_fs) {
1039 spin_lock(¤t->fs->lock);
1040 if (!current->fs->in_exec) {
1041 req->work.fs = current->fs;
1042 req->work.fs->users++;
1044 req->work.flags |= IO_WQ_WORK_CANCEL;
1046 spin_unlock(¤t->fs->lock);
1048 if (!req->work.task_pid)
1049 req->work.task_pid = task_pid_vnr(current);
1052 static inline void io_req_work_drop_env(struct io_kiocb *req)
1054 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1058 mmdrop(req->work.mm);
1059 req->work.mm = NULL;
1061 if (req->work.creds) {
1062 put_cred(req->work.creds);
1063 req->work.creds = NULL;
1066 struct fs_struct *fs = req->work.fs;
1068 spin_lock(&req->work.fs->lock);
1071 spin_unlock(&req->work.fs->lock);
1077 static inline void io_prep_async_work(struct io_kiocb *req,
1078 struct io_kiocb **link)
1080 const struct io_op_def *def = &io_op_defs[req->opcode];
1082 if (req->flags & REQ_F_ISREG) {
1083 if (def->hash_reg_file)
1084 io_wq_hash_work(&req->work, file_inode(req->file));
1086 if (def->unbound_nonreg_file)
1087 req->work.flags |= IO_WQ_WORK_UNBOUND;
1090 io_req_work_grab_env(req, def);
1092 *link = io_prep_linked_timeout(req);
1095 static inline void io_queue_async_work(struct io_kiocb *req)
1097 struct io_ring_ctx *ctx = req->ctx;
1098 struct io_kiocb *link;
1100 io_prep_async_work(req, &link);
1102 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1103 &req->work, req->flags);
1104 io_wq_enqueue(ctx->io_wq, &req->work);
1107 io_queue_linked_timeout(link);
1110 static void io_kill_timeout(struct io_kiocb *req)
1114 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1116 atomic_inc(&req->ctx->cq_timeouts);
1117 list_del_init(&req->list);
1118 req->flags |= REQ_F_COMP_LOCKED;
1119 io_cqring_fill_event(req, 0);
1124 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1126 struct io_kiocb *req, *tmp;
1128 spin_lock_irq(&ctx->completion_lock);
1129 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1130 io_kill_timeout(req);
1131 spin_unlock_irq(&ctx->completion_lock);
1134 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1137 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1138 struct io_kiocb, list);
1140 if (req_need_defer(req))
1142 list_del_init(&req->list);
1143 io_queue_async_work(req);
1144 } while (!list_empty(&ctx->defer_list));
1147 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1149 while (!list_empty(&ctx->timeout_list)) {
1150 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1151 struct io_kiocb, list);
1153 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
1155 if (req->timeout.target_seq != ctx->cached_cq_tail
1156 - atomic_read(&ctx->cq_timeouts))
1159 list_del_init(&req->list);
1160 io_kill_timeout(req);
1164 static void io_commit_cqring(struct io_ring_ctx *ctx)
1166 io_flush_timeouts(ctx);
1167 __io_commit_cqring(ctx);
1169 if (unlikely(!list_empty(&ctx->defer_list)))
1170 __io_queue_deferred(ctx);
1173 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1175 struct io_rings *rings = ctx->rings;
1178 tail = ctx->cached_cq_tail;
1180 * writes to the cq entry need to come after reading head; the
1181 * control dependency is enough as we're using WRITE_ONCE to
1184 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1187 ctx->cached_cq_tail++;
1188 return &rings->cqes[tail & ctx->cq_mask];
1191 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1195 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1197 if (!ctx->eventfd_async)
1199 return io_wq_current_is_worker();
1202 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1204 if (waitqueue_active(&ctx->wait))
1205 wake_up(&ctx->wait);
1206 if (waitqueue_active(&ctx->sqo_wait))
1207 wake_up(&ctx->sqo_wait);
1208 if (io_should_trigger_evfd(ctx))
1209 eventfd_signal(ctx->cq_ev_fd, 1);
1212 /* Returns true if there are no backlogged entries after the flush */
1213 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1215 struct io_rings *rings = ctx->rings;
1216 struct io_uring_cqe *cqe;
1217 struct io_kiocb *req;
1218 unsigned long flags;
1222 if (list_empty_careful(&ctx->cq_overflow_list))
1224 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1225 rings->cq_ring_entries))
1229 spin_lock_irqsave(&ctx->completion_lock, flags);
1231 /* if force is set, the ring is going away. always drop after that */
1233 ctx->cq_overflow_flushed = 1;
1236 while (!list_empty(&ctx->cq_overflow_list)) {
1237 cqe = io_get_cqring(ctx);
1241 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1243 list_move(&req->list, &list);
1244 req->flags &= ~REQ_F_OVERFLOW;
1246 WRITE_ONCE(cqe->user_data, req->user_data);
1247 WRITE_ONCE(cqe->res, req->result);
1248 WRITE_ONCE(cqe->flags, req->cflags);
1250 WRITE_ONCE(ctx->rings->cq_overflow,
1251 atomic_inc_return(&ctx->cached_cq_overflow));
1255 io_commit_cqring(ctx);
1257 clear_bit(0, &ctx->sq_check_overflow);
1258 clear_bit(0, &ctx->cq_check_overflow);
1260 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1261 io_cqring_ev_posted(ctx);
1263 while (!list_empty(&list)) {
1264 req = list_first_entry(&list, struct io_kiocb, list);
1265 list_del(&req->list);
1272 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1274 struct io_ring_ctx *ctx = req->ctx;
1275 struct io_uring_cqe *cqe;
1277 trace_io_uring_complete(ctx, req->user_data, res);
1280 * If we can't get a cq entry, userspace overflowed the
1281 * submission (by quite a lot). Increment the overflow count in
1284 cqe = io_get_cqring(ctx);
1286 WRITE_ONCE(cqe->user_data, req->user_data);
1287 WRITE_ONCE(cqe->res, res);
1288 WRITE_ONCE(cqe->flags, cflags);
1289 } else if (ctx->cq_overflow_flushed) {
1290 WRITE_ONCE(ctx->rings->cq_overflow,
1291 atomic_inc_return(&ctx->cached_cq_overflow));
1293 if (list_empty(&ctx->cq_overflow_list)) {
1294 set_bit(0, &ctx->sq_check_overflow);
1295 set_bit(0, &ctx->cq_check_overflow);
1297 req->flags |= REQ_F_OVERFLOW;
1298 refcount_inc(&req->refs);
1300 req->cflags = cflags;
1301 list_add_tail(&req->list, &ctx->cq_overflow_list);
1305 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1307 __io_cqring_fill_event(req, res, 0);
1310 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1312 struct io_ring_ctx *ctx = req->ctx;
1313 unsigned long flags;
1315 spin_lock_irqsave(&ctx->completion_lock, flags);
1316 __io_cqring_fill_event(req, res, cflags);
1317 io_commit_cqring(ctx);
1318 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1320 io_cqring_ev_posted(ctx);
1323 static void io_cqring_add_event(struct io_kiocb *req, long res)
1325 __io_cqring_add_event(req, res, 0);
1328 static inline bool io_is_fallback_req(struct io_kiocb *req)
1330 return req == (struct io_kiocb *)
1331 ((unsigned long) req->ctx->fallback_req & ~1UL);
1334 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1336 struct io_kiocb *req;
1338 req = ctx->fallback_req;
1339 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1345 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1346 struct io_submit_state *state)
1348 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1349 struct io_kiocb *req;
1352 req = kmem_cache_alloc(req_cachep, gfp);
1355 } else if (!state->free_reqs) {
1359 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1360 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1363 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1364 * retry single alloc to be on the safe side.
1366 if (unlikely(ret <= 0)) {
1367 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1368 if (!state->reqs[0])
1372 state->free_reqs = ret - 1;
1373 req = state->reqs[ret - 1];
1376 req = state->reqs[state->free_reqs];
1381 return io_get_fallback_req(ctx);
1384 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1388 percpu_ref_put(req->fixed_file_refs);
1393 static void __io_req_aux_free(struct io_kiocb *req)
1395 if (req->flags & REQ_F_NEED_CLEANUP)
1396 io_cleanup_req(req);
1400 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1402 put_task_struct(req->task);
1404 io_req_work_drop_env(req);
1407 static void __io_free_req(struct io_kiocb *req)
1409 __io_req_aux_free(req);
1411 if (req->flags & REQ_F_INFLIGHT) {
1412 struct io_ring_ctx *ctx = req->ctx;
1413 unsigned long flags;
1415 spin_lock_irqsave(&ctx->inflight_lock, flags);
1416 list_del(&req->inflight_entry);
1417 if (waitqueue_active(&ctx->inflight_wait))
1418 wake_up(&ctx->inflight_wait);
1419 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1422 percpu_ref_put(&req->ctx->refs);
1423 if (likely(!io_is_fallback_req(req)))
1424 kmem_cache_free(req_cachep, req);
1426 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1430 void *reqs[IO_IOPOLL_BATCH];
1435 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1439 if (rb->need_iter) {
1440 int i, inflight = 0;
1441 unsigned long flags;
1443 for (i = 0; i < rb->to_free; i++) {
1444 struct io_kiocb *req = rb->reqs[i];
1446 if (req->flags & REQ_F_INFLIGHT)
1448 __io_req_aux_free(req);
1453 spin_lock_irqsave(&ctx->inflight_lock, flags);
1454 for (i = 0; i < rb->to_free; i++) {
1455 struct io_kiocb *req = rb->reqs[i];
1457 if (req->flags & REQ_F_INFLIGHT) {
1458 list_del(&req->inflight_entry);
1463 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1465 if (waitqueue_active(&ctx->inflight_wait))
1466 wake_up(&ctx->inflight_wait);
1469 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1470 percpu_ref_put_many(&ctx->refs, rb->to_free);
1471 rb->to_free = rb->need_iter = 0;
1474 static bool io_link_cancel_timeout(struct io_kiocb *req)
1476 struct io_ring_ctx *ctx = req->ctx;
1479 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1481 io_cqring_fill_event(req, -ECANCELED);
1482 io_commit_cqring(ctx);
1483 req->flags &= ~REQ_F_LINK_HEAD;
1491 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1493 struct io_ring_ctx *ctx = req->ctx;
1494 bool wake_ev = false;
1496 /* Already got next link */
1497 if (req->flags & REQ_F_LINK_NEXT)
1501 * The list should never be empty when we are called here. But could
1502 * potentially happen if the chain is messed up, check to be on the
1505 while (!list_empty(&req->link_list)) {
1506 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1507 struct io_kiocb, link_list);
1509 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1510 (nxt->flags & REQ_F_TIMEOUT))) {
1511 list_del_init(&nxt->link_list);
1512 wake_ev |= io_link_cancel_timeout(nxt);
1513 req->flags &= ~REQ_F_LINK_TIMEOUT;
1517 list_del_init(&req->link_list);
1518 if (!list_empty(&nxt->link_list))
1519 nxt->flags |= REQ_F_LINK_HEAD;
1524 req->flags |= REQ_F_LINK_NEXT;
1526 io_cqring_ev_posted(ctx);
1530 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1532 static void io_fail_links(struct io_kiocb *req)
1534 struct io_ring_ctx *ctx = req->ctx;
1535 unsigned long flags;
1537 spin_lock_irqsave(&ctx->completion_lock, flags);
1539 while (!list_empty(&req->link_list)) {
1540 struct io_kiocb *link = list_first_entry(&req->link_list,
1541 struct io_kiocb, link_list);
1543 list_del_init(&link->link_list);
1544 trace_io_uring_fail_link(req, link);
1546 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1547 link->opcode == IORING_OP_LINK_TIMEOUT) {
1548 io_link_cancel_timeout(link);
1550 io_cqring_fill_event(link, -ECANCELED);
1551 __io_double_put_req(link);
1553 req->flags &= ~REQ_F_LINK_TIMEOUT;
1556 io_commit_cqring(ctx);
1557 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1558 io_cqring_ev_posted(ctx);
1561 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1563 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1567 * If LINK is set, we have dependent requests in this chain. If we
1568 * didn't fail this request, queue the first one up, moving any other
1569 * dependencies to the next request. In case of failure, fail the rest
1572 if (req->flags & REQ_F_FAIL_LINK) {
1574 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1575 REQ_F_LINK_TIMEOUT) {
1576 struct io_ring_ctx *ctx = req->ctx;
1577 unsigned long flags;
1580 * If this is a timeout link, we could be racing with the
1581 * timeout timer. Grab the completion lock for this case to
1582 * protect against that.
1584 spin_lock_irqsave(&ctx->completion_lock, flags);
1585 io_req_link_next(req, nxt);
1586 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1588 io_req_link_next(req, nxt);
1592 static void io_free_req(struct io_kiocb *req)
1594 struct io_kiocb *nxt = NULL;
1596 io_req_find_next(req, &nxt);
1600 io_queue_async_work(nxt);
1603 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1605 struct io_kiocb *link;
1606 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1608 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1609 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1611 *workptr = &nxt->work;
1612 link = io_prep_linked_timeout(nxt);
1614 nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1618 * Drop reference to request, return next in chain (if there is one) if this
1619 * was the last reference to this request.
1621 __attribute__((nonnull))
1622 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1624 if (refcount_dec_and_test(&req->refs)) {
1625 io_req_find_next(req, nxtptr);
1630 static void io_put_req(struct io_kiocb *req)
1632 if (refcount_dec_and_test(&req->refs))
1636 static void io_steal_work(struct io_kiocb *req,
1637 struct io_wq_work **workptr)
1640 * It's in an io-wq worker, so there always should be at least
1641 * one reference, which will be dropped in io_put_work() just
1642 * after the current handler returns.
1644 * It also means, that if the counter dropped to 1, then there is
1645 * no asynchronous users left, so it's safe to steal the next work.
1647 if (refcount_read(&req->refs) == 1) {
1648 struct io_kiocb *nxt = NULL;
1650 io_req_find_next(req, &nxt);
1652 io_wq_assign_next(workptr, nxt);
1657 * Must only be used if we don't need to care about links, usually from
1658 * within the completion handling itself.
1660 static void __io_double_put_req(struct io_kiocb *req)
1662 /* drop both submit and complete references */
1663 if (refcount_sub_and_test(2, &req->refs))
1667 static void io_double_put_req(struct io_kiocb *req)
1669 /* drop both submit and complete references */
1670 if (refcount_sub_and_test(2, &req->refs))
1674 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1676 struct io_rings *rings = ctx->rings;
1678 if (test_bit(0, &ctx->cq_check_overflow)) {
1680 * noflush == true is from the waitqueue handler, just ensure
1681 * we wake up the task, and the next invocation will flush the
1682 * entries. We cannot safely to it from here.
1684 if (noflush && !list_empty(&ctx->cq_overflow_list))
1687 io_cqring_overflow_flush(ctx, false);
1690 /* See comment at the top of this file */
1692 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1695 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1697 struct io_rings *rings = ctx->rings;
1699 /* make sure SQ entry isn't read before tail */
1700 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1703 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1705 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1708 if (req->file || req->io)
1711 rb->reqs[rb->to_free++] = req;
1712 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1713 io_free_req_many(req->ctx, rb);
1717 static int io_put_kbuf(struct io_kiocb *req)
1719 struct io_buffer *kbuf;
1722 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1723 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1724 cflags |= IORING_CQE_F_BUFFER;
1731 * Find and free completed poll iocbs
1733 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1734 struct list_head *done)
1736 struct req_batch rb;
1737 struct io_kiocb *req;
1739 rb.to_free = rb.need_iter = 0;
1740 while (!list_empty(done)) {
1743 req = list_first_entry(done, struct io_kiocb, list);
1744 list_del(&req->list);
1746 if (req->flags & REQ_F_BUFFER_SELECTED)
1747 cflags = io_put_kbuf(req);
1749 __io_cqring_fill_event(req, req->result, cflags);
1752 if (refcount_dec_and_test(&req->refs) &&
1753 !io_req_multi_free(&rb, req))
1757 io_commit_cqring(ctx);
1758 if (ctx->flags & IORING_SETUP_SQPOLL)
1759 io_cqring_ev_posted(ctx);
1760 io_free_req_many(ctx, &rb);
1763 static void io_iopoll_queue(struct list_head *again)
1765 struct io_kiocb *req;
1768 req = list_first_entry(again, struct io_kiocb, list);
1769 list_del(&req->list);
1770 refcount_inc(&req->refs);
1771 io_queue_async_work(req);
1772 } while (!list_empty(again));
1775 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1778 struct io_kiocb *req, *tmp;
1785 * Only spin for completions if we don't have multiple devices hanging
1786 * off our complete list, and we're under the requested amount.
1788 spin = !ctx->poll_multi_file && *nr_events < min;
1791 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1792 struct kiocb *kiocb = &req->rw.kiocb;
1795 * Move completed and retryable entries to our local lists.
1796 * If we find a request that requires polling, break out
1797 * and complete those lists first, if we have entries there.
1799 if (READ_ONCE(req->iopoll_completed)) {
1800 list_move_tail(&req->list, &done);
1803 if (!list_empty(&done))
1806 if (req->result == -EAGAIN) {
1807 list_move_tail(&req->list, &again);
1810 if (!list_empty(&again))
1813 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1822 if (!list_empty(&done))
1823 io_iopoll_complete(ctx, nr_events, &done);
1825 if (!list_empty(&again))
1826 io_iopoll_queue(&again);
1832 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1833 * non-spinning poll check - we'll still enter the driver poll loop, but only
1834 * as a non-spinning completion check.
1836 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1839 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1842 ret = io_do_iopoll(ctx, nr_events, min);
1845 if (!min || *nr_events >= min)
1853 * We can't just wait for polled events to come to us, we have to actively
1854 * find and complete them.
1856 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1858 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1861 mutex_lock(&ctx->uring_lock);
1862 while (!list_empty(&ctx->poll_list)) {
1863 unsigned int nr_events = 0;
1865 io_iopoll_getevents(ctx, &nr_events, 1);
1868 * Ensure we allow local-to-the-cpu processing to take place,
1869 * in this case we need to ensure that we reap all events.
1873 mutex_unlock(&ctx->uring_lock);
1876 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1879 int iters = 0, ret = 0;
1882 * We disallow the app entering submit/complete with polling, but we
1883 * still need to lock the ring to prevent racing with polled issue
1884 * that got punted to a workqueue.
