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>
81 #include <linux/pagemap.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int limit_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->iopoll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head iopoll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
399 struct list_head list;
403 /* NOTE: kiocb has the file as the first member, so don't do it here */
411 struct sockaddr __user *addr;
418 struct user_msghdr __user *umsg;
424 struct io_buffer *kbuf;
430 struct filename *filename;
432 unsigned long nofile;
435 struct io_files_update {
461 struct epoll_event event;
465 struct file *file_out;
466 struct file *file_in;
473 struct io_provide_buf {
487 const char __user *filename;
488 struct statx __user *buffer;
491 struct io_completion {
493 struct list_head list;
497 struct io_async_connect {
498 struct sockaddr_storage address;
501 struct io_async_msghdr {
502 struct iovec fast_iov[UIO_FASTIOV];
504 struct sockaddr __user *uaddr;
506 struct sockaddr_storage addr;
510 struct iovec fast_iov[UIO_FASTIOV];
514 struct wait_page_queue wpq;
517 struct io_async_ctx {
519 struct io_async_rw rw;
520 struct io_async_msghdr msg;
521 struct io_async_connect connect;
522 struct io_timeout_data timeout;
527 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
528 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
529 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
530 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
531 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
532 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
539 REQ_F_LINK_TIMEOUT_BIT,
541 REQ_F_COMP_LOCKED_BIT,
542 REQ_F_NEED_CLEANUP_BIT,
545 REQ_F_BUFFER_SELECTED_BIT,
546 REQ_F_NO_FILE_TABLE_BIT,
547 REQ_F_WORK_INITIALIZED_BIT,
548 REQ_F_TASK_PINNED_BIT,
550 /* not a real bit, just to check we're not overflowing the space */
556 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
557 /* drain existing IO first */
558 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
560 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
561 /* doesn't sever on completion < 0 */
562 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
564 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
565 /* IOSQE_BUFFER_SELECT */
566 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
569 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
570 /* fail rest of links */
571 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
572 /* on inflight list */
573 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
574 /* read/write uses file position */
575 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
576 /* must not punt to workers */
577 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
578 /* has linked timeout */
579 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
581 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
582 /* completion under lock */
583 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
585 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
586 /* in overflow list */
587 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
588 /* already went through poll handler */
589 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
590 /* buffer already selected */
591 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
592 /* doesn't need file table for this request */
593 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
594 /* io_wq_work is initialized */
595 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
596 /* req->task is refcounted */
597 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
601 struct io_poll_iocb poll;
602 struct io_poll_iocb *double_poll;
603 struct io_wq_work work;
607 * NOTE! Each of the iocb union members has the file pointer
608 * as the first entry in their struct definition. So you can
609 * access the file pointer through any of the sub-structs,
610 * or directly as just 'ki_filp' in this struct.
616 struct io_poll_iocb poll;
617 struct io_accept accept;
619 struct io_cancel cancel;
620 struct io_timeout timeout;
621 struct io_connect connect;
622 struct io_sr_msg sr_msg;
624 struct io_close close;
625 struct io_files_update files_update;
626 struct io_fadvise fadvise;
627 struct io_madvise madvise;
628 struct io_epoll epoll;
629 struct io_splice splice;
630 struct io_provide_buf pbuf;
631 struct io_statx statx;
632 /* use only after cleaning per-op data, see io_clean_op() */
633 struct io_completion compl;
636 struct io_async_ctx *io;
638 /* polled IO has completed */
644 struct io_ring_ctx *ctx;
647 struct task_struct *task;
650 struct list_head link_list;
653 * 1. used with ctx->iopoll_list with reads/writes
654 * 2. to track reqs with ->files (see io_op_def::file_table)
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 hlist_node hash_node;
669 struct async_poll *apoll;
671 struct io_wq_work work;
673 struct callback_head task_work;
676 struct io_defer_entry {
677 struct list_head list;
678 struct io_kiocb *req;
682 #define IO_IOPOLL_BATCH 8
684 struct io_comp_state {
686 struct list_head list;
687 struct io_ring_ctx *ctx;
690 struct io_submit_state {
691 struct blk_plug plug;
694 * io_kiocb alloc cache
696 void *reqs[IO_IOPOLL_BATCH];
697 unsigned int free_reqs;
700 * Batch completion logic
702 struct io_comp_state comp;
705 * File reference cache
709 unsigned int has_refs;
710 unsigned int ios_left;
714 /* needs req->io allocated for deferral/async */
715 unsigned async_ctx : 1;
716 /* needs current->mm setup, does mm access */
717 unsigned needs_mm : 1;
718 /* needs req->file assigned */
719 unsigned needs_file : 1;
720 /* don't fail if file grab fails */
721 unsigned needs_file_no_error : 1;
722 /* hash wq insertion if file is a regular file */
723 unsigned hash_reg_file : 1;
724 /* unbound wq insertion if file is a non-regular file */
725 unsigned unbound_nonreg_file : 1;
726 /* opcode is not supported by this kernel */
727 unsigned not_supported : 1;
728 /* needs file table */
729 unsigned file_table : 1;
731 unsigned needs_fs : 1;
732 /* set if opcode supports polled "wait" */
734 unsigned pollout : 1;
735 /* op supports buffer selection */
736 unsigned buffer_select : 1;
737 unsigned needs_fsize : 1;
740 static const struct io_op_def io_op_defs[] = {
741 [IORING_OP_NOP] = {},
742 [IORING_OP_READV] = {
746 .unbound_nonreg_file = 1,
750 [IORING_OP_WRITEV] = {
755 .unbound_nonreg_file = 1,
759 [IORING_OP_FSYNC] = {
762 [IORING_OP_READ_FIXED] = {
764 .unbound_nonreg_file = 1,
767 [IORING_OP_WRITE_FIXED] = {
770 .unbound_nonreg_file = 1,
774 [IORING_OP_POLL_ADD] = {
776 .unbound_nonreg_file = 1,
778 [IORING_OP_POLL_REMOVE] = {},
779 [IORING_OP_SYNC_FILE_RANGE] = {
782 [IORING_OP_SENDMSG] = {
786 .unbound_nonreg_file = 1,
790 [IORING_OP_RECVMSG] = {
794 .unbound_nonreg_file = 1,
799 [IORING_OP_TIMEOUT] = {
803 [IORING_OP_TIMEOUT_REMOVE] = {},
804 [IORING_OP_ACCEPT] = {
807 .unbound_nonreg_file = 1,
811 [IORING_OP_ASYNC_CANCEL] = {},
812 [IORING_OP_LINK_TIMEOUT] = {
816 [IORING_OP_CONNECT] = {
820 .unbound_nonreg_file = 1,
823 [IORING_OP_FALLOCATE] = {
827 [IORING_OP_OPENAT] = {
831 [IORING_OP_CLOSE] = {
833 .needs_file_no_error = 1,
836 [IORING_OP_FILES_UPDATE] = {
840 [IORING_OP_STATX] = {
848 .unbound_nonreg_file = 1,
852 [IORING_OP_WRITE] = {
855 .unbound_nonreg_file = 1,
859 [IORING_OP_FADVISE] = {
862 [IORING_OP_MADVISE] = {
868 .unbound_nonreg_file = 1,
874 .unbound_nonreg_file = 1,
878 [IORING_OP_OPENAT2] = {
882 [IORING_OP_EPOLL_CTL] = {
883 .unbound_nonreg_file = 1,
886 [IORING_OP_SPLICE] = {
889 .unbound_nonreg_file = 1,
891 [IORING_OP_PROVIDE_BUFFERS] = {},
892 [IORING_OP_REMOVE_BUFFERS] = {},
896 .unbound_nonreg_file = 1,
900 enum io_mem_account {
905 static bool io_rw_reissue(struct io_kiocb *req, long res);
906 static void io_cqring_fill_event(struct io_kiocb *req, long res);
907 static void io_put_req(struct io_kiocb *req);
908 static void io_double_put_req(struct io_kiocb *req);
909 static void __io_double_put_req(struct io_kiocb *req);
910 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
911 static void io_queue_linked_timeout(struct io_kiocb *req);
912 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
913 struct io_uring_files_update *ip,
915 static int io_grab_files(struct io_kiocb *req);
916 static void io_complete_rw_common(struct kiocb *kiocb, long res,
917 struct io_comp_state *cs);
918 static void __io_clean_op(struct io_kiocb *req);
919 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
920 int fd, struct file **out_file, bool fixed);
921 static void __io_queue_sqe(struct io_kiocb *req,
922 const struct io_uring_sqe *sqe,
923 struct io_comp_state *cs);
924 static void io_file_put_work(struct work_struct *work);
926 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
927 struct iovec **iovec, struct iov_iter *iter,
929 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
930 struct iovec *iovec, struct iovec *fast_iov,
931 struct iov_iter *iter);
933 static struct kmem_cache *req_cachep;
935 static const struct file_operations io_uring_fops;
937 struct sock *io_uring_get_socket(struct file *file)
939 #if defined(CONFIG_UNIX)
940 if (file->f_op == &io_uring_fops) {
941 struct io_ring_ctx *ctx = file->private_data;
943 return ctx->ring_sock->sk;
948 EXPORT_SYMBOL(io_uring_get_socket);
950 static void io_get_req_task(struct io_kiocb *req)
952 if (req->flags & REQ_F_TASK_PINNED)
954 get_task_struct(req->task);
955 req->flags |= REQ_F_TASK_PINNED;
958 static inline void io_clean_op(struct io_kiocb *req)
960 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
964 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
965 static void __io_put_req_task(struct io_kiocb *req)
967 if (req->flags & REQ_F_TASK_PINNED)
968 put_task_struct(req->task);
971 static void io_sq_thread_drop_mm(void)
973 struct mm_struct *mm = current->mm;
976 kthread_unuse_mm(mm);
981 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
984 if (unlikely(!ctx->sqo_mm || !mmget_not_zero(ctx->sqo_mm)))
986 kthread_use_mm(ctx->sqo_mm);
992 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
993 struct io_kiocb *req)
995 if (!io_op_defs[req->opcode].needs_mm)
997 return __io_sq_thread_acquire_mm(ctx);
1000 static inline void req_set_fail_links(struct io_kiocb *req)
1002 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1003 req->flags |= REQ_F_FAIL_LINK;
1007 * Note: must call io_req_init_async() for the first time you
1008 * touch any members of io_wq_work.
1010 static inline void io_req_init_async(struct io_kiocb *req)
1012 if (req->flags & REQ_F_WORK_INITIALIZED)
1015 memset(&req->work, 0, sizeof(req->work));
1016 req->flags |= REQ_F_WORK_INITIALIZED;
1019 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1021 return ctx->flags & IORING_SETUP_SQPOLL;
1024 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1026 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1028 complete(&ctx->ref_comp);
1031 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1033 return !req->timeout.off;
1036 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1038 struct io_ring_ctx *ctx;
1041 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1045 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1046 if (!ctx->fallback_req)
1050 * Use 5 bits less than the max cq entries, that should give us around
1051 * 32 entries per hash list if totally full and uniformly spread.
1053 hash_bits = ilog2(p->cq_entries);
1057 ctx->cancel_hash_bits = hash_bits;
1058 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1060 if (!ctx->cancel_hash)
1062 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1064 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1065 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1068 ctx->flags = p->flags;
1069 init_waitqueue_head(&ctx->sqo_wait);
1070 init_waitqueue_head(&ctx->cq_wait);
1071 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1072 init_completion(&ctx->ref_comp);
1073 init_completion(&ctx->sq_thread_comp);
1074 idr_init(&ctx->io_buffer_idr);
1075 idr_init(&ctx->personality_idr);
1076 mutex_init(&ctx->uring_lock);
1077 init_waitqueue_head(&ctx->wait);
1078 spin_lock_init(&ctx->completion_lock);
1079 INIT_LIST_HEAD(&ctx->iopoll_list);
1080 INIT_LIST_HEAD(&ctx->defer_list);
1081 INIT_LIST_HEAD(&ctx->timeout_list);
1082 init_waitqueue_head(&ctx->inflight_wait);
1083 spin_lock_init(&ctx->inflight_lock);
1084 INIT_LIST_HEAD(&ctx->inflight_list);
1085 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1086 init_llist_head(&ctx->file_put_llist);
1089 if (ctx->fallback_req)
1090 kmem_cache_free(req_cachep, ctx->fallback_req);
1091 kfree(ctx->cancel_hash);
1096 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1098 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1099 struct io_ring_ctx *ctx = req->ctx;
1101 return seq != ctx->cached_cq_tail
1102 + atomic_read(&ctx->cached_cq_overflow);
1108 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1110 struct io_rings *rings = ctx->rings;
1112 /* order cqe stores with ring update */
1113 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1115 if (wq_has_sleeper(&ctx->cq_wait)) {
1116 wake_up_interruptible(&ctx->cq_wait);
1117 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1121 static void io_req_clean_work(struct io_kiocb *req)
1123 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1127 mmdrop(req->work.mm);
1128 req->work.mm = NULL;
1130 if (req->work.creds) {
1131 put_cred(req->work.creds);
1132 req->work.creds = NULL;
1135 struct fs_struct *fs = req->work.fs;
1137 spin_lock(&req->work.fs->lock);
1140 spin_unlock(&req->work.fs->lock);
1146 static void io_prep_async_work(struct io_kiocb *req)
1148 const struct io_op_def *def = &io_op_defs[req->opcode];
1150 io_req_init_async(req);
1152 if (req->flags & REQ_F_ISREG) {
1153 if (def->hash_reg_file)
1154 io_wq_hash_work(&req->work, file_inode(req->file));
1156 if (def->unbound_nonreg_file)
1157 req->work.flags |= IO_WQ_WORK_UNBOUND;
1159 if (!req->work.mm && def->needs_mm) {
1160 mmgrab(current->mm);
1161 req->work.mm = current->mm;
1163 if (!req->work.creds)
1164 req->work.creds = get_current_cred();
1165 if (!req->work.fs && def->needs_fs) {
1166 spin_lock(¤t->fs->lock);
1167 if (!current->fs->in_exec) {
1168 req->work.fs = current->fs;
1169 req->work.fs->users++;
1171 req->work.flags |= IO_WQ_WORK_CANCEL;
1173 spin_unlock(¤t->fs->lock);
1175 if (def->needs_fsize)
1176 req->work.fsize = rlimit(RLIMIT_FSIZE);
1178 req->work.fsize = RLIM_INFINITY;
1181 static void io_prep_async_link(struct io_kiocb *req)
1183 struct io_kiocb *cur;
1185 io_prep_async_work(req);
1186 if (req->flags & REQ_F_LINK_HEAD)
1187 list_for_each_entry(cur, &req->link_list, link_list)
1188 io_prep_async_work(cur);
1191 static void __io_queue_async_work(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1194 struct io_kiocb *link = io_prep_linked_timeout(req);
1196 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1197 &req->work, req->flags);
1198 io_wq_enqueue(ctx->io_wq, &req->work);
1201 io_queue_linked_timeout(link);
1204 static void io_queue_async_work(struct io_kiocb *req)
1206 /* init ->work of the whole link before punting */
1207 io_prep_async_link(req);
1208 __io_queue_async_work(req);
1211 static void io_kill_timeout(struct io_kiocb *req)
1215 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1217 atomic_inc(&req->ctx->cq_timeouts);
1218 list_del_init(&req->timeout.list);
1219 req->flags |= REQ_F_COMP_LOCKED;
1220 io_cqring_fill_event(req, 0);
1225 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1227 struct io_kiocb *req, *tmp;
1229 spin_lock_irq(&ctx->completion_lock);
1230 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1231 io_kill_timeout(req);
1232 spin_unlock_irq(&ctx->completion_lock);
1235 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1238 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1239 struct io_defer_entry, list);
1241 if (req_need_defer(de->req, de->seq))
1243 list_del_init(&de->list);
1244 /* punt-init is done before queueing for defer */
1245 __io_queue_async_work(de->req);
1247 } while (!list_empty(&ctx->defer_list));
1250 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1252 while (!list_empty(&ctx->timeout_list)) {
1253 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1254 struct io_kiocb, timeout.list);
1256 if (io_is_timeout_noseq(req))
1258 if (req->timeout.target_seq != ctx->cached_cq_tail
1259 - atomic_read(&ctx->cq_timeouts))
1262 list_del_init(&req->timeout.list);
1263 io_kill_timeout(req);
1267 static void io_commit_cqring(struct io_ring_ctx *ctx)
1269 io_flush_timeouts(ctx);
1270 __io_commit_cqring(ctx);
1272 if (unlikely(!list_empty(&ctx->defer_list)))
1273 __io_queue_deferred(ctx);
1276 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1278 struct io_rings *rings = ctx->rings;
1281 tail = ctx->cached_cq_tail;
1283 * writes to the cq entry need to come after reading head; the
1284 * control dependency is enough as we're using WRITE_ONCE to
1287 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1290 ctx->cached_cq_tail++;
1291 return &rings->cqes[tail & ctx->cq_mask];
1294 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1298 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1300 if (!ctx->eventfd_async)
1302 return io_wq_current_is_worker();
1305 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1307 if (waitqueue_active(&ctx->wait))
1308 wake_up(&ctx->wait);
1309 if (waitqueue_active(&ctx->sqo_wait))
1310 wake_up(&ctx->sqo_wait);
1311 if (io_should_trigger_evfd(ctx))
1312 eventfd_signal(ctx->cq_ev_fd, 1);
1315 /* Returns true if there are no backlogged entries after the flush */
1316 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1318 struct io_rings *rings = ctx->rings;
1319 struct io_uring_cqe *cqe;
1320 struct io_kiocb *req;
1321 unsigned long flags;
1325 if (list_empty_careful(&ctx->cq_overflow_list))
1327 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1328 rings->cq_ring_entries))
1332 spin_lock_irqsave(&ctx->completion_lock, flags);
1334 /* if force is set, the ring is going away. always drop after that */
1336 ctx->cq_overflow_flushed = 1;
1339 while (!list_empty(&ctx->cq_overflow_list)) {
1340 cqe = io_get_cqring(ctx);
1344 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1346 list_move(&req->compl.list, &list);
1347 req->flags &= ~REQ_F_OVERFLOW;
1349 WRITE_ONCE(cqe->user_data, req->user_data);
1350 WRITE_ONCE(cqe->res, req->result);
1351 WRITE_ONCE(cqe->flags, req->compl.cflags);
1353 WRITE_ONCE(ctx->rings->cq_overflow,
1354 atomic_inc_return(&ctx->cached_cq_overflow));
1358 io_commit_cqring(ctx);
1360 clear_bit(0, &ctx->sq_check_overflow);
1361 clear_bit(0, &ctx->cq_check_overflow);
1362 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1364 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1365 io_cqring_ev_posted(ctx);
1367 while (!list_empty(&list)) {
1368 req = list_first_entry(&list, struct io_kiocb, compl.list);
1369 list_del(&req->compl.list);
1376 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1378 struct io_ring_ctx *ctx = req->ctx;
1379 struct io_uring_cqe *cqe;
1381 trace_io_uring_complete(ctx, req->user_data, res);
1384 * If we can't get a cq entry, userspace overflowed the
1385 * submission (by quite a lot). Increment the overflow count in
1388 cqe = io_get_cqring(ctx);
1390 WRITE_ONCE(cqe->user_data, req->user_data);
1391 WRITE_ONCE(cqe->res, res);
1392 WRITE_ONCE(cqe->flags, cflags);
1393 } else if (ctx->cq_overflow_flushed) {
1394 WRITE_ONCE(ctx->rings->cq_overflow,
1395 atomic_inc_return(&ctx->cached_cq_overflow));
1397 if (list_empty(&ctx->cq_overflow_list)) {
1398 set_bit(0, &ctx->sq_check_overflow);
1399 set_bit(0, &ctx->cq_check_overflow);
1400 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1403 req->flags |= REQ_F_OVERFLOW;
1405 req->compl.cflags = cflags;
1406 refcount_inc(&req->refs);
1407 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1411 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1413 __io_cqring_fill_event(req, res, 0);
1416 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1418 struct io_ring_ctx *ctx = req->ctx;
1419 unsigned long flags;
1421 spin_lock_irqsave(&ctx->completion_lock, flags);
1422 __io_cqring_fill_event(req, res, cflags);
1423 io_commit_cqring(ctx);
1424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1426 io_cqring_ev_posted(ctx);
1429 static void io_submit_flush_completions(struct io_comp_state *cs)
1431 struct io_ring_ctx *ctx = cs->ctx;
1433 spin_lock_irq(&ctx->completion_lock);
1434 while (!list_empty(&cs->list)) {
1435 struct io_kiocb *req;
1437 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1438 list_del(&req->compl.list);
1439 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1440 if (!(req->flags & REQ_F_LINK_HEAD)) {
1441 req->flags |= REQ_F_COMP_LOCKED;
1444 spin_unlock_irq(&ctx->completion_lock);
1446 spin_lock_irq(&ctx->completion_lock);
1449 io_commit_cqring(ctx);
1450 spin_unlock_irq(&ctx->completion_lock);
1452 io_cqring_ev_posted(ctx);
1456 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1457 struct io_comp_state *cs)
1460 io_cqring_add_event(req, res, cflags);
1465 req->compl.cflags = cflags;
1466 list_add_tail(&req->compl.list, &cs->list);
1468 io_submit_flush_completions(cs);
1472 static void io_req_complete(struct io_kiocb *req, long res)
1474 __io_req_complete(req, res, 0, NULL);
1477 static inline bool io_is_fallback_req(struct io_kiocb *req)
1479 return req == (struct io_kiocb *)
1480 ((unsigned long) req->ctx->fallback_req & ~1UL);
1483 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1485 struct io_kiocb *req;
1487 req = ctx->fallback_req;
1488 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1494 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1495 struct io_submit_state *state)
1497 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1498 struct io_kiocb *req;
1500 if (!state->free_reqs) {
1504 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1505 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1508 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1509 * retry single alloc to be on the safe side.
