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;
651 struct list_head link_list;
654 * 1. used with ctx->iopoll_list with reads/writes
655 * 2. to track reqs with ->files (see io_op_def::file_table)
657 struct list_head inflight_entry;
659 struct percpu_ref *fixed_file_refs;
663 * Only commands that never go async can use the below fields,
664 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
665 * async armed poll handlers for regular commands. The latter
666 * restore the work, if needed.
669 struct hlist_node hash_node;
670 struct async_poll *apoll;
672 struct io_wq_work work;
674 struct callback_head task_work;
677 struct io_defer_entry {
678 struct list_head list;
679 struct io_kiocb *req;
683 #define IO_IOPOLL_BATCH 8
685 struct io_comp_state {
687 struct list_head list;
688 struct io_ring_ctx *ctx;
691 struct io_submit_state {
692 struct blk_plug plug;
695 * io_kiocb alloc cache
697 void *reqs[IO_IOPOLL_BATCH];
698 unsigned int free_reqs;
701 * Batch completion logic
703 struct io_comp_state comp;
706 * File reference cache
710 unsigned int has_refs;
711 unsigned int used_refs;
712 unsigned int ios_left;
716 /* needs req->io allocated for deferral/async */
717 unsigned async_ctx : 1;
718 /* needs current->mm setup, does mm access */
719 unsigned needs_mm : 1;
720 /* needs req->file assigned */
721 unsigned needs_file : 1;
722 /* don't fail if file grab fails */
723 unsigned needs_file_no_error : 1;
724 /* hash wq insertion if file is a regular file */
725 unsigned hash_reg_file : 1;
726 /* unbound wq insertion if file is a non-regular file */
727 unsigned unbound_nonreg_file : 1;
728 /* opcode is not supported by this kernel */
729 unsigned not_supported : 1;
730 /* needs file table */
731 unsigned file_table : 1;
733 unsigned needs_fs : 1;
734 /* set if opcode supports polled "wait" */
736 unsigned pollout : 1;
737 /* op supports buffer selection */
738 unsigned buffer_select : 1;
741 static const struct io_op_def io_op_defs[] = {
742 [IORING_OP_NOP] = {},
743 [IORING_OP_READV] = {
747 .unbound_nonreg_file = 1,
751 [IORING_OP_WRITEV] = {
756 .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,
773 [IORING_OP_POLL_ADD] = {
775 .unbound_nonreg_file = 1,
777 [IORING_OP_POLL_REMOVE] = {},
778 [IORING_OP_SYNC_FILE_RANGE] = {
781 [IORING_OP_SENDMSG] = {
785 .unbound_nonreg_file = 1,
789 [IORING_OP_RECVMSG] = {
793 .unbound_nonreg_file = 1,
798 [IORING_OP_TIMEOUT] = {
802 [IORING_OP_TIMEOUT_REMOVE] = {},
803 [IORING_OP_ACCEPT] = {
806 .unbound_nonreg_file = 1,
810 [IORING_OP_ASYNC_CANCEL] = {},
811 [IORING_OP_LINK_TIMEOUT] = {
815 [IORING_OP_CONNECT] = {
819 .unbound_nonreg_file = 1,
822 [IORING_OP_FALLOCATE] = {
825 [IORING_OP_OPENAT] = {
829 [IORING_OP_CLOSE] = {
831 .needs_file_no_error = 1,
834 [IORING_OP_FILES_UPDATE] = {
838 [IORING_OP_STATX] = {
846 .unbound_nonreg_file = 1,
850 [IORING_OP_WRITE] = {
853 .unbound_nonreg_file = 1,
856 [IORING_OP_FADVISE] = {
859 [IORING_OP_MADVISE] = {
865 .unbound_nonreg_file = 1,
871 .unbound_nonreg_file = 1,
875 [IORING_OP_OPENAT2] = {
879 [IORING_OP_EPOLL_CTL] = {
880 .unbound_nonreg_file = 1,
883 [IORING_OP_SPLICE] = {
886 .unbound_nonreg_file = 1,
888 [IORING_OP_PROVIDE_BUFFERS] = {},
889 [IORING_OP_REMOVE_BUFFERS] = {},
893 .unbound_nonreg_file = 1,
897 enum io_mem_account {
902 static bool io_rw_reissue(struct io_kiocb *req, long res);
903 static void io_cqring_fill_event(struct io_kiocb *req, long res);
904 static void io_put_req(struct io_kiocb *req);
905 static void io_double_put_req(struct io_kiocb *req);
906 static void __io_double_put_req(struct io_kiocb *req);
907 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
908 static void io_queue_linked_timeout(struct io_kiocb *req);
909 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
910 struct io_uring_files_update *ip,
912 static int io_grab_files(struct io_kiocb *req);
913 static void io_complete_rw_common(struct kiocb *kiocb, long res,
914 struct io_comp_state *cs);
915 static void __io_clean_op(struct io_kiocb *req);
916 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
917 int fd, struct file **out_file, bool fixed);
918 static void __io_queue_sqe(struct io_kiocb *req,
919 const struct io_uring_sqe *sqe,
920 struct io_comp_state *cs);
921 static void io_file_put_work(struct work_struct *work);
923 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
924 struct iovec **iovec, struct iov_iter *iter,
926 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
927 struct iovec *iovec, struct iovec *fast_iov,
928 struct iov_iter *iter);
930 static struct kmem_cache *req_cachep;
932 static const struct file_operations io_uring_fops;
934 struct sock *io_uring_get_socket(struct file *file)
936 #if defined(CONFIG_UNIX)
937 if (file->f_op == &io_uring_fops) {
938 struct io_ring_ctx *ctx = file->private_data;
940 return ctx->ring_sock->sk;
945 EXPORT_SYMBOL(io_uring_get_socket);
947 static void io_get_req_task(struct io_kiocb *req)
949 if (req->flags & REQ_F_TASK_PINNED)
951 get_task_struct(req->task);
952 req->flags |= REQ_F_TASK_PINNED;
955 static inline void io_clean_op(struct io_kiocb *req)
957 if (req->flags & REQ_F_NEED_CLEANUP)
961 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
962 static void __io_put_req_task(struct io_kiocb *req)
964 if (req->flags & REQ_F_TASK_PINNED)
965 put_task_struct(req->task);
968 static void io_sq_thread_drop_mm(void)
970 struct mm_struct *mm = current->mm;
973 kthread_unuse_mm(mm);
978 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
981 if (unlikely(!ctx->sqo_mm || !mmget_not_zero(ctx->sqo_mm)))
983 kthread_use_mm(ctx->sqo_mm);
989 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
990 struct io_kiocb *req)
992 if (!io_op_defs[req->opcode].needs_mm)
994 return __io_sq_thread_acquire_mm(ctx);
997 static inline void req_set_fail_links(struct io_kiocb *req)
999 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1000 req->flags |= REQ_F_FAIL_LINK;
1004 * Note: must call io_req_init_async() for the first time you
1005 * touch any members of io_wq_work.