1886 mutex_lock(&ctx->uring_lock);
1891 * Don't enter poll loop if we already have events pending.
1892 * If we do, we can potentially be spinning for commands that
1893 * already triggered a CQE (eg in error).
1895 if (io_cqring_events(ctx, false))
1899 * If a submit got punted to a workqueue, we can have the
1900 * application entering polling for a command before it gets
1901 * issued. That app will hold the uring_lock for the duration
1902 * of the poll right here, so we need to take a breather every
1903 * now and then to ensure that the issue has a chance to add
1904 * the poll to the issued list. Otherwise we can spin here
1905 * forever, while the workqueue is stuck trying to acquire the
1908 if (!(++iters & 7)) {
1909 mutex_unlock(&ctx->uring_lock);
1910 mutex_lock(&ctx->uring_lock);
1913 if (*nr_events < min)
1914 tmin = min - *nr_events;
1916 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1920 } while (min && !*nr_events && !need_resched());
1922 mutex_unlock(&ctx->uring_lock);
1926 static void kiocb_end_write(struct io_kiocb *req)
1929 * Tell lockdep we inherited freeze protection from submission
1932 if (req->flags & REQ_F_ISREG) {
1933 struct inode *inode = file_inode(req->file);
1935 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1937 file_end_write(req->file);
1940 static inline void req_set_fail_links(struct io_kiocb *req)
1942 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1943 req->flags |= REQ_F_FAIL_LINK;
1946 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1948 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1951 if (kiocb->ki_flags & IOCB_WRITE)
1952 kiocb_end_write(req);
1954 if (res != req->result)
1955 req_set_fail_links(req);
1956 if (req->flags & REQ_F_BUFFER_SELECTED)
1957 cflags = io_put_kbuf(req);
1958 __io_cqring_add_event(req, res, cflags);
1961 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1963 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1965 io_complete_rw_common(kiocb, res);
1969 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1971 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1973 if (kiocb->ki_flags & IOCB_WRITE)
1974 kiocb_end_write(req);
1976 if (res != req->result)
1977 req_set_fail_links(req);
1980 WRITE_ONCE(req->iopoll_completed, 1);
1984 * After the iocb has been issued, it's safe to be found on the poll list.
1985 * Adding the kiocb to the list AFTER submission ensures that we don't
1986 * find it from a io_iopoll_getevents() thread before the issuer is done
1987 * accessing the kiocb cookie.
1989 static void io_iopoll_req_issued(struct io_kiocb *req)
1991 struct io_ring_ctx *ctx = req->ctx;
1994 * Track whether we have multiple files in our lists. This will impact
1995 * how we do polling eventually, not spinning if we're on potentially
1996 * different devices.
1998 if (list_empty(&ctx->poll_list)) {
1999 ctx->poll_multi_file = false;
2000 } else if (!ctx->poll_multi_file) {
2001 struct io_kiocb *list_req;
2003 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2005 if (list_req->file != req->file)
2006 ctx->poll_multi_file = true;
2010 * For fast devices, IO may have already completed. If it has, add
2011 * it to the front so we find it first.
2013 if (READ_ONCE(req->iopoll_completed))
2014 list_add(&req->list, &ctx->poll_list);
2016 list_add_tail(&req->list, &ctx->poll_list);
2018 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2019 wq_has_sleeper(&ctx->sqo_wait))
2020 wake_up(&ctx->sqo_wait);
2023 static void __io_state_file_put(struct io_submit_state *state)
2025 int diff = state->has_refs - state->used_refs;
2028 fput_many(state->file, diff);
2032 static inline void io_state_file_put(struct io_submit_state *state)
2035 __io_state_file_put(state);
2039 * Get as many references to a file as we have IOs left in this submission,
2040 * assuming most submissions are for one file, or at least that each file
2041 * has more than one submission.
2043 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2049 if (state->fd == fd) {
2054 __io_state_file_put(state);
2056 state->file = fget_many(fd, state->ios_left);
2061 state->has_refs = state->ios_left;
2062 state->used_refs = 1;
2068 * If we tracked the file through the SCM inflight mechanism, we could support
2069 * any file. For now, just ensure that anything potentially problematic is done
2072 static bool io_file_supports_async(struct file *file, int rw)
2074 umode_t mode = file_inode(file)->i_mode;
2076 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2078 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2081 /* any ->read/write should understand O_NONBLOCK */
2082 if (file->f_flags & O_NONBLOCK)
2085 if (!(file->f_mode & FMODE_NOWAIT))
2089 return file->f_op->read_iter != NULL;
2091 return file->f_op->write_iter != NULL;
2094 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2095 bool force_nonblock)
2097 struct io_ring_ctx *ctx = req->ctx;
2098 struct kiocb *kiocb = &req->rw.kiocb;
2102 if (S_ISREG(file_inode(req->file)->i_mode))
2103 req->flags |= REQ_F_ISREG;
2105 kiocb->ki_pos = READ_ONCE(sqe->off);
2106 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2107 req->flags |= REQ_F_CUR_POS;
2108 kiocb->ki_pos = req->file->f_pos;
2110 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2111 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2112 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2116 ioprio = READ_ONCE(sqe->ioprio);
2118 ret = ioprio_check_cap(ioprio);
2122 kiocb->ki_ioprio = ioprio;
2124 kiocb->ki_ioprio = get_current_ioprio();
2126 /* don't allow async punt if RWF_NOWAIT was requested */
2127 if (kiocb->ki_flags & IOCB_NOWAIT)
2128 req->flags |= REQ_F_NOWAIT;
2131 kiocb->ki_flags |= IOCB_NOWAIT;
2133 if (ctx->flags & IORING_SETUP_IOPOLL) {
2134 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2135 !kiocb->ki_filp->f_op->iopoll)
2138 kiocb->ki_flags |= IOCB_HIPRI;
2139 kiocb->ki_complete = io_complete_rw_iopoll;
2141 req->iopoll_completed = 0;
2143 if (kiocb->ki_flags & IOCB_HIPRI)
2145 kiocb->ki_complete = io_complete_rw;
2148 req->rw.addr = READ_ONCE(sqe->addr);
2149 req->rw.len = READ_ONCE(sqe->len);
2150 req->buf_index = READ_ONCE(sqe->buf_index);
2154 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2160 case -ERESTARTNOINTR:
2161 case -ERESTARTNOHAND:
2162 case -ERESTART_RESTARTBLOCK:
2164 * We can't just restart the syscall, since previously
2165 * submitted sqes may already be in progress. Just fail this
2171 kiocb->ki_complete(kiocb, ret, 0);
2175 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2177 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2179 if (req->flags & REQ_F_CUR_POS)
2180 req->file->f_pos = kiocb->ki_pos;
2181 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2182 io_complete_rw(kiocb, ret, 0);
2184 io_rw_done(kiocb, ret);
2187 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2188 struct iov_iter *iter)
2190 struct io_ring_ctx *ctx = req->ctx;
2191 size_t len = req->rw.len;
2192 struct io_mapped_ubuf *imu;
2193 u16 index, buf_index;
2197 /* attempt to use fixed buffers without having provided iovecs */
2198 if (unlikely(!ctx->user_bufs))
2201 buf_index = req->buf_index;
2202 if (unlikely(buf_index >= ctx->nr_user_bufs))
2205 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2206 imu = &ctx->user_bufs[index];
2207 buf_addr = req->rw.addr;
2210 if (buf_addr + len < buf_addr)
2212 /* not inside the mapped region */
2213 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2217 * May not be a start of buffer, set size appropriately
2218 * and advance us to the beginning.
2220 offset = buf_addr - imu->ubuf;
2221 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2225 * Don't use iov_iter_advance() here, as it's really slow for
2226 * using the latter parts of a big fixed buffer - it iterates
2227 * over each segment manually. We can cheat a bit here, because
2230 * 1) it's a BVEC iter, we set it up
2231 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2232 * first and last bvec
2234 * So just find our index, and adjust the iterator afterwards.
2235 * If the offset is within the first bvec (or the whole first
2236 * bvec, just use iov_iter_advance(). This makes it easier
2237 * since we can just skip the first segment, which may not
2238 * be PAGE_SIZE aligned.
2240 const struct bio_vec *bvec = imu->bvec;
2242 if (offset <= bvec->bv_len) {
2243 iov_iter_advance(iter, offset);
2245 unsigned long seg_skip;
2247 /* skip first vec */
2248 offset -= bvec->bv_len;
2249 seg_skip = 1 + (offset >> PAGE_SHIFT);
2251 iter->bvec = bvec + seg_skip;
2252 iter->nr_segs -= seg_skip;
2253 iter->count -= bvec->bv_len + offset;
2254 iter->iov_offset = offset & ~PAGE_MASK;
2261 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2264 mutex_unlock(&ctx->uring_lock);
2267 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2270 * "Normal" inline submissions always hold the uring_lock, since we
2271 * grab it from the system call. Same is true for the SQPOLL offload.
2272 * The only exception is when we've detached the request and issue it
2273 * from an async worker thread, grab the lock for that case.
2276 mutex_lock(&ctx->uring_lock);
2279 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2280 int bgid, struct io_buffer *kbuf,
2283 struct io_buffer *head;
2285 if (req->flags & REQ_F_BUFFER_SELECTED)
2288 io_ring_submit_lock(req->ctx, needs_lock);
2290 lockdep_assert_held(&req->ctx->uring_lock);
2292 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2294 if (!list_empty(&head->list)) {
2295 kbuf = list_last_entry(&head->list, struct io_buffer,
2297 list_del(&kbuf->list);
2300 idr_remove(&req->ctx->io_buffer_idr, bgid);
2302 if (*len > kbuf->len)
2305 kbuf = ERR_PTR(-ENOBUFS);
2308 io_ring_submit_unlock(req->ctx, needs_lock);
2313 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2316 struct io_buffer *kbuf;
2319 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2320 bgid = req->buf_index;
2321 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2324 req->rw.addr = (u64) (unsigned long) kbuf;
2325 req->flags |= REQ_F_BUFFER_SELECTED;
2326 return u64_to_user_ptr(kbuf->addr);
2329 #ifdef CONFIG_COMPAT
2330 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2333 struct compat_iovec __user *uiov;
2334 compat_ssize_t clen;
2338 uiov = u64_to_user_ptr(req->rw.addr);
2339 if (!access_ok(uiov, sizeof(*uiov)))
2341 if (__get_user(clen, &uiov->iov_len))
2347 buf = io_rw_buffer_select(req, &len, needs_lock);
2349 return PTR_ERR(buf);
2350 iov[0].iov_base = buf;
2351 iov[0].iov_len = (compat_size_t) len;
2356 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2359 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2363 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2366 len = iov[0].iov_len;
2369 buf = io_rw_buffer_select(req, &len, needs_lock);
2371 return PTR_ERR(buf);
2372 iov[0].iov_base = buf;
2373 iov[0].iov_len = len;
2377 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2380 if (req->flags & REQ_F_BUFFER_SELECTED) {
2381 struct io_buffer *kbuf;
2383 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2384 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2385 iov[0].iov_len = kbuf->len;
2390 else if (req->rw.len > 1)
2393 #ifdef CONFIG_COMPAT
2394 if (req->ctx->compat)
2395 return io_compat_import(req, iov, needs_lock);
2398 return __io_iov_buffer_select(req, iov, needs_lock);
2401 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2402 struct iovec **iovec, struct iov_iter *iter,
2405 void __user *buf = u64_to_user_ptr(req->rw.addr);
2406 size_t sqe_len = req->rw.len;
2410 opcode = req->opcode;
2411 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2413 return io_import_fixed(req, rw, iter);
2416 /* buffer index only valid with fixed read/write, or buffer select */
2417 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2420 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2421 if (req->flags & REQ_F_BUFFER_SELECT) {
2422 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2425 return PTR_ERR(buf);
2427 req->rw.len = sqe_len;
2430 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2432 return ret < 0 ? ret : sqe_len;
2436 struct io_async_rw *iorw = &req->io->rw;
2439 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2440 if (iorw->iov == iorw->fast_iov)
2445 if (req->flags & REQ_F_BUFFER_SELECT) {
2446 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2448 ret = (*iovec)->iov_len;
2449 iov_iter_init(iter, rw, *iovec, 1, ret);
2455 #ifdef CONFIG_COMPAT
2456 if (req->ctx->compat)
2457 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2461 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2465 * For files that don't have ->read_iter() and ->write_iter(), handle them
2466 * by looping over ->read() or ->write() manually.
2468 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2469 struct iov_iter *iter)
2474 * Don't support polled IO through this interface, and we can't
2475 * support non-blocking either. For the latter, this just causes
2476 * the kiocb to be handled from an async context.
2478 if (kiocb->ki_flags & IOCB_HIPRI)
2480 if (kiocb->ki_flags & IOCB_NOWAIT)
2483 while (iov_iter_count(iter)) {
2487 if (!iov_iter_is_bvec(iter)) {
2488 iovec = iov_iter_iovec(iter);
2490 /* fixed buffers import bvec */
2491 iovec.iov_base = kmap(iter->bvec->bv_page)
2493 iovec.iov_len = min(iter->count,
2494 iter->bvec->bv_len - iter->iov_offset);
2498 nr = file->f_op->read(file, iovec.iov_base,
2499 iovec.iov_len, &kiocb->ki_pos);
2501 nr = file->f_op->write(file, iovec.iov_base,
2502 iovec.iov_len, &kiocb->ki_pos);
2505 if (iov_iter_is_bvec(iter))
2506 kunmap(iter->bvec->bv_page);
2514 if (nr != iovec.iov_len)
2516 iov_iter_advance(iter, nr);
2522 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2523 struct iovec *iovec, struct iovec *fast_iov,
2524 struct iov_iter *iter)
2526 req->io->rw.nr_segs = iter->nr_segs;
2527 req->io->rw.size = io_size;
2528 req->io->rw.iov = iovec;
2529 if (!req->io->rw.iov) {
2530 req->io->rw.iov = req->io->rw.fast_iov;
2531 if (req->io->rw.iov != fast_iov)
2532 memcpy(req->io->rw.iov, fast_iov,
2533 sizeof(struct iovec) * iter->nr_segs);
2535 req->flags |= REQ_F_NEED_CLEANUP;
2539 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2541 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2542 return req->io == NULL;
2545 static int io_alloc_async_ctx(struct io_kiocb *req)
2547 if (!io_op_defs[req->opcode].async_ctx)
2550 return __io_alloc_async_ctx(req);
2553 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2554 struct iovec *iovec, struct iovec *fast_iov,
2555 struct iov_iter *iter)
2557 if (!io_op_defs[req->opcode].async_ctx)
2560 if (__io_alloc_async_ctx(req))
2563 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2568 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2569 bool force_nonblock)
2571 struct io_async_ctx *io;
2572 struct iov_iter iter;
2575 ret = io_prep_rw(req, sqe, force_nonblock);
2579 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2582 /* either don't need iovec imported or already have it */
2583 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2587 io->rw.iov = io->rw.fast_iov;
2589 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2594 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2598 static int io_read(struct io_kiocb *req, bool force_nonblock)
2600 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2601 struct kiocb *kiocb = &req->rw.kiocb;
2602 struct iov_iter iter;
2604 ssize_t io_size, ret;
2606 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2610 /* Ensure we clear previously set non-block flag */
2611 if (!force_nonblock)
2612 kiocb->ki_flags &= ~IOCB_NOWAIT;
2616 if (req->flags & REQ_F_LINK_HEAD)
2617 req->result = io_size;
2620 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2621 * we know to async punt it even if it was opened O_NONBLOCK
2623 if (force_nonblock && !io_file_supports_async(req->file, READ))
2626 iov_count = iov_iter_count(&iter);
2627 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2631 if (req->file->f_op->read_iter)
2632 ret2 = call_read_iter(req->file, kiocb, &iter);
2634 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2636 /* Catch -EAGAIN return for forced non-blocking submission */
2637 if (!force_nonblock || ret2 != -EAGAIN) {
2638 kiocb_done(kiocb, ret2);
2641 ret = io_setup_async_rw(req, io_size, iovec,
2642 inline_vecs, &iter);
2645 /* any defer here is final, must blocking retry */
2646 if (!(req->flags & REQ_F_NOWAIT) &&
2647 !file_can_poll(req->file))
2648 req->flags |= REQ_F_MUST_PUNT;
2654 req->flags &= ~REQ_F_NEED_CLEANUP;
2658 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2659 bool force_nonblock)
2661 struct io_async_ctx *io;
2662 struct iov_iter iter;
2665 ret = io_prep_rw(req, sqe, force_nonblock);
2669 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2672 req->fsize = rlimit(RLIMIT_FSIZE);
2674 /* either don't need iovec imported or already have it */
2675 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2679 io->rw.iov = io->rw.fast_iov;
2681 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2686 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2690 static int io_write(struct io_kiocb *req, bool force_nonblock)
2692 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2693 struct kiocb *kiocb = &req->rw.kiocb;
2694 struct iov_iter iter;
2696 ssize_t ret, io_size;
2698 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2702 /* Ensure we clear previously set non-block flag */
2703 if (!force_nonblock)
2704 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2708 if (req->flags & REQ_F_LINK_HEAD)
2709 req->result = io_size;
2712 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2713 * we know to async punt it even if it was opened O_NONBLOCK
2715 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2718 /* file path doesn't support NOWAIT for non-direct_IO */
2719 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2720 (req->flags & REQ_F_ISREG))
2723 iov_count = iov_iter_count(&iter);
2724 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2729 * Open-code file_start_write here to grab freeze protection,
2730 * which will be released by another thread in
2731 * io_complete_rw(). Fool lockdep by telling it the lock got
2732 * released so that it doesn't complain about the held lock when
2733 * we return to userspace.
2735 if (req->flags & REQ_F_ISREG) {
2736 __sb_start_write(file_inode(req->file)->i_sb,
2737 SB_FREEZE_WRITE, true);
2738 __sb_writers_release(file_inode(req->file)->i_sb,
2741 kiocb->ki_flags |= IOCB_WRITE;
2743 if (!force_nonblock)
2744 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2746 if (req->file->f_op->write_iter)
2747 ret2 = call_write_iter(req->file, kiocb, &iter);
2749 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2751 if (!force_nonblock)
2752 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2755 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2756 * retry them without IOCB_NOWAIT.