1511 if (unlikely(ret <= 0)) {
1512 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1513 if (!state->reqs[0])
1517 state->free_reqs = ret - 1;
1518 req = state->reqs[ret - 1];
1521 req = state->reqs[state->free_reqs];
1526 return io_get_fallback_req(ctx);
1529 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1533 percpu_ref_put(req->fixed_file_refs);
1538 static void io_dismantle_req(struct io_kiocb *req)
1545 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1546 __io_put_req_task(req);
1547 io_req_clean_work(req);
1549 if (req->flags & REQ_F_INFLIGHT) {
1550 struct io_ring_ctx *ctx = req->ctx;
1551 unsigned long flags;
1553 spin_lock_irqsave(&ctx->inflight_lock, flags);
1554 list_del(&req->inflight_entry);
1555 if (waitqueue_active(&ctx->inflight_wait))
1556 wake_up(&ctx->inflight_wait);
1557 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1561 static void __io_free_req(struct io_kiocb *req)
1563 struct io_ring_ctx *ctx;
1565 io_dismantle_req(req);
1567 if (likely(!io_is_fallback_req(req)))
1568 kmem_cache_free(req_cachep, req);
1570 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1571 percpu_ref_put(&ctx->refs);
1574 static bool io_link_cancel_timeout(struct io_kiocb *req)
1576 struct io_ring_ctx *ctx = req->ctx;
1579 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1581 io_cqring_fill_event(req, -ECANCELED);
1582 io_commit_cqring(ctx);
1583 req->flags &= ~REQ_F_LINK_HEAD;
1591 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1593 struct io_kiocb *link;
1596 if (list_empty(&req->link_list))
1598 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1599 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1602 list_del_init(&link->link_list);
1603 wake_ev = io_link_cancel_timeout(link);
1604 req->flags &= ~REQ_F_LINK_TIMEOUT;
1608 static void io_kill_linked_timeout(struct io_kiocb *req)
1610 struct io_ring_ctx *ctx = req->ctx;
1613 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1614 unsigned long flags;
1616 spin_lock_irqsave(&ctx->completion_lock, flags);
1617 wake_ev = __io_kill_linked_timeout(req);
1618 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1620 wake_ev = __io_kill_linked_timeout(req);
1624 io_cqring_ev_posted(ctx);
1627 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1629 struct io_kiocb *nxt;
1632 * The list should never be empty when we are called here. But could
1633 * potentially happen if the chain is messed up, check to be on the
1636 if (unlikely(list_empty(&req->link_list)))
1639 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1640 list_del_init(&req->link_list);
1641 if (!list_empty(&nxt->link_list))
1642 nxt->flags |= REQ_F_LINK_HEAD;
1647 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1649 static void __io_fail_links(struct io_kiocb *req)
1651 struct io_ring_ctx *ctx = req->ctx;
1653 while (!list_empty(&req->link_list)) {
1654 struct io_kiocb *link = list_first_entry(&req->link_list,
1655 struct io_kiocb, link_list);
1657 list_del_init(&link->link_list);
1658 trace_io_uring_fail_link(req, link);
1660 io_cqring_fill_event(link, -ECANCELED);
1661 __io_double_put_req(link);
1662 req->flags &= ~REQ_F_LINK_TIMEOUT;
1665 io_commit_cqring(ctx);
1666 io_cqring_ev_posted(ctx);
1669 static void io_fail_links(struct io_kiocb *req)
1671 struct io_ring_ctx *ctx = req->ctx;
1673 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1674 unsigned long flags;
1676 spin_lock_irqsave(&ctx->completion_lock, flags);
1677 __io_fail_links(req);
1678 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1680 __io_fail_links(req);
1683 io_cqring_ev_posted(ctx);
1686 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1688 req->flags &= ~REQ_F_LINK_HEAD;
1689 if (req->flags & REQ_F_LINK_TIMEOUT)
1690 io_kill_linked_timeout(req);
1693 * If LINK is set, we have dependent requests in this chain. If we
1694 * didn't fail this request, queue the first one up, moving any other
1695 * dependencies to the next request. In case of failure, fail the rest
1698 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1699 return io_req_link_next(req);
1704 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1706 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1708 return __io_req_find_next(req);
1711 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1713 struct task_struct *tsk = req->task;
1714 struct io_ring_ctx *ctx = req->ctx;
1715 int ret, notify = TWA_RESUME;
1718 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1719 * If we're not using an eventfd, then TWA_RESUME is always fine,
1720 * as we won't have dependencies between request completions for
1721 * other kernel wait conditions.
1723 if (ctx->flags & IORING_SETUP_SQPOLL)
1725 else if (ctx->cq_ev_fd)
1726 notify = TWA_SIGNAL;
1728 ret = task_work_add(tsk, cb, notify);
1730 wake_up_process(tsk);
1734 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1736 struct io_ring_ctx *ctx = req->ctx;
1738 spin_lock_irq(&ctx->completion_lock);
1739 io_cqring_fill_event(req, error);
1740 io_commit_cqring(ctx);
1741 spin_unlock_irq(&ctx->completion_lock);
1743 io_cqring_ev_posted(ctx);
1744 req_set_fail_links(req);
1745 io_double_put_req(req);
1748 static void io_req_task_cancel(struct callback_head *cb)
1750 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1752 __io_req_task_cancel(req, -ECANCELED);
1755 static void __io_req_task_submit(struct io_kiocb *req)
1757 struct io_ring_ctx *ctx = req->ctx;
1759 if (!__io_sq_thread_acquire_mm(ctx)) {
1760 mutex_lock(&ctx->uring_lock);
1761 __io_queue_sqe(req, NULL, NULL);
1762 mutex_unlock(&ctx->uring_lock);
1764 __io_req_task_cancel(req, -EFAULT);
1768 static void io_req_task_submit(struct callback_head *cb)
1770 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1772 __io_req_task_submit(req);
1775 static void io_req_task_queue(struct io_kiocb *req)
1779 init_task_work(&req->task_work, io_req_task_submit);
1781 ret = io_req_task_work_add(req, &req->task_work);
1782 if (unlikely(ret)) {
1783 struct task_struct *tsk;
1785 init_task_work(&req->task_work, io_req_task_cancel);
1786 tsk = io_wq_get_task(req->ctx->io_wq);
1787 task_work_add(tsk, &req->task_work, 0);
1788 wake_up_process(tsk);
1792 static void io_queue_next(struct io_kiocb *req)
1794 struct io_kiocb *nxt = io_req_find_next(req);
1797 io_req_task_queue(nxt);
1800 static void io_free_req(struct io_kiocb *req)
1807 void *reqs[IO_IOPOLL_BATCH];
1811 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1812 struct req_batch *rb)
1814 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1815 percpu_ref_put_many(&ctx->refs, rb->to_free);
1819 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1820 struct req_batch *rb)
1823 __io_req_free_batch_flush(ctx, rb);
1826 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1828 if (unlikely(io_is_fallback_req(req))) {
1832 if (req->flags & REQ_F_LINK_HEAD)
1835 io_dismantle_req(req);
1836 rb->reqs[rb->to_free++] = req;
1837 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1838 __io_req_free_batch_flush(req->ctx, rb);
1842 * Drop reference to request, return next in chain (if there is one) if this
1843 * was the last reference to this request.
1845 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1847 struct io_kiocb *nxt = NULL;
1849 if (refcount_dec_and_test(&req->refs)) {
1850 nxt = io_req_find_next(req);
1856 static void io_put_req(struct io_kiocb *req)
1858 if (refcount_dec_and_test(&req->refs))
1862 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1864 struct io_kiocb *nxt;
1867 * A ref is owned by io-wq in which context we're. So, if that's the
1868 * last one, it's safe to steal next work. False negatives are Ok,
1869 * it just will be re-punted async in io_put_work()
1871 if (refcount_read(&req->refs) != 1)
1874 nxt = io_req_find_next(req);
1875 return nxt ? &nxt->work : NULL;
1879 * Must only be used if we don't need to care about links, usually from
1880 * within the completion handling itself.
1882 static void __io_double_put_req(struct io_kiocb *req)
1884 /* drop both submit and complete references */
1885 if (refcount_sub_and_test(2, &req->refs))
1889 static void io_double_put_req(struct io_kiocb *req)
1891 /* drop both submit and complete references */
1892 if (refcount_sub_and_test(2, &req->refs))
1896 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1898 struct io_rings *rings = ctx->rings;
1900 if (test_bit(0, &ctx->cq_check_overflow)) {
1902 * noflush == true is from the waitqueue handler, just ensure
1903 * we wake up the task, and the next invocation will flush the
1904 * entries. We cannot safely to it from here.
1906 if (noflush && !list_empty(&ctx->cq_overflow_list))
1909 io_cqring_overflow_flush(ctx, false);
1912 /* See comment at the top of this file */
1914 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1917 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1919 struct io_rings *rings = ctx->rings;
1921 /* make sure SQ entry isn't read before tail */
1922 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1925 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1927 unsigned int cflags;
1929 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1930 cflags |= IORING_CQE_F_BUFFER;
1931 req->flags &= ~REQ_F_BUFFER_SELECTED;
1936 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
1938 struct io_buffer *kbuf;
1940 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1941 return io_put_kbuf(req, kbuf);
1944 static inline bool io_run_task_work(void)
1946 if (current->task_works) {
1947 __set_current_state(TASK_RUNNING);
1955 static void io_iopoll_queue(struct list_head *again)
1957 struct io_kiocb *req;
1960 req = list_first_entry(again, struct io_kiocb, inflight_entry);
1961 list_del(&req->inflight_entry);
1962 if (!io_rw_reissue(req, -EAGAIN))
1963 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1964 } while (!list_empty(again));
1968 * Find and free completed poll iocbs
1970 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1971 struct list_head *done)
1973 struct req_batch rb;
1974 struct io_kiocb *req;
1977 /* order with ->result store in io_complete_rw_iopoll() */
1981 while (!list_empty(done)) {
1984 req = list_first_entry(done, struct io_kiocb, inflight_entry);
1985 if (READ_ONCE(req->result) == -EAGAIN) {
1986 req->iopoll_completed = 0;
1987 list_move_tail(&req->inflight_entry, &again);
1990 list_del(&req->inflight_entry);
1992 if (req->flags & REQ_F_BUFFER_SELECTED)
1993 cflags = io_put_rw_kbuf(req);
1995 __io_cqring_fill_event(req, req->result, cflags);
1998 if (refcount_dec_and_test(&req->refs))
1999 io_req_free_batch(&rb, req);
2002 io_commit_cqring(ctx);
2003 if (ctx->flags & IORING_SETUP_SQPOLL)
2004 io_cqring_ev_posted(ctx);
2005 io_req_free_batch_finish(ctx, &rb);
2007 if (!list_empty(&again))
2008 io_iopoll_queue(&again);
2011 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2014 struct io_kiocb *req, *tmp;
2020 * Only spin for completions if we don't have multiple devices hanging
2021 * off our complete list, and we're under the requested amount.
2023 spin = !ctx->poll_multi_file && *nr_events < min;
2026 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2027 struct kiocb *kiocb = &req->rw.kiocb;
2030 * Move completed and retryable entries to our local lists.
2031 * If we find a request that requires polling, break out
2032 * and complete those lists first, if we have entries there.
2034 if (READ_ONCE(req->iopoll_completed)) {
2035 list_move_tail(&req->inflight_entry, &done);
2038 if (!list_empty(&done))
2041 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2045 /* iopoll may have completed current req */
2046 if (READ_ONCE(req->iopoll_completed))
2047 list_move_tail(&req->inflight_entry, &done);
2054 if (!list_empty(&done))
2055 io_iopoll_complete(ctx, nr_events, &done);
2061 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2062 * non-spinning poll check - we'll still enter the driver poll loop, but only
2063 * as a non-spinning completion check.
2065 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2068 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2071 ret = io_do_iopoll(ctx, nr_events, min);
2074 if (*nr_events >= min)
2082 * We can't just wait for polled events to come to us, we have to actively
2083 * find and complete them.
2085 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2087 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2090 mutex_lock(&ctx->uring_lock);
2091 while (!list_empty(&ctx->iopoll_list)) {
2092 unsigned int nr_events = 0;
2094 io_do_iopoll(ctx, &nr_events, 0);
2096 /* let it sleep and repeat later if can't complete a request */
2100 * Ensure we allow local-to-the-cpu processing to take place,
2101 * in this case we need to ensure that we reap all events.
2102 * Also let task_work, etc. to progress by releasing the mutex
2104 if (need_resched()) {
2105 mutex_unlock(&ctx->uring_lock);
2107 mutex_lock(&ctx->uring_lock);
2110 mutex_unlock(&ctx->uring_lock);
2113 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2115 unsigned int nr_events = 0;
2116 int iters = 0, ret = 0;
2119 * We disallow the app entering submit/complete with polling, but we
2120 * still need to lock the ring to prevent racing with polled issue
2121 * that got punted to a workqueue.
2123 mutex_lock(&ctx->uring_lock);
2126 * Don't enter poll loop if we already have events pending.
2127 * If we do, we can potentially be spinning for commands that
2128 * already triggered a CQE (eg in error).
2130 if (io_cqring_events(ctx, false))
2134 * If a submit got punted to a workqueue, we can have the
2135 * application entering polling for a command before it gets
2136 * issued. That app will hold the uring_lock for the duration
2137 * of the poll right here, so we need to take a breather every
2138 * now and then to ensure that the issue has a chance to add
2139 * the poll to the issued list. Otherwise we can spin here
2140 * forever, while the workqueue is stuck trying to acquire the
2143 if (!(++iters & 7)) {
2144 mutex_unlock(&ctx->uring_lock);
2146 mutex_lock(&ctx->uring_lock);
2149 ret = io_iopoll_getevents(ctx, &nr_events, min);
2153 } while (min && !nr_events && !need_resched());
2155 mutex_unlock(&ctx->uring_lock);
2159 static void kiocb_end_write(struct io_kiocb *req)
2162 * Tell lockdep we inherited freeze protection from submission
2165 if (req->flags & REQ_F_ISREG) {
2166 struct inode *inode = file_inode(req->file);
2168 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2170 file_end_write(req->file);
2173 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2174 struct io_comp_state *cs)
2176 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2179 if (kiocb->ki_flags & IOCB_WRITE)
2180 kiocb_end_write(req);
2182 if (res != req->result)
2183 req_set_fail_links(req);
2184 if (req->flags & REQ_F_BUFFER_SELECTED)
2185 cflags = io_put_rw_kbuf(req);
2186 __io_req_complete(req, res, cflags, cs);
2190 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2192 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2193 ssize_t ret = -ECANCELED;
2194 struct iov_iter iter;
2202 switch (req->opcode) {
2203 case IORING_OP_READV:
2204 case IORING_OP_READ_FIXED:
2205 case IORING_OP_READ:
2208 case IORING_OP_WRITEV:
2209 case IORING_OP_WRITE_FIXED:
2210 case IORING_OP_WRITE:
2214 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2219 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2222 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2227 req_set_fail_links(req);
2228 io_req_complete(req, ret);
2232 static void io_rw_resubmit(struct callback_head *cb)
2234 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2235 struct io_ring_ctx *ctx = req->ctx;
2238 err = io_sq_thread_acquire_mm(ctx, req);
2240 if (io_resubmit_prep(req, err)) {
2241 refcount_inc(&req->refs);
2242 io_queue_async_work(req);
2247 static bool io_rw_reissue(struct io_kiocb *req, long res)
2252 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2255 init_task_work(&req->task_work, io_rw_resubmit);
2256 ret = io_req_task_work_add(req, &req->task_work);
2263 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2264 struct io_comp_state *cs)
2266 if (!io_rw_reissue(req, res))
2267 io_complete_rw_common(&req->rw.kiocb, res, cs);
2270 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2272 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2274 __io_complete_rw(req, res, res2, NULL);
2277 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2279 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2281 if (kiocb->ki_flags & IOCB_WRITE)
2282 kiocb_end_write(req);
2284 if (res != -EAGAIN && res != req->result)
2285 req_set_fail_links(req);
2287 WRITE_ONCE(req->result, res);
2288 /* order with io_poll_complete() checking ->result */
2290 WRITE_ONCE(req->iopoll_completed, 1);
2294 * After the iocb has been issued, it's safe to be found on the poll list.
2295 * Adding the kiocb to the list AFTER submission ensures that we don't
2296 * find it from a io_iopoll_getevents() thread before the issuer is done
2297 * accessing the kiocb cookie.
2299 static void io_iopoll_req_issued(struct io_kiocb *req)
2301 struct io_ring_ctx *ctx = req->ctx;
2304 * Track whether we have multiple files in our lists. This will impact
2305 * how we do polling eventually, not spinning if we're on potentially
2306 * different devices.
2308 if (list_empty(&ctx->iopoll_list)) {
2309 ctx->poll_multi_file = false;
2310 } else if (!ctx->poll_multi_file) {
2311 struct io_kiocb *list_req;
2313 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2315 if (list_req->file != req->file)
2316 ctx->poll_multi_file = true;
2320 * For fast devices, IO may have already completed. If it has, add
2321 * it to the front so we find it first.
2323 if (READ_ONCE(req->iopoll_completed))
2324 list_add(&req->inflight_entry, &ctx->iopoll_list);
2326 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2328 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2329 wq_has_sleeper(&ctx->sqo_wait))
2330 wake_up(&ctx->sqo_wait);
2333 static void __io_state_file_put(struct io_submit_state *state)
2335 if (state->has_refs)
2336 fput_many(state->file, state->has_refs);
2340 static inline void io_state_file_put(struct io_submit_state *state)
2343 __io_state_file_put(state);
2347 * Get as many references to a file as we have IOs left in this submission,
2348 * assuming most submissions are for one file, or at least that each file
2349 * has more than one submission.
2351 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2357 if (state->fd == fd) {
2362 __io_state_file_put(state);
2364 state->file = fget_many(fd, state->ios_left);
2370 state->has_refs = state->ios_left;
2374 static bool io_bdev_nowait(struct block_device *bdev)
2377 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2384 * If we tracked the file through the SCM inflight mechanism, we could support
2385 * any file. For now, just ensure that anything potentially problematic is done
2388 static bool io_file_supports_async(struct file *file, int rw)
2390 umode_t mode = file_inode(file)->i_mode;
2392 if (S_ISBLK(mode)) {
2393 if (io_bdev_nowait(file->f_inode->i_bdev))
2397 if (S_ISCHR(mode) || S_ISSOCK(mode))
2399 if (S_ISREG(mode)) {
2400 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2401 file->f_op != &io_uring_fops)
2406 /* any ->read/write should understand O_NONBLOCK */
2407 if (file->f_flags & O_NONBLOCK)
2410 if (!(file->f_mode & FMODE_NOWAIT))
2414 return file->f_op->read_iter != NULL;
2416 return file->f_op->write_iter != NULL;
2419 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2420 bool force_nonblock)
2422 struct io_ring_ctx *ctx = req->ctx;
2423 struct kiocb *kiocb = &req->rw.kiocb;
2427 if (S_ISREG(file_inode(req->file)->i_mode))
2428 req->flags |= REQ_F_ISREG;
2430 kiocb->ki_pos = READ_ONCE(sqe->off);
2431 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2432 req->flags |= REQ_F_CUR_POS;
2433 kiocb->ki_pos = req->file->f_pos;
2435 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2436 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2437 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2441 ioprio = READ_ONCE(sqe->ioprio);
2443 ret = ioprio_check_cap(ioprio);
2447 kiocb->ki_ioprio = ioprio;
2449 kiocb->ki_ioprio = get_current_ioprio();
2451 /* don't allow async punt if RWF_NOWAIT was requested */
2452 if (kiocb->ki_flags & IOCB_NOWAIT)
2453 req->flags |= REQ_F_NOWAIT;
2455 if (kiocb->ki_flags & IOCB_DIRECT)
2456 io_get_req_task(req);
2459 kiocb->ki_flags |= IOCB_NOWAIT;
2461 if (ctx->flags & IORING_SETUP_IOPOLL) {
2462 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2463 !kiocb->ki_filp->f_op->iopoll)
2466 kiocb->ki_flags |= IOCB_HIPRI;
2467 kiocb->ki_complete = io_complete_rw_iopoll;
2468 req->iopoll_completed = 0;
2469 io_get_req_task(req);
2471 if (kiocb->ki_flags & IOCB_HIPRI)
2473 kiocb->ki_complete = io_complete_rw;
2476 req->rw.addr = READ_ONCE(sqe->addr);
2477 req->rw.len = READ_ONCE(sqe->len);
2478 req->buf_index = READ_ONCE(sqe->buf_index);
2482 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2488 case -ERESTARTNOINTR:
2489 case -ERESTARTNOHAND:
2490 case -ERESTART_RESTARTBLOCK:
2492 * We can't just restart the syscall, since previously
2493 * submitted sqes may already be in progress. Just fail this
2499 kiocb->ki_complete(kiocb, ret, 0);
2503 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2504 struct io_comp_state *cs)
2506 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2508 if (req->flags & REQ_F_CUR_POS)
2509 req->file->f_pos = kiocb->ki_pos;
2510 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2511 __io_complete_rw(req, ret, 0, cs);
2513 io_rw_done(kiocb, ret);
2516 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2517 struct iov_iter *iter)
2519 struct io_ring_ctx *ctx = req->ctx;
2520 size_t len = req->rw.len;
2521 struct io_mapped_ubuf *imu;
2522 u16 index, buf_index;
2526 /* attempt to use fixed buffers without having provided iovecs */
2527 if (unlikely(!ctx->user_bufs))
2530 buf_index = req->buf_index;
2531 if (unlikely(buf_index >= ctx->nr_user_bufs))
2534 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2535 imu = &ctx->user_bufs[index];
2536 buf_addr = req->rw.addr;
2539 if (buf_addr + len < buf_addr)
2541 /* not inside the mapped region */
2542 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2546 * May not be a start of buffer, set size appropriately
2547 * and advance us to the beginning.