1007 static inline void io_req_init_async(struct io_kiocb *req)
1009 if (req->flags & REQ_F_WORK_INITIALIZED)
1012 memset(&req->work, 0, sizeof(req->work));
1013 req->flags |= REQ_F_WORK_INITIALIZED;
1016 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1018 return ctx->flags & IORING_SETUP_SQPOLL;
1021 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1023 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1025 complete(&ctx->ref_comp);
1028 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1030 return !req->timeout.off;
1033 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1035 struct io_ring_ctx *ctx;
1038 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1042 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1043 if (!ctx->fallback_req)
1047 * Use 5 bits less than the max cq entries, that should give us around
1048 * 32 entries per hash list if totally full and uniformly spread.
1050 hash_bits = ilog2(p->cq_entries);
1054 ctx->cancel_hash_bits = hash_bits;
1055 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1057 if (!ctx->cancel_hash)
1059 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1061 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1062 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1065 ctx->flags = p->flags;
1066 init_waitqueue_head(&ctx->sqo_wait);
1067 init_waitqueue_head(&ctx->cq_wait);
1068 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1069 init_completion(&ctx->ref_comp);
1070 init_completion(&ctx->sq_thread_comp);
1071 idr_init(&ctx->io_buffer_idr);
1072 idr_init(&ctx->personality_idr);
1073 mutex_init(&ctx->uring_lock);
1074 init_waitqueue_head(&ctx->wait);
1075 spin_lock_init(&ctx->completion_lock);
1076 INIT_LIST_HEAD(&ctx->iopoll_list);
1077 INIT_LIST_HEAD(&ctx->defer_list);
1078 INIT_LIST_HEAD(&ctx->timeout_list);
1079 init_waitqueue_head(&ctx->inflight_wait);
1080 spin_lock_init(&ctx->inflight_lock);
1081 INIT_LIST_HEAD(&ctx->inflight_list);
1082 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1083 init_llist_head(&ctx->file_put_llist);
1086 if (ctx->fallback_req)
1087 kmem_cache_free(req_cachep, ctx->fallback_req);
1088 kfree(ctx->cancel_hash);
1093 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1095 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1096 struct io_ring_ctx *ctx = req->ctx;
1098 return seq != ctx->cached_cq_tail
1099 + atomic_read(&ctx->cached_cq_overflow);
1105 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1107 struct io_rings *rings = ctx->rings;
1109 /* order cqe stores with ring update */
1110 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1112 if (wq_has_sleeper(&ctx->cq_wait)) {
1113 wake_up_interruptible(&ctx->cq_wait);
1114 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1118 static void io_req_clean_work(struct io_kiocb *req)
1120 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1124 mmdrop(req->work.mm);
1125 req->work.mm = NULL;
1127 if (req->work.creds) {
1128 put_cred(req->work.creds);
1129 req->work.creds = NULL;
1132 struct fs_struct *fs = req->work.fs;
1134 spin_lock(&req->work.fs->lock);
1137 spin_unlock(&req->work.fs->lock);
1143 static void io_prep_async_work(struct io_kiocb *req)
1145 const struct io_op_def *def = &io_op_defs[req->opcode];
1147 io_req_init_async(req);
1149 if (req->flags & REQ_F_ISREG) {
1150 if (def->hash_reg_file)
1151 io_wq_hash_work(&req->work, file_inode(req->file));
1153 if (def->unbound_nonreg_file)
1154 req->work.flags |= IO_WQ_WORK_UNBOUND;
1156 if (!req->work.mm && def->needs_mm) {
1157 mmgrab(current->mm);
1158 req->work.mm = current->mm;
1160 if (!req->work.creds)
1161 req->work.creds = get_current_cred();
1162 if (!req->work.fs && def->needs_fs) {
1163 spin_lock(¤t->fs->lock);
1164 if (!current->fs->in_exec) {
1165 req->work.fs = current->fs;
1166 req->work.fs->users++;
1168 req->work.flags |= IO_WQ_WORK_CANCEL;
1170 spin_unlock(¤t->fs->lock);
1174 static void io_prep_async_link(struct io_kiocb *req)
1176 struct io_kiocb *cur;
1178 io_prep_async_work(req);
1179 if (req->flags & REQ_F_LINK_HEAD)
1180 list_for_each_entry(cur, &req->link_list, link_list)
1181 io_prep_async_work(cur);
1184 static void __io_queue_async_work(struct io_kiocb *req)
1186 struct io_ring_ctx *ctx = req->ctx;
1187 struct io_kiocb *link = io_prep_linked_timeout(req);
1189 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1190 &req->work, req->flags);
1191 io_wq_enqueue(ctx->io_wq, &req->work);
1194 io_queue_linked_timeout(link);
1197 static void io_queue_async_work(struct io_kiocb *req)
1199 /* init ->work of the whole link before punting */
1200 io_prep_async_link(req);
1201 __io_queue_async_work(req);
1204 static void io_kill_timeout(struct io_kiocb *req)
1208 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1210 atomic_inc(&req->ctx->cq_timeouts);
1211 list_del_init(&req->timeout.list);
1212 req->flags |= REQ_F_COMP_LOCKED;
1213 io_cqring_fill_event(req, 0);
1218 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1220 struct io_kiocb *req, *tmp;
1222 spin_lock_irq(&ctx->completion_lock);
1223 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1224 io_kill_timeout(req);
1225 spin_unlock_irq(&ctx->completion_lock);
1228 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1231 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1232 struct io_defer_entry, list);
1234 if (req_need_defer(de->req, de->seq))
1236 list_del_init(&de->list);
1237 /* punt-init is done before queueing for defer */
1238 __io_queue_async_work(de->req);
1240 } while (!list_empty(&ctx->defer_list));
1243 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1245 while (!list_empty(&ctx->timeout_list)) {
1246 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1247 struct io_kiocb, timeout.list);
1249 if (io_is_timeout_noseq(req))
1251 if (req->timeout.target_seq != ctx->cached_cq_tail
1252 - atomic_read(&ctx->cq_timeouts))
1255 list_del_init(&req->timeout.list);
1256 io_kill_timeout(req);
1260 static void io_commit_cqring(struct io_ring_ctx *ctx)
1262 io_flush_timeouts(ctx);
1263 __io_commit_cqring(ctx);
1265 if (unlikely(!list_empty(&ctx->defer_list)))
1266 __io_queue_deferred(ctx);
1269 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1271 struct io_rings *rings = ctx->rings;
1274 tail = ctx->cached_cq_tail;
1276 * writes to the cq entry need to come after reading head; the
1277 * control dependency is enough as we're using WRITE_ONCE to
1280 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1283 ctx->cached_cq_tail++;
1284 return &rings->cqes[tail & ctx->cq_mask];
1287 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1291 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1293 if (!ctx->eventfd_async)
1295 return io_wq_current_is_worker();
1298 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1300 if (waitqueue_active(&ctx->wait))
1301 wake_up(&ctx->wait);
1302 if (waitqueue_active(&ctx->sqo_wait))
1303 wake_up(&ctx->sqo_wait);
1304 if (io_should_trigger_evfd(ctx))
1305 eventfd_signal(ctx->cq_ev_fd, 1);
1308 /* Returns true if there are no backlogged entries after the flush */
1309 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1311 struct io_rings *rings = ctx->rings;
1312 struct io_uring_cqe *cqe;
1313 struct io_kiocb *req;
1314 unsigned long flags;
1318 if (list_empty_careful(&ctx->cq_overflow_list))
1320 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1321 rings->cq_ring_entries))
1325 spin_lock_irqsave(&ctx->completion_lock, flags);
1327 /* if force is set, the ring is going away. always drop after that */
1329 ctx->cq_overflow_flushed = 1;
1332 while (!list_empty(&ctx->cq_overflow_list)) {
1333 cqe = io_get_cqring(ctx);
1337 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1339 list_move(&req->compl.list, &list);
1340 req->flags &= ~REQ_F_OVERFLOW;
1342 WRITE_ONCE(cqe->user_data, req->user_data);
1343 WRITE_ONCE(cqe->res, req->result);
1344 WRITE_ONCE(cqe->flags, req->compl.cflags);
1346 WRITE_ONCE(ctx->rings->cq_overflow,
1347 atomic_inc_return(&ctx->cached_cq_overflow));
1351 io_commit_cqring(ctx);
1353 clear_bit(0, &ctx->sq_check_overflow);