2758 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2760 if (!force_nonblock || ret2 != -EAGAIN) {
2761 kiocb_done(kiocb, ret2);
2764 ret = io_setup_async_rw(req, io_size, iovec,
2765 inline_vecs, &iter);
2768 /* any defer here is final, must blocking retry */
2769 if (!(req->flags & REQ_F_NOWAIT) &&
2770 !file_can_poll(req->file))
2771 req->flags |= REQ_F_MUST_PUNT;
2776 req->flags &= ~REQ_F_NEED_CLEANUP;
2781 static int __io_splice_prep(struct io_kiocb *req,
2782 const struct io_uring_sqe *sqe)
2784 struct io_splice* sp = &req->splice;
2785 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2788 if (req->flags & REQ_F_NEED_CLEANUP)
2790 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2794 sp->len = READ_ONCE(sqe->len);
2795 sp->flags = READ_ONCE(sqe->splice_flags);
2797 if (unlikely(sp->flags & ~valid_flags))
2800 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2801 (sp->flags & SPLICE_F_FD_IN_FIXED));
2804 req->flags |= REQ_F_NEED_CLEANUP;
2806 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
2808 * Splice operation will be punted aync, and here need to
2809 * modify io_wq_work.flags, so initialize io_wq_work firstly.
2811 io_req_init_async(req);
2812 req->work.flags |= IO_WQ_WORK_UNBOUND;
2818 static int io_tee_prep(struct io_kiocb *req,
2819 const struct io_uring_sqe *sqe)
2821 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
2823 return __io_splice_prep(req, sqe);
2826 static int io_tee(struct io_kiocb *req, bool force_nonblock)
2828 struct io_splice *sp = &req->splice;
2829 struct file *in = sp->file_in;
2830 struct file *out = sp->file_out;
2831 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2837 ret = do_tee(in, out, sp->len, flags);
2839 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2840 req->flags &= ~REQ_F_NEED_CLEANUP;
2842 io_cqring_add_event(req, ret);
2844 req_set_fail_links(req);
2849 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2851 struct io_splice* sp = &req->splice;
2853 sp->off_in = READ_ONCE(sqe->splice_off_in);
2854 sp->off_out = READ_ONCE(sqe->off);
2855 return __io_splice_prep(req, sqe);
2858 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2860 struct io_splice *sp = &req->splice;
2861 struct file *in = sp->file_in;
2862 struct file *out = sp->file_out;
2863 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2864 loff_t *poff_in, *poff_out;
2870 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2871 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2874 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2876 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2877 req->flags &= ~REQ_F_NEED_CLEANUP;
2879 io_cqring_add_event(req, ret);
2881 req_set_fail_links(req);
2887 * IORING_OP_NOP just posts a completion event, nothing else.
2889 static int io_nop(struct io_kiocb *req)
2891 struct io_ring_ctx *ctx = req->ctx;
2893 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2896 io_cqring_add_event(req, 0);
2901 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2903 struct io_ring_ctx *ctx = req->ctx;
2908 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2910 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2913 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2914 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2917 req->sync.off = READ_ONCE(sqe->off);
2918 req->sync.len = READ_ONCE(sqe->len);
2922 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2924 loff_t end = req->sync.off + req->sync.len;
2927 /* fsync always requires a blocking context */
2931 ret = vfs_fsync_range(req->file, req->sync.off,
2932 end > 0 ? end : LLONG_MAX,
2933 req->sync.flags & IORING_FSYNC_DATASYNC);
2935 req_set_fail_links(req);
2936 io_cqring_add_event(req, ret);
2941 static int io_fallocate_prep(struct io_kiocb *req,
2942 const struct io_uring_sqe *sqe)
2944 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2946 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2949 req->sync.off = READ_ONCE(sqe->off);
2950 req->sync.len = READ_ONCE(sqe->addr);
2951 req->sync.mode = READ_ONCE(sqe->len);
2952 req->fsize = rlimit(RLIMIT_FSIZE);
2956 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2960 /* fallocate always requiring blocking context */
2964 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2965 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2967 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2969 req_set_fail_links(req);
2970 io_cqring_add_event(req, ret);
2975 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2977 const char __user *fname;
2980 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2982 if (unlikely(sqe->ioprio || sqe->buf_index))
2984 if (unlikely(req->flags & REQ_F_FIXED_FILE))
2987 /* open.how should be already initialised */
2988 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2989 req->open.how.flags |= O_LARGEFILE;
2991 req->open.dfd = READ_ONCE(sqe->fd);
2992 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2993 req->open.filename = getname(fname);
2994 if (IS_ERR(req->open.filename)) {
2995 ret = PTR_ERR(req->open.filename);
2996 req->open.filename = NULL;
2999 req->open.nofile = rlimit(RLIMIT_NOFILE);
3000 req->flags |= REQ_F_NEED_CLEANUP;
3004 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3008 if (req->flags & REQ_F_NEED_CLEANUP)
3010 mode = READ_ONCE(sqe->len);
3011 flags = READ_ONCE(sqe->open_flags);
3012 req->open.how = build_open_how(flags, mode);
3013 return __io_openat_prep(req, sqe);
3016 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3018 struct open_how __user *how;
3022 if (req->flags & REQ_F_NEED_CLEANUP)
3024 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3025 len = READ_ONCE(sqe->len);
3026 if (len < OPEN_HOW_SIZE_VER0)
3029 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3034 return __io_openat_prep(req, sqe);
3037 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3039 struct open_flags op;
3046 ret = build_open_flags(&req->open.how, &op);
3050 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3054 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3057 ret = PTR_ERR(file);
3059 fsnotify_open(file);
3060 fd_install(ret, file);
3063 putname(req->open.filename);
3064 req->flags &= ~REQ_F_NEED_CLEANUP;
3066 req_set_fail_links(req);
3067 io_cqring_add_event(req, ret);
3072 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3074 return io_openat2(req, force_nonblock);
3077 static int io_remove_buffers_prep(struct io_kiocb *req,
3078 const struct io_uring_sqe *sqe)
3080 struct io_provide_buf *p = &req->pbuf;
3083 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3086 tmp = READ_ONCE(sqe->fd);
3087 if (!tmp || tmp > USHRT_MAX)
3090 memset(p, 0, sizeof(*p));
3092 p->bgid = READ_ONCE(sqe->buf_group);
3096 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3097 int bgid, unsigned nbufs)
3101 /* shouldn't happen */
3105 /* the head kbuf is the list itself */
3106 while (!list_empty(&buf->list)) {
3107 struct io_buffer *nxt;
3109 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3110 list_del(&nxt->list);
3117 idr_remove(&ctx->io_buffer_idr, bgid);
3122 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3124 struct io_provide_buf *p = &req->pbuf;
3125 struct io_ring_ctx *ctx = req->ctx;
3126 struct io_buffer *head;
3129 io_ring_submit_lock(ctx, !force_nonblock);
3131 lockdep_assert_held(&ctx->uring_lock);
3134 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3136 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3138 io_ring_submit_lock(ctx, !force_nonblock);
3140 req_set_fail_links(req);
3141 io_cqring_add_event(req, ret);
3146 static int io_provide_buffers_prep(struct io_kiocb *req,
3147 const struct io_uring_sqe *sqe)
3149 struct io_provide_buf *p = &req->pbuf;
3152 if (sqe->ioprio || sqe->rw_flags)
3155 tmp = READ_ONCE(sqe->fd);
3156 if (!tmp || tmp > USHRT_MAX)
3159 p->addr = READ_ONCE(sqe->addr);
3160 p->len = READ_ONCE(sqe->len);
3162 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3165 p->bgid = READ_ONCE(sqe->buf_group);
3166 tmp = READ_ONCE(sqe->off);
3167 if (tmp > USHRT_MAX)
3173 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3175 struct io_buffer *buf;
3176 u64 addr = pbuf->addr;
3177 int i, bid = pbuf->bid;
3179 for (i = 0; i < pbuf->nbufs; i++) {
3180 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3185 buf->len = pbuf->len;
3190 INIT_LIST_HEAD(&buf->list);
3193 list_add_tail(&buf->list, &(*head)->list);
3197 return i ? i : -ENOMEM;
3200 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3202 struct io_provide_buf *p = &req->pbuf;
3203 struct io_ring_ctx *ctx = req->ctx;
3204 struct io_buffer *head, *list;
3207 io_ring_submit_lock(ctx, !force_nonblock);
3209 lockdep_assert_held(&ctx->uring_lock);
3211 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3213 ret = io_add_buffers(p, &head);
3218 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3221 __io_remove_buffers(ctx, head, p->bgid, -1U);
3226 io_ring_submit_unlock(ctx, !force_nonblock);
3228 req_set_fail_links(req);
3229 io_cqring_add_event(req, ret);
3234 static int io_epoll_ctl_prep(struct io_kiocb *req,
3235 const struct io_uring_sqe *sqe)
3237 #if defined(CONFIG_EPOLL)
3238 if (sqe->ioprio || sqe->buf_index)
3240 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3243 req->epoll.epfd = READ_ONCE(sqe->fd);
3244 req->epoll.op = READ_ONCE(sqe->len);
3245 req->epoll.fd = READ_ONCE(sqe->off);
3247 if (ep_op_has_event(req->epoll.op)) {
3248 struct epoll_event __user *ev;
3250 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3251 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3261 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3263 #if defined(CONFIG_EPOLL)
3264 struct io_epoll *ie = &req->epoll;
3267 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3268 if (force_nonblock && ret == -EAGAIN)
3272 req_set_fail_links(req);
3273 io_cqring_add_event(req, ret);
3281 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3283 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3284 if (sqe->ioprio || sqe->buf_index || sqe->off)
3286 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3289 req->madvise.addr = READ_ONCE(sqe->addr);
3290 req->madvise.len = READ_ONCE(sqe->len);
3291 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3298 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3300 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3301 struct io_madvise *ma = &req->madvise;
3307 ret = do_madvise(ma->addr, ma->len, ma->advice);
3309 req_set_fail_links(req);
3310 io_cqring_add_event(req, ret);
3318 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3320 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3322 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3325 req->fadvise.offset = READ_ONCE(sqe->off);
3326 req->fadvise.len = READ_ONCE(sqe->len);
3327 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3331 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3333 struct io_fadvise *fa = &req->fadvise;
3336 if (force_nonblock) {
3337 switch (fa->advice) {
3338 case POSIX_FADV_NORMAL:
3339 case POSIX_FADV_RANDOM:
3340 case POSIX_FADV_SEQUENTIAL:
3347 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3349 req_set_fail_links(req);
3350 io_cqring_add_event(req, ret);
3355 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3357 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3359 if (sqe->ioprio || sqe->buf_index)
3361 if (req->flags & REQ_F_FIXED_FILE)
3364 req->statx.dfd = READ_ONCE(sqe->fd);
3365 req->statx.mask = READ_ONCE(sqe->len);
3366 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3367 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3368 req->statx.flags = READ_ONCE(sqe->statx_flags);
3373 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3375 struct io_statx *ctx = &req->statx;
3378 if (force_nonblock) {
3379 /* only need file table for an actual valid fd */
3380 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3381 req->flags |= REQ_F_NO_FILE_TABLE;
3385 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3389 req_set_fail_links(req);
3390 io_cqring_add_event(req, ret);
3395 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3398 * If we queue this for async, it must not be cancellable. That would
3399 * leave the 'file' in an undeterminate state, and here need to modify
3400 * io_wq_work.flags, so initialize io_wq_work firstly.
3402 io_req_init_async(req);
3403 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3405 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3407 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3408 sqe->rw_flags || sqe->buf_index)
3410 if (req->flags & REQ_F_FIXED_FILE)
3413 req->close.fd = READ_ONCE(sqe->fd);
3414 if ((req->file && req->file->f_op == &io_uring_fops) ||
3415 req->close.fd == req->ctx->ring_fd)
3418 req->close.put_file = NULL;
3422 static int io_close(struct io_kiocb *req, bool force_nonblock)
3424 struct io_close *close = &req->close;
3427 /* might be already done during nonblock submission */
3428 if (!close->put_file) {
3429 ret = __close_fd_get_file(close->fd, &close->put_file);
3431 return (ret == -ENOENT) ? -EBADF : ret;
3434 /* if the file has a flush method, be safe and punt to async */
3435 if (close->put_file->f_op->flush && force_nonblock) {
3436 /* avoid grabbing files - we don't need the files */
3437 req->flags |= REQ_F_NO_FILE_TABLE | REQ_F_MUST_PUNT;
3441 /* No ->flush() or already async, safely close from here */
3442 ret = filp_close(close->put_file, req->work.files);
3444 req_set_fail_links(req);
3445 io_cqring_add_event(req, ret);
3446 fput(close->put_file);
3447 close->put_file = NULL;
3452 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3454 struct io_ring_ctx *ctx = req->ctx;
3459 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3461 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3464 req->sync.off = READ_ONCE(sqe->off);
3465 req->sync.len = READ_ONCE(sqe->len);
3466 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3470 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3474 /* sync_file_range always requires a blocking context */
3478 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3481 req_set_fail_links(req);
3482 io_cqring_add_event(req, ret);
3487 #if defined(CONFIG_NET)
3488 static int io_setup_async_msg(struct io_kiocb *req,
3489 struct io_async_msghdr *kmsg)
3493 if (io_alloc_async_ctx(req)) {
3494 if (kmsg->iov != kmsg->fast_iov)
3498 req->flags |= REQ_F_NEED_CLEANUP;
3499 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3503 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3505 struct io_sr_msg *sr = &req->sr_msg;
3506 struct io_async_ctx *io = req->io;
3509 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3512 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3513 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3514 sr->len = READ_ONCE(sqe->len);
3516 #ifdef CONFIG_COMPAT
3517 if (req->ctx->compat)
3518 sr->msg_flags |= MSG_CMSG_COMPAT;
3521 if (!io || req->opcode == IORING_OP_SEND)
3523 /* iovec is already imported */
3524 if (req->flags & REQ_F_NEED_CLEANUP)
3527 io->msg.iov = io->msg.fast_iov;
3528 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3531 req->flags |= REQ_F_NEED_CLEANUP;
3535 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3537 struct io_async_msghdr *kmsg = NULL;
3538 struct socket *sock;
3541 sock = sock_from_file(req->file, &ret);
3543 struct io_async_ctx io;
3547 kmsg = &req->io->msg;
3548 kmsg->msg.msg_name = &req->io->msg.addr;
3549 /* if iov is set, it's allocated already */
3551 kmsg->iov = kmsg->fast_iov;
3552 kmsg->msg.msg_iter.iov = kmsg->iov;
3554 struct io_sr_msg *sr = &req->sr_msg;
3557 kmsg->msg.msg_name = &io.msg.addr;
3559 io.msg.iov = io.msg.fast_iov;
3560 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3561 sr->msg_flags, &io.msg.iov);
3566 flags = req->sr_msg.msg_flags;
3567 if (flags & MSG_DONTWAIT)
3568 req->flags |= REQ_F_NOWAIT;
3569 else if (force_nonblock)
3570 flags |= MSG_DONTWAIT;
3572 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3573 if (force_nonblock && ret == -EAGAIN)
3574 return io_setup_async_msg(req, kmsg);
3575 if (ret == -ERESTARTSYS)
3579 if (kmsg && kmsg->iov != kmsg->fast_iov)
3581 req->flags &= ~REQ_F_NEED_CLEANUP;
3582 io_cqring_add_event(req, ret);
3584 req_set_fail_links(req);
3589 static int io_send(struct io_kiocb *req, bool force_nonblock)
3591 struct socket *sock;
3594 sock = sock_from_file(req->file, &ret);
3596 struct io_sr_msg *sr = &req->sr_msg;
3601 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3606 msg.msg_name = NULL;
3607 msg.msg_control = NULL;
3608 msg.msg_controllen = 0;
3609 msg.msg_namelen = 0;
3611 flags = req->sr_msg.msg_flags;
3612 if (flags & MSG_DONTWAIT)
3613 req->flags |= REQ_F_NOWAIT;
3614 else if (force_nonblock)
3615 flags |= MSG_DONTWAIT;
3617 msg.msg_flags = flags;
3618 ret = sock_sendmsg(sock, &msg);
3619 if (force_nonblock && ret == -EAGAIN)
3621 if (ret == -ERESTARTSYS)
3625 io_cqring_add_event(req, ret);
3627 req_set_fail_links(req);
3632 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3634 struct io_sr_msg *sr = &req->sr_msg;
3635 struct iovec __user *uiov;
3639 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3644 if (req->flags & REQ_F_BUFFER_SELECT) {
3647 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3649 sr->len = io->msg.iov[0].iov_len;
3650 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3654 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3655 &io->msg.iov, &io->msg.msg.msg_iter);
3663 #ifdef CONFIG_COMPAT
3664 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3665 struct io_async_ctx *io)
3667 struct compat_msghdr __user *msg_compat;
3668 struct io_sr_msg *sr = &req->sr_msg;
3669 struct compat_iovec __user *uiov;
3674 msg_compat = (struct compat_msghdr __user *) sr->msg;
3675 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3680 uiov = compat_ptr(ptr);
3681 if (req->flags & REQ_F_BUFFER_SELECT) {
3682 compat_ssize_t clen;
3686 if (!access_ok(uiov, sizeof(*uiov)))
3688 if (__get_user(clen, &uiov->iov_len))
3692 sr->len = io->msg.iov[0].iov_len;
3695 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3697 &io->msg.msg.msg_iter);
3706 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3708 io->msg.iov = io->msg.fast_iov;
3710 #ifdef CONFIG_COMPAT
3711 if (req->ctx->compat)
3712 return __io_compat_recvmsg_copy_hdr(req, io);
3715 return __io_recvmsg_copy_hdr(req, io);
3718 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3719 int *cflags, bool needs_lock)
3721 struct io_sr_msg *sr = &req->sr_msg;
3722 struct io_buffer *kbuf;