2549 offset = buf_addr - imu->ubuf;
2550 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2554 * Don't use iov_iter_advance() here, as it's really slow for
2555 * using the latter parts of a big fixed buffer - it iterates
2556 * over each segment manually. We can cheat a bit here, because
2559 * 1) it's a BVEC iter, we set it up
2560 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2561 * first and last bvec
2563 * So just find our index, and adjust the iterator afterwards.
2564 * If the offset is within the first bvec (or the whole first
2565 * bvec, just use iov_iter_advance(). This makes it easier
2566 * since we can just skip the first segment, which may not
2567 * be PAGE_SIZE aligned.
2569 const struct bio_vec *bvec = imu->bvec;
2571 if (offset <= bvec->bv_len) {
2572 iov_iter_advance(iter, offset);
2574 unsigned long seg_skip;
2576 /* skip first vec */
2577 offset -= bvec->bv_len;
2578 seg_skip = 1 + (offset >> PAGE_SHIFT);
2580 iter->bvec = bvec + seg_skip;
2581 iter->nr_segs -= seg_skip;
2582 iter->count -= bvec->bv_len + offset;
2583 iter->iov_offset = offset & ~PAGE_MASK;
2590 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2593 mutex_unlock(&ctx->uring_lock);
2596 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2599 * "Normal" inline submissions always hold the uring_lock, since we
2600 * grab it from the system call. Same is true for the SQPOLL offload.
2601 * The only exception is when we've detached the request and issue it
2602 * from an async worker thread, grab the lock for that case.
2605 mutex_lock(&ctx->uring_lock);
2608 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2609 int bgid, struct io_buffer *kbuf,
2612 struct io_buffer *head;
2614 if (req->flags & REQ_F_BUFFER_SELECTED)
2617 io_ring_submit_lock(req->ctx, needs_lock);
2619 lockdep_assert_held(&req->ctx->uring_lock);
2621 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2623 if (!list_empty(&head->list)) {
2624 kbuf = list_last_entry(&head->list, struct io_buffer,
2626 list_del(&kbuf->list);
2629 idr_remove(&req->ctx->io_buffer_idr, bgid);
2631 if (*len > kbuf->len)
2634 kbuf = ERR_PTR(-ENOBUFS);
2637 io_ring_submit_unlock(req->ctx, needs_lock);
2642 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2645 struct io_buffer *kbuf;
2648 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2649 bgid = req->buf_index;
2650 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2653 req->rw.addr = (u64) (unsigned long) kbuf;
2654 req->flags |= REQ_F_BUFFER_SELECTED;
2655 return u64_to_user_ptr(kbuf->addr);
2658 #ifdef CONFIG_COMPAT
2659 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2662 struct compat_iovec __user *uiov;
2663 compat_ssize_t clen;
2667 uiov = u64_to_user_ptr(req->rw.addr);
2668 if (!access_ok(uiov, sizeof(*uiov)))
2670 if (__get_user(clen, &uiov->iov_len))
2676 buf = io_rw_buffer_select(req, &len, needs_lock);
2678 return PTR_ERR(buf);
2679 iov[0].iov_base = buf;
2680 iov[0].iov_len = (compat_size_t) len;
2685 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2688 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2692 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2695 len = iov[0].iov_len;
2698 buf = io_rw_buffer_select(req, &len, needs_lock);
2700 return PTR_ERR(buf);
2701 iov[0].iov_base = buf;
2702 iov[0].iov_len = len;
2706 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2709 if (req->flags & REQ_F_BUFFER_SELECTED) {
2710 struct io_buffer *kbuf;
2712 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2713 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2714 iov[0].iov_len = kbuf->len;
2719 else if (req->rw.len > 1)
2722 #ifdef CONFIG_COMPAT
2723 if (req->ctx->compat)
2724 return io_compat_import(req, iov, needs_lock);
2727 return __io_iov_buffer_select(req, iov, needs_lock);
2730 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2731 struct iovec **iovec, struct iov_iter *iter,
2734 void __user *buf = u64_to_user_ptr(req->rw.addr);
2735 size_t sqe_len = req->rw.len;
2739 opcode = req->opcode;
2740 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2742 return io_import_fixed(req, rw, iter);
2745 /* buffer index only valid with fixed read/write, or buffer select */
2746 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2749 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2750 if (req->flags & REQ_F_BUFFER_SELECT) {
2751 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2754 return PTR_ERR(buf);
2756 req->rw.len = sqe_len;
2759 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2761 return ret < 0 ? ret : sqe_len;
2765 struct io_async_rw *iorw = &req->io->rw;
2767 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2772 if (req->flags & REQ_F_BUFFER_SELECT) {
2773 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2775 ret = (*iovec)->iov_len;
2776 iov_iter_init(iter, rw, *iovec, 1, ret);
2782 #ifdef CONFIG_COMPAT
2783 if (req->ctx->compat)
2784 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2788 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2792 * For files that don't have ->read_iter() and ->write_iter(), handle them
2793 * by looping over ->read() or ->write() manually.
2795 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2796 struct iov_iter *iter)
2801 * Don't support polled IO through this interface, and we can't
2802 * support non-blocking either. For the latter, this just causes
2803 * the kiocb to be handled from an async context.
2805 if (kiocb->ki_flags & IOCB_HIPRI)
2807 if (kiocb->ki_flags & IOCB_NOWAIT)
2810 while (iov_iter_count(iter)) {
2814 if (!iov_iter_is_bvec(iter)) {
2815 iovec = iov_iter_iovec(iter);
2817 /* fixed buffers import bvec */
2818 iovec.iov_base = kmap(iter->bvec->bv_page)
2820 iovec.iov_len = min(iter->count,
2821 iter->bvec->bv_len - iter->iov_offset);
2825 nr = file->f_op->read(file, iovec.iov_base,
2826 iovec.iov_len, &kiocb->ki_pos);
2828 nr = file->f_op->write(file, iovec.iov_base,
2829 iovec.iov_len, &kiocb->ki_pos);
2832 if (iov_iter_is_bvec(iter))
2833 kunmap(iter->bvec->bv_page);
2841 if (nr != iovec.iov_len)
2843 iov_iter_advance(iter, nr);
2849 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2850 struct iovec *iovec, struct iovec *fast_iov,
2851 struct iov_iter *iter)
2853 struct io_async_rw *rw = &req->io->rw;
2855 rw->nr_segs = iter->nr_segs;
2858 rw->iov = rw->fast_iov;
2859 if (rw->iov != fast_iov)
2860 memcpy(rw->iov, fast_iov,
2861 sizeof(struct iovec) * iter->nr_segs);
2864 req->flags |= REQ_F_NEED_CLEANUP;
2868 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2870 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2871 return req->io == NULL;
2874 static int io_alloc_async_ctx(struct io_kiocb *req)
2876 if (!io_op_defs[req->opcode].async_ctx)
2879 return __io_alloc_async_ctx(req);
2882 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2883 struct iovec *iovec, struct iovec *fast_iov,
2884 struct iov_iter *iter)
2886 if (!io_op_defs[req->opcode].async_ctx)
2889 if (__io_alloc_async_ctx(req))
2892 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2897 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2898 bool force_nonblock)
2900 struct io_async_ctx *io = req->io;
2901 struct iov_iter iter;
2904 io->rw.iov = io->rw.fast_iov;
2906 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2908 if (unlikely(ret < 0))
2911 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2915 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2916 bool force_nonblock)
2920 ret = io_prep_rw(req, sqe, force_nonblock);
2924 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2927 /* either don't need iovec imported or already have it */
2928 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2930 return io_rw_prep_async(req, READ, force_nonblock);
2933 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2934 int sync, void *arg)
2936 struct wait_page_queue *wpq;
2937 struct io_kiocb *req = wait->private;
2938 struct wait_page_key *key = arg;
2941 wpq = container_of(wait, struct wait_page_queue, wait);
2943 ret = wake_page_match(wpq, key);
2947 list_del_init(&wait->entry);
2949 init_task_work(&req->task_work, io_req_task_submit);
2950 /* submit ref gets dropped, acquire a new one */
2951 refcount_inc(&req->refs);
2952 ret = io_req_task_work_add(req, &req->task_work);
2953 if (unlikely(ret)) {
2954 struct task_struct *tsk;
2956 /* queue just for cancelation */
2957 init_task_work(&req->task_work, io_req_task_cancel);
2958 tsk = io_wq_get_task(req->ctx->io_wq);
2959 task_work_add(tsk, &req->task_work, 0);
2960 wake_up_process(tsk);
2965 static bool io_rw_should_retry(struct io_kiocb *req)
2967 struct kiocb *kiocb = &req->rw.kiocb;
2970 /* never retry for NOWAIT, we just complete with -EAGAIN */
2971 if (req->flags & REQ_F_NOWAIT)
2974 /* already tried, or we're doing O_DIRECT */
2975 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2978 * just use poll if we can, and don't attempt if the fs doesn't
2979 * support callback based unlocks
2981 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2985 * If request type doesn't require req->io to defer in general,
2986 * we need to allocate it here
2988 if (!req->io && __io_alloc_async_ctx(req))
2991 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2992 io_async_buf_func, req);
2994 io_get_req_task(req);
3001 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3003 if (req->file->f_op->read_iter)
3004 return call_read_iter(req->file, &req->rw.kiocb, iter);
3005 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3008 static int io_read(struct io_kiocb *req, bool force_nonblock,
3009 struct io_comp_state *cs)
3011 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3012 struct kiocb *kiocb = &req->rw.kiocb;
3013 struct iov_iter iter;
3015 ssize_t io_size, ret;
3017 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3021 /* Ensure we clear previously set non-block flag */
3022 if (!force_nonblock)
3023 kiocb->ki_flags &= ~IOCB_NOWAIT;
3026 req->result = io_size;
3028 /* If the file doesn't support async, just async punt */
3029 if (force_nonblock && !io_file_supports_async(req->file, READ))
3032 iov_count = iov_iter_count(&iter);
3033 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3035 unsigned long nr_segs = iter.nr_segs;
3038 ret2 = io_iter_do_read(req, &iter);
3040 /* Catch -EAGAIN return for forced non-blocking submission */
3041 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3042 kiocb_done(kiocb, ret2, cs);
3044 iter.count = iov_count;
3045 iter.nr_segs = nr_segs;
3047 ret = io_setup_async_rw(req, io_size, iovec,
3048 inline_vecs, &iter);
3051 /* it's copied and will be cleaned with ->io */
3053 /* if we can retry, do so with the callbacks armed */
3054 if (io_rw_should_retry(req)) {
3055 ret2 = io_iter_do_read(req, &iter);
3056 if (ret2 == -EIOCBQUEUED) {
3058 } else if (ret2 != -EAGAIN) {
3059 kiocb_done(kiocb, ret2, cs);
3063 kiocb->ki_flags &= ~IOCB_WAITQ;
3073 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3074 bool force_nonblock)
3078 ret = io_prep_rw(req, sqe, force_nonblock);
3082 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3085 /* either don't need iovec imported or already have it */
3086 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3088 return io_rw_prep_async(req, WRITE, force_nonblock);
3091 static int io_write(struct io_kiocb *req, bool force_nonblock,
3092 struct io_comp_state *cs)
3094 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3095 struct kiocb *kiocb = &req->rw.kiocb;
3096 struct iov_iter iter;
3098 ssize_t ret, io_size;
3100 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3104 /* Ensure we clear previously set non-block flag */
3105 if (!force_nonblock)
3106 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3109 req->result = io_size;
3111 /* If the file doesn't support async, just async punt */
3112 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3115 /* file path doesn't support NOWAIT for non-direct_IO */
3116 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3117 (req->flags & REQ_F_ISREG))
3120 iov_count = iov_iter_count(&iter);
3121 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3123 unsigned long nr_segs = iter.nr_segs;
3127 * Open-code file_start_write here to grab freeze protection,
3128 * which will be released by another thread in
3129 * io_complete_rw(). Fool lockdep by telling it the lock got
3130 * released so that it doesn't complain about the held lock when
3131 * we return to userspace.
3133 if (req->flags & REQ_F_ISREG) {
3134 __sb_start_write(file_inode(req->file)->i_sb,
3135 SB_FREEZE_WRITE, true);
3136 __sb_writers_release(file_inode(req->file)->i_sb,
3139 kiocb->ki_flags |= IOCB_WRITE;
3141 if (req->file->f_op->write_iter)
3142 ret2 = call_write_iter(req->file, kiocb, &iter);
3144 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3147 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3148 * retry them without IOCB_NOWAIT.
3150 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3152 if (!force_nonblock || ret2 != -EAGAIN) {
3153 kiocb_done(kiocb, ret2, cs);
3155 iter.count = iov_count;
3156 iter.nr_segs = nr_segs;
3158 ret = io_setup_async_rw(req, io_size, iovec,
3159 inline_vecs, &iter);
3162 /* it's copied and will be cleaned with ->io */
3173 static int __io_splice_prep(struct io_kiocb *req,
3174 const struct io_uring_sqe *sqe)
3176 struct io_splice* sp = &req->splice;
3177 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3180 if (req->flags & REQ_F_NEED_CLEANUP)
3182 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3186 sp->len = READ_ONCE(sqe->len);
3187 sp->flags = READ_ONCE(sqe->splice_flags);
3189 if (unlikely(sp->flags & ~valid_flags))
3192 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3193 (sp->flags & SPLICE_F_FD_IN_FIXED));
3196 req->flags |= REQ_F_NEED_CLEANUP;
3198 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3200 * Splice operation will be punted aync, and here need to
3201 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3203 io_req_init_async(req);
3204 req->work.flags |= IO_WQ_WORK_UNBOUND;
3210 static int io_tee_prep(struct io_kiocb *req,
3211 const struct io_uring_sqe *sqe)
3213 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3215 return __io_splice_prep(req, sqe);
3218 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3220 struct io_splice *sp = &req->splice;
3221 struct file *in = sp->file_in;
3222 struct file *out = sp->file_out;
3223 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3229 ret = do_tee(in, out, sp->len, flags);
3231 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3232 req->flags &= ~REQ_F_NEED_CLEANUP;
3235 req_set_fail_links(req);
3236 io_req_complete(req, ret);
3240 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3242 struct io_splice* sp = &req->splice;
3244 sp->off_in = READ_ONCE(sqe->splice_off_in);
3245 sp->off_out = READ_ONCE(sqe->off);
3246 return __io_splice_prep(req, sqe);
3249 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3251 struct io_splice *sp = &req->splice;
3252 struct file *in = sp->file_in;
3253 struct file *out = sp->file_out;
3254 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3255 loff_t *poff_in, *poff_out;
3261 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3262 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3265 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3267 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3268 req->flags &= ~REQ_F_NEED_CLEANUP;
3271 req_set_fail_links(req);
3272 io_req_complete(req, ret);
3277 * IORING_OP_NOP just posts a completion event, nothing else.
3279 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3281 struct io_ring_ctx *ctx = req->ctx;
3283 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3286 __io_req_complete(req, 0, 0, cs);
3290 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3292 struct io_ring_ctx *ctx = req->ctx;
3297 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3299 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3302 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3303 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3306 req->sync.off = READ_ONCE(sqe->off);
3307 req->sync.len = READ_ONCE(sqe->len);
3311 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3313 loff_t end = req->sync.off + req->sync.len;
3316 /* fsync always requires a blocking context */
3320 ret = vfs_fsync_range(req->file, req->sync.off,
3321 end > 0 ? end : LLONG_MAX,
3322 req->sync.flags & IORING_FSYNC_DATASYNC);
3324 req_set_fail_links(req);
3325 io_req_complete(req, ret);
3329 static int io_fallocate_prep(struct io_kiocb *req,
3330 const struct io_uring_sqe *sqe)
3332 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3334 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3337 req->sync.off = READ_ONCE(sqe->off);
3338 req->sync.len = READ_ONCE(sqe->addr);
3339 req->sync.mode = READ_ONCE(sqe->len);
3343 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3347 /* fallocate always requiring blocking context */
3350 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3353 req_set_fail_links(req);
3354 io_req_complete(req, ret);
3358 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3360 const char __user *fname;
3363 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3365 if (unlikely(sqe->ioprio || sqe->buf_index))
3367 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3370 /* open.how should be already initialised */
3371 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3372 req->open.how.flags |= O_LARGEFILE;
3374 req->open.dfd = READ_ONCE(sqe->fd);
3375 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3376 req->open.filename = getname(fname);
3377 if (IS_ERR(req->open.filename)) {
3378 ret = PTR_ERR(req->open.filename);
3379 req->open.filename = NULL;
3382 req->open.nofile = rlimit(RLIMIT_NOFILE);
3383 req->flags |= REQ_F_NEED_CLEANUP;
3387 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3391 if (req->flags & REQ_F_NEED_CLEANUP)
3393 mode = READ_ONCE(sqe->len);
3394 flags = READ_ONCE(sqe->open_flags);
3395 req->open.how = build_open_how(flags, mode);
3396 return __io_openat_prep(req, sqe);
3399 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3401 struct open_how __user *how;
3405 if (req->flags & REQ_F_NEED_CLEANUP)
3407 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3408 len = READ_ONCE(sqe->len);
3409 if (len < OPEN_HOW_SIZE_VER0)
3412 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3417 return __io_openat_prep(req, sqe);
3420 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3422 struct open_flags op;
3429 ret = build_open_flags(&req->open.how, &op);
3433 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3437 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3440 ret = PTR_ERR(file);
3442 fsnotify_open(file);
3443 fd_install(ret, file);
3446 putname(req->open.filename);
3447 req->flags &= ~REQ_F_NEED_CLEANUP;
3449 req_set_fail_links(req);
3450 io_req_complete(req, ret);
3454 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3456 return io_openat2(req, force_nonblock);
3459 static int io_remove_buffers_prep(struct io_kiocb *req,
3460 const struct io_uring_sqe *sqe)
3462 struct io_provide_buf *p = &req->pbuf;
3465 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3468 tmp = READ_ONCE(sqe->fd);
3469 if (!tmp || tmp > USHRT_MAX)
3472 memset(p, 0, sizeof(*p));
3474 p->bgid = READ_ONCE(sqe->buf_group);
3478 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3479 int bgid, unsigned nbufs)
3483 /* shouldn't happen */
3487 /* the head kbuf is the list itself */
3488 while (!list_empty(&buf->list)) {
3489 struct io_buffer *nxt;
3491 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3492 list_del(&nxt->list);
3499 idr_remove(&ctx->io_buffer_idr, bgid);
3504 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3505 struct io_comp_state *cs)
3507 struct io_provide_buf *p = &req->pbuf;
3508 struct io_ring_ctx *ctx = req->ctx;
3509 struct io_buffer *head;
3512 io_ring_submit_lock(ctx, !force_nonblock);
3514 lockdep_assert_held(&ctx->uring_lock);
3517 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3519 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3521 io_ring_submit_lock(ctx, !force_nonblock);
3523 req_set_fail_links(req);
3524 __io_req_complete(req, ret, 0, cs);
3528 static int io_provide_buffers_prep(struct io_kiocb *req,
3529 const struct io_uring_sqe *sqe)
3531 struct io_provide_buf *p = &req->pbuf;
3534 if (sqe->ioprio || sqe->rw_flags)
3537 tmp = READ_ONCE(sqe->fd);
3538 if (!tmp || tmp > USHRT_MAX)
3541 p->addr = READ_ONCE(sqe->addr);
3542 p->len = READ_ONCE(sqe->len);
3544 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3547 p->bgid = READ_ONCE(sqe->buf_group);
3548 tmp = READ_ONCE(sqe->off);
3549 if (tmp > USHRT_MAX)
3555 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3557 struct io_buffer *buf;
3558 u64 addr = pbuf->addr;
3559 int i, bid = pbuf->bid;
3561 for (i = 0; i < pbuf->nbufs; i++) {
3562 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3567 buf->len = pbuf->len;
3572 INIT_LIST_HEAD(&buf->list);
3575 list_add_tail(&buf->list, &(*head)->list);
3579 return i ? i : -ENOMEM;
3582 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3583 struct io_comp_state *cs)
3585 struct io_provide_buf *p = &req->pbuf;
3586 struct io_ring_ctx *ctx = req->ctx;
3587 struct io_buffer *head, *list;
3590 io_ring_submit_lock(ctx, !force_nonblock);
3592 lockdep_assert_held(&ctx->uring_lock);
3594 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3596 ret = io_add_buffers(p, &head);
3601 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3604 __io_remove_buffers(ctx, head, p->bgid, -1U);
3609 io_ring_submit_unlock(ctx, !force_nonblock);
3611 req_set_fail_links(req);
3612 __io_req_complete(req, ret, 0, cs);
3616 static int io_epoll_ctl_prep(struct io_kiocb *req,
3617 const struct io_uring_sqe *sqe)
3619 #if defined(CONFIG_EPOLL)
3620 if (sqe->ioprio || sqe->buf_index)
3622 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3625 req->epoll.epfd = READ_ONCE(sqe->fd);
3626 req->epoll.op = READ_ONCE(sqe->len);
3627 req->epoll.fd = READ_ONCE(sqe->off);
3629 if (ep_op_has_event(req->epoll.op)) {
3630 struct epoll_event __user *ev;
3632 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3633 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3643 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3644 struct io_comp_state *cs)
3646 #if defined(CONFIG_EPOLL)
3647 struct io_epoll *ie = &req->epoll;
3650 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3651 if (force_nonblock && ret == -EAGAIN)
3655 req_set_fail_links(req);
3656 __io_req_complete(req, ret, 0, cs);
3663 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3665 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3666 if (sqe->ioprio || sqe->buf_index || sqe->off)
3668 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3671 req->madvise.addr = READ_ONCE(sqe->addr);
3672 req->madvise.len = READ_ONCE(sqe->len);
3673 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3680 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3682 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3683 struct io_madvise *ma = &req->madvise;
3689 ret = do_madvise(ma->addr, ma->len, ma->advice);
3691 req_set_fail_links(req);
3692 io_req_complete(req, ret);
3699 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3701 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3703 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3706 req->fadvise.offset = READ_ONCE(sqe->off);
3707 req->fadvise.len = READ_ONCE(sqe->len);
3708 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3712 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3714 struct io_fadvise *fa = &req->fadvise;
3717 if (force_nonblock) {
3718 switch (fa->advice) {
3719 case POSIX_FADV_NORMAL:
3720 case POSIX_FADV_RANDOM:
3721 case POSIX_FADV_SEQUENTIAL:
3728 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3730 req_set_fail_links(req);
3731 io_req_complete(req, ret);
3735 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3737 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3739 if (sqe->ioprio || sqe->buf_index)
3741 if (req->flags & REQ_F_FIXED_FILE)
3744 req->statx.dfd = READ_ONCE(sqe->fd);
3745 req->statx.mask = READ_ONCE(sqe->len);
3746 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3747 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3748 req->statx.flags = READ_ONCE(sqe->statx_flags);
3753 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3755 struct io_statx *ctx = &req->statx;
3758 if (force_nonblock) {
3759 /* only need file table for an actual valid fd */
3760 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3761 req->flags |= REQ_F_NO_FILE_TABLE;
3765 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3769 req_set_fail_links(req);
3770 io_req_complete(req, ret);
3774 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3777 * If we queue this for async, it must not be cancellable. That would
3778 * leave the 'file' in an undeterminate state, and here need to modify
3779 * io_wq_work.flags, so initialize io_wq_work firstly.