1354 clear_bit(0, &ctx->cq_check_overflow);
1355 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1357 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1358 io_cqring_ev_posted(ctx);
1360 while (!list_empty(&list)) {
1361 req = list_first_entry(&list, struct io_kiocb, compl.list);
1362 list_del(&req->compl.list);
1369 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1371 struct io_ring_ctx *ctx = req->ctx;
1372 struct io_uring_cqe *cqe;
1374 trace_io_uring_complete(ctx, req->user_data, res);
1377 * If we can't get a cq entry, userspace overflowed the
1378 * submission (by quite a lot). Increment the overflow count in
1381 cqe = io_get_cqring(ctx);
1383 WRITE_ONCE(cqe->user_data, req->user_data);
1384 WRITE_ONCE(cqe->res, res);
1385 WRITE_ONCE(cqe->flags, cflags);
1386 } else if (ctx->cq_overflow_flushed) {
1387 WRITE_ONCE(ctx->rings->cq_overflow,
1388 atomic_inc_return(&ctx->cached_cq_overflow));
1390 if (list_empty(&ctx->cq_overflow_list)) {
1391 set_bit(0, &ctx->sq_check_overflow);
1392 set_bit(0, &ctx->cq_check_overflow);
1393 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1396 req->flags |= REQ_F_OVERFLOW;
1398 req->compl.cflags = cflags;
1399 refcount_inc(&req->refs);
1400 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1404 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1406 __io_cqring_fill_event(req, res, 0);
1409 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1411 struct io_ring_ctx *ctx = req->ctx;
1412 unsigned long flags;
1414 spin_lock_irqsave(&ctx->completion_lock, flags);
1415 __io_cqring_fill_event(req, res, cflags);
1416 io_commit_cqring(ctx);
1417 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1419 io_cqring_ev_posted(ctx);
1422 static void io_submit_flush_completions(struct io_comp_state *cs)
1424 struct io_ring_ctx *ctx = cs->ctx;
1426 spin_lock_irq(&ctx->completion_lock);
1427 while (!list_empty(&cs->list)) {
1428 struct io_kiocb *req;
1430 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1431 list_del(&req->compl.list);
1432 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1433 if (!(req->flags & REQ_F_LINK_HEAD)) {
1434 req->flags |= REQ_F_COMP_LOCKED;
1437 spin_unlock_irq(&ctx->completion_lock);
1439 spin_lock_irq(&ctx->completion_lock);
1442 io_commit_cqring(ctx);
1443 spin_unlock_irq(&ctx->completion_lock);
1445 io_cqring_ev_posted(ctx);
1449 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1450 struct io_comp_state *cs)
1453 io_cqring_add_event(req, res, cflags);
1458 req->compl.cflags = cflags;
1459 list_add_tail(&req->compl.list, &cs->list);
1461 io_submit_flush_completions(cs);
1465 static void io_req_complete(struct io_kiocb *req, long res)
1467 __io_req_complete(req, res, 0, NULL);
1470 static inline bool io_is_fallback_req(struct io_kiocb *req)
1472 return req == (struct io_kiocb *)
1473 ((unsigned long) req->ctx->fallback_req & ~1UL);
1476 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1478 struct io_kiocb *req;
1480 req = ctx->fallback_req;
1481 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1487 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1488 struct io_submit_state *state)
1490 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1491 struct io_kiocb *req;
1493 if (!state->free_reqs) {
1497 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1498 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1501 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1502 * retry single alloc to be on the safe side.
1504 if (unlikely(ret <= 0)) {
1505 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1506 if (!state->reqs[0])
1510 state->free_reqs = ret - 1;
1511 req = state->reqs[ret - 1];
1514 req = state->reqs[state->free_reqs];
1519 return io_get_fallback_req(ctx);
1522 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1526 percpu_ref_put(req->fixed_file_refs);
1531 static void io_dismantle_req(struct io_kiocb *req)
1538 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1539 __io_put_req_task(req);
1540 io_req_clean_work(req);
1542 if (req->flags & REQ_F_INFLIGHT) {
1543 struct io_ring_ctx *ctx = req->ctx;
1544 unsigned long flags;
1546 spin_lock_irqsave(&ctx->inflight_lock, flags);
1547 list_del(&req->inflight_entry);
1548 if (waitqueue_active(&ctx->inflight_wait))
1549 wake_up(&ctx->inflight_wait);
1550 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1554 static void __io_free_req(struct io_kiocb *req)
1556 struct io_ring_ctx *ctx;
1558 io_dismantle_req(req);
1560 if (likely(!io_is_fallback_req(req)))
1561 kmem_cache_free(req_cachep, req);
1563 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1564 percpu_ref_put(&ctx->refs);
1567 static bool io_link_cancel_timeout(struct io_kiocb *req)
1569 struct io_ring_ctx *ctx = req->ctx;
1572 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1574 io_cqring_fill_event(req, -ECANCELED);
1575 io_commit_cqring(ctx);
1576 req->flags &= ~REQ_F_LINK_HEAD;
1584 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1586 struct io_kiocb *link;
1589 if (list_empty(&req->link_list))
1591 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1592 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1595 list_del_init(&link->link_list);
1596 wake_ev = io_link_cancel_timeout(link);
1597 req->flags &= ~REQ_F_LINK_TIMEOUT;
1601 static void io_kill_linked_timeout(struct io_kiocb *req)
1603 struct io_ring_ctx *ctx = req->ctx;
1606 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1607 unsigned long flags;
1609 spin_lock_irqsave(&ctx->completion_lock, flags);
1610 wake_ev = __io_kill_linked_timeout(req);
1611 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1613 wake_ev = __io_kill_linked_timeout(req);
1617 io_cqring_ev_posted(ctx);
1620 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1622 struct io_kiocb *nxt;
1625 * The list should never be empty when we are called here. But could
1626 * potentially happen if the chain is messed up, check to be on the
1629 if (unlikely(list_empty(&req->link_list)))
1632 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1633 list_del_init(&req->link_list);
1634 if (!list_empty(&nxt->link_list))
1635 nxt->flags |= REQ_F_LINK_HEAD;
1640 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1642 static void __io_fail_links(struct io_kiocb *req)
1644 struct io_ring_ctx *ctx = req->ctx;
1646 while (!list_empty(&req->link_list)) {
1647 struct io_kiocb *link = list_first_entry(&req->link_list,
1648 struct io_kiocb, link_list);
1650 list_del_init(&link->link_list);
1651 trace_io_uring_fail_link(req, link);
1653 io_cqring_fill_event(link, -ECANCELED);
1654 __io_double_put_req(link);
1655 req->flags &= ~REQ_F_LINK_TIMEOUT;
1658 io_commit_cqring(ctx);
1659 io_cqring_ev_posted(ctx);
1662 static void io_fail_links(struct io_kiocb *req)
1664 struct io_ring_ctx *ctx = req->ctx;
1666 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1667 unsigned long flags;
1669 spin_lock_irqsave(&ctx->completion_lock, flags);
1670 __io_fail_links(req);
1671 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1673 __io_fail_links(req);
1676 io_cqring_ev_posted(ctx);
1679 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1681 req->flags &= ~REQ_F_LINK_HEAD;
1682 if (req->flags & REQ_F_LINK_TIMEOUT)
1683 io_kill_linked_timeout(req);
1686 * If LINK is set, we have dependent requests in this chain. If we
1687 * didn't fail this request, queue the first one up, moving any other
1688 * dependencies to the next request. In case of failure, fail the rest
1691 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1692 return io_req_link_next(req);
1697 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1699 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1701 return __io_req_find_next(req);
1704 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1706 struct task_struct *tsk = req->task;
1707 struct io_ring_ctx *ctx = req->ctx;
1708 int ret, notify = TWA_RESUME;
1711 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1712 * If we're not using an eventfd, then TWA_RESUME is always fine,
1713 * as we won't have dependencies between request completions for
1714 * other kernel wait conditions.