3724 if (!(req->flags & REQ_F_BUFFER_SELECT))
3727 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3732 req->flags |= REQ_F_BUFFER_SELECTED;
3734 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3735 *cflags |= IORING_CQE_F_BUFFER;
3739 static int io_recvmsg_prep(struct io_kiocb *req,
3740 const struct io_uring_sqe *sqe)
3742 struct io_sr_msg *sr = &req->sr_msg;
3743 struct io_async_ctx *io = req->io;
3746 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3749 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3750 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3751 sr->len = READ_ONCE(sqe->len);
3752 sr->bgid = READ_ONCE(sqe->buf_group);
3754 #ifdef CONFIG_COMPAT
3755 if (req->ctx->compat)
3756 sr->msg_flags |= MSG_CMSG_COMPAT;
3759 if (!io || req->opcode == IORING_OP_RECV)
3761 /* iovec is already imported */
3762 if (req->flags & REQ_F_NEED_CLEANUP)
3765 ret = io_recvmsg_copy_hdr(req, io);
3767 req->flags |= REQ_F_NEED_CLEANUP;
3771 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3773 struct io_async_msghdr *kmsg = NULL;
3774 struct socket *sock;
3775 int ret, cflags = 0;
3777 sock = sock_from_file(req->file, &ret);
3779 struct io_buffer *kbuf;
3780 struct io_async_ctx io;
3784 kmsg = &req->io->msg;
3785 kmsg->msg.msg_name = &req->io->msg.addr;
3786 /* if iov is set, it's allocated already */
3788 kmsg->iov = kmsg->fast_iov;
3789 kmsg->msg.msg_iter.iov = kmsg->iov;
3792 kmsg->msg.msg_name = &io.msg.addr;
3794 ret = io_recvmsg_copy_hdr(req, &io);
3799 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3801 return PTR_ERR(kbuf);
3803 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3804 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3805 1, req->sr_msg.len);
3808 flags = req->sr_msg.msg_flags;
3809 if (flags & MSG_DONTWAIT)
3810 req->flags |= REQ_F_NOWAIT;
3811 else if (force_nonblock)
3812 flags |= MSG_DONTWAIT;
3814 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3815 kmsg->uaddr, flags);
3816 if (force_nonblock && ret == -EAGAIN)
3817 return io_setup_async_msg(req, kmsg);
3818 if (ret == -ERESTARTSYS)
3822 if (kmsg && kmsg->iov != kmsg->fast_iov)
3824 req->flags &= ~REQ_F_NEED_CLEANUP;
3825 __io_cqring_add_event(req, ret, cflags);
3827 req_set_fail_links(req);
3832 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3834 struct io_buffer *kbuf = NULL;
3835 struct socket *sock;
3836 int ret, cflags = 0;
3838 sock = sock_from_file(req->file, &ret);
3840 struct io_sr_msg *sr = &req->sr_msg;
3841 void __user *buf = sr->buf;
3846 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3848 return PTR_ERR(kbuf);
3850 buf = u64_to_user_ptr(kbuf->addr);
3852 ret = import_single_range(READ, buf, sr->len, &iov,
3859 req->flags |= REQ_F_NEED_CLEANUP;
3860 msg.msg_name = NULL;
3861 msg.msg_control = NULL;
3862 msg.msg_controllen = 0;
3863 msg.msg_namelen = 0;
3864 msg.msg_iocb = NULL;
3867 flags = req->sr_msg.msg_flags;
3868 if (flags & MSG_DONTWAIT)
3869 req->flags |= REQ_F_NOWAIT;
3870 else if (force_nonblock)
3871 flags |= MSG_DONTWAIT;
3873 ret = sock_recvmsg(sock, &msg, flags);
3874 if (force_nonblock && ret == -EAGAIN)
3876 if (ret == -ERESTARTSYS)
3881 req->flags &= ~REQ_F_NEED_CLEANUP;
3882 __io_cqring_add_event(req, ret, cflags);
3884 req_set_fail_links(req);
3889 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3891 struct io_accept *accept = &req->accept;
3893 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3895 if (sqe->ioprio || sqe->len || sqe->buf_index)
3898 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3899 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3900 accept->flags = READ_ONCE(sqe->accept_flags);
3901 accept->nofile = rlimit(RLIMIT_NOFILE);
3905 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3907 struct io_accept *accept = &req->accept;
3908 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
3911 if (req->file->f_flags & O_NONBLOCK)
3912 req->flags |= REQ_F_NOWAIT;
3914 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3915 accept->addr_len, accept->flags,
3917 if (ret == -EAGAIN && force_nonblock)
3920 if (ret == -ERESTARTSYS)
3922 req_set_fail_links(req);
3924 io_cqring_add_event(req, ret);
3929 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3931 struct io_connect *conn = &req->connect;
3932 struct io_async_ctx *io = req->io;
3934 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3936 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3939 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3940 conn->addr_len = READ_ONCE(sqe->addr2);
3945 return move_addr_to_kernel(conn->addr, conn->addr_len,
3946 &io->connect.address);
3949 static int io_connect(struct io_kiocb *req, bool force_nonblock)
3951 struct io_async_ctx __io, *io;
3952 unsigned file_flags;
3958 ret = move_addr_to_kernel(req->connect.addr,
3959 req->connect.addr_len,
3960 &__io.connect.address);
3966 file_flags = force_nonblock ? O_NONBLOCK : 0;
3968 ret = __sys_connect_file(req->file, &io->connect.address,
3969 req->connect.addr_len, file_flags);
3970 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3973 if (io_alloc_async_ctx(req)) {
3977 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
3980 if (ret == -ERESTARTSYS)
3984 req_set_fail_links(req);
3985 io_cqring_add_event(req, ret);
3989 #else /* !CONFIG_NET */
3990 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3995 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4000 static int io_send(struct io_kiocb *req, bool force_nonblock)
4005 static int io_recvmsg_prep(struct io_kiocb *req,
4006 const struct io_uring_sqe *sqe)
4011 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4016 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4021 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4026 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4031 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4036 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4040 #endif /* CONFIG_NET */
4042 struct io_poll_table {
4043 struct poll_table_struct pt;
4044 struct io_kiocb *req;
4048 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4049 __poll_t mask, task_work_func_t func)
4051 struct task_struct *tsk;
4054 /* for instances that support it check for an event match first: */
4055 if (mask && !(mask & poll->events))
4058 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4060 list_del_init(&poll->wait.entry);
4064 init_task_work(&req->task_work, func);
4066 * If this fails, then the task is exiting. When a task exits, the
4067 * work gets canceled, so just cancel this request as well instead
4068 * of executing it. We can't safely execute it anyway, as we may not
4069 * have the needed state needed for it anyway.
4071 ret = task_work_add(tsk, &req->task_work, true);
4072 if (unlikely(ret)) {
4073 WRITE_ONCE(poll->canceled, true);
4074 tsk = io_wq_get_task(req->ctx->io_wq);
4075 task_work_add(tsk, &req->task_work, true);
4077 wake_up_process(tsk);
4081 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4082 __acquires(&req->ctx->completion_lock)
4084 struct io_ring_ctx *ctx = req->ctx;
4086 if (!req->result && !READ_ONCE(poll->canceled)) {
4087 struct poll_table_struct pt = { ._key = poll->events };
4089 req->result = vfs_poll(req->file, &pt) & poll->events;
4092 spin_lock_irq(&ctx->completion_lock);
4093 if (!req->result && !READ_ONCE(poll->canceled)) {
4094 add_wait_queue(poll->head, &poll->wait);
4101 static void io_poll_remove_double(struct io_kiocb *req)
4103 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4105 lockdep_assert_held(&req->ctx->completion_lock);
4107 if (poll && poll->head) {
4108 struct wait_queue_head *head = poll->head;
4110 spin_lock(&head->lock);
4111 list_del_init(&poll->wait.entry);
4112 if (poll->wait.private)
4113 refcount_dec(&req->refs);
4115 spin_unlock(&head->lock);
4119 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4121 struct io_ring_ctx *ctx = req->ctx;
4123 io_poll_remove_double(req);
4124 req->poll.done = true;
4125 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4126 io_commit_cqring(ctx);
4129 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4131 struct io_ring_ctx *ctx = req->ctx;
4133 if (io_poll_rewait(req, &req->poll)) {
4134 spin_unlock_irq(&ctx->completion_lock);
4138 hash_del(&req->hash_node);
4139 io_poll_complete(req, req->result, 0);
4140 req->flags |= REQ_F_COMP_LOCKED;
4141 io_put_req_find_next(req, nxt);
4142 spin_unlock_irq(&ctx->completion_lock);
4144 io_cqring_ev_posted(ctx);
4147 static void io_poll_task_func(struct callback_head *cb)
4149 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4150 struct io_kiocb *nxt = NULL;
4152 io_poll_task_handler(req, &nxt);
4154 struct io_ring_ctx *ctx = nxt->ctx;
4156 mutex_lock(&ctx->uring_lock);
4157 __io_queue_sqe(nxt, NULL);
4158 mutex_unlock(&ctx->uring_lock);
4162 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4163 int sync, void *key)
4165 struct io_kiocb *req = wait->private;
4166 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4167 __poll_t mask = key_to_poll(key);
4169 /* for instances that support it check for an event match first: */
4170 if (mask && !(mask & poll->events))
4173 if (req->poll.head) {
4176 spin_lock(&req->poll.head->lock);
4177 done = list_empty(&req->poll.wait.entry);
4179 list_del_init(&req->poll.wait.entry);
4180 spin_unlock(&req->poll.head->lock);
4182 __io_async_wake(req, poll, mask, io_poll_task_func);
4184 refcount_dec(&req->refs);
4188 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4189 wait_queue_func_t wake_func)
4193 poll->canceled = false;
4194 poll->events = events;
4195 INIT_LIST_HEAD(&poll->wait.entry);
4196 init_waitqueue_func_entry(&poll->wait, wake_func);
4199 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4200 struct wait_queue_head *head)
4202 struct io_kiocb *req = pt->req;
4205 * If poll->head is already set, it's because the file being polled
4206 * uses multiple waitqueues for poll handling (eg one for read, one
4207 * for write). Setup a separate io_poll_iocb if this happens.
4209 if (unlikely(poll->head)) {
4210 /* already have a 2nd entry, fail a third attempt */
4212 pt->error = -EINVAL;
4215 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4217 pt->error = -ENOMEM;
4220 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4221 refcount_inc(&req->refs);
4222 poll->wait.private = req;
4223 req->io = (void *) poll;
4228 add_wait_queue(head, &poll->wait);
4231 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4232 struct poll_table_struct *p)
4234 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4236 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4239 static void io_async_task_func(struct callback_head *cb)
4241 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4242 struct async_poll *apoll = req->apoll;
4243 struct io_ring_ctx *ctx = req->ctx;
4244 bool canceled = false;
4246 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4248 if (io_poll_rewait(req, &apoll->poll)) {
4249 spin_unlock_irq(&ctx->completion_lock);
4253 /* If req is still hashed, it cannot have been canceled. Don't check. */
4254 if (hash_hashed(&req->hash_node)) {
4255 hash_del(&req->hash_node);
4257 canceled = READ_ONCE(apoll->poll.canceled);
4259 io_cqring_fill_event(req, -ECANCELED);
4260 io_commit_cqring(ctx);
4264 spin_unlock_irq(&ctx->completion_lock);
4266 /* restore ->work in case we need to retry again */
4267 if (req->flags & REQ_F_WORK_INITIALIZED)
4268 memcpy(&req->work, &apoll->work, sizeof(req->work));
4272 __set_current_state(TASK_RUNNING);
4273 mutex_lock(&ctx->uring_lock);
4274 __io_queue_sqe(req, NULL);
4275 mutex_unlock(&ctx->uring_lock);
4277 io_cqring_ev_posted(ctx);
4278 req_set_fail_links(req);
4279 io_double_put_req(req);
4283 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4286 struct io_kiocb *req = wait->private;
4287 struct io_poll_iocb *poll = &req->apoll->poll;
4289 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4292 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4295 static void io_poll_req_insert(struct io_kiocb *req)
4297 struct io_ring_ctx *ctx = req->ctx;
4298 struct hlist_head *list;
4300 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4301 hlist_add_head(&req->hash_node, list);
4304 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4305 struct io_poll_iocb *poll,
4306 struct io_poll_table *ipt, __poll_t mask,
4307 wait_queue_func_t wake_func)
4308 __acquires(&ctx->completion_lock)
4310 struct io_ring_ctx *ctx = req->ctx;
4311 bool cancel = false;
4313 poll->file = req->file;
4314 io_init_poll_iocb(poll, mask, wake_func);
4315 poll->wait.private = req;
4317 ipt->pt._key = mask;
4319 ipt->error = -EINVAL;
4321 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4323 spin_lock_irq(&ctx->completion_lock);
4324 if (likely(poll->head)) {
4325 spin_lock(&poll->head->lock);
4326 if (unlikely(list_empty(&poll->wait.entry))) {
4332 if (mask || ipt->error)
4333 list_del_init(&poll->wait.entry);
4335 WRITE_ONCE(poll->canceled, true);
4336 else if (!poll->done) /* actually waiting for an event */
4337 io_poll_req_insert(req);
4338 spin_unlock(&poll->head->lock);
4344 static bool io_arm_poll_handler(struct io_kiocb *req)
4346 const struct io_op_def *def = &io_op_defs[req->opcode];
4347 struct io_ring_ctx *ctx = req->ctx;
4348 struct async_poll *apoll;
4349 struct io_poll_table ipt;
4353 if (!req->file || !file_can_poll(req->file))
4355 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4357 if (!def->pollin && !def->pollout)
4360 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4361 if (unlikely(!apoll))
4364 req->flags |= REQ_F_POLLED;
4365 if (req->flags & REQ_F_WORK_INITIALIZED)
4366 memcpy(&apoll->work, &req->work, sizeof(req->work));
4367 had_io = req->io != NULL;
4369 get_task_struct(current);
4370 req->task = current;
4372 INIT_HLIST_NODE(&req->hash_node);
4376 mask |= POLLIN | POLLRDNORM;
4378 mask |= POLLOUT | POLLWRNORM;
4379 mask |= POLLERR | POLLPRI;
4381 ipt.pt._qproc = io_async_queue_proc;
4383 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4387 /* only remove double add if we did it here */
4389 io_poll_remove_double(req);
4390 spin_unlock_irq(&ctx->completion_lock);
4391 if (req->flags & REQ_F_WORK_INITIALIZED)
4392 memcpy(&req->work, &apoll->work, sizeof(req->work));
4396 spin_unlock_irq(&ctx->completion_lock);
4397 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4398 apoll->poll.events);
4402 static bool __io_poll_remove_one(struct io_kiocb *req,
4403 struct io_poll_iocb *poll)
4405 bool do_complete = false;
4407 spin_lock(&poll->head->lock);
4408 WRITE_ONCE(poll->canceled, true);
4409 if (!list_empty(&poll->wait.entry)) {
4410 list_del_init(&poll->wait.entry);
4413 spin_unlock(&poll->head->lock);
4414 hash_del(&req->hash_node);
4418 static bool io_poll_remove_one(struct io_kiocb *req)
4422 if (req->opcode == IORING_OP_POLL_ADD) {
4423 io_poll_remove_double(req);
4424 do_complete = __io_poll_remove_one(req, &req->poll);
4426 struct async_poll *apoll = req->apoll;
4428 /* non-poll requests have submit ref still */
4429 do_complete = __io_poll_remove_one(req, &apoll->poll);
4433 * restore ->work because we will call
4434 * io_req_work_drop_env below when dropping the
4437 if (req->flags & REQ_F_WORK_INITIALIZED)
4438 memcpy(&req->work, &apoll->work,
4445 io_cqring_fill_event(req, -ECANCELED);
4446 io_commit_cqring(req->ctx);
4447 req->flags |= REQ_F_COMP_LOCKED;
4454 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4456 struct hlist_node *tmp;
4457 struct io_kiocb *req;
4460 spin_lock_irq(&ctx->completion_lock);
4461 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4462 struct hlist_head *list;
4464 list = &ctx->cancel_hash[i];
4465 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4466 posted += io_poll_remove_one(req);
4468 spin_unlock_irq(&ctx->completion_lock);
4471 io_cqring_ev_posted(ctx);
4474 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4476 struct hlist_head *list;
4477 struct io_kiocb *req;
4479 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4480 hlist_for_each_entry(req, list, hash_node) {
4481 if (sqe_addr != req->user_data)
4483 if (io_poll_remove_one(req))
4491 static int io_poll_remove_prep(struct io_kiocb *req,
4492 const struct io_uring_sqe *sqe)
4494 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4496 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4500 req->poll.addr = READ_ONCE(sqe->addr);
4505 * Find a running poll command that matches one specified in sqe->addr,
4506 * and remove it if found.
4508 static int io_poll_remove(struct io_kiocb *req)
4510 struct io_ring_ctx *ctx = req->ctx;
4514 addr = req->poll.addr;
4515 spin_lock_irq(&ctx->completion_lock);
4516 ret = io_poll_cancel(ctx, addr);
4517 spin_unlock_irq(&ctx->completion_lock);
4519 io_cqring_add_event(req, ret);
4521 req_set_fail_links(req);
4526 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4529 struct io_kiocb *req = wait->private;
4530 struct io_poll_iocb *poll = &req->poll;
4532 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4535 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4536 struct poll_table_struct *p)
4538 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4540 __io_queue_proc(&pt->req->poll, pt, head);
4543 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4545 struct io_poll_iocb *poll = &req->poll;
4548 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4550 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4555 events = READ_ONCE(sqe->poll_events);
4556 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4558 get_task_struct(current);
4559 req->task = current;
4563 static int io_poll_add(struct io_kiocb *req)
4565 struct io_poll_iocb *poll = &req->poll;
4566 struct io_ring_ctx *ctx = req->ctx;
4567 struct io_poll_table ipt;
4570 INIT_HLIST_NODE(&req->hash_node);
4571 INIT_LIST_HEAD(&req->list);
4572 ipt.pt._qproc = io_poll_queue_proc;
4574 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4577 if (mask) { /* no async, we'd stolen it */
4579 io_poll_complete(req, mask, 0);
4581 spin_unlock_irq(&ctx->completion_lock);
4584 io_cqring_ev_posted(ctx);
4590 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4592 struct io_timeout_data *data = container_of(timer,
4593 struct io_timeout_data, timer);
4594 struct io_kiocb *req = data->req;
4595 struct io_ring_ctx *ctx = req->ctx;
4596 unsigned long flags;
4598 atomic_inc(&ctx->cq_timeouts);
4600 spin_lock_irqsave(&ctx->completion_lock, flags);
4602 * We could be racing with timeout deletion. If the list is empty,
4603 * then timeout lookup already found it and will be handling it.