3781 io_req_init_async(req);
3782 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3784 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3786 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3787 sqe->rw_flags || sqe->buf_index)
3789 if (req->flags & REQ_F_FIXED_FILE)
3792 req->close.fd = READ_ONCE(sqe->fd);
3793 if ((req->file && req->file->f_op == &io_uring_fops) ||
3794 req->close.fd == req->ctx->ring_fd)
3797 req->close.put_file = NULL;
3801 static int io_close(struct io_kiocb *req, bool force_nonblock,
3802 struct io_comp_state *cs)
3804 struct io_close *close = &req->close;
3807 /* might be already done during nonblock submission */
3808 if (!close->put_file) {
3809 ret = __close_fd_get_file(close->fd, &close->put_file);
3811 return (ret == -ENOENT) ? -EBADF : ret;
3814 /* if the file has a flush method, be safe and punt to async */
3815 if (close->put_file->f_op->flush && force_nonblock) {
3816 /* was never set, but play safe */
3817 req->flags &= ~REQ_F_NOWAIT;
3818 /* avoid grabbing files - we don't need the files */
3819 req->flags |= REQ_F_NO_FILE_TABLE;
3823 /* No ->flush() or already async, safely close from here */
3824 ret = filp_close(close->put_file, req->work.files);
3826 req_set_fail_links(req);
3827 fput(close->put_file);
3828 close->put_file = NULL;
3829 __io_req_complete(req, ret, 0, cs);
3833 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3835 struct io_ring_ctx *ctx = req->ctx;
3840 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3842 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3845 req->sync.off = READ_ONCE(sqe->off);
3846 req->sync.len = READ_ONCE(sqe->len);
3847 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3851 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3855 /* sync_file_range always requires a blocking context */
3859 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3862 req_set_fail_links(req);
3863 io_req_complete(req, ret);
3867 #if defined(CONFIG_NET)
3868 static int io_setup_async_msg(struct io_kiocb *req,
3869 struct io_async_msghdr *kmsg)
3873 if (io_alloc_async_ctx(req)) {
3874 if (kmsg->iov != kmsg->fast_iov)
3878 req->flags |= REQ_F_NEED_CLEANUP;
3879 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3883 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3884 struct io_async_msghdr *iomsg)
3886 iomsg->iov = iomsg->fast_iov;
3887 iomsg->msg.msg_name = &iomsg->addr;
3888 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3889 req->sr_msg.msg_flags, &iomsg->iov);
3892 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3894 struct io_sr_msg *sr = &req->sr_msg;
3895 struct io_async_ctx *io = req->io;
3898 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3901 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3902 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3903 sr->len = READ_ONCE(sqe->len);
3905 #ifdef CONFIG_COMPAT
3906 if (req->ctx->compat)
3907 sr->msg_flags |= MSG_CMSG_COMPAT;
3910 if (!io || req->opcode == IORING_OP_SEND)
3912 /* iovec is already imported */
3913 if (req->flags & REQ_F_NEED_CLEANUP)
3916 ret = io_sendmsg_copy_hdr(req, &io->msg);
3918 req->flags |= REQ_F_NEED_CLEANUP;
3922 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3923 struct io_comp_state *cs)
3925 struct io_async_msghdr iomsg, *kmsg;
3926 struct socket *sock;
3930 sock = sock_from_file(req->file, &ret);
3931 if (unlikely(!sock))
3935 kmsg = &req->io->msg;
3936 kmsg->msg.msg_name = &req->io->msg.addr;
3937 /* if iov is set, it's allocated already */
3939 kmsg->iov = kmsg->fast_iov;
3940 kmsg->msg.msg_iter.iov = kmsg->iov;
3942 ret = io_sendmsg_copy_hdr(req, &iomsg);
3948 flags = req->sr_msg.msg_flags;
3949 if (flags & MSG_DONTWAIT)
3950 req->flags |= REQ_F_NOWAIT;
3951 else if (force_nonblock)
3952 flags |= MSG_DONTWAIT;
3954 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3955 if (force_nonblock && ret == -EAGAIN)
3956 return io_setup_async_msg(req, kmsg);
3957 if (ret == -ERESTARTSYS)
3960 if (kmsg->iov != kmsg->fast_iov)
3962 req->flags &= ~REQ_F_NEED_CLEANUP;
3964 req_set_fail_links(req);
3965 __io_req_complete(req, ret, 0, cs);
3969 static int io_send(struct io_kiocb *req, bool force_nonblock,
3970 struct io_comp_state *cs)
3972 struct io_sr_msg *sr = &req->sr_msg;
3975 struct socket *sock;
3979 sock = sock_from_file(req->file, &ret);
3980 if (unlikely(!sock))
3983 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
3987 msg.msg_name = NULL;
3988 msg.msg_control = NULL;
3989 msg.msg_controllen = 0;
3990 msg.msg_namelen = 0;
3992 flags = req->sr_msg.msg_flags;
3993 if (flags & MSG_DONTWAIT)
3994 req->flags |= REQ_F_NOWAIT;
3995 else if (force_nonblock)
3996 flags |= MSG_DONTWAIT;
3998 msg.msg_flags = flags;
3999 ret = sock_sendmsg(sock, &msg);
4000 if (force_nonblock && ret == -EAGAIN)
4002 if (ret == -ERESTARTSYS)
4006 req_set_fail_links(req);
4007 __io_req_complete(req, ret, 0, cs);
4011 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4012 struct io_async_msghdr *iomsg)
4014 struct io_sr_msg *sr = &req->sr_msg;
4015 struct iovec __user *uiov;
4019 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4020 &iomsg->uaddr, &uiov, &iov_len);
4024 if (req->flags & REQ_F_BUFFER_SELECT) {
4027 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4029 sr->len = iomsg->iov[0].iov_len;
4030 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4034 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4035 &iomsg->iov, &iomsg->msg.msg_iter);
4043 #ifdef CONFIG_COMPAT
4044 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4045 struct io_async_msghdr *iomsg)
4047 struct compat_msghdr __user *msg_compat;
4048 struct io_sr_msg *sr = &req->sr_msg;
4049 struct compat_iovec __user *uiov;
4054 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4055 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4060 uiov = compat_ptr(ptr);
4061 if (req->flags & REQ_F_BUFFER_SELECT) {
4062 compat_ssize_t clen;
4066 if (!access_ok(uiov, sizeof(*uiov)))
4068 if (__get_user(clen, &uiov->iov_len))
4072 sr->len = iomsg->iov[0].iov_len;
4075 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4077 &iomsg->msg.msg_iter);
4086 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4087 struct io_async_msghdr *iomsg)
4089 iomsg->msg.msg_name = &iomsg->addr;
4090 iomsg->iov = iomsg->fast_iov;
4092 #ifdef CONFIG_COMPAT
4093 if (req->ctx->compat)
4094 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4097 return __io_recvmsg_copy_hdr(req, iomsg);
4100 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4103 struct io_sr_msg *sr = &req->sr_msg;
4104 struct io_buffer *kbuf;
4106 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4111 req->flags |= REQ_F_BUFFER_SELECTED;
4115 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4117 return io_put_kbuf(req, req->sr_msg.kbuf);
4120 static int io_recvmsg_prep(struct io_kiocb *req,
4121 const struct io_uring_sqe *sqe)
4123 struct io_sr_msg *sr = &req->sr_msg;
4124 struct io_async_ctx *io = req->io;
4127 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4130 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4131 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4132 sr->len = READ_ONCE(sqe->len);
4133 sr->bgid = READ_ONCE(sqe->buf_group);
4135 #ifdef CONFIG_COMPAT
4136 if (req->ctx->compat)
4137 sr->msg_flags |= MSG_CMSG_COMPAT;
4140 if (!io || req->opcode == IORING_OP_RECV)
4142 /* iovec is already imported */
4143 if (req->flags & REQ_F_NEED_CLEANUP)
4146 ret = io_recvmsg_copy_hdr(req, &io->msg);
4148 req->flags |= REQ_F_NEED_CLEANUP;
4152 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4153 struct io_comp_state *cs)
4155 struct io_async_msghdr iomsg, *kmsg;
4156 struct socket *sock;
4157 struct io_buffer *kbuf;
4159 int ret, cflags = 0;
4161 sock = sock_from_file(req->file, &ret);
4162 if (unlikely(!sock))
4166 kmsg = &req->io->msg;
4167 kmsg->msg.msg_name = &req->io->msg.addr;
4168 /* if iov is set, it's allocated already */
4170 kmsg->iov = kmsg->fast_iov;
4171 kmsg->msg.msg_iter.iov = kmsg->iov;
4173 ret = io_recvmsg_copy_hdr(req, &iomsg);
4179 if (req->flags & REQ_F_BUFFER_SELECT) {
4180 kbuf = io_recv_buffer_select(req, !force_nonblock);
4182 return PTR_ERR(kbuf);
4183 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4184 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4185 1, req->sr_msg.len);
4188 flags = req->sr_msg.msg_flags;
4189 if (flags & MSG_DONTWAIT)
4190 req->flags |= REQ_F_NOWAIT;
4191 else if (force_nonblock)
4192 flags |= MSG_DONTWAIT;
4194 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4195 kmsg->uaddr, flags);
4196 if (force_nonblock && ret == -EAGAIN)
4197 return io_setup_async_msg(req, kmsg);
4198 if (ret == -ERESTARTSYS)
4201 if (req->flags & REQ_F_BUFFER_SELECTED)
4202 cflags = io_put_recv_kbuf(req);
4203 if (kmsg->iov != kmsg->fast_iov)
4205 req->flags &= ~REQ_F_NEED_CLEANUP;
4207 req_set_fail_links(req);
4208 __io_req_complete(req, ret, cflags, cs);
4212 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4213 struct io_comp_state *cs)
4215 struct io_buffer *kbuf;
4216 struct io_sr_msg *sr = &req->sr_msg;
4218 void __user *buf = sr->buf;
4219 struct socket *sock;
4222 int ret, cflags = 0;
4224 sock = sock_from_file(req->file, &ret);
4225 if (unlikely(!sock))
4228 if (req->flags & REQ_F_BUFFER_SELECT) {
4229 kbuf = io_recv_buffer_select(req, !force_nonblock);
4231 return PTR_ERR(kbuf);
4232 buf = u64_to_user_ptr(kbuf->addr);
4235 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4239 msg.msg_name = NULL;
4240 msg.msg_control = NULL;
4241 msg.msg_controllen = 0;
4242 msg.msg_namelen = 0;
4243 msg.msg_iocb = NULL;
4246 flags = req->sr_msg.msg_flags;
4247 if (flags & MSG_DONTWAIT)
4248 req->flags |= REQ_F_NOWAIT;
4249 else if (force_nonblock)
4250 flags |= MSG_DONTWAIT;
4252 ret = sock_recvmsg(sock, &msg, flags);
4253 if (force_nonblock && ret == -EAGAIN)
4255 if (ret == -ERESTARTSYS)
4258 if (req->flags & REQ_F_BUFFER_SELECTED)
4259 cflags = io_put_recv_kbuf(req);
4261 req_set_fail_links(req);
4262 __io_req_complete(req, ret, cflags, cs);
4266 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4268 struct io_accept *accept = &req->accept;
4270 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4272 if (sqe->ioprio || sqe->len || sqe->buf_index)
4275 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4276 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4277 accept->flags = READ_ONCE(sqe->accept_flags);
4278 accept->nofile = rlimit(RLIMIT_NOFILE);
4282 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4283 struct io_comp_state *cs)
4285 struct io_accept *accept = &req->accept;
4286 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4289 if (req->file->f_flags & O_NONBLOCK)
4290 req->flags |= REQ_F_NOWAIT;
4292 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4293 accept->addr_len, accept->flags,
4295 if (ret == -EAGAIN && force_nonblock)
4298 if (ret == -ERESTARTSYS)
4300 req_set_fail_links(req);
4302 __io_req_complete(req, ret, 0, cs);
4306 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4308 struct io_connect *conn = &req->connect;
4309 struct io_async_ctx *io = req->io;
4311 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4313 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4316 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4317 conn->addr_len = READ_ONCE(sqe->addr2);
4322 return move_addr_to_kernel(conn->addr, conn->addr_len,
4323 &io->connect.address);
4326 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4327 struct io_comp_state *cs)
4329 struct io_async_ctx __io, *io;
4330 unsigned file_flags;
4336 ret = move_addr_to_kernel(req->connect.addr,
4337 req->connect.addr_len,
4338 &__io.connect.address);
4344 file_flags = force_nonblock ? O_NONBLOCK : 0;
4346 ret = __sys_connect_file(req->file, &io->connect.address,
4347 req->connect.addr_len, file_flags);
4348 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4351 if (io_alloc_async_ctx(req)) {
4355 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4358 if (ret == -ERESTARTSYS)
4362 req_set_fail_links(req);
4363 __io_req_complete(req, ret, 0, cs);
4366 #else /* !CONFIG_NET */
4367 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4372 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4373 struct io_comp_state *cs)
4378 static int io_send(struct io_kiocb *req, bool force_nonblock,
4379 struct io_comp_state *cs)
4384 static int io_recvmsg_prep(struct io_kiocb *req,
4385 const struct io_uring_sqe *sqe)
4390 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4391 struct io_comp_state *cs)
4396 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4397 struct io_comp_state *cs)
4402 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4407 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4408 struct io_comp_state *cs)
4413 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4418 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4419 struct io_comp_state *cs)
4423 #endif /* CONFIG_NET */
4425 struct io_poll_table {
4426 struct poll_table_struct pt;
4427 struct io_kiocb *req;
4431 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4432 __poll_t mask, task_work_func_t func)
4436 /* for instances that support it check for an event match first: */
4437 if (mask && !(mask & poll->events))
4440 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4442 list_del_init(&poll->wait.entry);
4445 init_task_work(&req->task_work, func);
4447 * If this fails, then the task is exiting. When a task exits, the
4448 * work gets canceled, so just cancel this request as well instead
4449 * of executing it. We can't safely execute it anyway, as we may not
4450 * have the needed state needed for it anyway.
4452 ret = io_req_task_work_add(req, &req->task_work);
4453 if (unlikely(ret)) {
4454 struct task_struct *tsk;
4456 WRITE_ONCE(poll->canceled, true);
4457 tsk = io_wq_get_task(req->ctx->io_wq);
4458 task_work_add(tsk, &req->task_work, 0);
4459 wake_up_process(tsk);
4464 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4465 __acquires(&req->ctx->completion_lock)
4467 struct io_ring_ctx *ctx = req->ctx;
4469 if (!req->result && !READ_ONCE(poll->canceled)) {
4470 struct poll_table_struct pt = { ._key = poll->events };
4472 req->result = vfs_poll(req->file, &pt) & poll->events;
4475 spin_lock_irq(&ctx->completion_lock);
4476 if (!req->result && !READ_ONCE(poll->canceled)) {
4477 add_wait_queue(poll->head, &poll->wait);
4484 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4486 struct io_poll_iocb *poll = data;
4488 lockdep_assert_held(&req->ctx->completion_lock);
4490 if (poll && poll->head) {
4491 struct wait_queue_head *head = poll->head;
4493 spin_lock(&head->lock);
4494 list_del_init(&poll->wait.entry);
4495 if (poll->wait.private)
4496 refcount_dec(&req->refs);
4498 spin_unlock(&head->lock);
4502 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4504 struct io_ring_ctx *ctx = req->ctx;
4506 io_poll_remove_double(req, req->io);
4507 req->poll.done = true;
4508 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4509 io_commit_cqring(ctx);
4512 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4514 struct io_ring_ctx *ctx = req->ctx;
4516 if (io_poll_rewait(req, &req->poll)) {
4517 spin_unlock_irq(&ctx->completion_lock);
4521 hash_del(&req->hash_node);
4522 io_poll_complete(req, req->result, 0);
4523 req->flags |= REQ_F_COMP_LOCKED;
4524 *nxt = io_put_req_find_next(req);
4525 spin_unlock_irq(&ctx->completion_lock);
4527 io_cqring_ev_posted(ctx);
4530 static void io_poll_task_func(struct callback_head *cb)
4532 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4533 struct io_kiocb *nxt = NULL;
4535 io_poll_task_handler(req, &nxt);
4537 __io_req_task_submit(nxt);
4540 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4541 int sync, void *key)
4543 struct io_kiocb *req = wait->private;
4544 struct io_poll_iocb *poll = req->apoll->double_poll;
4545 __poll_t mask = key_to_poll(key);
4547 /* for instances that support it check for an event match first: */
4548 if (mask && !(mask & poll->events))
4551 if (poll && poll->head) {
4554 spin_lock(&poll->head->lock);
4555 done = list_empty(&poll->wait.entry);
4557 list_del_init(&poll->wait.entry);
4558 spin_unlock(&poll->head->lock);
4560 __io_async_wake(req, poll, mask, io_poll_task_func);
4562 refcount_dec(&req->refs);
4566 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4567 wait_queue_func_t wake_func)
4571 poll->canceled = false;
4572 poll->events = events;
4573 INIT_LIST_HEAD(&poll->wait.entry);
4574 init_waitqueue_func_entry(&poll->wait, wake_func);
4577 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4578 struct wait_queue_head *head,
4579 struct io_poll_iocb **poll_ptr)
4581 struct io_kiocb *req = pt->req;
4584 * If poll->head is already set, it's because the file being polled
4585 * uses multiple waitqueues for poll handling (eg one for read, one
4586 * for write). Setup a separate io_poll_iocb if this happens.