1716 if (ctx->flags & IORING_SETUP_SQPOLL)
1718 else if (ctx->cq_ev_fd)
1719 notify = TWA_SIGNAL;
1721 ret = task_work_add(tsk, cb, notify);
1723 wake_up_process(tsk);
1727 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1729 struct io_ring_ctx *ctx = req->ctx;
1731 spin_lock_irq(&ctx->completion_lock);
1732 io_cqring_fill_event(req, error);
1733 io_commit_cqring(ctx);
1734 spin_unlock_irq(&ctx->completion_lock);
1736 io_cqring_ev_posted(ctx);
1737 req_set_fail_links(req);
1738 io_double_put_req(req);
1741 static void io_req_task_cancel(struct callback_head *cb)
1743 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1745 __io_req_task_cancel(req, -ECANCELED);
1748 static void __io_req_task_submit(struct io_kiocb *req)
1750 struct io_ring_ctx *ctx = req->ctx;
1752 if (!__io_sq_thread_acquire_mm(ctx)) {
1753 mutex_lock(&ctx->uring_lock);
1754 __io_queue_sqe(req, NULL, NULL);
1755 mutex_unlock(&ctx->uring_lock);
1757 __io_req_task_cancel(req, -EFAULT);
1761 static void io_req_task_submit(struct callback_head *cb)
1763 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1765 __io_req_task_submit(req);
1768 static void io_req_task_queue(struct io_kiocb *req)
1772 init_task_work(&req->task_work, io_req_task_submit);
1774 ret = io_req_task_work_add(req, &req->task_work);
1775 if (unlikely(ret)) {
1776 struct task_struct *tsk;
1778 init_task_work(&req->task_work, io_req_task_cancel);
1779 tsk = io_wq_get_task(req->ctx->io_wq);
1780 task_work_add(tsk, &req->task_work, 0);
1781 wake_up_process(tsk);
1785 static void io_queue_next(struct io_kiocb *req)
1787 struct io_kiocb *nxt = io_req_find_next(req);
1790 io_req_task_queue(nxt);
1793 static void io_free_req(struct io_kiocb *req)
1800 void *reqs[IO_IOPOLL_BATCH];
1804 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1805 struct req_batch *rb)
1807 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1808 percpu_ref_put_many(&ctx->refs, rb->to_free);
1812 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1813 struct req_batch *rb)
1816 __io_req_free_batch_flush(ctx, rb);
1819 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1821 if (unlikely(io_is_fallback_req(req))) {
1825 if (req->flags & REQ_F_LINK_HEAD)
1828 io_dismantle_req(req);
1829 rb->reqs[rb->to_free++] = req;
1830 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1831 __io_req_free_batch_flush(req->ctx, rb);
1835 * Drop reference to request, return next in chain (if there is one) if this
1836 * was the last reference to this request.
1838 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1840 struct io_kiocb *nxt = NULL;
1842 if (refcount_dec_and_test(&req->refs)) {
1843 nxt = io_req_find_next(req);
1849 static void io_put_req(struct io_kiocb *req)
1851 if (refcount_dec_and_test(&req->refs))
1855 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1857 struct io_kiocb *nxt;
1860 * A ref is owned by io-wq in which context we're. So, if that's the
1861 * last one, it's safe to steal next work. False negatives are Ok,
1862 * it just will be re-punted async in io_put_work()
1864 if (refcount_read(&req->refs) != 1)
1867 nxt = io_req_find_next(req);
1868 return nxt ? &nxt->work : NULL;
1872 * Must only be used if we don't need to care about links, usually from
1873 * within the completion handling itself.
1875 static void __io_double_put_req(struct io_kiocb *req)
1877 /* drop both submit and complete references */
1878 if (refcount_sub_and_test(2, &req->refs))
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 unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1891 struct io_rings *rings = ctx->rings;
1893 if (test_bit(0, &ctx->cq_check_overflow)) {
1895 * noflush == true is from the waitqueue handler, just ensure
1896 * we wake up the task, and the next invocation will flush the
1897 * entries. We cannot safely to it from here.