4605 if (!list_empty(&req->list))
4606 list_del_init(&req->list);
4608 io_cqring_fill_event(req, -ETIME);
4609 io_commit_cqring(ctx);
4610 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4612 io_cqring_ev_posted(ctx);
4613 req_set_fail_links(req);
4615 return HRTIMER_NORESTART;
4618 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4620 struct io_kiocb *req;
4623 list_for_each_entry(req, &ctx->timeout_list, list) {
4624 if (user_data == req->user_data) {
4625 list_del_init(&req->list);
4634 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4638 req_set_fail_links(req);
4639 io_cqring_fill_event(req, -ECANCELED);
4644 static int io_timeout_remove_prep(struct io_kiocb *req,
4645 const struct io_uring_sqe *sqe)
4647 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4649 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4652 req->timeout.addr = READ_ONCE(sqe->addr);
4653 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4654 if (req->timeout.flags)
4661 * Remove or update an existing timeout command
4663 static int io_timeout_remove(struct io_kiocb *req)
4665 struct io_ring_ctx *ctx = req->ctx;
4668 spin_lock_irq(&ctx->completion_lock);
4669 ret = io_timeout_cancel(ctx, req->timeout.addr);
4671 io_cqring_fill_event(req, ret);
4672 io_commit_cqring(ctx);
4673 spin_unlock_irq(&ctx->completion_lock);
4674 io_cqring_ev_posted(ctx);
4676 req_set_fail_links(req);
4681 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4682 bool is_timeout_link)
4684 struct io_timeout_data *data;
4686 u32 off = READ_ONCE(sqe->off);
4688 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4690 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4692 if (off && is_timeout_link)
4694 flags = READ_ONCE(sqe->timeout_flags);
4695 if (flags & ~IORING_TIMEOUT_ABS)
4698 req->timeout.off = off;
4700 if (!req->io && io_alloc_async_ctx(req))
4703 data = &req->io->timeout;
4705 req->flags |= REQ_F_TIMEOUT;
4707 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4710 if (flags & IORING_TIMEOUT_ABS)
4711 data->mode = HRTIMER_MODE_ABS;
4713 data->mode = HRTIMER_MODE_REL;
4715 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4719 static int io_timeout(struct io_kiocb *req)
4721 struct io_ring_ctx *ctx = req->ctx;
4722 struct io_timeout_data *data = &req->io->timeout;
4723 struct list_head *entry;
4724 u32 tail, off = req->timeout.off;
4726 spin_lock_irq(&ctx->completion_lock);
4729 * sqe->off holds how many events that need to occur for this
4730 * timeout event to be satisfied. If it isn't set, then this is
4731 * a pure timeout request, sequence isn't used.
4734 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4735 entry = ctx->timeout_list.prev;
4739 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
4740 req->timeout.target_seq = tail + off;
4743 * Insertion sort, ensuring the first entry in the list is always
4744 * the one we need first.
4746 list_for_each_prev(entry, &ctx->timeout_list) {
4747 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4749 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4751 /* nxt.seq is behind @tail, otherwise would've been completed */
4752 if (off >= nxt->timeout.target_seq - tail)
4756 list_add(&req->list, entry);
4757 data->timer.function = io_timeout_fn;
4758 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4759 spin_unlock_irq(&ctx->completion_lock);
4763 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4765 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4767 return req->user_data == (unsigned long) data;
4770 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4772 enum io_wq_cancel cancel_ret;
4775 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4776 switch (cancel_ret) {
4777 case IO_WQ_CANCEL_OK:
4780 case IO_WQ_CANCEL_RUNNING:
4783 case IO_WQ_CANCEL_NOTFOUND:
4791 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4792 struct io_kiocb *req, __u64 sqe_addr,
4795 unsigned long flags;
4798 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4799 if (ret != -ENOENT) {
4800 spin_lock_irqsave(&ctx->completion_lock, flags);
4804 spin_lock_irqsave(&ctx->completion_lock, flags);
4805 ret = io_timeout_cancel(ctx, sqe_addr);
4808 ret = io_poll_cancel(ctx, sqe_addr);
4812 io_cqring_fill_event(req, ret);
4813 io_commit_cqring(ctx);
4814 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4815 io_cqring_ev_posted(ctx);
4818 req_set_fail_links(req);
4822 static int io_async_cancel_prep(struct io_kiocb *req,
4823 const struct io_uring_sqe *sqe)
4825 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4827 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4831 req->cancel.addr = READ_ONCE(sqe->addr);
4835 static int io_async_cancel(struct io_kiocb *req)
4837 struct io_ring_ctx *ctx = req->ctx;
4839 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4843 static int io_files_update_prep(struct io_kiocb *req,
4844 const struct io_uring_sqe *sqe)
4846 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4849 req->files_update.offset = READ_ONCE(sqe->off);
4850 req->files_update.nr_args = READ_ONCE(sqe->len);
4851 if (!req->files_update.nr_args)
4853 req->files_update.arg = READ_ONCE(sqe->addr);
4857 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4859 struct io_ring_ctx *ctx = req->ctx;
4860 struct io_uring_files_update up;
4866 up.offset = req->files_update.offset;
4867 up.fds = req->files_update.arg;
4869 mutex_lock(&ctx->uring_lock);
4870 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4871 mutex_unlock(&ctx->uring_lock);
4874 req_set_fail_links(req);
4875 io_cqring_add_event(req, ret);
4880 static int io_req_defer_prep(struct io_kiocb *req,
4881 const struct io_uring_sqe *sqe)
4888 io_req_init_async(req);
4890 if (io_op_defs[req->opcode].file_table) {
4891 ret = io_grab_files(req);
4896 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4898 switch (req->opcode) {
4901 case IORING_OP_READV:
4902 case IORING_OP_READ_FIXED:
4903 case IORING_OP_READ:
4904 ret = io_read_prep(req, sqe, true);
4906 case IORING_OP_WRITEV:
4907 case IORING_OP_WRITE_FIXED:
4908 case IORING_OP_WRITE:
4909 ret = io_write_prep(req, sqe, true);
4911 case IORING_OP_POLL_ADD:
4912 ret = io_poll_add_prep(req, sqe);
4914 case IORING_OP_POLL_REMOVE:
4915 ret = io_poll_remove_prep(req, sqe);
4917 case IORING_OP_FSYNC:
4918 ret = io_prep_fsync(req, sqe);
4920 case IORING_OP_SYNC_FILE_RANGE:
4921 ret = io_prep_sfr(req, sqe);
4923 case IORING_OP_SENDMSG:
4924 case IORING_OP_SEND:
4925 ret = io_sendmsg_prep(req, sqe);
4927 case IORING_OP_RECVMSG:
4928 case IORING_OP_RECV:
4929 ret = io_recvmsg_prep(req, sqe);
4931 case IORING_OP_CONNECT:
4932 ret = io_connect_prep(req, sqe);
4934 case IORING_OP_TIMEOUT:
4935 ret = io_timeout_prep(req, sqe, false);
4937 case IORING_OP_TIMEOUT_REMOVE:
4938 ret = io_timeout_remove_prep(req, sqe);
4940 case IORING_OP_ASYNC_CANCEL:
4941 ret = io_async_cancel_prep(req, sqe);
4943 case IORING_OP_LINK_TIMEOUT:
4944 ret = io_timeout_prep(req, sqe, true);
4946 case IORING_OP_ACCEPT:
4947 ret = io_accept_prep(req, sqe);
4949 case IORING_OP_FALLOCATE:
4950 ret = io_fallocate_prep(req, sqe);
4952 case IORING_OP_OPENAT:
4953 ret = io_openat_prep(req, sqe);
4955 case IORING_OP_CLOSE:
4956 ret = io_close_prep(req, sqe);
4958 case IORING_OP_FILES_UPDATE:
4959 ret = io_files_update_prep(req, sqe);
4961 case IORING_OP_STATX:
4962 ret = io_statx_prep(req, sqe);
4964 case IORING_OP_FADVISE:
4965 ret = io_fadvise_prep(req, sqe);
4967 case IORING_OP_MADVISE:
4968 ret = io_madvise_prep(req, sqe);
4970 case IORING_OP_OPENAT2:
4971 ret = io_openat2_prep(req, sqe);
4973 case IORING_OP_EPOLL_CTL:
4974 ret = io_epoll_ctl_prep(req, sqe);
4976 case IORING_OP_SPLICE:
4977 ret = io_splice_prep(req, sqe);
4979 case IORING_OP_PROVIDE_BUFFERS:
4980 ret = io_provide_buffers_prep(req, sqe);
4982 case IORING_OP_REMOVE_BUFFERS:
4983 ret = io_remove_buffers_prep(req, sqe);
4986 ret = io_tee_prep(req, sqe);
4989 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4998 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5000 struct io_ring_ctx *ctx = req->ctx;
5003 /* Still need defer if there is pending req in defer list. */
5004 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5008 if (io_alloc_async_ctx(req))
5010 ret = io_req_defer_prep(req, sqe);
5015 spin_lock_irq(&ctx->completion_lock);
5016 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5017 spin_unlock_irq(&ctx->completion_lock);
5021 trace_io_uring_defer(ctx, req, req->user_data);
5022 list_add_tail(&req->list, &ctx->defer_list);
5023 spin_unlock_irq(&ctx->completion_lock);
5024 return -EIOCBQUEUED;
5027 static void io_cleanup_req(struct io_kiocb *req)
5029 struct io_async_ctx *io = req->io;
5031 switch (req->opcode) {
5032 case IORING_OP_READV:
5033 case IORING_OP_READ_FIXED:
5034 case IORING_OP_READ:
5035 if (req->flags & REQ_F_BUFFER_SELECTED)
5036 kfree((void *)(unsigned long)req->rw.addr);
5038 case IORING_OP_WRITEV:
5039 case IORING_OP_WRITE_FIXED:
5040 case IORING_OP_WRITE:
5041 if (io->rw.iov != io->rw.fast_iov)
5044 case IORING_OP_RECVMSG:
5045 if (req->flags & REQ_F_BUFFER_SELECTED)
5046 kfree(req->sr_msg.kbuf);
5048 case IORING_OP_SENDMSG:
5049 if (io->msg.iov != io->msg.fast_iov)
5052 case IORING_OP_RECV:
5053 if (req->flags & REQ_F_BUFFER_SELECTED)
5054 kfree(req->sr_msg.kbuf);
5056 case IORING_OP_OPENAT:
5057 case IORING_OP_OPENAT2:
5059 case IORING_OP_SPLICE:
5061 io_put_file(req, req->splice.file_in,
5062 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5066 req->flags &= ~REQ_F_NEED_CLEANUP;
5069 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5070 bool force_nonblock)
5072 struct io_ring_ctx *ctx = req->ctx;
5075 switch (req->opcode) {
5079 case IORING_OP_READV:
5080 case IORING_OP_READ_FIXED:
5081 case IORING_OP_READ:
5083 ret = io_read_prep(req, sqe, force_nonblock);
5087 ret = io_read(req, force_nonblock);
5089 case IORING_OP_WRITEV:
5090 case IORING_OP_WRITE_FIXED:
5091 case IORING_OP_WRITE:
5093 ret = io_write_prep(req, sqe, force_nonblock);
5097 ret = io_write(req, force_nonblock);
5099 case IORING_OP_FSYNC:
5101 ret = io_prep_fsync(req, sqe);
5105 ret = io_fsync(req, force_nonblock);
5107 case IORING_OP_POLL_ADD:
5109 ret = io_poll_add_prep(req, sqe);
5113 ret = io_poll_add(req);
5115 case IORING_OP_POLL_REMOVE:
5117 ret = io_poll_remove_prep(req, sqe);
5121 ret = io_poll_remove(req);
5123 case IORING_OP_SYNC_FILE_RANGE:
5125 ret = io_prep_sfr(req, sqe);
5129 ret = io_sync_file_range(req, force_nonblock);
5131 case IORING_OP_SENDMSG:
5132 case IORING_OP_SEND:
5134 ret = io_sendmsg_prep(req, sqe);
5138 if (req->opcode == IORING_OP_SENDMSG)
5139 ret = io_sendmsg(req, force_nonblock);
5141 ret = io_send(req, force_nonblock);
5143 case IORING_OP_RECVMSG:
5144 case IORING_OP_RECV:
5146 ret = io_recvmsg_prep(req, sqe);
5150 if (req->opcode == IORING_OP_RECVMSG)
5151 ret = io_recvmsg(req, force_nonblock);
5153 ret = io_recv(req, force_nonblock);
5155 case IORING_OP_TIMEOUT:
5157 ret = io_timeout_prep(req, sqe, false);
5161 ret = io_timeout(req);
5163 case IORING_OP_TIMEOUT_REMOVE:
5165 ret = io_timeout_remove_prep(req, sqe);
5169 ret = io_timeout_remove(req);
5171 case IORING_OP_ACCEPT:
5173 ret = io_accept_prep(req, sqe);
5177 ret = io_accept(req, force_nonblock);
5179 case IORING_OP_CONNECT:
5181 ret = io_connect_prep(req, sqe);
5185 ret = io_connect(req, force_nonblock);
5187 case IORING_OP_ASYNC_CANCEL:
5189 ret = io_async_cancel_prep(req, sqe);
5193 ret = io_async_cancel(req);
5195 case IORING_OP_FALLOCATE:
5197 ret = io_fallocate_prep(req, sqe);
5201 ret = io_fallocate(req, force_nonblock);
5203 case IORING_OP_OPENAT:
5205 ret = io_openat_prep(req, sqe);
5209 ret = io_openat(req, force_nonblock);
5211 case IORING_OP_CLOSE:
5213 ret = io_close_prep(req, sqe);
5217 ret = io_close(req, force_nonblock);
5219 case IORING_OP_FILES_UPDATE:
5221 ret = io_files_update_prep(req, sqe);
5225 ret = io_files_update(req, force_nonblock);
5227 case IORING_OP_STATX:
5229 ret = io_statx_prep(req, sqe);
5233 ret = io_statx(req, force_nonblock);
5235 case IORING_OP_FADVISE:
5237 ret = io_fadvise_prep(req, sqe);
5241 ret = io_fadvise(req, force_nonblock);
5243 case IORING_OP_MADVISE:
5245 ret = io_madvise_prep(req, sqe);
5249 ret = io_madvise(req, force_nonblock);
5251 case IORING_OP_OPENAT2:
5253 ret = io_openat2_prep(req, sqe);
5257 ret = io_openat2(req, force_nonblock);
5259 case IORING_OP_EPOLL_CTL:
5261 ret = io_epoll_ctl_prep(req, sqe);
5265 ret = io_epoll_ctl(req, force_nonblock);
5267 case IORING_OP_SPLICE:
5269 ret = io_splice_prep(req, sqe);
5273 ret = io_splice(req, force_nonblock);
5275 case IORING_OP_PROVIDE_BUFFERS:
5277 ret = io_provide_buffers_prep(req, sqe);
5281 ret = io_provide_buffers(req, force_nonblock);
5283 case IORING_OP_REMOVE_BUFFERS:
5285 ret = io_remove_buffers_prep(req, sqe);
5289 ret = io_remove_buffers(req, force_nonblock);
5293 ret = io_tee_prep(req, sqe);
5297 ret = io_tee(req, force_nonblock);
5307 /* If the op doesn't have a file, we're not polling for it */
5308 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5309 const bool in_async = io_wq_current_is_worker();
5311 if (req->result == -EAGAIN)
5314 /* workqueue context doesn't hold uring_lock, grab it now */
5316 mutex_lock(&ctx->uring_lock);
5318 io_iopoll_req_issued(req);
5321 mutex_unlock(&ctx->uring_lock);
5327 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5329 struct io_kiocb *link;
5331 /* link head's timeout is queued in io_queue_async_work() */
5332 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5335 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5336 io_queue_linked_timeout(link);
5339 static void io_wq_submit_work(struct io_wq_work **workptr)
5341 struct io_wq_work *work = *workptr;
5342 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5345 io_arm_async_linked_timeout(req);
5347 /* if NO_CANCEL is set, we must still run the work */
5348 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5349 IO_WQ_WORK_CANCEL) {
5355 ret = io_issue_sqe(req, NULL, false);
5357 * We can get EAGAIN for polled IO even though we're
5358 * forcing a sync submission from here, since we can't
5359 * wait for request slots on the block side.
5368 req_set_fail_links(req);
5369 io_cqring_add_event(req, ret);
5373 io_steal_work(req, workptr);
5376 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5379 struct fixed_file_table *table;
5381 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5382 return table->files[index & IORING_FILE_TABLE_MASK];
5385 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5386 int fd, struct file **out_file, bool fixed)
5388 struct io_ring_ctx *ctx = req->ctx;
5392 if (unlikely(!ctx->file_data ||
5393 (unsigned) fd >= ctx->nr_user_files))
5395 fd = array_index_nospec(fd, ctx->nr_user_files);
5396 file = io_file_from_index(ctx, fd);
5398 req->fixed_file_refs = ctx->file_data->cur_refs;
5399 percpu_ref_get(req->fixed_file_refs);
5402 trace_io_uring_file_get(ctx, fd);
5403 file = __io_file_get(state, fd);
5406 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5413 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5418 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5419 if (unlikely(!fixed && io_async_submit(req->ctx)))
5422 return io_file_get(state, req, fd, &req->file, fixed);
5425 static int io_grab_files(struct io_kiocb *req)
5428 struct io_ring_ctx *ctx = req->ctx;
5430 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5432 if (!ctx->ring_file)
5436 spin_lock_irq(&ctx->inflight_lock);
5438 * We use the f_ops->flush() handler to ensure that we can flush
5439 * out work accessing these files if the fd is closed. Check if
5440 * the fd has changed since we started down this path, and disallow
5441 * this operation if it has.