4588 if (unlikely(poll->head)) {
4589 /* already have a 2nd entry, fail a third attempt */
4591 pt->error = -EINVAL;
4594 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4596 pt->error = -ENOMEM;
4599 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4600 refcount_inc(&req->refs);
4601 poll->wait.private = req;
4608 if (poll->events & EPOLLEXCLUSIVE)
4609 add_wait_queue_exclusive(head, &poll->wait);
4611 add_wait_queue(head, &poll->wait);
4614 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4615 struct poll_table_struct *p)
4617 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4618 struct async_poll *apoll = pt->req->apoll;
4620 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4623 static void io_async_task_func(struct callback_head *cb)
4625 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4626 struct async_poll *apoll = req->apoll;
4627 struct io_ring_ctx *ctx = req->ctx;
4629 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4631 if (io_poll_rewait(req, &apoll->poll)) {
4632 spin_unlock_irq(&ctx->completion_lock);
4636 /* If req is still hashed, it cannot have been canceled. Don't check. */
4637 if (hash_hashed(&req->hash_node))
4638 hash_del(&req->hash_node);
4640 io_poll_remove_double(req, apoll->double_poll);
4641 spin_unlock_irq(&ctx->completion_lock);
4643 /* restore ->work in case we need to retry again */
4644 if (req->flags & REQ_F_WORK_INITIALIZED)
4645 memcpy(&req->work, &apoll->work, sizeof(req->work));
4647 if (!READ_ONCE(apoll->poll.canceled))
4648 __io_req_task_submit(req);
4650 __io_req_task_cancel(req, -ECANCELED);
4652 kfree(apoll->double_poll);
4656 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4659 struct io_kiocb *req = wait->private;
4660 struct io_poll_iocb *poll = &req->apoll->poll;
4662 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4665 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4668 static void io_poll_req_insert(struct io_kiocb *req)
4670 struct io_ring_ctx *ctx = req->ctx;
4671 struct hlist_head *list;
4673 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4674 hlist_add_head(&req->hash_node, list);
4677 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4678 struct io_poll_iocb *poll,
4679 struct io_poll_table *ipt, __poll_t mask,
4680 wait_queue_func_t wake_func)
4681 __acquires(&ctx->completion_lock)
4683 struct io_ring_ctx *ctx = req->ctx;
4684 bool cancel = false;
4686 io_init_poll_iocb(poll, mask, wake_func);
4687 poll->file = req->file;
4688 poll->wait.private = req;
4690 ipt->pt._key = mask;
4692 ipt->error = -EINVAL;
4694 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4696 spin_lock_irq(&ctx->completion_lock);
4697 if (likely(poll->head)) {
4698 spin_lock(&poll->head->lock);
4699 if (unlikely(list_empty(&poll->wait.entry))) {
4705 if (mask || ipt->error)
4706 list_del_init(&poll->wait.entry);
4708 WRITE_ONCE(poll->canceled, true);
4709 else if (!poll->done) /* actually waiting for an event */
4710 io_poll_req_insert(req);
4711 spin_unlock(&poll->head->lock);
4717 static bool io_arm_poll_handler(struct io_kiocb *req)
4719 const struct io_op_def *def = &io_op_defs[req->opcode];
4720 struct io_ring_ctx *ctx = req->ctx;
4721 struct async_poll *apoll;
4722 struct io_poll_table ipt;
4725 if (!req->file || !file_can_poll(req->file))
4727 if (req->flags & REQ_F_POLLED)
4729 if (!def->pollin && !def->pollout)
4732 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4733 if (unlikely(!apoll))
4735 apoll->double_poll = NULL;
4737 req->flags |= REQ_F_POLLED;
4738 if (req->flags & REQ_F_WORK_INITIALIZED)
4739 memcpy(&apoll->work, &req->work, sizeof(req->work));
4741 io_get_req_task(req);
4743 INIT_HLIST_NODE(&req->hash_node);
4747 mask |= POLLIN | POLLRDNORM;
4749 mask |= POLLOUT | POLLWRNORM;
4750 mask |= POLLERR | POLLPRI;
4752 ipt.pt._qproc = io_async_queue_proc;
4754 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4757 io_poll_remove_double(req, apoll->double_poll);
4758 spin_unlock_irq(&ctx->completion_lock);
4759 if (req->flags & REQ_F_WORK_INITIALIZED)
4760 memcpy(&req->work, &apoll->work, sizeof(req->work));
4761 kfree(apoll->double_poll);
4765 spin_unlock_irq(&ctx->completion_lock);
4766 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4767 apoll->poll.events);
4771 static bool __io_poll_remove_one(struct io_kiocb *req,
4772 struct io_poll_iocb *poll)
4774 bool do_complete = false;
4776 spin_lock(&poll->head->lock);
4777 WRITE_ONCE(poll->canceled, true);
4778 if (!list_empty(&poll->wait.entry)) {
4779 list_del_init(&poll->wait.entry);
4782 spin_unlock(&poll->head->lock);
4783 hash_del(&req->hash_node);
4787 static bool io_poll_remove_one(struct io_kiocb *req)
4791 if (req->opcode == IORING_OP_POLL_ADD) {
4792 io_poll_remove_double(req, req->io);
4793 do_complete = __io_poll_remove_one(req, &req->poll);
4795 struct async_poll *apoll = req->apoll;
4797 io_poll_remove_double(req, apoll->double_poll);
4799 /* non-poll requests have submit ref still */
4800 do_complete = __io_poll_remove_one(req, &apoll->poll);
4804 * restore ->work because we will call
4805 * io_req_clean_work below when dropping the
4808 if (req->flags & REQ_F_WORK_INITIALIZED)
4809 memcpy(&req->work, &apoll->work,
4811 kfree(apoll->double_poll);
4817 io_cqring_fill_event(req, -ECANCELED);
4818 io_commit_cqring(req->ctx);
4819 req->flags |= REQ_F_COMP_LOCKED;
4826 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4828 struct hlist_node *tmp;
4829 struct io_kiocb *req;
4832 spin_lock_irq(&ctx->completion_lock);
4833 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4834 struct hlist_head *list;
4836 list = &ctx->cancel_hash[i];
4837 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4838 posted += io_poll_remove_one(req);
4840 spin_unlock_irq(&ctx->completion_lock);
4843 io_cqring_ev_posted(ctx);
4846 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4848 struct hlist_head *list;
4849 struct io_kiocb *req;
4851 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4852 hlist_for_each_entry(req, list, hash_node) {
4853 if (sqe_addr != req->user_data)
4855 if (io_poll_remove_one(req))
4863 static int io_poll_remove_prep(struct io_kiocb *req,
4864 const struct io_uring_sqe *sqe)
4866 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4868 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4872 req->poll.addr = READ_ONCE(sqe->addr);
4877 * Find a running poll command that matches one specified in sqe->addr,
4878 * and remove it if found.
4880 static int io_poll_remove(struct io_kiocb *req)
4882 struct io_ring_ctx *ctx = req->ctx;
4886 addr = req->poll.addr;
4887 spin_lock_irq(&ctx->completion_lock);
4888 ret = io_poll_cancel(ctx, addr);
4889 spin_unlock_irq(&ctx->completion_lock);
4892 req_set_fail_links(req);
4893 io_req_complete(req, ret);
4897 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4900 struct io_kiocb *req = wait->private;
4901 struct io_poll_iocb *poll = &req->poll;
4903 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4906 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4907 struct poll_table_struct *p)
4909 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4911 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4914 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4916 struct io_poll_iocb *poll = &req->poll;
4919 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4921 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4926 events = READ_ONCE(sqe->poll32_events);
4928 events = swahw32(events);
4930 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4931 (events & EPOLLEXCLUSIVE);
4933 io_get_req_task(req);
4937 static int io_poll_add(struct io_kiocb *req)
4939 struct io_poll_iocb *poll = &req->poll;
4940 struct io_ring_ctx *ctx = req->ctx;
4941 struct io_poll_table ipt;
4944 /* ->work is in union with hash_node and others */
4945 io_req_clean_work(req);
4946 req->flags &= ~REQ_F_WORK_INITIALIZED;
4948 INIT_HLIST_NODE(&req->hash_node);
4949 ipt.pt._qproc = io_poll_queue_proc;
4951 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4954 if (mask) { /* no async, we'd stolen it */
4956 io_poll_complete(req, mask, 0);
4958 spin_unlock_irq(&ctx->completion_lock);
4961 io_cqring_ev_posted(ctx);
4967 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4969 struct io_timeout_data *data = container_of(timer,
4970 struct io_timeout_data, timer);
4971 struct io_kiocb *req = data->req;
4972 struct io_ring_ctx *ctx = req->ctx;
4973 unsigned long flags;
4975 atomic_inc(&ctx->cq_timeouts);
4977 spin_lock_irqsave(&ctx->completion_lock, flags);
4979 * We could be racing with timeout deletion. If the list is empty,
4980 * then timeout lookup already found it and will be handling it.
4982 if (!list_empty(&req->timeout.list))
4983 list_del_init(&req->timeout.list);
4985 io_cqring_fill_event(req, -ETIME);
4986 io_commit_cqring(ctx);
4987 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4989 io_cqring_ev_posted(ctx);
4990 req_set_fail_links(req);
4992 return HRTIMER_NORESTART;
4995 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4997 struct io_kiocb *req;
5000 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5001 if (user_data == req->user_data) {
5002 list_del_init(&req->timeout.list);
5011 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5015 req_set_fail_links(req);
5016 io_cqring_fill_event(req, -ECANCELED);
5021 static int io_timeout_remove_prep(struct io_kiocb *req,
5022 const struct io_uring_sqe *sqe)
5024 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5026 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5028 if (sqe->ioprio || sqe->buf_index || sqe->len)
5031 req->timeout.addr = READ_ONCE(sqe->addr);
5032 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5033 if (req->timeout.flags)
5040 * Remove or update an existing timeout command
5042 static int io_timeout_remove(struct io_kiocb *req)
5044 struct io_ring_ctx *ctx = req->ctx;
5047 spin_lock_irq(&ctx->completion_lock);
5048 ret = io_timeout_cancel(ctx, req->timeout.addr);
5050 io_cqring_fill_event(req, ret);
5051 io_commit_cqring(ctx);
5052 spin_unlock_irq(&ctx->completion_lock);
5053 io_cqring_ev_posted(ctx);
5055 req_set_fail_links(req);
5060 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5061 bool is_timeout_link)
5063 struct io_timeout_data *data;
5065 u32 off = READ_ONCE(sqe->off);
5067 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5069 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5071 if (off && is_timeout_link)
5073 flags = READ_ONCE(sqe->timeout_flags);
5074 if (flags & ~IORING_TIMEOUT_ABS)
5077 req->timeout.off = off;
5079 if (!req->io && io_alloc_async_ctx(req))
5082 data = &req->io->timeout;
5085 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5088 if (flags & IORING_TIMEOUT_ABS)
5089 data->mode = HRTIMER_MODE_ABS;
5091 data->mode = HRTIMER_MODE_REL;
5093 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5097 static int io_timeout(struct io_kiocb *req)
5099 struct io_ring_ctx *ctx = req->ctx;
5100 struct io_timeout_data *data = &req->io->timeout;
5101 struct list_head *entry;
5102 u32 tail, off = req->timeout.off;
5104 spin_lock_irq(&ctx->completion_lock);
5107 * sqe->off holds how many events that need to occur for this
5108 * timeout event to be satisfied. If it isn't set, then this is
5109 * a pure timeout request, sequence isn't used.
5111 if (io_is_timeout_noseq(req)) {
5112 entry = ctx->timeout_list.prev;
5116 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5117 req->timeout.target_seq = tail + off;
5120 * Insertion sort, ensuring the first entry in the list is always
5121 * the one we need first.
5123 list_for_each_prev(entry, &ctx->timeout_list) {
5124 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5127 if (io_is_timeout_noseq(nxt))
5129 /* nxt.seq is behind @tail, otherwise would've been completed */
5130 if (off >= nxt->timeout.target_seq - tail)
5134 list_add(&req->timeout.list, entry);
5135 data->timer.function = io_timeout_fn;
5136 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5137 spin_unlock_irq(&ctx->completion_lock);
5141 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5143 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5145 return req->user_data == (unsigned long) data;
5148 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5150 enum io_wq_cancel cancel_ret;
5153 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5154 switch (cancel_ret) {
5155 case IO_WQ_CANCEL_OK:
5158 case IO_WQ_CANCEL_RUNNING:
5161 case IO_WQ_CANCEL_NOTFOUND:
5169 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5170 struct io_kiocb *req, __u64 sqe_addr,
5173 unsigned long flags;
5176 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5177 if (ret != -ENOENT) {
5178 spin_lock_irqsave(&ctx->completion_lock, flags);
5182 spin_lock_irqsave(&ctx->completion_lock, flags);
5183 ret = io_timeout_cancel(ctx, sqe_addr);
5186 ret = io_poll_cancel(ctx, sqe_addr);
5190 io_cqring_fill_event(req, ret);
5191 io_commit_cqring(ctx);
5192 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5193 io_cqring_ev_posted(ctx);
5196 req_set_fail_links(req);
5200 static int io_async_cancel_prep(struct io_kiocb *req,
5201 const struct io_uring_sqe *sqe)
5203 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5205 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5207 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5210 req->cancel.addr = READ_ONCE(sqe->addr);
5214 static int io_async_cancel(struct io_kiocb *req)
5216 struct io_ring_ctx *ctx = req->ctx;
5218 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5222 static int io_files_update_prep(struct io_kiocb *req,
5223 const struct io_uring_sqe *sqe)
5225 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5227 if (sqe->ioprio || sqe->rw_flags)
5230 req->files_update.offset = READ_ONCE(sqe->off);
5231 req->files_update.nr_args = READ_ONCE(sqe->len);
5232 if (!req->files_update.nr_args)
5234 req->files_update.arg = READ_ONCE(sqe->addr);
5238 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5239 struct io_comp_state *cs)
5241 struct io_ring_ctx *ctx = req->ctx;
5242 struct io_uring_files_update up;
5248 up.offset = req->files_update.offset;
5249 up.fds = req->files_update.arg;
5251 mutex_lock(&ctx->uring_lock);
5252 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5253 mutex_unlock(&ctx->uring_lock);
5256 req_set_fail_links(req);
5257 __io_req_complete(req, ret, 0, cs);
5261 static int io_req_defer_prep(struct io_kiocb *req,
5262 const struct io_uring_sqe *sqe)
5269 if (io_alloc_async_ctx(req))
5272 if (io_op_defs[req->opcode].file_table) {
5273 io_req_init_async(req);
5274 ret = io_grab_files(req);
5279 switch (req->opcode) {
5282 case IORING_OP_READV:
5283 case IORING_OP_READ_FIXED:
5284 case IORING_OP_READ:
5285 ret = io_read_prep(req, sqe, true);
5287 case IORING_OP_WRITEV:
5288 case IORING_OP_WRITE_FIXED:
5289 case IORING_OP_WRITE:
5290 ret = io_write_prep(req, sqe, true);
5292 case IORING_OP_POLL_ADD:
5293 ret = io_poll_add_prep(req, sqe);
5295 case IORING_OP_POLL_REMOVE:
5296 ret = io_poll_remove_prep(req, sqe);
5298 case IORING_OP_FSYNC:
5299 ret = io_prep_fsync(req, sqe);
5301 case IORING_OP_SYNC_FILE_RANGE:
5302 ret = io_prep_sfr(req, sqe);
5304 case IORING_OP_SENDMSG:
5305 case IORING_OP_SEND:
5306 ret = io_sendmsg_prep(req, sqe);
5308 case IORING_OP_RECVMSG:
5309 case IORING_OP_RECV:
5310 ret = io_recvmsg_prep(req, sqe);
5312 case IORING_OP_CONNECT:
5313 ret = io_connect_prep(req, sqe);
5315 case IORING_OP_TIMEOUT:
5316 ret = io_timeout_prep(req, sqe, false);
5318 case IORING_OP_TIMEOUT_REMOVE:
5319 ret = io_timeout_remove_prep(req, sqe);
5321 case IORING_OP_ASYNC_CANCEL:
5322 ret = io_async_cancel_prep(req, sqe);
5324 case IORING_OP_LINK_TIMEOUT:
5325 ret = io_timeout_prep(req, sqe, true);
5327 case IORING_OP_ACCEPT:
5328 ret = io_accept_prep(req, sqe);
5330 case IORING_OP_FALLOCATE:
5331 ret = io_fallocate_prep(req, sqe);
5333 case IORING_OP_OPENAT:
5334 ret = io_openat_prep(req, sqe);
5336 case IORING_OP_CLOSE:
5337 ret = io_close_prep(req, sqe);
5339 case IORING_OP_FILES_UPDATE:
5340 ret = io_files_update_prep(req, sqe);
5342 case IORING_OP_STATX:
5343 ret = io_statx_prep(req, sqe);
5345 case IORING_OP_FADVISE:
5346 ret = io_fadvise_prep(req, sqe);
5348 case IORING_OP_MADVISE:
5349 ret = io_madvise_prep(req, sqe);
5351 case IORING_OP_OPENAT2:
5352 ret = io_openat2_prep(req, sqe);
5354 case IORING_OP_EPOLL_CTL:
5355 ret = io_epoll_ctl_prep(req, sqe);
5357 case IORING_OP_SPLICE:
5358 ret = io_splice_prep(req, sqe);
5360 case IORING_OP_PROVIDE_BUFFERS:
5361 ret = io_provide_buffers_prep(req, sqe);
5363 case IORING_OP_REMOVE_BUFFERS:
5364 ret = io_remove_buffers_prep(req, sqe);
5367 ret = io_tee_prep(req, sqe);
5370 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5379 static u32 io_get_sequence(struct io_kiocb *req)
5381 struct io_kiocb *pos;
5382 struct io_ring_ctx *ctx = req->ctx;
5383 u32 total_submitted, nr_reqs = 1;
5385 if (req->flags & REQ_F_LINK_HEAD)
5386 list_for_each_entry(pos, &req->link_list, link_list)
5389 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5390 return total_submitted - nr_reqs;
5393 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5395 struct io_ring_ctx *ctx = req->ctx;
5396 struct io_defer_entry *de;
5400 /* Still need defer if there is pending req in defer list. */
5401 if (likely(list_empty_careful(&ctx->defer_list) &&
5402 !(req->flags & REQ_F_IO_DRAIN)))
5405 seq = io_get_sequence(req);
5406 /* Still a chance to pass the sequence check */
5407 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5411 ret = io_req_defer_prep(req, sqe);
5415 io_prep_async_link(req);
5416 de = kmalloc(sizeof(*de), GFP_KERNEL);
5420 spin_lock_irq(&ctx->completion_lock);
5421 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5422 spin_unlock_irq(&ctx->completion_lock);
5427 trace_io_uring_defer(ctx, req, req->user_data);
5430 list_add_tail(&de->list, &ctx->defer_list);
5431 spin_unlock_irq(&ctx->completion_lock);
5432 return -EIOCBQUEUED;
5435 static void __io_clean_op(struct io_kiocb *req)
5437 struct io_async_ctx *io = req->io;
5439 if (req->flags & REQ_F_BUFFER_SELECTED) {
5440 switch (req->opcode) {
5441 case IORING_OP_READV:
5442 case IORING_OP_READ_FIXED:
5443 case IORING_OP_READ:
5444 kfree((void *)(unsigned long)req->rw.addr);
5446 case IORING_OP_RECVMSG:
5447 case IORING_OP_RECV:
5448 kfree(req->sr_msg.kbuf);
5451 req->flags &= ~REQ_F_BUFFER_SELECTED;
5454 if (req->flags & REQ_F_NEED_CLEANUP) {
5455 switch (req->opcode) {
5456 case IORING_OP_READV:
5457 case IORING_OP_READ_FIXED:
5458 case IORING_OP_READ:
5459 case IORING_OP_WRITEV:
5460 case IORING_OP_WRITE_FIXED:
5461 case IORING_OP_WRITE:
5462 if (io->rw.iov != io->rw.fast_iov)
5465 case IORING_OP_RECVMSG:
5466 case IORING_OP_SENDMSG:
5467 if (io->msg.iov != io->msg.fast_iov)
5470 case IORING_OP_SPLICE:
5472 io_put_file(req, req->splice.file_in,
5473 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5476 req->flags &= ~REQ_F_NEED_CLEANUP;
5480 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5481 bool force_nonblock, struct io_comp_state *cs)
5483 struct io_ring_ctx *ctx = req->ctx;
5486 switch (req->opcode) {
5488 ret = io_nop(req, cs);
5490 case IORING_OP_READV:
5491 case IORING_OP_READ_FIXED:
5492 case IORING_OP_READ:
5494 ret = io_read_prep(req, sqe, force_nonblock);
5498 ret = io_read(req, force_nonblock, cs);
5500 case IORING_OP_WRITEV:
5501 case IORING_OP_WRITE_FIXED:
5502 case IORING_OP_WRITE:
5504 ret = io_write_prep(req, sqe, force_nonblock);
5508 ret = io_write(req, force_nonblock, cs);
5510 case IORING_OP_FSYNC:
5512 ret = io_prep_fsync(req, sqe);
5516 ret = io_fsync(req, force_nonblock);
5518 case IORING_OP_POLL_ADD:
5520 ret = io_poll_add_prep(req, sqe);
5524 ret = io_poll_add(req);
5526 case IORING_OP_POLL_REMOVE:
5528 ret = io_poll_remove_prep(req, sqe);
5532 ret = io_poll_remove(req);
5534 case IORING_OP_SYNC_FILE_RANGE:
5536 ret = io_prep_sfr(req, sqe);
5540 ret = io_sync_file_range(req, force_nonblock);
5542 case IORING_OP_SENDMSG:
5543 case IORING_OP_SEND:
5545 ret = io_sendmsg_prep(req, sqe);
5549 if (req->opcode == IORING_OP_SENDMSG)
5550 ret = io_sendmsg(req, force_nonblock, cs);
5552 ret = io_send(req, force_nonblock, cs);
5554 case IORING_OP_RECVMSG:
5555 case IORING_OP_RECV:
5557 ret = io_recvmsg_prep(req, sqe);
5561 if (req->opcode == IORING_OP_RECVMSG)
5562 ret = io_recvmsg(req, force_nonblock, cs);
5564 ret = io_recv(req, force_nonblock, cs);
5566 case IORING_OP_TIMEOUT:
5568 ret = io_timeout_prep(req, sqe, false);
5572 ret = io_timeout(req);
5574 case IORING_OP_TIMEOUT_REMOVE:
5576 ret = io_timeout_remove_prep(req, sqe);
5580 ret = io_timeout_remove(req);
5582 case IORING_OP_ACCEPT:
5584 ret = io_accept_prep(req, sqe);
5588 ret = io_accept(req, force_nonblock, cs);
5590 case IORING_OP_CONNECT:
5592 ret = io_connect_prep(req, sqe);
5596 ret = io_connect(req, force_nonblock, cs);
5598 case IORING_OP_ASYNC_CANCEL:
5600 ret = io_async_cancel_prep(req, sqe);
5604 ret = io_async_cancel(req);
5606 case IORING_OP_FALLOCATE:
5608 ret = io_fallocate_prep(req, sqe);
5612 ret = io_fallocate(req, force_nonblock);
5614 case IORING_OP_OPENAT:
5616 ret = io_openat_prep(req, sqe);
5620 ret = io_openat(req, force_nonblock);
5622 case IORING_OP_CLOSE:
5624 ret = io_close_prep(req, sqe);
5628 ret = io_close(req, force_nonblock, cs);
5630 case IORING_OP_FILES_UPDATE:
5632 ret = io_files_update_prep(req, sqe);
5636 ret = io_files_update(req, force_nonblock, cs);
5638 case IORING_OP_STATX:
5640 ret = io_statx_prep(req, sqe);
5644 ret = io_statx(req, force_nonblock);
5646 case IORING_OP_FADVISE:
5648 ret = io_fadvise_prep(req, sqe);
5652 ret = io_fadvise(req, force_nonblock);
5654 case IORING_OP_MADVISE:
5656 ret = io_madvise_prep(req, sqe);
5660 ret = io_madvise(req, force_nonblock);
5662 case IORING_OP_OPENAT2:
5664 ret = io_openat2_prep(req, sqe);
5668 ret = io_openat2(req, force_nonblock);
5670 case IORING_OP_EPOLL_CTL:
5672 ret = io_epoll_ctl_prep(req, sqe);
5676 ret = io_epoll_ctl(req, force_nonblock, cs);
5678 case IORING_OP_SPLICE:
5680 ret = io_splice_prep(req, sqe);
5684 ret = io_splice(req, force_nonblock);
5686 case IORING_OP_PROVIDE_BUFFERS:
5688 ret = io_provide_buffers_prep(req, sqe);
5692 ret = io_provide_buffers(req, force_nonblock, cs);
5694 case IORING_OP_REMOVE_BUFFERS:
5696 ret = io_remove_buffers_prep(req, sqe);
5700 ret = io_remove_buffers(req, force_nonblock, cs);
5704 ret = io_tee_prep(req, sqe);
5708 ret = io_tee(req, force_nonblock);
5718 /* If the op doesn't have a file, we're not polling for it */
5719 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5720 const bool in_async = io_wq_current_is_worker();
5722 /* workqueue context doesn't hold uring_lock, grab it now */
5724 mutex_lock(&ctx->uring_lock);
5726 io_iopoll_req_issued(req);
5729 mutex_unlock(&ctx->uring_lock);
5735 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5737 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5738 struct io_kiocb *timeout;
5741 timeout = io_prep_linked_timeout(req);
5743 io_queue_linked_timeout(timeout);
5745 /* if NO_CANCEL is set, we must still run the work */
5746 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5747 IO_WQ_WORK_CANCEL) {
5753 ret = io_issue_sqe(req, NULL, false, NULL);
5755 * We can get EAGAIN for polled IO even though we're
5756 * forcing a sync submission from here, since we can't
5757 * wait for request slots on the block side.