1899 if (noflush && !list_empty(&ctx->cq_overflow_list))
1902 io_cqring_overflow_flush(ctx, false);
1905 /* See comment at the top of this file */
1907 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1910 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1912 struct io_rings *rings = ctx->rings;
1914 /* make sure SQ entry isn't read before tail */
1915 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1918 static int io_put_kbuf(struct io_kiocb *req)
1920 struct io_buffer *kbuf;
1923 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1924 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1925 cflags |= IORING_CQE_F_BUFFER;
1931 static inline bool io_run_task_work(void)
1933 if (current->task_works) {
1934 __set_current_state(TASK_RUNNING);
1942 static void io_iopoll_queue(struct list_head *again)
1944 struct io_kiocb *req;
1947 req = list_first_entry(again, struct io_kiocb, inflight_entry);
1948 list_del(&req->inflight_entry);
1949 if (!io_rw_reissue(req, -EAGAIN))
1950 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1951 } while (!list_empty(again));
1955 * Find and free completed poll iocbs
1957 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1958 struct list_head *done)
1960 struct req_batch rb;
1961 struct io_kiocb *req;
1964 /* order with ->result store in io_complete_rw_iopoll() */
1968 while (!list_empty(done)) {
1971 req = list_first_entry(done, struct io_kiocb, inflight_entry);
1972 if (READ_ONCE(req->result) == -EAGAIN) {
1973 req->iopoll_completed = 0;
1974 list_move_tail(&req->inflight_entry, &again);
1977 list_del(&req->inflight_entry);
1979 if (req->flags & REQ_F_BUFFER_SELECTED)
1980 cflags = io_put_kbuf(req);
1982 __io_cqring_fill_event(req, req->result, cflags);
1985 if (refcount_dec_and_test(&req->refs))
1986 io_req_free_batch(&rb, req);
1989 io_commit_cqring(ctx);
1990 if (ctx->flags & IORING_SETUP_SQPOLL)
1991 io_cqring_ev_posted(ctx);
1992 io_req_free_batch_finish(ctx, &rb);
1994 if (!list_empty(&again))
1995 io_iopoll_queue(&again);
1998 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2001 struct io_kiocb *req, *tmp;
2007 * Only spin for completions if we don't have multiple devices hanging
2008 * off our complete list, and we're under the requested amount.
2010 spin = !ctx->poll_multi_file && *nr_events < min;
2013 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2014 struct kiocb *kiocb = &req->rw.kiocb;
2017 * Move completed and retryable entries to our local lists.
2018 * If we find a request that requires polling, break out
2019 * and complete those lists first, if we have entries there.
2021 if (READ_ONCE(req->iopoll_completed)) {
2022 list_move_tail(&req->inflight_entry, &done);
2025 if (!list_empty(&done))
2028 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2032 /* iopoll may have completed current req */
2033 if (READ_ONCE(req->iopoll_completed))
2034 list_move_tail(&req->inflight_entry, &done);
2041 if (!list_empty(&done))
2042 io_iopoll_complete(ctx, nr_events, &done);
2048 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2049 * non-spinning poll check - we'll still enter the driver poll loop, but only
2050 * as a non-spinning completion check.
2052 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2055 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2058 ret = io_do_iopoll(ctx, nr_events, min);
2061 if (*nr_events >= min)
2069 * We can't just wait for polled events to come to us, we have to actively
2070 * find and complete them.
2072 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2074 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2077 mutex_lock(&ctx->uring_lock);
2078 while (!list_empty(&ctx->iopoll_list)) {
2079 unsigned int nr_events = 0;
2081 io_do_iopoll(ctx, &nr_events, 0);
2083 /* let it sleep and repeat later if can't complete a request */
2087 * Ensure we allow local-to-the-cpu processing to take place,
2088 * in this case we need to ensure that we reap all events.
2089 * Also let task_work, etc. to progress by releasing the mutex
2091 if (need_resched()) {
2092 mutex_unlock(&ctx->uring_lock);
2094 mutex_lock(&ctx->uring_lock);
2097 mutex_unlock(&ctx->uring_lock);
2100 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2102 unsigned int nr_events = 0;
2103 int iters = 0, ret = 0;
2106 * We disallow the app entering submit/complete with polling, but we
2107 * still need to lock the ring to prevent racing with polled issue
2108 * that got punted to a workqueue.
2110 mutex_lock(&ctx->uring_lock);
2113 * Don't enter poll loop if we already have events pending.
2114 * If we do, we can potentially be spinning for commands that
2115 * already triggered a CQE (eg in error).
2117 if (io_cqring_events(ctx, false))
2121 * If a submit got punted to a workqueue, we can have the
2122 * application entering polling for a command before it gets
2123 * issued. That app will hold the uring_lock for the duration
2124 * of the poll right here, so we need to take a breather every
2125 * now and then to ensure that the issue has a chance to add
2126 * the poll to the issued list. Otherwise we can spin here
2127 * forever, while the workqueue is stuck trying to acquire the
2130 if (!(++iters & 7)) {
2131 mutex_unlock(&ctx->uring_lock);
2133 mutex_lock(&ctx->uring_lock);
2136 ret = io_iopoll_getevents(ctx, &nr_events, min);
2140 } while (min && !nr_events && !need_resched());
2142 mutex_unlock(&ctx->uring_lock);
2146 static void kiocb_end_write(struct io_kiocb *req)
2149 * Tell lockdep we inherited freeze protection from submission
2152 if (req->flags & REQ_F_ISREG) {
2153 struct inode *inode = file_inode(req->file);
2155 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2157 file_end_write(req->file);
2160 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2161 struct io_comp_state *cs)
2163 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2166 if (kiocb->ki_flags & IOCB_WRITE)
2167 kiocb_end_write(req);
2169 if (res != req->result)
2170 req_set_fail_links(req);
2171 if (req->flags & REQ_F_BUFFER_SELECTED)
2172 cflags = io_put_kbuf(req);
2173 __io_req_complete(req, res, cflags, cs);
2177 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2179 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2180 ssize_t ret = -ECANCELED;
2181 struct iov_iter iter;
2189 switch (req->opcode) {
2190 case IORING_OP_READV:
2191 case IORING_OP_READ_FIXED:
2192 case IORING_OP_READ:
2195 case IORING_OP_WRITEV:
2196 case IORING_OP_WRITE_FIXED:
2197 case IORING_OP_WRITE:
2201 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2206 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2209 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2214 req_set_fail_links(req);
2215 io_req_complete(req, ret);
2219 static void io_rw_resubmit(struct callback_head *cb)
2221 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2222 struct io_ring_ctx *ctx = req->ctx;
2225 err = io_sq_thread_acquire_mm(ctx, req);
2227 if (io_resubmit_prep(req, err)) {
2228 refcount_inc(&req->refs);
2229 io_queue_async_work(req);
2234 static bool io_rw_reissue(struct io_kiocb *req, long res)
2239 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2242 init_task_work(&req->task_work, io_rw_resubmit);
2243 ret = io_req_task_work_add(req, &req->task_work);
2250 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2251 struct io_comp_state *cs)
2253 if (!io_rw_reissue(req, res))
2254 io_complete_rw_common(&req->rw.kiocb, res, cs);
2257 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2259 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2261 __io_complete_rw(req, res, res2, NULL);
2264 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2266 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2268 if (kiocb->ki_flags & IOCB_WRITE)
2269 kiocb_end_write(req);
2271 if (res != -EAGAIN && res != req->result)
2272 req_set_fail_links(req);
2274 WRITE_ONCE(req->result, res);
2275 /* order with io_poll_complete() checking ->result */
2277 WRITE_ONCE(req->iopoll_completed, 1);
2281 * After the iocb has been issued, it's safe to be found on the poll list.
2282 * Adding the kiocb to the list AFTER submission ensures that we don't
2283 * find it from a io_iopoll_getevents() thread before the issuer is done
2284 * accessing the kiocb cookie.
2286 static void io_iopoll_req_issued(struct io_kiocb *req)
2288 struct io_ring_ctx *ctx = req->ctx;
2291 * Track whether we have multiple files in our lists. This will impact
2292 * how we do polling eventually, not spinning if we're on potentially
2293 * different devices.