5443 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5444 list_add(&req->inflight_entry, &ctx->inflight_list);
5445 req->flags |= REQ_F_INFLIGHT;
5446 req->work.files = current->files;
5449 spin_unlock_irq(&ctx->inflight_lock);
5455 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5457 struct io_timeout_data *data = container_of(timer,
5458 struct io_timeout_data, timer);
5459 struct io_kiocb *req = data->req;
5460 struct io_ring_ctx *ctx = req->ctx;
5461 struct io_kiocb *prev = NULL;
5462 unsigned long flags;
5464 spin_lock_irqsave(&ctx->completion_lock, flags);
5467 * We don't expect the list to be empty, that will only happen if we
5468 * race with the completion of the linked work.
5470 if (!list_empty(&req->link_list)) {
5471 prev = list_entry(req->link_list.prev, struct io_kiocb,
5473 if (refcount_inc_not_zero(&prev->refs)) {
5474 list_del_init(&req->link_list);
5475 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5480 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5483 req_set_fail_links(prev);
5484 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5487 io_cqring_add_event(req, -ETIME);
5490 return HRTIMER_NORESTART;
5493 static void io_queue_linked_timeout(struct io_kiocb *req)
5495 struct io_ring_ctx *ctx = req->ctx;
5498 * If the list is now empty, then our linked request finished before
5499 * we got a chance to setup the timer
5501 spin_lock_irq(&ctx->completion_lock);
5502 if (!list_empty(&req->link_list)) {
5503 struct io_timeout_data *data = &req->io->timeout;
5505 data->timer.function = io_link_timeout_fn;
5506 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5509 spin_unlock_irq(&ctx->completion_lock);
5511 /* drop submission reference */
5515 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5517 struct io_kiocb *nxt;
5519 if (!(req->flags & REQ_F_LINK_HEAD))
5521 /* for polled retry, if flag is set, we already went through here */
5522 if (req->flags & REQ_F_POLLED)
5525 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5527 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5530 req->flags |= REQ_F_LINK_TIMEOUT;
5534 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5536 struct io_kiocb *linked_timeout;
5537 struct io_kiocb *nxt;
5538 const struct cred *old_creds = NULL;
5542 linked_timeout = io_prep_linked_timeout(req);
5544 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5545 req->work.creds != current_cred()) {
5547 revert_creds(old_creds);
5548 if (old_creds == req->work.creds)
5549 old_creds = NULL; /* restored original creds */
5551 old_creds = override_creds(req->work.creds);
5554 ret = io_issue_sqe(req, sqe, true);
5557 * We async punt it if the file wasn't marked NOWAIT, or if the file
5558 * doesn't support non-blocking read/write attempts
5560 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5561 (req->flags & REQ_F_MUST_PUNT))) {
5562 if (io_arm_poll_handler(req)) {
5564 io_queue_linked_timeout(linked_timeout);
5568 io_req_init_async(req);
5570 if (io_op_defs[req->opcode].file_table) {
5571 ret = io_grab_files(req);
5577 * Queued up for async execution, worker will release
5578 * submit reference when the iocb is actually submitted.
5580 io_queue_async_work(req);
5586 /* drop submission reference */
5587 io_put_req_find_next(req, &nxt);
5589 if (linked_timeout) {
5591 io_queue_linked_timeout(linked_timeout);
5593 io_put_req(linked_timeout);
5596 /* and drop final reference, if we failed */
5598 io_cqring_add_event(req, ret);
5599 req_set_fail_links(req);
5605 if (req->flags & REQ_F_FORCE_ASYNC)
5611 revert_creds(old_creds);
5614 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5618 ret = io_req_defer(req, sqe);
5620 if (ret != -EIOCBQUEUED) {
5622 io_cqring_add_event(req, ret);
5623 req_set_fail_links(req);
5624 io_double_put_req(req);
5626 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5629 if (io_alloc_async_ctx(req))
5631 ret = io_req_defer_prep(req, sqe);
5632 if (unlikely(ret < 0))
5637 * Never try inline submit of IOSQE_ASYNC is set, go straight
5638 * to async execution.
5640 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5641 io_queue_async_work(req);
5643 __io_queue_sqe(req, sqe);
5647 static inline void io_queue_link_head(struct io_kiocb *req)
5649 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5650 io_cqring_add_event(req, -ECANCELED);
5651 io_double_put_req(req);
5653 io_queue_sqe(req, NULL);
5656 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5657 struct io_kiocb **link)
5659 struct io_ring_ctx *ctx = req->ctx;
5663 * If we already have a head request, queue this one for async
5664 * submittal once the head completes. If we don't have a head but
5665 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5666 * submitted sync once the chain is complete. If none of those
5667 * conditions are true (normal request), then just queue it.
5670 struct io_kiocb *head = *link;
5673 * Taking sequential execution of a link, draining both sides
5674 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5675 * requests in the link. So, it drains the head and the
5676 * next after the link request. The last one is done via
5677 * drain_next flag to persist the effect across calls.
5679 if (req->flags & REQ_F_IO_DRAIN) {
5680 head->flags |= REQ_F_IO_DRAIN;
5681 ctx->drain_next = 1;
5683 if (io_alloc_async_ctx(req))
5686 ret = io_req_defer_prep(req, sqe);
5688 /* fail even hard links since we don't submit */
5689 head->flags |= REQ_F_FAIL_LINK;
5692 trace_io_uring_link(ctx, req, head);
5693 list_add_tail(&req->link_list, &head->link_list);
5695 /* last request of a link, enqueue the link */
5696 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5697 io_queue_link_head(head);
5701 if (unlikely(ctx->drain_next)) {
5702 req->flags |= REQ_F_IO_DRAIN;
5703 ctx->drain_next = 0;
5705 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5706 req->flags |= REQ_F_LINK_HEAD;
5707 INIT_LIST_HEAD(&req->link_list);
5709 if (io_alloc_async_ctx(req))
5712 ret = io_req_defer_prep(req, sqe);
5714 req->flags |= REQ_F_FAIL_LINK;
5717 io_queue_sqe(req, sqe);
5725 * Batched submission is done, ensure local IO is flushed out.
5727 static void io_submit_state_end(struct io_submit_state *state)
5729 blk_finish_plug(&state->plug);
5730 io_state_file_put(state);
5731 if (state->free_reqs)
5732 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5736 * Start submission side cache.
5738 static void io_submit_state_start(struct io_submit_state *state,
5739 unsigned int max_ios)
5741 blk_start_plug(&state->plug);
5742 state->free_reqs = 0;
5744 state->ios_left = max_ios;
5747 static void io_commit_sqring(struct io_ring_ctx *ctx)
5749 struct io_rings *rings = ctx->rings;
5752 * Ensure any loads from the SQEs are done at this point,
5753 * since once we write the new head, the application could
5754 * write new data to them.
5756 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5760 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5761 * that is mapped by userspace. This means that care needs to be taken to
5762 * ensure that reads are stable, as we cannot rely on userspace always
5763 * being a good citizen. If members of the sqe are validated and then later
5764 * used, it's important that those reads are done through READ_ONCE() to
5765 * prevent a re-load down the line.
5767 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5769 u32 *sq_array = ctx->sq_array;
5773 * The cached sq head (or cq tail) serves two purposes:
5775 * 1) allows us to batch the cost of updating the user visible
5777 * 2) allows the kernel side to track the head on its own, even
5778 * though the application is the one updating it.
5780 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5781 if (likely(head < ctx->sq_entries))
5782 return &ctx->sq_sqes[head];
5784 /* drop invalid entries */
5785 ctx->cached_sq_dropped++;
5786 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5790 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5792 ctx->cached_sq_head++;
5795 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5796 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5797 IOSQE_BUFFER_SELECT)
5799 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5800 const struct io_uring_sqe *sqe,
5801 struct io_submit_state *state)
5803 unsigned int sqe_flags;
5807 * All io need record the previous position, if LINK vs DARIN,
5808 * it can be used to mark the position of the first IO in the
5811 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5812 req->opcode = READ_ONCE(sqe->opcode);
5813 req->user_data = READ_ONCE(sqe->user_data);
5818 /* one is dropped after submission, the other at completion */
5819 refcount_set(&req->refs, 2);
5823 if (unlikely(req->opcode >= IORING_OP_LAST))
5826 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5827 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5829 kthread_use_mm(ctx->sqo_mm);
5832 sqe_flags = READ_ONCE(sqe->flags);
5833 /* enforce forwards compatibility on users */
5834 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5837 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5838 !io_op_defs[req->opcode].buffer_select)
5841 id = READ_ONCE(sqe->personality);
5843 io_req_init_async(req);
5844 req->work.creds = idr_find(&ctx->personality_idr, id);
5845 if (unlikely(!req->work.creds))
5847 get_cred(req->work.creds);
5850 /* same numerical values with corresponding REQ_F_*, safe to copy */
5851 req->flags |= sqe_flags;
5853 if (!io_op_defs[req->opcode].needs_file)
5856 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
5859 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5860 struct file *ring_file, int ring_fd)
5862 struct io_submit_state state, *statep = NULL;
5863 struct io_kiocb *link = NULL;
5864 int i, submitted = 0;
5866 /* if we have a backlog and couldn't flush it all, return BUSY */
5867 if (test_bit(0, &ctx->sq_check_overflow)) {
5868 if (!list_empty(&ctx->cq_overflow_list) &&
5869 !io_cqring_overflow_flush(ctx, false))
5873 /* make sure SQ entry isn't read before tail */
5874 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5876 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5879 if (nr > IO_PLUG_THRESHOLD) {
5880 io_submit_state_start(&state, nr);
5884 ctx->ring_fd = ring_fd;
5885 ctx->ring_file = ring_file;
5887 for (i = 0; i < nr; i++) {
5888 const struct io_uring_sqe *sqe;
5889 struct io_kiocb *req;
5892 sqe = io_get_sqe(ctx);
5893 if (unlikely(!sqe)) {
5894 io_consume_sqe(ctx);
5897 req = io_alloc_req(ctx, statep);
5898 if (unlikely(!req)) {
5900 submitted = -EAGAIN;
5904 err = io_init_req(ctx, req, sqe, statep);
5905 io_consume_sqe(ctx);
5906 /* will complete beyond this point, count as submitted */
5909 if (unlikely(err)) {
5911 io_cqring_add_event(req, err);
5912 io_double_put_req(req);
5916 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5917 true, io_async_submit(ctx));
5918 err = io_submit_sqe(req, sqe, &link);
5923 if (unlikely(submitted != nr)) {
5924 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5926 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5929 io_queue_link_head(link);
5931 io_submit_state_end(&state);
5933 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5934 io_commit_sqring(ctx);
5939 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5941 struct mm_struct *mm = current->mm;
5944 kthread_unuse_mm(mm);
5949 static int io_sq_thread(void *data)
5951 struct io_ring_ctx *ctx = data;
5952 const struct cred *old_cred;
5954 unsigned long timeout;
5957 complete(&ctx->sq_thread_comp);
5959 old_cred = override_creds(ctx->creds);
5961 timeout = jiffies + ctx->sq_thread_idle;
5962 while (!kthread_should_park()) {
5963 unsigned int to_submit;
5965 if (!list_empty(&ctx->poll_list)) {
5966 unsigned nr_events = 0;
5968 mutex_lock(&ctx->uring_lock);
5969 if (!list_empty(&ctx->poll_list))
5970 io_iopoll_getevents(ctx, &nr_events, 0);
5972 timeout = jiffies + ctx->sq_thread_idle;
5973 mutex_unlock(&ctx->uring_lock);
5976 to_submit = io_sqring_entries(ctx);
5979 * If submit got -EBUSY, flag us as needing the application
5980 * to enter the kernel to reap and flush events.
5982 if (!to_submit || ret == -EBUSY) {
5984 * Drop cur_mm before scheduling, we can't hold it for
5985 * long periods (or over schedule()). Do this before
5986 * adding ourselves to the waitqueue, as the unuse/drop
5989 io_sq_thread_drop_mm(ctx);
5992 * We're polling. If we're within the defined idle
5993 * period, then let us spin without work before going
5994 * to sleep. The exception is if we got EBUSY doing
5995 * more IO, we should wait for the application to
5996 * reap events and wake us up.
5998 if (!list_empty(&ctx->poll_list) ||
5999 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6000 !percpu_ref_is_dying(&ctx->refs))) {
6001 if (current->task_works)
6007 prepare_to_wait(&ctx->sqo_wait, &wait,
6008 TASK_INTERRUPTIBLE);
6011 * While doing polled IO, before going to sleep, we need
6012 * to check if there are new reqs added to poll_list, it
6013 * is because reqs may have been punted to io worker and
6014 * will be added to poll_list later, hence check the
6017 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6018 !list_empty_careful(&ctx->poll_list)) {
6019 finish_wait(&ctx->sqo_wait, &wait);
6023 /* Tell userspace we may need a wakeup call */
6024 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6025 /* make sure to read SQ tail after writing flags */
6028 to_submit = io_sqring_entries(ctx);
6029 if (!to_submit || ret == -EBUSY) {
6030 if (kthread_should_park()) {
6031 finish_wait(&ctx->sqo_wait, &wait);
6034 if (current->task_works) {
6036 finish_wait(&ctx->sqo_wait, &wait);
6039 if (signal_pending(current))
6040 flush_signals(current);
6042 finish_wait(&ctx->sqo_wait, &wait);
6044 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6048 finish_wait(&ctx->sqo_wait, &wait);
6050 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6053 mutex_lock(&ctx->uring_lock);
6054 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6055 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6056 mutex_unlock(&ctx->uring_lock);
6057 timeout = jiffies + ctx->sq_thread_idle;
6060 if (current->task_works)
6063 io_sq_thread_drop_mm(ctx);
6064 revert_creds(old_cred);
6071 struct io_wait_queue {
6072 struct wait_queue_entry wq;
6073 struct io_ring_ctx *ctx;
6075 unsigned nr_timeouts;
6078 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6080 struct io_ring_ctx *ctx = iowq->ctx;
6083 * Wake up if we have enough events, or if a timeout occurred since we
6084 * started waiting. For timeouts, we always want to return to userspace,
6085 * regardless of event count.
6087 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6088 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6091 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6092 int wake_flags, void *key)
6094 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6097 /* use noflush == true, as we can't safely rely on locking context */
6098 if (!io_should_wake(iowq, true))
6101 return autoremove_wake_function(curr, mode, wake_flags, key);
6105 * Wait until events become available, if we don't already have some. The
6106 * application must reap them itself, as they reside on the shared cq ring.
6108 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6109 const sigset_t __user *sig, size_t sigsz)
6111 struct io_wait_queue iowq = {
6114 .func = io_wake_function,
6115 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6118 .to_wait = min_events,
6120 struct io_rings *rings = ctx->rings;
6124 if (io_cqring_events(ctx, false) >= min_events)
6126 if (!current->task_works)
6132 #ifdef CONFIG_COMPAT
6133 if (in_compat_syscall())
6134 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6138 ret = set_user_sigmask(sig, sigsz);
6144 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6145 trace_io_uring_cqring_wait(ctx, min_events);
6147 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6148 TASK_INTERRUPTIBLE);
6149 if (current->task_works)
6151 if (io_should_wake(&iowq, false))
6154 if (signal_pending(current)) {
6159 finish_wait(&ctx->wait, &iowq.wq);
6161 restore_saved_sigmask_unless(ret == -EINTR);
6163 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6166 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6168 #if defined(CONFIG_UNIX)
6169 if (ctx->ring_sock) {
6170 struct sock *sock = ctx->ring_sock->sk;
6171 struct sk_buff *skb;
6173 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6179 for (i = 0; i < ctx->nr_user_files; i++) {
6182 file = io_file_from_index(ctx, i);
6189 static void io_file_ref_kill(struct percpu_ref *ref)
6191 struct fixed_file_data *data;
6193 data = container_of(ref, struct fixed_file_data, refs);
6194 complete(&data->done);
6197 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6199 struct fixed_file_data *data = ctx->file_data;
6200 struct fixed_file_ref_node *ref_node = NULL;
6201 unsigned nr_tables, i;
6206 spin_lock(&data->lock);
6207 if (!list_empty(&data->ref_list))
6208 ref_node = list_first_entry(&data->ref_list,
6209 struct fixed_file_ref_node, node);
6210 spin_unlock(&data->lock);
6212 percpu_ref_kill(&ref_node->refs);
6214 percpu_ref_kill(&data->refs);
6216 /* wait for all refs nodes to complete */
6217 flush_delayed_work(&ctx->file_put_work);
6218 wait_for_completion(&data->done);
6220 __io_sqe_files_unregister(ctx);
6221 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6222 for (i = 0; i < nr_tables; i++)
6223 kfree(data->table[i].files);
6225 percpu_ref_exit(&data->refs);
6227 ctx->file_data = NULL;
6228 ctx->nr_user_files = 0;
6232 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6234 if (ctx->sqo_thread) {
6235 wait_for_completion(&ctx->sq_thread_comp);
6237 * The park is a bit of a work-around, without it we get
6238 * warning spews on shutdown with SQPOLL set and affinity
6239 * set to a single CPU.
6241 kthread_park(ctx->sqo_thread);
6242 kthread_stop(ctx->sqo_thread);
6243 ctx->sqo_thread = NULL;
6247 static void io_finish_async(struct io_ring_ctx *ctx)
6249 io_sq_thread_stop(ctx);
6252 io_wq_destroy(ctx->io_wq);
6257 #if defined(CONFIG_UNIX)
6259 * Ensure the UNIX gc is aware of our file set, so we are certain that
6260 * the io_uring can be safely unregistered on process exit, even if we have
6261 * loops in the file referencing.
6263 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6265 struct sock *sk = ctx->ring_sock->sk;
6266 struct scm_fp_list *fpl;
6267 struct sk_buff *skb;
6270 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6274 skb = alloc_skb(0, GFP_KERNEL);
6283 fpl->user = get_uid(ctx->user);
6284 for (i = 0; i < nr; i++) {
6285 struct file *file = io_file_from_index(ctx, i + offset);
6289 fpl->fp[nr_files] = get_file(file);
6290 unix_inflight(fpl->user, fpl->fp[nr_files]);
6295 fpl->max = SCM_MAX_FD;
6296 fpl->count = nr_files;
6297 UNIXCB(skb).fp = fpl;
6298 skb->destructor = unix_destruct_scm;
6299 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6300 skb_queue_head(&sk->sk_receive_queue, skb);
6302 for (i = 0; i < nr_files; i++)
6313 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6314 * causes regular reference counting to break down. We rely on the UNIX
6315 * garbage collection to take care of this problem for us.