5766 req_set_fail_links(req);
5767 io_req_complete(req, ret);
5770 return io_steal_work(req);
5773 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5776 struct fixed_file_table *table;
5778 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5779 return table->files[index & IORING_FILE_TABLE_MASK];
5782 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5783 int fd, struct file **out_file, bool fixed)
5785 struct io_ring_ctx *ctx = req->ctx;
5789 if (unlikely(!ctx->file_data ||
5790 (unsigned) fd >= ctx->nr_user_files))
5792 fd = array_index_nospec(fd, ctx->nr_user_files);
5793 file = io_file_from_index(ctx, fd);
5795 req->fixed_file_refs = ctx->file_data->cur_refs;
5796 percpu_ref_get(req->fixed_file_refs);
5799 trace_io_uring_file_get(ctx, fd);
5800 file = __io_file_get(state, fd);
5803 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5810 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5815 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5816 if (unlikely(!fixed && io_async_submit(req->ctx)))
5819 return io_file_get(state, req, fd, &req->file, fixed);
5822 static int io_grab_files(struct io_kiocb *req)
5825 struct io_ring_ctx *ctx = req->ctx;
5827 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5829 if (!ctx->ring_file)
5833 spin_lock_irq(&ctx->inflight_lock);
5835 * We use the f_ops->flush() handler to ensure that we can flush
5836 * out work accessing these files if the fd is closed. Check if
5837 * the fd has changed since we started down this path, and disallow
5838 * this operation if it has.
5840 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5841 list_add(&req->inflight_entry, &ctx->inflight_list);
5842 req->flags |= REQ_F_INFLIGHT;
5843 req->work.files = current->files;
5846 spin_unlock_irq(&ctx->inflight_lock);
5852 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5854 struct io_timeout_data *data = container_of(timer,
5855 struct io_timeout_data, timer);
5856 struct io_kiocb *req = data->req;
5857 struct io_ring_ctx *ctx = req->ctx;
5858 struct io_kiocb *prev = NULL;
5859 unsigned long flags;
5861 spin_lock_irqsave(&ctx->completion_lock, flags);
5864 * We don't expect the list to be empty, that will only happen if we
5865 * race with the completion of the linked work.
5867 if (!list_empty(&req->link_list)) {
5868 prev = list_entry(req->link_list.prev, struct io_kiocb,
5870 if (refcount_inc_not_zero(&prev->refs)) {
5871 list_del_init(&req->link_list);
5872 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5877 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5880 req_set_fail_links(prev);
5881 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5884 io_req_complete(req, -ETIME);
5886 return HRTIMER_NORESTART;
5889 static void io_queue_linked_timeout(struct io_kiocb *req)
5891 struct io_ring_ctx *ctx = req->ctx;
5894 * If the list is now empty, then our linked request finished before
5895 * we got a chance to setup the timer
5897 spin_lock_irq(&ctx->completion_lock);
5898 if (!list_empty(&req->link_list)) {
5899 struct io_timeout_data *data = &req->io->timeout;
5901 data->timer.function = io_link_timeout_fn;
5902 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5905 spin_unlock_irq(&ctx->completion_lock);
5907 /* drop submission reference */
5911 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5913 struct io_kiocb *nxt;
5915 if (!(req->flags & REQ_F_LINK_HEAD))
5917 if (req->flags & REQ_F_LINK_TIMEOUT)
5920 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5922 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5925 req->flags |= REQ_F_LINK_TIMEOUT;
5929 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5930 struct io_comp_state *cs)
5932 struct io_kiocb *linked_timeout;
5933 struct io_kiocb *nxt;
5934 const struct cred *old_creds = NULL;
5938 linked_timeout = io_prep_linked_timeout(req);
5940 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5941 req->work.creds != current_cred()) {
5943 revert_creds(old_creds);
5944 if (old_creds == req->work.creds)
5945 old_creds = NULL; /* restored original creds */
5947 old_creds = override_creds(req->work.creds);
5950 ret = io_issue_sqe(req, sqe, true, cs);
5953 * We async punt it if the file wasn't marked NOWAIT, or if the file
5954 * doesn't support non-blocking read/write attempts
5956 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5957 if (io_arm_poll_handler(req)) {
5959 io_queue_linked_timeout(linked_timeout);
5963 io_req_init_async(req);
5965 if (io_op_defs[req->opcode].file_table) {
5966 ret = io_grab_files(req);
5972 * Queued up for async execution, worker will release
5973 * submit reference when the iocb is actually submitted.
5975 io_queue_async_work(req);
5979 if (unlikely(ret)) {
5981 /* un-prep timeout, so it'll be killed as any other linked */
5982 req->flags &= ~REQ_F_LINK_TIMEOUT;
5983 req_set_fail_links(req);
5985 io_req_complete(req, ret);
5989 /* drop submission reference */
5990 nxt = io_put_req_find_next(req);
5992 io_queue_linked_timeout(linked_timeout);
5997 if (req->flags & REQ_F_FORCE_ASYNC)
6003 revert_creds(old_creds);
6006 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6007 struct io_comp_state *cs)
6011 ret = io_req_defer(req, sqe);
6013 if (ret != -EIOCBQUEUED) {
6015 req_set_fail_links(req);
6017 io_req_complete(req, ret);
6019 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6021 ret = io_req_defer_prep(req, sqe);
6027 * Never try inline submit of IOSQE_ASYNC is set, go straight
6028 * to async execution.
6030 io_req_init_async(req);
6031 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6032 io_queue_async_work(req);
6034 __io_queue_sqe(req, sqe, cs);
6038 static inline void io_queue_link_head(struct io_kiocb *req,
6039 struct io_comp_state *cs)
6041 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6043 io_req_complete(req, -ECANCELED);
6045 io_queue_sqe(req, NULL, cs);
6048 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6049 struct io_kiocb **link, struct io_comp_state *cs)
6051 struct io_ring_ctx *ctx = req->ctx;
6055 * If we already have a head request, queue this one for async
6056 * submittal once the head completes. If we don't have a head but
6057 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6058 * submitted sync once the chain is complete. If none of those
6059 * conditions are true (normal request), then just queue it.
6062 struct io_kiocb *head = *link;
6065 * Taking sequential execution of a link, draining both sides
6066 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6067 * requests in the link. So, it drains the head and the
6068 * next after the link request. The last one is done via
6069 * drain_next flag to persist the effect across calls.
6071 if (req->flags & REQ_F_IO_DRAIN) {
6072 head->flags |= REQ_F_IO_DRAIN;
6073 ctx->drain_next = 1;
6075 ret = io_req_defer_prep(req, sqe);
6076 if (unlikely(ret)) {
6077 /* fail even hard links since we don't submit */
6078 head->flags |= REQ_F_FAIL_LINK;
6081 trace_io_uring_link(ctx, req, head);
6082 io_get_req_task(req);
6083 list_add_tail(&req->link_list, &head->link_list);
6085 /* last request of a link, enqueue the link */
6086 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6087 io_queue_link_head(head, cs);
6091 if (unlikely(ctx->drain_next)) {
6092 req->flags |= REQ_F_IO_DRAIN;
6093 ctx->drain_next = 0;
6095 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6096 req->flags |= REQ_F_LINK_HEAD;
6097 INIT_LIST_HEAD(&req->link_list);
6099 ret = io_req_defer_prep(req, sqe);
6101 req->flags |= REQ_F_FAIL_LINK;
6104 io_queue_sqe(req, sqe, cs);
6112 * Batched submission is done, ensure local IO is flushed out.
6114 static void io_submit_state_end(struct io_submit_state *state)
6116 if (!list_empty(&state->comp.list))
6117 io_submit_flush_completions(&state->comp);
6118 blk_finish_plug(&state->plug);
6119 io_state_file_put(state);
6120 if (state->free_reqs)
6121 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6125 * Start submission side cache.
6127 static void io_submit_state_start(struct io_submit_state *state,
6128 struct io_ring_ctx *ctx, unsigned int max_ios)
6130 blk_start_plug(&state->plug);
6132 state->plug.nowait = true;
6135 INIT_LIST_HEAD(&state->comp.list);
6136 state->comp.ctx = ctx;
6137 state->free_reqs = 0;
6139 state->ios_left = max_ios;
6142 static void io_commit_sqring(struct io_ring_ctx *ctx)
6144 struct io_rings *rings = ctx->rings;
6147 * Ensure any loads from the SQEs are done at this point,
6148 * since once we write the new head, the application could
6149 * write new data to them.
6151 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6155 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6156 * that is mapped by userspace. This means that care needs to be taken to
6157 * ensure that reads are stable, as we cannot rely on userspace always
6158 * being a good citizen. If members of the sqe are validated and then later
6159 * used, it's important that those reads are done through READ_ONCE() to
6160 * prevent a re-load down the line.
6162 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6164 u32 *sq_array = ctx->sq_array;
6168 * The cached sq head (or cq tail) serves two purposes:
6170 * 1) allows us to batch the cost of updating the user visible
6172 * 2) allows the kernel side to track the head on its own, even
6173 * though the application is the one updating it.
6175 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6176 if (likely(head < ctx->sq_entries))
6177 return &ctx->sq_sqes[head];
6179 /* drop invalid entries */
6180 ctx->cached_sq_dropped++;
6181 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6185 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6187 ctx->cached_sq_head++;
6190 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6191 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6192 IOSQE_BUFFER_SELECT)
6194 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6195 const struct io_uring_sqe *sqe,
6196 struct io_submit_state *state)
6198 unsigned int sqe_flags;
6201 req->opcode = READ_ONCE(sqe->opcode);
6202 req->user_data = READ_ONCE(sqe->user_data);
6207 /* one is dropped after submission, the other at completion */
6208 refcount_set(&req->refs, 2);
6209 req->task = current;
6212 if (unlikely(req->opcode >= IORING_OP_LAST))
6215 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6218 sqe_flags = READ_ONCE(sqe->flags);
6219 /* enforce forwards compatibility on users */
6220 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6223 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6224 !io_op_defs[req->opcode].buffer_select)
6227 id = READ_ONCE(sqe->personality);
6229 io_req_init_async(req);
6230 req->work.creds = idr_find(&ctx->personality_idr, id);
6231 if (unlikely(!req->work.creds))
6233 get_cred(req->work.creds);
6236 /* same numerical values with corresponding REQ_F_*, safe to copy */
6237 req->flags |= sqe_flags;
6239 if (!io_op_defs[req->opcode].needs_file)
6242 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6245 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6246 struct file *ring_file, int ring_fd)
6248 struct io_submit_state state;
6249 struct io_kiocb *link = NULL;
6250 int i, submitted = 0;
6252 /* if we have a backlog and couldn't flush it all, return BUSY */
6253 if (test_bit(0, &ctx->sq_check_overflow)) {
6254 if (!list_empty(&ctx->cq_overflow_list) &&
6255 !io_cqring_overflow_flush(ctx, false))
6259 /* make sure SQ entry isn't read before tail */
6260 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6262 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6265 io_submit_state_start(&state, ctx, nr);
6267 ctx->ring_fd = ring_fd;
6268 ctx->ring_file = ring_file;
6270 for (i = 0; i < nr; i++) {
6271 const struct io_uring_sqe *sqe;
6272 struct io_kiocb *req;
6275 sqe = io_get_sqe(ctx);
6276 if (unlikely(!sqe)) {
6277 io_consume_sqe(ctx);
6280 req = io_alloc_req(ctx, &state);
6281 if (unlikely(!req)) {
6283 submitted = -EAGAIN;
6287 err = io_init_req(ctx, req, sqe, &state);
6288 io_consume_sqe(ctx);
6289 /* will complete beyond this point, count as submitted */
6292 if (unlikely(err)) {
6295 io_req_complete(req, err);
6299 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6300 true, io_async_submit(ctx));
6301 err = io_submit_sqe(req, sqe, &link, &state.comp);
6306 if (unlikely(submitted != nr)) {
6307 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6309 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6312 io_queue_link_head(link, &state.comp);
6313 io_submit_state_end(&state);
6315 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6316 io_commit_sqring(ctx);
6321 static int io_sq_thread(void *data)
6323 struct io_ring_ctx *ctx = data;
6324 const struct cred *old_cred;
6326 unsigned long timeout;
6329 complete(&ctx->sq_thread_comp);
6331 old_cred = override_creds(ctx->creds);
6333 timeout = jiffies + ctx->sq_thread_idle;
6334 while (!kthread_should_park()) {
6335 unsigned int to_submit;
6337 if (!list_empty(&ctx->iopoll_list)) {
6338 unsigned nr_events = 0;
6340 mutex_lock(&ctx->uring_lock);
6341 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6342 io_do_iopoll(ctx, &nr_events, 0);
6344 timeout = jiffies + ctx->sq_thread_idle;
6345 mutex_unlock(&ctx->uring_lock);
6348 to_submit = io_sqring_entries(ctx);
6351 * If submit got -EBUSY, flag us as needing the application
6352 * to enter the kernel to reap and flush events.
6354 if (!to_submit || ret == -EBUSY || need_resched()) {
6356 * Drop cur_mm before scheduling, we can't hold it for
6357 * long periods (or over schedule()). Do this before
6358 * adding ourselves to the waitqueue, as the unuse/drop
6361 io_sq_thread_drop_mm();
6364 * We're polling. If we're within the defined idle
6365 * period, then let us spin without work before going
6366 * to sleep. The exception is if we got EBUSY doing
6367 * more IO, we should wait for the application to
6368 * reap events and wake us up.
6370 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6371 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6372 !percpu_ref_is_dying(&ctx->refs))) {
6378 prepare_to_wait(&ctx->sqo_wait, &wait,
6379 TASK_INTERRUPTIBLE);
6382 * While doing polled IO, before going to sleep, we need
6383 * to check if there are new reqs added to iopoll_list,
6384 * it is because reqs may have been punted to io worker
6385 * and will be added to iopoll_list later, hence check
6386 * the iopoll_list again.
6388 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6389 !list_empty_careful(&ctx->iopoll_list)) {
6390 finish_wait(&ctx->sqo_wait, &wait);
6394 /* Tell userspace we may need a wakeup call */
6395 spin_lock_irq(&ctx->completion_lock);
6396 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6397 spin_unlock_irq(&ctx->completion_lock);
6399 to_submit = io_sqring_entries(ctx);
6400 if (!to_submit || ret == -EBUSY) {
6401 if (kthread_should_park()) {
6402 finish_wait(&ctx->sqo_wait, &wait);
6405 if (io_run_task_work()) {
6406 finish_wait(&ctx->sqo_wait, &wait);
6409 if (signal_pending(current))
6410 flush_signals(current);
6412 finish_wait(&ctx->sqo_wait, &wait);
6414 spin_lock_irq(&ctx->completion_lock);
6415 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6416 spin_unlock_irq(&ctx->completion_lock);
6420 finish_wait(&ctx->sqo_wait, &wait);
6422 spin_lock_irq(&ctx->completion_lock);
6423 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6424 spin_unlock_irq(&ctx->completion_lock);
6427 mutex_lock(&ctx->uring_lock);
6428 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6429 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6430 mutex_unlock(&ctx->uring_lock);
6431 timeout = jiffies + ctx->sq_thread_idle;
6436 io_sq_thread_drop_mm();
6437 revert_creds(old_cred);
6444 struct io_wait_queue {
6445 struct wait_queue_entry wq;
6446 struct io_ring_ctx *ctx;
6448 unsigned nr_timeouts;
6451 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6453 struct io_ring_ctx *ctx = iowq->ctx;
6456 * Wake up if we have enough events, or if a timeout occurred since we
6457 * started waiting. For timeouts, we always want to return to userspace,
6458 * regardless of event count.