2295 if (list_empty(&ctx->iopoll_list)) {
2296 ctx->poll_multi_file = false;
2297 } else if (!ctx->poll_multi_file) {
2298 struct io_kiocb *list_req;
2300 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2302 if (list_req->file != req->file)
2303 ctx->poll_multi_file = true;
2307 * For fast devices, IO may have already completed. If it has, add
2308 * it to the front so we find it first.
2310 if (READ_ONCE(req->iopoll_completed))
2311 list_add(&req->inflight_entry, &ctx->iopoll_list);
2313 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2315 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2316 wq_has_sleeper(&ctx->sqo_wait))
2317 wake_up(&ctx->sqo_wait);
2320 static void __io_state_file_put(struct io_submit_state *state)
2322 int diff = state->has_refs - state->used_refs;
2325 fput_many(state->file, diff);
2329 static inline void io_state_file_put(struct io_submit_state *state)
2332 __io_state_file_put(state);
2336 * Get as many references to a file as we have IOs left in this submission,
2337 * assuming most submissions are for one file, or at least that each file
2338 * has more than one submission.
2340 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2346 if (state->fd == fd) {
2351 __io_state_file_put(state);
2353 state->file = fget_many(fd, state->ios_left);
2358 state->has_refs = state->ios_left;
2359 state->used_refs = 1;
2364 static bool io_bdev_nowait(struct block_device *bdev)
2367 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2374 * If we tracked the file through the SCM inflight mechanism, we could support
2375 * any file. For now, just ensure that anything potentially problematic is done
2378 static bool io_file_supports_async(struct file *file, int rw)
2380 umode_t mode = file_inode(file)->i_mode;
2382 if (S_ISBLK(mode)) {
2383 if (io_bdev_nowait(file->f_inode->i_bdev))
2387 if (S_ISCHR(mode) || S_ISSOCK(mode))
2389 if (S_ISREG(mode)) {
2390 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2391 file->f_op != &io_uring_fops)
2396 /* any ->read/write should understand O_NONBLOCK */
2397 if (file->f_flags & O_NONBLOCK)
2400 if (!(file->f_mode & FMODE_NOWAIT))
2404 return file->f_op->read_iter != NULL;
2406 return file->f_op->write_iter != NULL;
2409 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2410 bool force_nonblock)
2412 struct io_ring_ctx *ctx = req->ctx;
2413 struct kiocb *kiocb = &req->rw.kiocb;
2417 if (S_ISREG(file_inode(req->file)->i_mode))
2418 req->flags |= REQ_F_ISREG;
2420 kiocb->ki_pos = READ_ONCE(sqe->off);
2421 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2422 req->flags |= REQ_F_CUR_POS;
2423 kiocb->ki_pos = req->file->f_pos;
2425 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2426 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2427 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2431 ioprio = READ_ONCE(sqe->ioprio);
2433 ret = ioprio_check_cap(ioprio);
2437 kiocb->ki_ioprio = ioprio;
2439 kiocb->ki_ioprio = get_current_ioprio();
2441 /* don't allow async punt if RWF_NOWAIT was requested */
2442 if (kiocb->ki_flags & IOCB_NOWAIT)
2443 req->flags |= REQ_F_NOWAIT;
2445 if (kiocb->ki_flags & IOCB_DIRECT)
2446 io_get_req_task(req);
2449 kiocb->ki_flags |= IOCB_NOWAIT;
2451 if (ctx->flags & IORING_SETUP_IOPOLL) {
2452 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2453 !kiocb->ki_filp->f_op->iopoll)
2456 kiocb->ki_flags |= IOCB_HIPRI;
2457 kiocb->ki_complete = io_complete_rw_iopoll;
2458 req->iopoll_completed = 0;
2459 io_get_req_task(req);
2461 if (kiocb->ki_flags & IOCB_HIPRI)
2463 kiocb->ki_complete = io_complete_rw;
2466 req->rw.addr = READ_ONCE(sqe->addr);
2467 req->rw.len = READ_ONCE(sqe->len);
2468 req->buf_index = READ_ONCE(sqe->buf_index);
2472 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2478 case -ERESTARTNOINTR:
2479 case -ERESTARTNOHAND:
2480 case -ERESTART_RESTARTBLOCK:
2482 * We can't just restart the syscall, since previously
2483 * submitted sqes may already be in progress. Just fail this
2489 kiocb->ki_complete(kiocb, ret, 0);
2493 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2494 struct io_comp_state *cs)
2496 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2498 if (req->flags & REQ_F_CUR_POS)
2499 req->file->f_pos = kiocb->ki_pos;
2500 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2501 __io_complete_rw(req, ret, 0, cs);
2503 io_rw_done(kiocb, ret);
2506 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2507 struct iov_iter *iter)
2509 struct io_ring_ctx *ctx = req->ctx;
2510 size_t len = req->rw.len;
2511 struct io_mapped_ubuf *imu;
2512 u16 index, buf_index;
2516 /* attempt to use fixed buffers without having provided iovecs */
2517 if (unlikely(!ctx->user_bufs))
2520 buf_index = req->buf_index;
2521 if (unlikely(buf_index >= ctx->nr_user_bufs))
2524 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2525 imu = &ctx->user_bufs[index];
2526 buf_addr = req->rw.addr;
2529 if (buf_addr + len < buf_addr)
2531 /* not inside the mapped region */
2532 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2536 * May not be a start of buffer, set size appropriately
2537 * and advance us to the beginning.
2539 offset = buf_addr - imu->ubuf;
2540 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2544 * Don't use iov_iter_advance() here, as it's really slow for
2545 * using the latter parts of a big fixed buffer - it iterates
2546 * over each segment manually. We can cheat a bit here, because
2549 * 1) it's a BVEC iter, we set it up
2550 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2551 * first and last bvec
2553 * So just find our index, and adjust the iterator afterwards.
2554 * If the offset is within the first bvec (or the whole first
2555 * bvec, just use iov_iter_advance(). This makes it easier
2556 * since we can just skip the first segment, which may not
2557 * be PAGE_SIZE aligned.
2559 const struct bio_vec *bvec = imu->bvec;
2561 if (offset <= bvec->bv_len) {
2562 iov_iter_advance(iter, offset);
2564 unsigned long seg_skip;
2566 /* skip first vec */
2567 offset -= bvec->bv_len;
2568 seg_skip = 1 + (offset >> PAGE_SHIFT);
2570 iter->bvec = bvec + seg_skip;
2571 iter->nr_segs -= seg_skip;
2572 iter->count -= bvec->bv_len + offset;
2573 iter->iov_offset = offset & ~PAGE_MASK;
2580 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2583 mutex_unlock(&ctx->uring_lock);
2586 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2589 * "Normal" inline submissions always hold the uring_lock, since we
2590 * grab it from the system call. Same is true for the SQPOLL offload.
2591 * The only exception is when we've detached the request and issue it
2592 * from an async worker thread, grab the lock for that case.