6317 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6319 unsigned left, total;
6323 left = ctx->nr_user_files;
6325 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6327 ret = __io_sqe_files_scm(ctx, this_files, total);
6331 total += this_files;
6337 while (total < ctx->nr_user_files) {
6338 struct file *file = io_file_from_index(ctx, total);
6348 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6354 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6359 for (i = 0; i < nr_tables; i++) {
6360 struct fixed_file_table *table = &ctx->file_data->table[i];
6361 unsigned this_files;
6363 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6364 table->files = kcalloc(this_files, sizeof(struct file *),
6368 nr_files -= this_files;
6374 for (i = 0; i < nr_tables; i++) {
6375 struct fixed_file_table *table = &ctx->file_data->table[i];
6376 kfree(table->files);
6381 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6383 #if defined(CONFIG_UNIX)
6384 struct sock *sock = ctx->ring_sock->sk;
6385 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6386 struct sk_buff *skb;
6389 __skb_queue_head_init(&list);
6392 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6393 * remove this entry and rearrange the file array.
6395 skb = skb_dequeue(head);
6397 struct scm_fp_list *fp;
6399 fp = UNIXCB(skb).fp;
6400 for (i = 0; i < fp->count; i++) {
6403 if (fp->fp[i] != file)
6406 unix_notinflight(fp->user, fp->fp[i]);
6407 left = fp->count - 1 - i;
6409 memmove(&fp->fp[i], &fp->fp[i + 1],
6410 left * sizeof(struct file *));
6417 __skb_queue_tail(&list, skb);
6427 __skb_queue_tail(&list, skb);
6429 skb = skb_dequeue(head);
6432 if (skb_peek(&list)) {
6433 spin_lock_irq(&head->lock);
6434 while ((skb = __skb_dequeue(&list)) != NULL)
6435 __skb_queue_tail(head, skb);
6436 spin_unlock_irq(&head->lock);
6443 struct io_file_put {
6444 struct list_head list;
6448 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6450 struct fixed_file_data *file_data = ref_node->file_data;
6451 struct io_ring_ctx *ctx = file_data->ctx;
6452 struct io_file_put *pfile, *tmp;
6454 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6455 list_del(&pfile->list);
6456 io_ring_file_put(ctx, pfile->file);
6460 spin_lock(&file_data->lock);
6461 list_del(&ref_node->node);
6462 spin_unlock(&file_data->lock);
6464 percpu_ref_exit(&ref_node->refs);
6466 percpu_ref_put(&file_data->refs);
6469 static void io_file_put_work(struct work_struct *work)
6471 struct io_ring_ctx *ctx;
6472 struct llist_node *node;
6474 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6475 node = llist_del_all(&ctx->file_put_llist);
6478 struct fixed_file_ref_node *ref_node;
6479 struct llist_node *next = node->next;
6481 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6482 __io_file_put_work(ref_node);
6487 static void io_file_data_ref_zero(struct percpu_ref *ref)
6489 struct fixed_file_ref_node *ref_node;
6490 struct io_ring_ctx *ctx;
6494 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6495 ctx = ref_node->file_data->ctx;
6497 if (percpu_ref_is_dying(&ctx->file_data->refs))
6500 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6502 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6504 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6507 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6508 struct io_ring_ctx *ctx)
6510 struct fixed_file_ref_node *ref_node;
6512 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6514 return ERR_PTR(-ENOMEM);
6516 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6519 return ERR_PTR(-ENOMEM);
6521 INIT_LIST_HEAD(&ref_node->node);
6522 INIT_LIST_HEAD(&ref_node->file_list);
6523 ref_node->file_data = ctx->file_data;
6527 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6529 percpu_ref_exit(&ref_node->refs);
6533 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6536 __s32 __user *fds = (__s32 __user *) arg;
6541 struct fixed_file_ref_node *ref_node;
6547 if (nr_args > IORING_MAX_FIXED_FILES)
6550 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6551 if (!ctx->file_data)
6553 ctx->file_data->ctx = ctx;
6554 init_completion(&ctx->file_data->done);
6555 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6556 spin_lock_init(&ctx->file_data->lock);
6558 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6559 ctx->file_data->table = kcalloc(nr_tables,
6560 sizeof(struct fixed_file_table),
6562 if (!ctx->file_data->table) {
6563 kfree(ctx->file_data);
6564 ctx->file_data = NULL;
6568 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6569 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6570 kfree(ctx->file_data->table);
6571 kfree(ctx->file_data);
6572 ctx->file_data = NULL;
6576 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6577 percpu_ref_exit(&ctx->file_data->refs);
6578 kfree(ctx->file_data->table);
6579 kfree(ctx->file_data);
6580 ctx->file_data = NULL;
6584 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6585 struct fixed_file_table *table;
6589 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6591 /* allow sparse sets */
6597 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6598 index = i & IORING_FILE_TABLE_MASK;
6606 * Don't allow io_uring instances to be registered. If UNIX
6607 * isn't enabled, then this causes a reference cycle and this
6608 * instance can never get freed. If UNIX is enabled we'll
6609 * handle it just fine, but there's still no point in allowing
6610 * a ring fd as it doesn't support regular read/write anyway.
6612 if (file->f_op == &io_uring_fops) {
6617 table->files[index] = file;
6621 for (i = 0; i < ctx->nr_user_files; i++) {
6622 file = io_file_from_index(ctx, i);
6626 for (i = 0; i < nr_tables; i++)
6627 kfree(ctx->file_data->table[i].files);
6629 kfree(ctx->file_data->table);
6630 kfree(ctx->file_data);
6631 ctx->file_data = NULL;
6632 ctx->nr_user_files = 0;
6636 ret = io_sqe_files_scm(ctx);
6638 io_sqe_files_unregister(ctx);
6642 ref_node = alloc_fixed_file_ref_node(ctx);
6643 if (IS_ERR(ref_node)) {
6644 io_sqe_files_unregister(ctx);
6645 return PTR_ERR(ref_node);
6648 ctx->file_data->cur_refs = &ref_node->refs;
6649 spin_lock(&ctx->file_data->lock);
6650 list_add(&ref_node->node, &ctx->file_data->ref_list);
6651 spin_unlock(&ctx->file_data->lock);
6652 percpu_ref_get(&ctx->file_data->refs);
6656 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6659 #if defined(CONFIG_UNIX)
6660 struct sock *sock = ctx->ring_sock->sk;
6661 struct sk_buff_head *head = &sock->sk_receive_queue;
6662 struct sk_buff *skb;
6665 * See if we can merge this file into an existing skb SCM_RIGHTS
6666 * file set. If there's no room, fall back to allocating a new skb
6667 * and filling it in.
6669 spin_lock_irq(&head->lock);
6670 skb = skb_peek(head);
6672 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6674 if (fpl->count < SCM_MAX_FD) {
6675 __skb_unlink(skb, head);
6676 spin_unlock_irq(&head->lock);
6677 fpl->fp[fpl->count] = get_file(file);
6678 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6680 spin_lock_irq(&head->lock);
6681 __skb_queue_head(head, skb);
6686 spin_unlock_irq(&head->lock);
6693 return __io_sqe_files_scm(ctx, 1, index);
6699 static int io_queue_file_removal(struct fixed_file_data *data,
6702 struct io_file_put *pfile;
6703 struct percpu_ref *refs = data->cur_refs;
6704 struct fixed_file_ref_node *ref_node;
6706 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6710 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6712 list_add(&pfile->list, &ref_node->file_list);
6717 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6718 struct io_uring_files_update *up,
6721 struct fixed_file_data *data = ctx->file_data;
6722 struct fixed_file_ref_node *ref_node;
6727 bool needs_switch = false;
6729 if (check_add_overflow(up->offset, nr_args, &done))
6731 if (done > ctx->nr_user_files)
6734 ref_node = alloc_fixed_file_ref_node(ctx);
6735 if (IS_ERR(ref_node))
6736 return PTR_ERR(ref_node);
6739 fds = u64_to_user_ptr(up->fds);
6741 struct fixed_file_table *table;
6745 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6749 i = array_index_nospec(up->offset, ctx->nr_user_files);
6750 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6751 index = i & IORING_FILE_TABLE_MASK;
6752 if (table->files[index]) {
6753 file = io_file_from_index(ctx, index);
6754 err = io_queue_file_removal(data, file);
6757 table->files[index] = NULL;
6758 needs_switch = true;
6767 * Don't allow io_uring instances to be registered. If
6768 * UNIX isn't enabled, then this causes a reference
6769 * cycle and this instance can never get freed. If UNIX
6770 * is enabled we'll handle it just fine, but there's
6771 * still no point in allowing a ring fd as it doesn't
6772 * support regular read/write anyway.
6774 if (file->f_op == &io_uring_fops) {
6779 table->files[index] = file;
6780 err = io_sqe_file_register(ctx, file, i);
6790 percpu_ref_kill(data->cur_refs);
6791 spin_lock(&data->lock);
6792 list_add(&ref_node->node, &data->ref_list);
6793 data->cur_refs = &ref_node->refs;
6794 spin_unlock(&data->lock);
6795 percpu_ref_get(&ctx->file_data->refs);
6797 destroy_fixed_file_ref_node(ref_node);
6799 return done ? done : err;
6802 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6805 struct io_uring_files_update up;
6807 if (!ctx->file_data)
6811 if (copy_from_user(&up, arg, sizeof(up)))
6816 return __io_sqe_files_update(ctx, &up, nr_args);
6819 static void io_free_work(struct io_wq_work *work)
6821 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6823 /* Consider that io_steal_work() relies on this ref */
6827 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6828 struct io_uring_params *p)
6830 struct io_wq_data data;
6832 struct io_ring_ctx *ctx_attach;
6833 unsigned int concurrency;
6836 data.user = ctx->user;
6837 data.free_work = io_free_work;
6838 data.do_work = io_wq_submit_work;
6840 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6841 /* Do QD, or 4 * CPUS, whatever is smallest */
6842 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6844 ctx->io_wq = io_wq_create(concurrency, &data);
6845 if (IS_ERR(ctx->io_wq)) {
6846 ret = PTR_ERR(ctx->io_wq);
6852 f = fdget(p->wq_fd);
6856 if (f.file->f_op != &io_uring_fops) {
6861 ctx_attach = f.file->private_data;
6862 /* @io_wq is protected by holding the fd */
6863 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6868 ctx->io_wq = ctx_attach->io_wq;
6874 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6875 struct io_uring_params *p)
6879 mmgrab(current->mm);
6880 ctx->sqo_mm = current->mm;
6882 if (ctx->flags & IORING_SETUP_SQPOLL) {
6884 if (!capable(CAP_SYS_ADMIN))
6887 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6888 if (!ctx->sq_thread_idle)
6889 ctx->sq_thread_idle = HZ;
6891 if (p->flags & IORING_SETUP_SQ_AFF) {
6892 int cpu = p->sq_thread_cpu;
6895 if (cpu >= nr_cpu_ids)
6897 if (!cpu_online(cpu))
6900 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6904 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6907 if (IS_ERR(ctx->sqo_thread)) {
6908 ret = PTR_ERR(ctx->sqo_thread);
6909 ctx->sqo_thread = NULL;
6912 wake_up_process(ctx->sqo_thread);
6913 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6914 /* Can't have SQ_AFF without SQPOLL */
6919 ret = io_init_wq_offload(ctx, p);
6925 io_finish_async(ctx);
6926 mmdrop(ctx->sqo_mm);
6931 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6933 atomic_long_sub(nr_pages, &user->locked_vm);
6936 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6938 unsigned long page_limit, cur_pages, new_pages;
6940 /* Don't allow more pages than we can safely lock */
6941 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6944 cur_pages = atomic_long_read(&user->locked_vm);
6945 new_pages = cur_pages + nr_pages;
6946 if (new_pages > page_limit)
6948 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6949 new_pages) != cur_pages);
6954 static void io_mem_free(void *ptr)
6961 page = virt_to_head_page(ptr);
6962 if (put_page_testzero(page))
6963 free_compound_page(page);
6966 static void *io_mem_alloc(size_t size)
6968 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6971 return (void *) __get_free_pages(gfp_flags, get_order(size));
6974 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6977 struct io_rings *rings;
6978 size_t off, sq_array_size;
6980 off = struct_size(rings, cqes, cq_entries);
6981 if (off == SIZE_MAX)
6985 off = ALIGN(off, SMP_CACHE_BYTES);
6990 sq_array_size = array_size(sizeof(u32), sq_entries);
6991 if (sq_array_size == SIZE_MAX)
6994 if (check_add_overflow(off, sq_array_size, &off))
7003 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7007 pages = (size_t)1 << get_order(
7008 rings_size(sq_entries, cq_entries, NULL));
7009 pages += (size_t)1 << get_order(
7010 array_size(sizeof(struct io_uring_sqe), sq_entries));
7015 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7019 if (!ctx->user_bufs)
7022 for (i = 0; i < ctx->nr_user_bufs; i++) {
7023 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7025 for (j = 0; j < imu->nr_bvecs; j++)
7026 unpin_user_page(imu->bvec[j].bv_page);
7028 if (ctx->account_mem)
7029 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7034 kfree(ctx->user_bufs);
7035 ctx->user_bufs = NULL;
7036 ctx->nr_user_bufs = 0;
7040 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7041 void __user *arg, unsigned index)
7043 struct iovec __user *src;
7045 #ifdef CONFIG_COMPAT
7047 struct compat_iovec __user *ciovs;
7048 struct compat_iovec ciov;
7050 ciovs = (struct compat_iovec __user *) arg;
7051 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7054 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7055 dst->iov_len = ciov.iov_len;
7059 src = (struct iovec __user *) arg;
7060 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7065 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7068 struct vm_area_struct **vmas = NULL;
7069 struct page **pages = NULL;
7070 int i, j, got_pages = 0;
7075 if (!nr_args || nr_args > UIO_MAXIOV)
7078 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7080 if (!ctx->user_bufs)
7083 for (i = 0; i < nr_args; i++) {
7084 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7085 unsigned long off, start, end, ubuf;
7090 ret = io_copy_iov(ctx, &iov, arg, i);
7095 * Don't impose further limits on the size and buffer
7096 * constraints here, we'll -EINVAL later when IO is
7097 * submitted if they are wrong.
7100 if (!iov.iov_base || !iov.iov_len)
7103 /* arbitrary limit, but we need something */
7104 if (iov.iov_len > SZ_1G)
7107 ubuf = (unsigned long) iov.iov_base;
7108 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7109 start = ubuf >> PAGE_SHIFT;
7110 nr_pages = end - start;
7112 if (ctx->account_mem) {
7113 ret = io_account_mem(ctx->user, nr_pages);
7119 if (!pages || nr_pages > got_pages) {
7122 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7124 vmas = kvmalloc_array(nr_pages,
7125 sizeof(struct vm_area_struct *),
7127 if (!pages || !vmas) {
7129 if (ctx->account_mem)
7130 io_unaccount_mem(ctx->user, nr_pages);
7133 got_pages = nr_pages;
7136 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7140 if (ctx->account_mem)
7141 io_unaccount_mem(ctx->user, nr_pages);
7146 mmap_read_lock(current->mm);
7147 pret = pin_user_pages(ubuf, nr_pages,
7148 FOLL_WRITE | FOLL_LONGTERM,
7150 if (pret == nr_pages) {
7151 /* don't support file backed memory */
7152 for (j = 0; j < nr_pages; j++) {
7153 struct vm_area_struct *vma = vmas[j];
7156 !is_file_hugepages(vma->vm_file)) {
7162 ret = pret < 0 ? pret : -EFAULT;
7164 mmap_read_unlock(current->mm);
7167 * if we did partial map, or found file backed vmas,
7168 * release any pages we did get
7171 unpin_user_pages(pages, pret);
7172 if (ctx->account_mem)
7173 io_unaccount_mem(ctx->user, nr_pages);
7178 off = ubuf & ~PAGE_MASK;
7180 for (j = 0; j < nr_pages; j++) {
7183 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7184 imu->bvec[j].bv_page = pages[j];
7185 imu->bvec[j].bv_len = vec_len;
7186 imu->bvec[j].bv_offset = off;
7190 /* store original address for later verification */
7192 imu->len = iov.iov_len;
7193 imu->nr_bvecs = nr_pages;
7195 ctx->nr_user_bufs++;
7203 io_sqe_buffer_unregister(ctx);
7207 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7209 __s32 __user *fds = arg;
7215 if (copy_from_user(&fd, fds, sizeof(*fds)))
7218 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7219 if (IS_ERR(ctx->cq_ev_fd)) {
7220 int ret = PTR_ERR(ctx->cq_ev_fd);
7221 ctx->cq_ev_fd = NULL;
7228 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7230 if (ctx->cq_ev_fd) {
7231 eventfd_ctx_put(ctx->cq_ev_fd);
7232 ctx->cq_ev_fd = NULL;
7239 static int __io_destroy_buffers(int id, void *p, void *data)
7241 struct io_ring_ctx *ctx = data;
7242 struct io_buffer *buf = p;
7244 __io_remove_buffers(ctx, buf, id, -1U);
7248 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7250 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7251 idr_destroy(&ctx->io_buffer_idr);
7254 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7256 io_finish_async(ctx);
7258 mmdrop(ctx->sqo_mm);
7260 io_iopoll_reap_events(ctx);
7261 io_sqe_buffer_unregister(ctx);
7262 io_sqe_files_unregister(ctx);
7263 io_eventfd_unregister(ctx);
7264 io_destroy_buffers(ctx);
7265 idr_destroy(&ctx->personality_idr);
7267 #if defined(CONFIG_UNIX)
7268 if (ctx->ring_sock) {
7269 ctx->ring_sock->file = NULL; /* so that iput() is called */
7270 sock_release(ctx->ring_sock);
7274 io_mem_free(ctx->rings);
7275 io_mem_free(ctx->sq_sqes);
7277 percpu_ref_exit(&ctx->refs);
7278 if (ctx->account_mem)
7279 io_unaccount_mem(ctx->user,
7280 ring_pages(ctx->sq_entries, ctx->cq_entries));
7281 free_uid(ctx->user);
7282 put_cred(ctx->creds);
7283 kfree(ctx->cancel_hash);
7284 kmem_cache_free(req_cachep, ctx->fallback_req);
7288 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7290 struct io_ring_ctx *ctx = file->private_data;
7293 poll_wait(file, &ctx->cq_wait, wait);
7295 * synchronizes with barrier from wq_has_sleeper call in
7299 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7300 ctx->rings->sq_ring_entries)
7301 mask |= EPOLLOUT | EPOLLWRNORM;
7302 if (io_cqring_events(ctx, false))
7303 mask |= EPOLLIN | EPOLLRDNORM;
7308 static int io_uring_fasync(int fd, struct file *file, int on)
7310 struct io_ring_ctx *ctx = file->private_data;
7312 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7315 static int io_remove_personalities(int id, void *p, void *data)
7317 struct io_ring_ctx *ctx = data;
7318 const struct cred *cred;
7320 cred = idr_remove(&ctx->personality_idr, id);
7326 static void io_ring_exit_work(struct work_struct *work)
7328 struct io_ring_ctx *ctx;
7330 ctx = container_of(work, struct io_ring_ctx, exit_work);
7332 io_cqring_overflow_flush(ctx, true);
7334 wait_for_completion(&ctx->ref_comp);
7335 io_ring_ctx_free(ctx);
7338 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7340 mutex_lock(&ctx->uring_lock);
7341 percpu_ref_kill(&ctx->refs);
7342 mutex_unlock(&ctx->uring_lock);
7344 io_kill_timeouts(ctx);
7345 io_poll_remove_all(ctx);
7348 io_wq_cancel_all(ctx->io_wq);
7350 io_iopoll_reap_events(ctx);
7351 /* if we failed setting up the ctx, we might not have any rings */
7353 io_cqring_overflow_flush(ctx, true);
7354 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7355 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7356 queue_work(system_wq, &ctx->exit_work);
7359 static int io_uring_release(struct inode *inode, struct file *file)
7361 struct io_ring_ctx *ctx = file->private_data;
7363 file->private_data = NULL;
7364 io_ring_ctx_wait_and_kill(ctx);
7368 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7369 struct files_struct *files)
7371 while (!list_empty_careful(&ctx->inflight_list)) {
7372 struct io_kiocb *cancel_req = NULL, *req;
7375 spin_lock_irq(&ctx->inflight_lock);
7376 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7377 if (req->work.files != files)
7379 /* req is being completed, ignore */
7380 if (!refcount_inc_not_zero(&req->refs))
7386 prepare_to_wait(&ctx->inflight_wait, &wait,
7387 TASK_UNINTERRUPTIBLE);
7388 spin_unlock_irq(&ctx->inflight_lock);
7390 /* We need to keep going until we don't find a matching req */
7394 if (cancel_req->flags & REQ_F_OVERFLOW) {
7395 spin_lock_irq(&ctx->completion_lock);
7396 list_del(&cancel_req->list);
7397 cancel_req->flags &= ~REQ_F_OVERFLOW;
7398 if (list_empty(&ctx->cq_overflow_list)) {
7399 clear_bit(0, &ctx->sq_check_overflow);
7400 clear_bit(0, &ctx->cq_check_overflow);
7402 spin_unlock_irq(&ctx->completion_lock);
7404 WRITE_ONCE(ctx->rings->cq_overflow,
7405 atomic_inc_return(&ctx->cached_cq_overflow));
7408 * Put inflight ref and overflow ref. If that's
7409 * all we had, then we're done with this request.