6460 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6461 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6464 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6465 int wake_flags, void *key)
6467 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6470 /* use noflush == true, as we can't safely rely on locking context */
6471 if (!io_should_wake(iowq, true))
6474 return autoremove_wake_function(curr, mode, wake_flags, key);
6478 * Wait until events become available, if we don't already have some. The
6479 * application must reap them itself, as they reside on the shared cq ring.
6481 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6482 const sigset_t __user *sig, size_t sigsz)
6484 struct io_wait_queue iowq = {
6487 .func = io_wake_function,
6488 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6491 .to_wait = min_events,
6493 struct io_rings *rings = ctx->rings;
6497 if (io_cqring_events(ctx, false) >= min_events)
6499 if (!io_run_task_work())
6504 #ifdef CONFIG_COMPAT
6505 if (in_compat_syscall())
6506 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6510 ret = set_user_sigmask(sig, sigsz);
6516 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6517 trace_io_uring_cqring_wait(ctx, min_events);
6519 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6520 TASK_INTERRUPTIBLE);
6521 /* make sure we run task_work before checking for signals */
6522 if (io_run_task_work())
6524 if (signal_pending(current)) {
6525 if (current->jobctl & JOBCTL_TASK_WORK) {
6526 spin_lock_irq(¤t->sighand->siglock);
6527 current->jobctl &= ~JOBCTL_TASK_WORK;
6528 recalc_sigpending();
6529 spin_unlock_irq(¤t->sighand->siglock);
6535 if (io_should_wake(&iowq, false))
6539 finish_wait(&ctx->wait, &iowq.wq);
6541 restore_saved_sigmask_unless(ret == -EINTR);
6543 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6546 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6548 #if defined(CONFIG_UNIX)
6549 if (ctx->ring_sock) {
6550 struct sock *sock = ctx->ring_sock->sk;
6551 struct sk_buff *skb;
6553 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6559 for (i = 0; i < ctx->nr_user_files; i++) {
6562 file = io_file_from_index(ctx, i);
6569 static void io_file_ref_kill(struct percpu_ref *ref)
6571 struct fixed_file_data *data;
6573 data = container_of(ref, struct fixed_file_data, refs);
6574 complete(&data->done);
6577 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6579 struct fixed_file_data *data = ctx->file_data;
6580 struct fixed_file_ref_node *ref_node = NULL;
6581 unsigned nr_tables, i;
6586 spin_lock(&data->lock);
6587 if (!list_empty(&data->ref_list))
6588 ref_node = list_first_entry(&data->ref_list,
6589 struct fixed_file_ref_node, node);
6590 spin_unlock(&data->lock);
6592 percpu_ref_kill(&ref_node->refs);
6594 percpu_ref_kill(&data->refs);
6596 /* wait for all refs nodes to complete */
6597 flush_delayed_work(&ctx->file_put_work);
6598 wait_for_completion(&data->done);
6600 __io_sqe_files_unregister(ctx);
6601 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6602 for (i = 0; i < nr_tables; i++)
6603 kfree(data->table[i].files);
6605 percpu_ref_exit(&data->refs);
6607 ctx->file_data = NULL;
6608 ctx->nr_user_files = 0;
6612 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6614 if (ctx->sqo_thread) {
6615 wait_for_completion(&ctx->sq_thread_comp);
6617 * The park is a bit of a work-around, without it we get
6618 * warning spews on shutdown with SQPOLL set and affinity
6619 * set to a single CPU.
6621 kthread_park(ctx->sqo_thread);
6622 kthread_stop(ctx->sqo_thread);
6623 ctx->sqo_thread = NULL;
6627 static void io_finish_async(struct io_ring_ctx *ctx)
6629 io_sq_thread_stop(ctx);
6632 io_wq_destroy(ctx->io_wq);
6637 #if defined(CONFIG_UNIX)
6639 * Ensure the UNIX gc is aware of our file set, so we are certain that
6640 * the io_uring can be safely unregistered on process exit, even if we have
6641 * loops in the file referencing.
6643 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6645 struct sock *sk = ctx->ring_sock->sk;
6646 struct scm_fp_list *fpl;
6647 struct sk_buff *skb;
6650 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6654 skb = alloc_skb(0, GFP_KERNEL);
6663 fpl->user = get_uid(ctx->user);
6664 for (i = 0; i < nr; i++) {
6665 struct file *file = io_file_from_index(ctx, i + offset);
6669 fpl->fp[nr_files] = get_file(file);
6670 unix_inflight(fpl->user, fpl->fp[nr_files]);
6675 fpl->max = SCM_MAX_FD;
6676 fpl->count = nr_files;
6677 UNIXCB(skb).fp = fpl;
6678 skb->destructor = unix_destruct_scm;
6679 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6680 skb_queue_head(&sk->sk_receive_queue, skb);
6682 for (i = 0; i < nr_files; i++)
6693 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6694 * causes regular reference counting to break down. We rely on the UNIX
6695 * garbage collection to take care of this problem for us.
6697 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6699 unsigned left, total;
6703 left = ctx->nr_user_files;
6705 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6707 ret = __io_sqe_files_scm(ctx, this_files, total);
6711 total += this_files;
6717 while (total < ctx->nr_user_files) {
6718 struct file *file = io_file_from_index(ctx, total);
6728 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6734 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6739 for (i = 0; i < nr_tables; i++) {
6740 struct fixed_file_table *table = &ctx->file_data->table[i];
6741 unsigned this_files;
6743 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6744 table->files = kcalloc(this_files, sizeof(struct file *),
6748 nr_files -= this_files;
6754 for (i = 0; i < nr_tables; i++) {
6755 struct fixed_file_table *table = &ctx->file_data->table[i];
6756 kfree(table->files);
6761 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6763 #if defined(CONFIG_UNIX)
6764 struct sock *sock = ctx->ring_sock->sk;
6765 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6766 struct sk_buff *skb;
6769 __skb_queue_head_init(&list);
6772 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6773 * remove this entry and rearrange the file array.
6775 skb = skb_dequeue(head);
6777 struct scm_fp_list *fp;
6779 fp = UNIXCB(skb).fp;
6780 for (i = 0; i < fp->count; i++) {
6783 if (fp->fp[i] != file)
6786 unix_notinflight(fp->user, fp->fp[i]);
6787 left = fp->count - 1 - i;
6789 memmove(&fp->fp[i], &fp->fp[i + 1],
6790 left * sizeof(struct file *));
6797 __skb_queue_tail(&list, skb);
6807 __skb_queue_tail(&list, skb);
6809 skb = skb_dequeue(head);
6812 if (skb_peek(&list)) {
6813 spin_lock_irq(&head->lock);
6814 while ((skb = __skb_dequeue(&list)) != NULL)
6815 __skb_queue_tail(head, skb);
6816 spin_unlock_irq(&head->lock);
6823 struct io_file_put {
6824 struct list_head list;
6828 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6830 struct fixed_file_data *file_data = ref_node->file_data;
6831 struct io_ring_ctx *ctx = file_data->ctx;
6832 struct io_file_put *pfile, *tmp;
6834 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6835 list_del(&pfile->list);
6836 io_ring_file_put(ctx, pfile->file);
6840 spin_lock(&file_data->lock);
6841 list_del(&ref_node->node);
6842 spin_unlock(&file_data->lock);
6844 percpu_ref_exit(&ref_node->refs);
6846 percpu_ref_put(&file_data->refs);
6849 static void io_file_put_work(struct work_struct *work)
6851 struct io_ring_ctx *ctx;
6852 struct llist_node *node;
6854 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6855 node = llist_del_all(&ctx->file_put_llist);
6858 struct fixed_file_ref_node *ref_node;
6859 struct llist_node *next = node->next;
6861 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6862 __io_file_put_work(ref_node);
6867 static void io_file_data_ref_zero(struct percpu_ref *ref)
6869 struct fixed_file_ref_node *ref_node;
6870 struct io_ring_ctx *ctx;
6874 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6875 ctx = ref_node->file_data->ctx;
6877 if (percpu_ref_is_dying(&ctx->file_data->refs))
6880 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6882 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6884 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6887 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6888 struct io_ring_ctx *ctx)
6890 struct fixed_file_ref_node *ref_node;
6892 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6894 return ERR_PTR(-ENOMEM);
6896 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6899 return ERR_PTR(-ENOMEM);
6901 INIT_LIST_HEAD(&ref_node->node);
6902 INIT_LIST_HEAD(&ref_node->file_list);
6903 ref_node->file_data = ctx->file_data;
6907 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6909 percpu_ref_exit(&ref_node->refs);
6913 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6916 __s32 __user *fds = (__s32 __user *) arg;
6921 struct fixed_file_ref_node *ref_node;
6927 if (nr_args > IORING_MAX_FIXED_FILES)
6930 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6931 if (!ctx->file_data)
6933 ctx->file_data->ctx = ctx;
6934 init_completion(&ctx->file_data->done);
6935 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6936 spin_lock_init(&ctx->file_data->lock);
6938 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6939 ctx->file_data->table = kcalloc(nr_tables,
6940 sizeof(struct fixed_file_table),
6942 if (!ctx->file_data->table) {
6943 kfree(ctx->file_data);
6944 ctx->file_data = NULL;
6948 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6949 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6950 kfree(ctx->file_data->table);
6951 kfree(ctx->file_data);
6952 ctx->file_data = NULL;
6956 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6957 percpu_ref_exit(&ctx->file_data->refs);
6958 kfree(ctx->file_data->table);
6959 kfree(ctx->file_data);
6960 ctx->file_data = NULL;
6964 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6965 struct fixed_file_table *table;
6969 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6971 /* allow sparse sets */
6977 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6978 index = i & IORING_FILE_TABLE_MASK;
6986 * Don't allow io_uring instances to be registered. If UNIX
6987 * isn't enabled, then this causes a reference cycle and this
6988 * instance can never get freed. If UNIX is enabled we'll
6989 * handle it just fine, but there's still no point in allowing
6990 * a ring fd as it doesn't support regular read/write anyway.
6992 if (file->f_op == &io_uring_fops) {
6997 table->files[index] = file;
7001 for (i = 0; i < ctx->nr_user_files; i++) {
7002 file = io_file_from_index(ctx, i);
7006 for (i = 0; i < nr_tables; i++)
7007 kfree(ctx->file_data->table[i].files);
7009 percpu_ref_exit(&ctx->file_data->refs);
7010 kfree(ctx->file_data->table);
7011 kfree(ctx->file_data);
7012 ctx->file_data = NULL;
7013 ctx->nr_user_files = 0;
7017 ret = io_sqe_files_scm(ctx);
7019 io_sqe_files_unregister(ctx);
7023 ref_node = alloc_fixed_file_ref_node(ctx);
7024 if (IS_ERR(ref_node)) {
7025 io_sqe_files_unregister(ctx);
7026 return PTR_ERR(ref_node);
7029 ctx->file_data->cur_refs = &ref_node->refs;
7030 spin_lock(&ctx->file_data->lock);
7031 list_add(&ref_node->node, &ctx->file_data->ref_list);
7032 spin_unlock(&ctx->file_data->lock);
7033 percpu_ref_get(&ctx->file_data->refs);
7037 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7040 #if defined(CONFIG_UNIX)
7041 struct sock *sock = ctx->ring_sock->sk;
7042 struct sk_buff_head *head = &sock->sk_receive_queue;
7043 struct sk_buff *skb;
7046 * See if we can merge this file into an existing skb SCM_RIGHTS
7047 * file set. If there's no room, fall back to allocating a new skb
7048 * and filling it in.
7050 spin_lock_irq(&head->lock);
7051 skb = skb_peek(head);
7053 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7055 if (fpl->count < SCM_MAX_FD) {
7056 __skb_unlink(skb, head);
7057 spin_unlock_irq(&head->lock);
7058 fpl->fp[fpl->count] = get_file(file);
7059 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7061 spin_lock_irq(&head->lock);
7062 __skb_queue_head(head, skb);
7067 spin_unlock_irq(&head->lock);
7074 return __io_sqe_files_scm(ctx, 1, index);
7080 static int io_queue_file_removal(struct fixed_file_data *data,
7083 struct io_file_put *pfile;
7084 struct percpu_ref *refs = data->cur_refs;
7085 struct fixed_file_ref_node *ref_node;
7087 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7091 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7093 list_add(&pfile->list, &ref_node->file_list);
7098 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7099 struct io_uring_files_update *up,
7102 struct fixed_file_data *data = ctx->file_data;
7103 struct fixed_file_ref_node *ref_node;
7108 bool needs_switch = false;
7110 if (check_add_overflow(up->offset, nr_args, &done))
7112 if (done > ctx->nr_user_files)
7115 ref_node = alloc_fixed_file_ref_node(ctx);
7116 if (IS_ERR(ref_node))
7117 return PTR_ERR(ref_node);
7120 fds = u64_to_user_ptr(up->fds);
7122 struct fixed_file_table *table;
7126 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7130 i = array_index_nospec(up->offset, ctx->nr_user_files);
7131 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7132 index = i & IORING_FILE_TABLE_MASK;
7133 if (table->files[index]) {
7134 file = io_file_from_index(ctx, index);
7135 err = io_queue_file_removal(data, file);
7138 table->files[index] = NULL;
7139 needs_switch = true;
7148 * Don't allow io_uring instances to be registered. If
7149 * UNIX isn't enabled, then this causes a reference
7150 * cycle and this instance can never get freed. If UNIX
7151 * is enabled we'll handle it just fine, but there's
7152 * still no point in allowing a ring fd as it doesn't
7153 * support regular read/write anyway.
7155 if (file->f_op == &io_uring_fops) {
7160 table->files[index] = file;
7161 err = io_sqe_file_register(ctx, file, i);
7173 percpu_ref_kill(data->cur_refs);
7174 spin_lock(&data->lock);
7175 list_add(&ref_node->node, &data->ref_list);
7176 data->cur_refs = &ref_node->refs;
7177 spin_unlock(&data->lock);
7178 percpu_ref_get(&ctx->file_data->refs);
7180 destroy_fixed_file_ref_node(ref_node);
7182 return done ? done : err;
7185 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7188 struct io_uring_files_update up;
7190 if (!ctx->file_data)
7194 if (copy_from_user(&up, arg, sizeof(up)))
7199 return __io_sqe_files_update(ctx, &up, nr_args);
7202 static void io_free_work(struct io_wq_work *work)
7204 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7206 /* Consider that io_steal_work() relies on this ref */
7210 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7211 struct io_uring_params *p)
7213 struct io_wq_data data;
7215 struct io_ring_ctx *ctx_attach;
7216 unsigned int concurrency;
7219 data.user = ctx->user;
7220 data.free_work = io_free_work;
7221 data.do_work = io_wq_submit_work;
7223 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7224 /* Do QD, or 4 * CPUS, whatever is smallest */
7225 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7227 ctx->io_wq = io_wq_create(concurrency, &data);
7228 if (IS_ERR(ctx->io_wq)) {
7229 ret = PTR_ERR(ctx->io_wq);
7235 f = fdget(p->wq_fd);
7239 if (f.file->f_op != &io_uring_fops) {
7244 ctx_attach = f.file->private_data;
7245 /* @io_wq is protected by holding the fd */
7246 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7251 ctx->io_wq = ctx_attach->io_wq;
7257 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7258 struct io_uring_params *p)
7262 if (ctx->flags & IORING_SETUP_SQPOLL) {
7263 mmgrab(current->mm);
7264 ctx->sqo_mm = current->mm;
7267 if (!capable(CAP_SYS_ADMIN))
7270 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7271 if (!ctx->sq_thread_idle)
7272 ctx->sq_thread_idle = HZ;
7274 if (p->flags & IORING_SETUP_SQ_AFF) {
7275 int cpu = p->sq_thread_cpu;
7278 if (cpu >= nr_cpu_ids)
7280 if (!cpu_online(cpu))
7283 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7287 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7290 if (IS_ERR(ctx->sqo_thread)) {
7291 ret = PTR_ERR(ctx->sqo_thread);
7292 ctx->sqo_thread = NULL;
7295 wake_up_process(ctx->sqo_thread);
7296 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7297 /* Can't have SQ_AFF without SQPOLL */
7302 ret = io_init_wq_offload(ctx, p);
7308 io_finish_async(ctx);
7310 mmdrop(ctx->sqo_mm);
7316 static inline void __io_unaccount_mem(struct user_struct *user,
7317 unsigned long nr_pages)
7319 atomic_long_sub(nr_pages, &user->locked_vm);
7322 static inline int __io_account_mem(struct user_struct *user,
7323 unsigned long nr_pages)
7325 unsigned long page_limit, cur_pages, new_pages;
7327 /* Don't allow more pages than we can safely lock */
7328 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7331 cur_pages = atomic_long_read(&user->locked_vm);
7332 new_pages = cur_pages + nr_pages;
7333 if (new_pages > page_limit)
7335 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7336 new_pages) != cur_pages);
7341 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7342 enum io_mem_account acct)
7345 __io_unaccount_mem(ctx->user, nr_pages);
7348 if (acct == ACCT_LOCKED)
7349 ctx->sqo_mm->locked_vm -= nr_pages;
7350 else if (acct == ACCT_PINNED)
7351 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7355 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7356 enum io_mem_account acct)
7360 if (ctx->limit_mem) {
7361 ret = __io_account_mem(ctx->user, nr_pages);
7367 if (acct == ACCT_LOCKED)
7368 ctx->sqo_mm->locked_vm += nr_pages;
7369 else if (acct == ACCT_PINNED)
7370 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7376 static void io_mem_free(void *ptr)
7383 page = virt_to_head_page(ptr);
7384 if (put_page_testzero(page))
7385 free_compound_page(page);
7388 static void *io_mem_alloc(size_t size)
7390 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7393 return (void *) __get_free_pages(gfp_flags, get_order(size));
7396 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7399 struct io_rings *rings;
7400 size_t off, sq_array_size;
7402 off = struct_size(rings, cqes, cq_entries);
7403 if (off == SIZE_MAX)
7407 off = ALIGN(off, SMP_CACHE_BYTES);
7415 sq_array_size = array_size(sizeof(u32), sq_entries);
7416 if (sq_array_size == SIZE_MAX)
7419 if (check_add_overflow(off, sq_array_size, &off))
7425 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7429 pages = (size_t)1 << get_order(
7430 rings_size(sq_entries, cq_entries, NULL));
7431 pages += (size_t)1 << get_order(
7432 array_size(sizeof(struct io_uring_sqe), sq_entries));
7437 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7441 if (!ctx->user_bufs)
7444 for (i = 0; i < ctx->nr_user_bufs; i++) {
7445 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7447 for (j = 0; j < imu->nr_bvecs; j++)
7448 unpin_user_page(imu->bvec[j].bv_page);
7450 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7455 kfree(ctx->user_bufs);
7456 ctx->user_bufs = NULL;
7457 ctx->nr_user_bufs = 0;
7461 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7462 void __user *arg, unsigned index)
7464 struct iovec __user *src;
7466 #ifdef CONFIG_COMPAT
7468 struct compat_iovec __user *ciovs;
7469 struct compat_iovec ciov;
7471 ciovs = (struct compat_iovec __user *) arg;
7472 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7475 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7476 dst->iov_len = ciov.iov_len;
7480 src = (struct iovec __user *) arg;
7481 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7486 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7489 struct vm_area_struct **vmas = NULL;
7490 struct page **pages = NULL;
7491 int i, j, got_pages = 0;
7496 if (!nr_args || nr_args > UIO_MAXIOV)
7499 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7501 if (!ctx->user_bufs)
7504 for (i = 0; i < nr_args; i++) {
7505 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7506 unsigned long off, start, end, ubuf;
7511 ret = io_copy_iov(ctx, &iov, arg, i);
7516 * Don't impose further limits on the size and buffer
7517 * constraints here, we'll -EINVAL later when IO is
7518 * submitted if they are wrong.