2595 mutex_lock(&ctx->uring_lock);
2598 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2599 int bgid, struct io_buffer *kbuf,
2602 struct io_buffer *head;
2604 if (req->flags & REQ_F_BUFFER_SELECTED)
2607 io_ring_submit_lock(req->ctx, needs_lock);
2609 lockdep_assert_held(&req->ctx->uring_lock);
2611 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2613 if (!list_empty(&head->list)) {
2614 kbuf = list_last_entry(&head->list, struct io_buffer,
2616 list_del(&kbuf->list);
2619 idr_remove(&req->ctx->io_buffer_idr, bgid);
2621 if (*len > kbuf->len)
2624 kbuf = ERR_PTR(-ENOBUFS);
2627 io_ring_submit_unlock(req->ctx, needs_lock);
2632 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2635 struct io_buffer *kbuf;
2638 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2639 bgid = req->buf_index;
2640 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2643 req->rw.addr = (u64) (unsigned long) kbuf;
2644 req->flags |= REQ_F_BUFFER_SELECTED;
2645 return u64_to_user_ptr(kbuf->addr);
2648 #ifdef CONFIG_COMPAT
2649 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2652 struct compat_iovec __user *uiov;
2653 compat_ssize_t clen;
2657 uiov = u64_to_user_ptr(req->rw.addr);
2658 if (!access_ok(uiov, sizeof(*uiov)))
2660 if (__get_user(clen, &uiov->iov_len))
2666 buf = io_rw_buffer_select(req, &len, needs_lock);
2668 return PTR_ERR(buf);
2669 iov[0].iov_base = buf;
2670 iov[0].iov_len = (compat_size_t) len;
2675 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2678 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2682 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2685 len = iov[0].iov_len;
2688 buf = io_rw_buffer_select(req, &len, needs_lock);
2690 return PTR_ERR(buf);
2691 iov[0].iov_base = buf;
2692 iov[0].iov_len = len;
2696 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2699 if (req->flags & REQ_F_BUFFER_SELECTED) {
2700 struct io_buffer *kbuf;
2702 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2703 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2704 iov[0].iov_len = kbuf->len;
2709 else if (req->rw.len > 1)
2712 #ifdef CONFIG_COMPAT
2713 if (req->ctx->compat)
2714 return io_compat_import(req, iov, needs_lock);
2717 return __io_iov_buffer_select(req, iov, needs_lock);
2720 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2721 struct iovec **iovec, struct iov_iter *iter,
2724 void __user *buf = u64_to_user_ptr(req->rw.addr);
2725 size_t sqe_len = req->rw.len;
2729 opcode = req->opcode;
2730 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2732 return io_import_fixed(req, rw, iter);
2735 /* buffer index only valid with fixed read/write, or buffer select */
2736 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2739 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2740 if (req->flags & REQ_F_BUFFER_SELECT) {
2741 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2744 return PTR_ERR(buf);
2746 req->rw.len = sqe_len;
2749 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2751 return ret < 0 ? ret : sqe_len;
2755 struct io_async_rw *iorw = &req->io->rw;
2757 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2762 if (req->flags & REQ_F_BUFFER_SELECT) {
2763 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2765 ret = (*iovec)->iov_len;
2766 iov_iter_init(iter, rw, *iovec, 1, ret);
2772 #ifdef CONFIG_COMPAT
2773 if (req->ctx->compat)
2774 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2778 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2782 * For files that don't have ->read_iter() and ->write_iter(), handle them
2783 * by looping over ->read() or ->write() manually.
2785 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2786 struct iov_iter *iter)
2791 * Don't support polled IO through this interface, and we can't
2792 * support non-blocking either. For the latter, this just causes
2793 * the kiocb to be handled from an async context.
2795 if (kiocb->ki_flags & IOCB_HIPRI)
2797 if (kiocb->ki_flags & IOCB_NOWAIT)
2800 while (iov_iter_count(iter)) {
2804 if (!iov_iter_is_bvec(iter)) {
2805 iovec = iov_iter_iovec(iter);
2807 /* fixed buffers import bvec */
2808 iovec.iov_base = kmap(iter->bvec->bv_page)
2810 iovec.iov_len = min(iter->count,
2811 iter->bvec->bv_len - iter->iov_offset);
2815 nr = file->f_op->read(file, iovec.iov_base,
2816 iovec.iov_len, &kiocb->ki_pos);
2818 nr = file->f_op->write(file, iovec.iov_base,
2819 iovec.iov_len, &kiocb->ki_pos);
2822 if (iov_iter_is_bvec(iter))
2823 kunmap(iter->bvec->bv_page);
2831 if (nr != iovec.iov_len)
2833 iov_iter_advance(iter, nr);
2839 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2840 struct iovec *iovec, struct iovec *fast_iov,
2841 struct iov_iter *iter)
2843 struct io_async_rw *rw = &req->io->rw;
2845 rw->nr_segs = iter->nr_segs;
2848 rw->iov = rw->fast_iov;
2849 if (rw->iov != fast_iov)
2850 memcpy(rw->iov, fast_iov,
2851 sizeof(struct iovec) * iter->nr_segs);
2854 req->flags |= REQ_F_NEED_CLEANUP;
2858 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2860 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2861 return req->io == NULL;
2864 static int io_alloc_async_ctx(struct io_kiocb *req)
2866 if (!io_op_defs[req->opcode].async_ctx)
2869 return __io_alloc_async_ctx(req);
2872 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2873 struct iovec *iovec, struct iovec *fast_iov,
2874 struct iov_iter *iter)
2876 if (!io_op_defs[req->opcode].async_ctx)
2879 if (__io_alloc_async_ctx(req))
2882 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2887 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2888 bool force_nonblock)
2890 struct io_async_ctx *io = req->io;
2891 struct iov_iter iter;
2894 io->rw.iov = io->rw.fast_iov;
2896 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2898 if (unlikely(ret < 0))
2901 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2905 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2906 bool force_nonblock)
2910 ret = io_prep_rw(req, sqe, force_nonblock);
2914 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2917 /* either don't need iovec imported or already have it */
2918 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2920 return io_rw_prep_async(req, READ, force_nonblock);
2923 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2924 int sync, void *arg)
2926 struct wait_page_queue *wpq;
2927 struct io_kiocb *req = wait->private;
2928 struct wait_page_key *key = arg;
2931 wpq = container_of(wait, struct wait_page_queue, wait);
2933 ret = wake_page_match(wpq, key);
2937 list_del_init(&wait->entry);
2939 init_task_work(&req->task_work, io_req_task_submit);
2940 /* submit ref gets dropped, acquire a new one */
2941 refcount_inc(&req->refs);
2942 ret = io_req_task_work_add(req, &req->task_work);
2943 if (unlikely(ret)) {
2944 struct task_struct *tsk;
2946 /* queue just for cancelation */
2947 init_task_work(&req->task_work, io_req_task_cancel);
2948 tsk = io_wq_get_task(req->ctx->io_wq);
2949 task_work_add(tsk, &req->task_work, 0);
2950 wake_up_process(tsk);
2955 static bool io_rw_should_retry(struct io_kiocb *req)
2957 struct kiocb *kiocb = &req->rw.kiocb;
2960 /* never retry for NOWAIT, we just complete with -EAGAIN */
2961 if (req->flags & REQ_F_NOWAIT)
2964 /* already tried, or we're doing O_DIRECT */
2965 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2968 * just use poll if we can, and don't attempt if the fs doesn't
2969 * support callback based unlocks
2971 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2975 * If request type doesn't require req->io to defer in general,