7411 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7412 io_free_req(cancel_req);
7413 finish_wait(&ctx->inflight_wait, &wait);
7417 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7418 io_put_req(cancel_req);
7422 finish_wait(&ctx->inflight_wait, &wait);
7426 static int io_uring_flush(struct file *file, void *data)
7428 struct io_ring_ctx *ctx = file->private_data;
7430 io_uring_cancel_files(ctx, data);
7433 * If the task is going away, cancel work it may have pending
7435 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7436 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7441 static void *io_uring_validate_mmap_request(struct file *file,
7442 loff_t pgoff, size_t sz)
7444 struct io_ring_ctx *ctx = file->private_data;
7445 loff_t offset = pgoff << PAGE_SHIFT;
7450 case IORING_OFF_SQ_RING:
7451 case IORING_OFF_CQ_RING:
7454 case IORING_OFF_SQES:
7458 return ERR_PTR(-EINVAL);
7461 page = virt_to_head_page(ptr);
7462 if (sz > page_size(page))
7463 return ERR_PTR(-EINVAL);
7470 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7472 size_t sz = vma->vm_end - vma->vm_start;
7476 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7478 return PTR_ERR(ptr);
7480 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7481 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7484 #else /* !CONFIG_MMU */
7486 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7488 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7491 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7493 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7496 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7497 unsigned long addr, unsigned long len,
7498 unsigned long pgoff, unsigned long flags)
7502 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7504 return PTR_ERR(ptr);
7506 return (unsigned long) ptr;
7509 #endif /* !CONFIG_MMU */
7511 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7512 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7515 struct io_ring_ctx *ctx;
7520 if (current->task_works)
7523 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7531 if (f.file->f_op != &io_uring_fops)
7535 ctx = f.file->private_data;
7536 if (!percpu_ref_tryget(&ctx->refs))
7540 * For SQ polling, the thread will do all submissions and completions.
7541 * Just return the requested submit count, and wake the thread if
7545 if (ctx->flags & IORING_SETUP_SQPOLL) {
7546 if (!list_empty_careful(&ctx->cq_overflow_list))
7547 io_cqring_overflow_flush(ctx, false);
7548 if (flags & IORING_ENTER_SQ_WAKEUP)
7549 wake_up(&ctx->sqo_wait);
7550 submitted = to_submit;
7551 } else if (to_submit) {
7552 mutex_lock(&ctx->uring_lock);
7553 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7554 mutex_unlock(&ctx->uring_lock);
7556 if (submitted != to_submit)
7559 if (flags & IORING_ENTER_GETEVENTS) {
7560 unsigned nr_events = 0;
7562 min_complete = min(min_complete, ctx->cq_entries);
7565 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7566 * space applications don't need to do io completion events
7567 * polling again, they can rely on io_sq_thread to do polling
7568 * work, which can reduce cpu usage and uring_lock contention.
7570 if (ctx->flags & IORING_SETUP_IOPOLL &&
7571 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7572 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7574 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7579 percpu_ref_put(&ctx->refs);
7582 return submitted ? submitted : ret;
7585 #ifdef CONFIG_PROC_FS
7586 static int io_uring_show_cred(int id, void *p, void *data)
7588 const struct cred *cred = p;
7589 struct seq_file *m = data;
7590 struct user_namespace *uns = seq_user_ns(m);
7591 struct group_info *gi;
7596 seq_printf(m, "%5d\n", id);
7597 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7598 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7599 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7600 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7601 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7602 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7603 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7604 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7605 seq_puts(m, "\n\tGroups:\t");
7606 gi = cred->group_info;
7607 for (g = 0; g < gi->ngroups; g++) {
7608 seq_put_decimal_ull(m, g ? " " : "",
7609 from_kgid_munged(uns, gi->gid[g]));
7611 seq_puts(m, "\n\tCapEff:\t");
7612 cap = cred->cap_effective;
7613 CAP_FOR_EACH_U32(__capi)
7614 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7619 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7623 mutex_lock(&ctx->uring_lock);
7624 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7625 for (i = 0; i < ctx->nr_user_files; i++) {
7626 struct fixed_file_table *table;
7629 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7630 f = table->files[i & IORING_FILE_TABLE_MASK];
7632 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7634 seq_printf(m, "%5u: <none>\n", i);
7636 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7637 for (i = 0; i < ctx->nr_user_bufs; i++) {
7638 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7640 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7641 (unsigned int) buf->len);
7643 if (!idr_is_empty(&ctx->personality_idr)) {
7644 seq_printf(m, "Personalities:\n");
7645 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7647 seq_printf(m, "PollList:\n");
7648 spin_lock_irq(&ctx->completion_lock);
7649 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7650 struct hlist_head *list = &ctx->cancel_hash[i];
7651 struct io_kiocb *req;
7653 hlist_for_each_entry(req, list, hash_node)
7654 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7655 req->task->task_works != NULL);
7657 spin_unlock_irq(&ctx->completion_lock);
7658 mutex_unlock(&ctx->uring_lock);
7661 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7663 struct io_ring_ctx *ctx = f->private_data;
7665 if (percpu_ref_tryget(&ctx->refs)) {
7666 __io_uring_show_fdinfo(ctx, m);
7667 percpu_ref_put(&ctx->refs);
7672 static const struct file_operations io_uring_fops = {
7673 .release = io_uring_release,
7674 .flush = io_uring_flush,
7675 .mmap = io_uring_mmap,
7677 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7678 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7680 .poll = io_uring_poll,
7681 .fasync = io_uring_fasync,
7682 #ifdef CONFIG_PROC_FS
7683 .show_fdinfo = io_uring_show_fdinfo,
7687 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7688 struct io_uring_params *p)
7690 struct io_rings *rings;
7691 size_t size, sq_array_offset;
7693 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7694 if (size == SIZE_MAX)
7697 rings = io_mem_alloc(size);
7702 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7703 rings->sq_ring_mask = p->sq_entries - 1;
7704 rings->cq_ring_mask = p->cq_entries - 1;
7705 rings->sq_ring_entries = p->sq_entries;
7706 rings->cq_ring_entries = p->cq_entries;
7707 ctx->sq_mask = rings->sq_ring_mask;
7708 ctx->cq_mask = rings->cq_ring_mask;
7709 ctx->sq_entries = rings->sq_ring_entries;
7710 ctx->cq_entries = rings->cq_ring_entries;
7712 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7713 if (size == SIZE_MAX) {
7714 io_mem_free(ctx->rings);
7719 ctx->sq_sqes = io_mem_alloc(size);
7720 if (!ctx->sq_sqes) {
7721 io_mem_free(ctx->rings);
7730 * Allocate an anonymous fd, this is what constitutes the application
7731 * visible backing of an io_uring instance. The application mmaps this
7732 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7733 * we have to tie this fd to a socket for file garbage collection purposes.
7735 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7740 #if defined(CONFIG_UNIX)
7741 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7747 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7751 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7752 O_RDWR | O_CLOEXEC);
7755 ret = PTR_ERR(file);
7759 #if defined(CONFIG_UNIX)
7760 ctx->ring_sock->file = file;
7762 fd_install(ret, file);
7765 #if defined(CONFIG_UNIX)
7766 sock_release(ctx->ring_sock);
7767 ctx->ring_sock = NULL;
7772 static int io_uring_create(unsigned entries, struct io_uring_params *p,
7773 struct io_uring_params __user *params)
7775 struct user_struct *user = NULL;
7776 struct io_ring_ctx *ctx;
7782 if (entries > IORING_MAX_ENTRIES) {
7783 if (!(p->flags & IORING_SETUP_CLAMP))
7785 entries = IORING_MAX_ENTRIES;
7789 * Use twice as many entries for the CQ ring. It's possible for the
7790 * application to drive a higher depth than the size of the SQ ring,
7791 * since the sqes are only used at submission time. This allows for
7792 * some flexibility in overcommitting a bit. If the application has
7793 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7794 * of CQ ring entries manually.
7796 p->sq_entries = roundup_pow_of_two(entries);
7797 if (p->flags & IORING_SETUP_CQSIZE) {
7799 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7800 * to a power-of-two, if it isn't already. We do NOT impose
7801 * any cq vs sq ring sizing.
7803 if (p->cq_entries < p->sq_entries)
7805 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7806 if (!(p->flags & IORING_SETUP_CLAMP))
7808 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7810 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7812 p->cq_entries = 2 * p->sq_entries;
7815 user = get_uid(current_user());
7816 account_mem = !capable(CAP_IPC_LOCK);
7819 ret = io_account_mem(user,
7820 ring_pages(p->sq_entries, p->cq_entries));
7827 ctx = io_ring_ctx_alloc(p);
7830 io_unaccount_mem(user, ring_pages(p->sq_entries,
7835 ctx->compat = in_compat_syscall();
7836 ctx->account_mem = account_mem;
7838 ctx->creds = get_current_cred();
7840 ret = io_allocate_scq_urings(ctx, p);
7844 ret = io_sq_offload_start(ctx, p);
7848 memset(&p->sq_off, 0, sizeof(p->sq_off));
7849 p->sq_off.head = offsetof(struct io_rings, sq.head);
7850 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7851 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7852 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7853 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7854 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7855 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7857 memset(&p->cq_off, 0, sizeof(p->cq_off));
7858 p->cq_off.head = offsetof(struct io_rings, cq.head);
7859 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7860 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7861 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7862 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7863 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7864 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
7866 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7867 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7868 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7870 if (copy_to_user(params, p, sizeof(*p))) {
7875 * Install ring fd as the very last thing, so we don't risk someone
7876 * having closed it before we finish setup
7878 ret = io_uring_get_fd(ctx);
7882 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7885 io_ring_ctx_wait_and_kill(ctx);
7890 * Sets up an aio uring context, and returns the fd. Applications asks for a
7891 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7892 * params structure passed in.
7894 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7896 struct io_uring_params p;
7899 if (copy_from_user(&p, params, sizeof(p)))
7901 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7906 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7907 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7908 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7911 return io_uring_create(entries, &p, params);
7914 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7915 struct io_uring_params __user *, params)
7917 return io_uring_setup(entries, params);
7920 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7922 struct io_uring_probe *p;
7926 size = struct_size(p, ops, nr_args);
7927 if (size == SIZE_MAX)
7929 p = kzalloc(size, GFP_KERNEL);
7934 if (copy_from_user(p, arg, size))
7937 if (memchr_inv(p, 0, size))
7940 p->last_op = IORING_OP_LAST - 1;
7941 if (nr_args > IORING_OP_LAST)
7942 nr_args = IORING_OP_LAST;
7944 for (i = 0; i < nr_args; i++) {
7946 if (!io_op_defs[i].not_supported)
7947 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7952 if (copy_to_user(arg, p, size))
7959 static int io_register_personality(struct io_ring_ctx *ctx)
7961 const struct cred *creds = get_current_cred();
7964 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7965 USHRT_MAX, GFP_KERNEL);
7971 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7973 const struct cred *old_creds;
7975 old_creds = idr_remove(&ctx->personality_idr, id);
7977 put_cred(old_creds);
7984 static bool io_register_op_must_quiesce(int op)
7987 case IORING_UNREGISTER_FILES:
7988 case IORING_REGISTER_FILES_UPDATE:
7989 case IORING_REGISTER_PROBE:
7990 case IORING_REGISTER_PERSONALITY:
7991 case IORING_UNREGISTER_PERSONALITY:
7998 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7999 void __user *arg, unsigned nr_args)
8000 __releases(ctx->uring_lock)
8001 __acquires(ctx->uring_lock)
8006 * We're inside the ring mutex, if the ref is already dying, then
8007 * someone else killed the ctx or is already going through
8008 * io_uring_register().
8010 if (percpu_ref_is_dying(&ctx->refs))
8013 if (io_register_op_must_quiesce(opcode)) {
8014 percpu_ref_kill(&ctx->refs);
8017 * Drop uring mutex before waiting for references to exit. If
8018 * another thread is currently inside io_uring_enter() it might
8019 * need to grab the uring_lock to make progress. If we hold it
8020 * here across the drain wait, then we can deadlock. It's safe
8021 * to drop the mutex here, since no new references will come in
8022 * after we've killed the percpu ref.
8024 mutex_unlock(&ctx->uring_lock);
8025 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8026 mutex_lock(&ctx->uring_lock);
8028 percpu_ref_resurrect(&ctx->refs);
8035 case IORING_REGISTER_BUFFERS:
8036 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8038 case IORING_UNREGISTER_BUFFERS:
8042 ret = io_sqe_buffer_unregister(ctx);
8044 case IORING_REGISTER_FILES:
8045 ret = io_sqe_files_register(ctx, arg, nr_args);
8047 case IORING_UNREGISTER_FILES:
8051 ret = io_sqe_files_unregister(ctx);
8053 case IORING_REGISTER_FILES_UPDATE:
8054 ret = io_sqe_files_update(ctx, arg, nr_args);
8056 case IORING_REGISTER_EVENTFD:
8057 case IORING_REGISTER_EVENTFD_ASYNC:
8061 ret = io_eventfd_register(ctx, arg);
8064 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8065 ctx->eventfd_async = 1;
8067 ctx->eventfd_async = 0;
8069 case IORING_UNREGISTER_EVENTFD:
8073 ret = io_eventfd_unregister(ctx);
8075 case IORING_REGISTER_PROBE:
8077 if (!arg || nr_args > 256)
8079 ret = io_probe(ctx, arg, nr_args);
8081 case IORING_REGISTER_PERSONALITY:
8085 ret = io_register_personality(ctx);
8087 case IORING_UNREGISTER_PERSONALITY:
8091 ret = io_unregister_personality(ctx, nr_args);
8098 if (io_register_op_must_quiesce(opcode)) {
8099 /* bring the ctx back to life */
8100 percpu_ref_reinit(&ctx->refs);
8102 reinit_completion(&ctx->ref_comp);
8107 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8108 void __user *, arg, unsigned int, nr_args)
8110 struct io_ring_ctx *ctx;
8119 if (f.file->f_op != &io_uring_fops)
8122 ctx = f.file->private_data;
8124 mutex_lock(&ctx->uring_lock);
8125 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8126 mutex_unlock(&ctx->uring_lock);
8127 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8128 ctx->cq_ev_fd != NULL, ret);
8134 static int __init io_uring_init(void)
8136 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8137 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8138 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8141 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8142 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8143 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8144 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8145 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8146 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8147 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8148 BUILD_BUG_SQE_ELEM(8, __u64, off);
8149 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8150 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8151 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8152 BUILD_BUG_SQE_ELEM(24, __u32, len);
8153 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8154 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8155 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8156 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8157 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8158 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8159 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8160 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8161 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8162 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8163 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8164 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8165 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8166 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8167 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8168 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8169 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8170 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8172 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8173 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8174 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8177 __initcall(io_uring_init);