7521 if (!iov.iov_base || !iov.iov_len)
7524 /* arbitrary limit, but we need something */
7525 if (iov.iov_len > SZ_1G)
7528 ubuf = (unsigned long) iov.iov_base;
7529 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7530 start = ubuf >> PAGE_SHIFT;
7531 nr_pages = end - start;
7533 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7538 if (!pages || nr_pages > got_pages) {
7541 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7543 vmas = kvmalloc_array(nr_pages,
7544 sizeof(struct vm_area_struct *),
7546 if (!pages || !vmas) {
7548 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7551 got_pages = nr_pages;
7554 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7558 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7563 mmap_read_lock(current->mm);
7564 pret = pin_user_pages(ubuf, nr_pages,
7565 FOLL_WRITE | FOLL_LONGTERM,
7567 if (pret == nr_pages) {
7568 /* don't support file backed memory */
7569 for (j = 0; j < nr_pages; j++) {
7570 struct vm_area_struct *vma = vmas[j];
7573 !is_file_hugepages(vma->vm_file)) {
7579 ret = pret < 0 ? pret : -EFAULT;
7581 mmap_read_unlock(current->mm);
7584 * if we did partial map, or found file backed vmas,
7585 * release any pages we did get
7588 unpin_user_pages(pages, pret);
7589 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7594 off = ubuf & ~PAGE_MASK;
7596 for (j = 0; j < nr_pages; j++) {
7599 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7600 imu->bvec[j].bv_page = pages[j];
7601 imu->bvec[j].bv_len = vec_len;
7602 imu->bvec[j].bv_offset = off;
7606 /* store original address for later verification */
7608 imu->len = iov.iov_len;
7609 imu->nr_bvecs = nr_pages;
7611 ctx->nr_user_bufs++;
7619 io_sqe_buffer_unregister(ctx);
7623 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7625 __s32 __user *fds = arg;
7631 if (copy_from_user(&fd, fds, sizeof(*fds)))
7634 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7635 if (IS_ERR(ctx->cq_ev_fd)) {
7636 int ret = PTR_ERR(ctx->cq_ev_fd);
7637 ctx->cq_ev_fd = NULL;
7644 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7646 if (ctx->cq_ev_fd) {
7647 eventfd_ctx_put(ctx->cq_ev_fd);
7648 ctx->cq_ev_fd = NULL;
7655 static int __io_destroy_buffers(int id, void *p, void *data)
7657 struct io_ring_ctx *ctx = data;
7658 struct io_buffer *buf = p;
7660 __io_remove_buffers(ctx, buf, id, -1U);
7664 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7666 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7667 idr_destroy(&ctx->io_buffer_idr);
7670 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7672 io_finish_async(ctx);
7674 mmdrop(ctx->sqo_mm);
7678 io_sqe_buffer_unregister(ctx);
7679 io_sqe_files_unregister(ctx);
7680 io_eventfd_unregister(ctx);
7681 io_destroy_buffers(ctx);
7682 idr_destroy(&ctx->personality_idr);
7684 #if defined(CONFIG_UNIX)
7685 if (ctx->ring_sock) {
7686 ctx->ring_sock->file = NULL; /* so that iput() is called */
7687 sock_release(ctx->ring_sock);
7691 io_mem_free(ctx->rings);
7692 io_mem_free(ctx->sq_sqes);
7694 percpu_ref_exit(&ctx->refs);
7695 free_uid(ctx->user);
7696 put_cred(ctx->creds);
7697 kfree(ctx->cancel_hash);
7698 kmem_cache_free(req_cachep, ctx->fallback_req);
7702 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7704 struct io_ring_ctx *ctx = file->private_data;
7707 poll_wait(file, &ctx->cq_wait, wait);
7709 * synchronizes with barrier from wq_has_sleeper call in
7713 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7714 ctx->rings->sq_ring_entries)
7715 mask |= EPOLLOUT | EPOLLWRNORM;
7716 if (io_cqring_events(ctx, false))
7717 mask |= EPOLLIN | EPOLLRDNORM;
7722 static int io_uring_fasync(int fd, struct file *file, int on)
7724 struct io_ring_ctx *ctx = file->private_data;
7726 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7729 static int io_remove_personalities(int id, void *p, void *data)
7731 struct io_ring_ctx *ctx = data;
7732 const struct cred *cred;
7734 cred = idr_remove(&ctx->personality_idr, id);
7740 static void io_ring_exit_work(struct work_struct *work)
7742 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7746 * If we're doing polled IO and end up having requests being
7747 * submitted async (out-of-line), then completions can come in while
7748 * we're waiting for refs to drop. We need to reap these manually,
7749 * as nobody else will be looking for them.
7753 io_cqring_overflow_flush(ctx, true);
7754 io_iopoll_try_reap_events(ctx);
7755 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7756 io_ring_ctx_free(ctx);
7759 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7761 mutex_lock(&ctx->uring_lock);
7762 percpu_ref_kill(&ctx->refs);
7763 mutex_unlock(&ctx->uring_lock);
7765 io_kill_timeouts(ctx);
7766 io_poll_remove_all(ctx);
7769 io_wq_cancel_all(ctx->io_wq);
7771 /* if we failed setting up the ctx, we might not have any rings */
7773 io_cqring_overflow_flush(ctx, true);
7774 io_iopoll_try_reap_events(ctx);
7775 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7778 * Do this upfront, so we won't have a grace period where the ring
7779 * is closed but resources aren't reaped yet. This can cause
7780 * spurious failure in setting up a new ring.
7782 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7785 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7786 queue_work(system_wq, &ctx->exit_work);
7789 static int io_uring_release(struct inode *inode, struct file *file)
7791 struct io_ring_ctx *ctx = file->private_data;
7793 file->private_data = NULL;
7794 io_ring_ctx_wait_and_kill(ctx);
7798 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7800 struct files_struct *files = data;
7802 return work->files == files;
7805 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7806 struct files_struct *files)
7808 if (list_empty_careful(&ctx->inflight_list))
7811 /* cancel all at once, should be faster than doing it one by one*/
7812 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7814 while (!list_empty_careful(&ctx->inflight_list)) {
7815 struct io_kiocb *cancel_req = NULL, *req;
7818 spin_lock_irq(&ctx->inflight_lock);
7819 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7820 if (req->work.files != files)
7822 /* req is being completed, ignore */
7823 if (!refcount_inc_not_zero(&req->refs))
7829 prepare_to_wait(&ctx->inflight_wait, &wait,
7830 TASK_UNINTERRUPTIBLE);
7831 spin_unlock_irq(&ctx->inflight_lock);
7833 /* We need to keep going until we don't find a matching req */
7837 if (cancel_req->flags & REQ_F_OVERFLOW) {
7838 spin_lock_irq(&ctx->completion_lock);
7839 list_del(&cancel_req->compl.list);
7840 cancel_req->flags &= ~REQ_F_OVERFLOW;
7841 if (list_empty(&ctx->cq_overflow_list)) {
7842 clear_bit(0, &ctx->sq_check_overflow);
7843 clear_bit(0, &ctx->cq_check_overflow);
7844 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
7846 spin_unlock_irq(&ctx->completion_lock);
7848 WRITE_ONCE(ctx->rings->cq_overflow,
7849 atomic_inc_return(&ctx->cached_cq_overflow));
7852 * Put inflight ref and overflow ref. If that's
7853 * all we had, then we're done with this request.
7855 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7856 io_free_req(cancel_req);
7857 finish_wait(&ctx->inflight_wait, &wait);
7861 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7862 io_put_req(cancel_req);
7866 finish_wait(&ctx->inflight_wait, &wait);
7870 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7872 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7873 struct task_struct *task = data;
7875 return req->task == task;
7878 static int io_uring_flush(struct file *file, void *data)
7880 struct io_ring_ctx *ctx = file->private_data;
7882 io_uring_cancel_files(ctx, data);
7885 * If the task is going away, cancel work it may have pending
7887 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7888 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7893 static void *io_uring_validate_mmap_request(struct file *file,
7894 loff_t pgoff, size_t sz)
7896 struct io_ring_ctx *ctx = file->private_data;
7897 loff_t offset = pgoff << PAGE_SHIFT;
7902 case IORING_OFF_SQ_RING:
7903 case IORING_OFF_CQ_RING:
7906 case IORING_OFF_SQES:
7910 return ERR_PTR(-EINVAL);
7913 page = virt_to_head_page(ptr);
7914 if (sz > page_size(page))
7915 return ERR_PTR(-EINVAL);
7922 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7924 size_t sz = vma->vm_end - vma->vm_start;
7928 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7930 return PTR_ERR(ptr);
7932 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7933 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7936 #else /* !CONFIG_MMU */
7938 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7940 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7943 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7945 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7948 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7949 unsigned long addr, unsigned long len,
7950 unsigned long pgoff, unsigned long flags)
7954 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7956 return PTR_ERR(ptr);
7958 return (unsigned long) ptr;
7961 #endif /* !CONFIG_MMU */
7963 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7964 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7967 struct io_ring_ctx *ctx;
7974 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7982 if (f.file->f_op != &io_uring_fops)
7986 ctx = f.file->private_data;
7987 if (!percpu_ref_tryget(&ctx->refs))
7991 * For SQ polling, the thread will do all submissions and completions.
7992 * Just return the requested submit count, and wake the thread if
7996 if (ctx->flags & IORING_SETUP_SQPOLL) {
7997 if (!list_empty_careful(&ctx->cq_overflow_list))
7998 io_cqring_overflow_flush(ctx, false);
7999 if (flags & IORING_ENTER_SQ_WAKEUP)
8000 wake_up(&ctx->sqo_wait);
8001 submitted = to_submit;
8002 } else if (to_submit) {
8003 mutex_lock(&ctx->uring_lock);
8004 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8005 mutex_unlock(&ctx->uring_lock);
8007 if (submitted != to_submit)
8010 if (flags & IORING_ENTER_GETEVENTS) {
8011 min_complete = min(min_complete, ctx->cq_entries);
8014 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8015 * space applications don't need to do io completion events
8016 * polling again, they can rely on io_sq_thread to do polling
8017 * work, which can reduce cpu usage and uring_lock contention.
8019 if (ctx->flags & IORING_SETUP_IOPOLL &&
8020 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8021 ret = io_iopoll_check(ctx, min_complete);
8023 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8028 percpu_ref_put(&ctx->refs);
8031 return submitted ? submitted : ret;
8034 #ifdef CONFIG_PROC_FS
8035 static int io_uring_show_cred(int id, void *p, void *data)
8037 const struct cred *cred = p;
8038 struct seq_file *m = data;
8039 struct user_namespace *uns = seq_user_ns(m);
8040 struct group_info *gi;
8045 seq_printf(m, "%5d\n", id);
8046 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8047 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8048 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8049 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8050 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8051 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8052 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8053 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8054 seq_puts(m, "\n\tGroups:\t");
8055 gi = cred->group_info;
8056 for (g = 0; g < gi->ngroups; g++) {
8057 seq_put_decimal_ull(m, g ? " " : "",
8058 from_kgid_munged(uns, gi->gid[g]));
8060 seq_puts(m, "\n\tCapEff:\t");
8061 cap = cred->cap_effective;
8062 CAP_FOR_EACH_U32(__capi)
8063 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8068 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8072 mutex_lock(&ctx->uring_lock);
8073 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8074 for (i = 0; i < ctx->nr_user_files; i++) {
8075 struct fixed_file_table *table;
8078 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8079 f = table->files[i & IORING_FILE_TABLE_MASK];
8081 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8083 seq_printf(m, "%5u: <none>\n", i);
8085 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8086 for (i = 0; i < ctx->nr_user_bufs; i++) {
8087 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8089 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8090 (unsigned int) buf->len);
8092 if (!idr_is_empty(&ctx->personality_idr)) {
8093 seq_printf(m, "Personalities:\n");
8094 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8096 seq_printf(m, "PollList:\n");
8097 spin_lock_irq(&ctx->completion_lock);
8098 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8099 struct hlist_head *list = &ctx->cancel_hash[i];
8100 struct io_kiocb *req;
8102 hlist_for_each_entry(req, list, hash_node)
8103 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8104 req->task->task_works != NULL);
8106 spin_unlock_irq(&ctx->completion_lock);
8107 mutex_unlock(&ctx->uring_lock);
8110 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8112 struct io_ring_ctx *ctx = f->private_data;
8114 if (percpu_ref_tryget(&ctx->refs)) {
8115 __io_uring_show_fdinfo(ctx, m);
8116 percpu_ref_put(&ctx->refs);
8121 static const struct file_operations io_uring_fops = {
8122 .release = io_uring_release,
8123 .flush = io_uring_flush,
8124 .mmap = io_uring_mmap,
8126 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8127 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8129 .poll = io_uring_poll,
8130 .fasync = io_uring_fasync,
8131 #ifdef CONFIG_PROC_FS
8132 .show_fdinfo = io_uring_show_fdinfo,
8136 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8137 struct io_uring_params *p)
8139 struct io_rings *rings;
8140 size_t size, sq_array_offset;
8142 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8143 if (size == SIZE_MAX)
8146 rings = io_mem_alloc(size);
8151 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8152 rings->sq_ring_mask = p->sq_entries - 1;
8153 rings->cq_ring_mask = p->cq_entries - 1;
8154 rings->sq_ring_entries = p->sq_entries;
8155 rings->cq_ring_entries = p->cq_entries;
8156 ctx->sq_mask = rings->sq_ring_mask;
8157 ctx->cq_mask = rings->cq_ring_mask;
8158 ctx->sq_entries = rings->sq_ring_entries;
8159 ctx->cq_entries = rings->cq_ring_entries;
8161 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8162 if (size == SIZE_MAX) {
8163 io_mem_free(ctx->rings);
8168 ctx->sq_sqes = io_mem_alloc(size);
8169 if (!ctx->sq_sqes) {
8170 io_mem_free(ctx->rings);
8179 * Allocate an anonymous fd, this is what constitutes the application
8180 * visible backing of an io_uring instance. The application mmaps this
8181 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8182 * we have to tie this fd to a socket for file garbage collection purposes.
8184 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8189 #if defined(CONFIG_UNIX)
8190 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8196 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8200 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8201 O_RDWR | O_CLOEXEC);
8204 ret = PTR_ERR(file);
8208 #if defined(CONFIG_UNIX)
8209 ctx->ring_sock->file = file;
8211 fd_install(ret, file);
8214 #if defined(CONFIG_UNIX)
8215 sock_release(ctx->ring_sock);
8216 ctx->ring_sock = NULL;
8221 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8222 struct io_uring_params __user *params)
8224 struct user_struct *user = NULL;
8225 struct io_ring_ctx *ctx;
8231 if (entries > IORING_MAX_ENTRIES) {
8232 if (!(p->flags & IORING_SETUP_CLAMP))
8234 entries = IORING_MAX_ENTRIES;
8238 * Use twice as many entries for the CQ ring. It's possible for the
8239 * application to drive a higher depth than the size of the SQ ring,
8240 * since the sqes are only used at submission time. This allows for
8241 * some flexibility in overcommitting a bit. If the application has
8242 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8243 * of CQ ring entries manually.
8245 p->sq_entries = roundup_pow_of_two(entries);
8246 if (p->flags & IORING_SETUP_CQSIZE) {
8248 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8249 * to a power-of-two, if it isn't already. We do NOT impose
8250 * any cq vs sq ring sizing.
8252 if (p->cq_entries < p->sq_entries)
8254 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8255 if (!(p->flags & IORING_SETUP_CLAMP))
8257 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8259 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8261 p->cq_entries = 2 * p->sq_entries;
8264 user = get_uid(current_user());
8265 limit_mem = !capable(CAP_IPC_LOCK);
8268 ret = __io_account_mem(user,
8269 ring_pages(p->sq_entries, p->cq_entries));
8276 ctx = io_ring_ctx_alloc(p);
8279 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8284 ctx->compat = in_compat_syscall();
8286 ctx->creds = get_current_cred();
8288 ret = io_allocate_scq_urings(ctx, p);
8292 ret = io_sq_offload_start(ctx, p);
8296 memset(&p->sq_off, 0, sizeof(p->sq_off));
8297 p->sq_off.head = offsetof(struct io_rings, sq.head);
8298 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8299 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8300 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8301 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8302 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8303 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8305 memset(&p->cq_off, 0, sizeof(p->cq_off));
8306 p->cq_off.head = offsetof(struct io_rings, cq.head);
8307 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8308 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8309 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8310 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8311 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8312 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8314 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8315 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8316 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8317 IORING_FEAT_POLL_32BITS;
8319 if (copy_to_user(params, p, sizeof(*p))) {
8324 * Install ring fd as the very last thing, so we don't risk someone
8325 * having closed it before we finish setup
8327 ret = io_uring_get_fd(ctx);
8331 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8332 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8334 ctx->limit_mem = limit_mem;
8337 io_ring_ctx_wait_and_kill(ctx);
8342 * Sets up an aio uring context, and returns the fd. Applications asks for a
8343 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8344 * params structure passed in.
8346 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8348 struct io_uring_params p;
8351 if (copy_from_user(&p, params, sizeof(p)))
8353 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8358 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8359 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8360 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8363 return io_uring_create(entries, &p, params);
8366 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8367 struct io_uring_params __user *, params)
8369 return io_uring_setup(entries, params);
8372 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8374 struct io_uring_probe *p;
8378 size = struct_size(p, ops, nr_args);
8379 if (size == SIZE_MAX)
8381 p = kzalloc(size, GFP_KERNEL);
8386 if (copy_from_user(p, arg, size))
8389 if (memchr_inv(p, 0, size))
8392 p->last_op = IORING_OP_LAST - 1;
8393 if (nr_args > IORING_OP_LAST)
8394 nr_args = IORING_OP_LAST;
8396 for (i = 0; i < nr_args; i++) {
8398 if (!io_op_defs[i].not_supported)
8399 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8404 if (copy_to_user(arg, p, size))
8411 static int io_register_personality(struct io_ring_ctx *ctx)
8413 const struct cred *creds = get_current_cred();
8416 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8417 USHRT_MAX, GFP_KERNEL);
8423 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8425 const struct cred *old_creds;
8427 old_creds = idr_remove(&ctx->personality_idr, id);
8429 put_cred(old_creds);
8436 static bool io_register_op_must_quiesce(int op)
8439 case IORING_UNREGISTER_FILES:
8440 case IORING_REGISTER_FILES_UPDATE:
8441 case IORING_REGISTER_PROBE:
8442 case IORING_REGISTER_PERSONALITY:
8443 case IORING_UNREGISTER_PERSONALITY:
8450 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8451 void __user *arg, unsigned nr_args)
8452 __releases(ctx->uring_lock)
8453 __acquires(ctx->uring_lock)
8458 * We're inside the ring mutex, if the ref is already dying, then
8459 * someone else killed the ctx or is already going through
8460 * io_uring_register().
8462 if (percpu_ref_is_dying(&ctx->refs))
8465 if (io_register_op_must_quiesce(opcode)) {
8466 percpu_ref_kill(&ctx->refs);
8469 * Drop uring mutex before waiting for references to exit. If
8470 * another thread is currently inside io_uring_enter() it might
8471 * need to grab the uring_lock to make progress. If we hold it
8472 * here across the drain wait, then we can deadlock. It's safe
8473 * to drop the mutex here, since no new references will come in
8474 * after we've killed the percpu ref.
8476 mutex_unlock(&ctx->uring_lock);
8477 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8478 mutex_lock(&ctx->uring_lock);
8480 percpu_ref_resurrect(&ctx->refs);
8487 case IORING_REGISTER_BUFFERS:
8488 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8490 case IORING_UNREGISTER_BUFFERS:
8494 ret = io_sqe_buffer_unregister(ctx);
8496 case IORING_REGISTER_FILES:
8497 ret = io_sqe_files_register(ctx, arg, nr_args);
8499 case IORING_UNREGISTER_FILES:
8503 ret = io_sqe_files_unregister(ctx);
8505 case IORING_REGISTER_FILES_UPDATE:
8506 ret = io_sqe_files_update(ctx, arg, nr_args);
8508 case IORING_REGISTER_EVENTFD:
8509 case IORING_REGISTER_EVENTFD_ASYNC:
8513 ret = io_eventfd_register(ctx, arg);
8516 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8517 ctx->eventfd_async = 1;
8519 ctx->eventfd_async = 0;
8521 case IORING_UNREGISTER_EVENTFD:
8525 ret = io_eventfd_unregister(ctx);
8527 case IORING_REGISTER_PROBE:
8529 if (!arg || nr_args > 256)
8531 ret = io_probe(ctx, arg, nr_args);
8533 case IORING_REGISTER_PERSONALITY:
8537 ret = io_register_personality(ctx);
8539 case IORING_UNREGISTER_PERSONALITY:
8543 ret = io_unregister_personality(ctx, nr_args);
8550 if (io_register_op_must_quiesce(opcode)) {
8551 /* bring the ctx back to life */
8552 percpu_ref_reinit(&ctx->refs);
8554 reinit_completion(&ctx->ref_comp);
8559 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8560 void __user *, arg, unsigned int, nr_args)
8562 struct io_ring_ctx *ctx;
8571 if (f.file->f_op != &io_uring_fops)
8574 ctx = f.file->private_data;
8576 mutex_lock(&ctx->uring_lock);
8577 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8578 mutex_unlock(&ctx->uring_lock);
8579 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8580 ctx->cq_ev_fd != NULL, ret);
8586 static int __init io_uring_init(void)
8588 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8589 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8590 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8593 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8594 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8595 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8596 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8597 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8598 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8599 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8600 BUILD_BUG_SQE_ELEM(8, __u64, off);
8601 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8602 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8603 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8604 BUILD_BUG_SQE_ELEM(24, __u32, len);
8605 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8606 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8607 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8608 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8609 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8610 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8611 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8612 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8613 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8614 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8615 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8616 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8617 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8618 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8619 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8620 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8621 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8622 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8623 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8625 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8626 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8627 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8630 __initcall(io_uring_init);