2976 * we need to allocate it here
2978 if (!req->io && __io_alloc_async_ctx(req))
2981 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2982 io_async_buf_func, req);
2984 io_get_req_task(req);
2991 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2993 if (req->file->f_op->read_iter)
2994 return call_read_iter(req->file, &req->rw.kiocb, iter);
2995 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
2998 static int io_read(struct io_kiocb *req, bool force_nonblock,
2999 struct io_comp_state *cs)
3001 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3002 struct kiocb *kiocb = &req->rw.kiocb;
3003 struct iov_iter iter;
3005 ssize_t io_size, ret;
3007 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3011 /* Ensure we clear previously set non-block flag */
3012 if (!force_nonblock)
3013 kiocb->ki_flags &= ~IOCB_NOWAIT;
3016 req->result = io_size;
3018 /* If the file doesn't support async, just async punt */
3019 if (force_nonblock && !io_file_supports_async(req->file, READ))
3022 iov_count = iov_iter_count(&iter);
3023 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3025 unsigned long nr_segs = iter.nr_segs;
3028 ret2 = io_iter_do_read(req, &iter);
3030 /* Catch -EAGAIN return for forced non-blocking submission */
3031 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3032 kiocb_done(kiocb, ret2, cs);
3034 iter.count = iov_count;
3035 iter.nr_segs = nr_segs;
3037 ret = io_setup_async_rw(req, io_size, iovec,
3038 inline_vecs, &iter);
3041 /* it's copied and will be cleaned with ->io */
3043 /* if we can retry, do so with the callbacks armed */
3044 if (io_rw_should_retry(req)) {
3045 ret2 = io_iter_do_read(req, &iter);
3046 if (ret2 == -EIOCBQUEUED) {
3048 } else if (ret2 != -EAGAIN) {
3049 kiocb_done(kiocb, ret2, cs);
3053 kiocb->ki_flags &= ~IOCB_WAITQ;
3063 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3064 bool force_nonblock)
3068 ret = io_prep_rw(req, sqe, force_nonblock);
3072 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3075 req->fsize = rlimit(RLIMIT_FSIZE);
3077 /* either don't need iovec imported or already have it */
3078 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3080 return io_rw_prep_async(req, WRITE, force_nonblock);
3083 static int io_write(struct io_kiocb *req, bool force_nonblock,
3084 struct io_comp_state *cs)
3086 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3087 struct kiocb *kiocb = &req->rw.kiocb;
3088 struct iov_iter iter;
3090 ssize_t ret, io_size;
3092 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3096 /* Ensure we clear previously set non-block flag */
3097 if (!force_nonblock)
3098 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3101 req->result = io_size;
3103 /* If the file doesn't support async, just async punt */
3104 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3107 /* file path doesn't support NOWAIT for non-direct_IO */
3108 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3109 (req->flags & REQ_F_ISREG))
3112 iov_count = iov_iter_count(&iter);
3113 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3115 unsigned long nr_segs = iter.nr_segs;
3119 * Open-code file_start_write here to grab freeze protection,
3120 * which will be released by another thread in
3121 * io_complete_rw(). Fool lockdep by telling it the lock got
3122 * released so that it doesn't complain about the held lock when
3123 * we return to userspace.
3125 if (req->flags & REQ_F_ISREG) {
3126 __sb_start_write(file_inode(req->file)->i_sb,
3127 SB_FREEZE_WRITE, true);
3128 __sb_writers_release(file_inode(req->file)->i_sb,
3131 kiocb->ki_flags |= IOCB_WRITE;
3133 if (!force_nonblock)
3134 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3136 if (req->file->f_op->write_iter)
3137 ret2 = call_write_iter(req->file, kiocb, &iter);
3139 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3141 if (!force_nonblock)
3142 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3145 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3146 * retry them without IOCB_NOWAIT.
3148 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3150 if (!force_nonblock || ret2 != -EAGAIN) {
3151 kiocb_done(kiocb, ret2, cs);
3153 iter.count = iov_count;
3154 iter.nr_segs = nr_segs;
3156 ret = io_setup_async_rw(req, io_size, iovec,
3157 inline_vecs, &iter);
3160 /* it's copied and will be cleaned with ->io */
3171 static int __io_splice_prep(struct io_kiocb *req,
3172 const struct io_uring_sqe *sqe)
3174 struct io_splice* sp = &req->splice;
3175 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3178 if (req->flags & REQ_F_NEED_CLEANUP)
3180 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3184 sp->len = READ_ONCE(sqe->len);
3185 sp->flags = READ_ONCE(sqe->splice_flags);
3187 if (unlikely(sp->flags & ~valid_flags))
3190 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3191 (sp->flags & SPLICE_F_FD_IN_FIXED));
3194 req->flags |= REQ_F_NEED_CLEANUP;
3196 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3198 * Splice operation will be punted aync, and here need to
3199 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3201 io_req_init_async(req);
3202 req->work.flags |= IO_WQ_WORK_UNBOUND;
3208 static int io_tee_prep(struct io_kiocb *req,
3209 const struct io_uring_sqe *sqe)
3211 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3213 return __io_splice_prep(req, sqe);
3216 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3218 struct io_splice *sp = &req->splice;
3219 struct file *in = sp->file_in;
3220 struct file *out = sp->file_out;
3221 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3227 ret = do_tee(in, out, sp->len, flags);
3229 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3230 req->flags &= ~REQ_F_NEED_CLEANUP;
3233 req_set_fail_links(req);
3234 io_req_complete(req, ret);
3238 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3240 struct io_splice* sp = &req->splice;
3242 sp->off_in = READ_ONCE(sqe->splice_off_in);
3243 sp->off_out = READ_ONCE(sqe->off);
3244 return __io_splice_prep(req, sqe);
3247 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3249 struct io_splice *sp = &req->splice;
3250 struct file *in = sp->file_in;
3251 struct file *out = sp->file_out;
3252 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3253 loff_t *poff_in, *poff_out;
3259 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3260 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3263 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3265 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3266 req->flags &= ~REQ_F_NEED_CLEANUP;
3269 req_set_fail_links(req);
3270 io_req_complete(req, ret);
3275 * IORING_OP_NOP just posts a completion event, nothing else.
3277 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3279 struct io_ring_ctx *ctx = req->ctx;
3281 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3284 __io_req_complete(req, 0, 0, cs);
3288 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3290 struct io_ring_ctx *ctx = req->ctx;
3295 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3297 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3300 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3301 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3304 req->sync.off = READ_ONCE(sqe->off);
3305 req->sync.len = READ_ONCE(sqe->len);
3309 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3311 loff_t end = req->sync.off + req->sync.len;
3314 /* fsync always requires a blocking context */
3318 ret = vfs_fsync_range(req->file, req->sync.off,
3319 end > 0 ? end : LLONG_MAX,
3320 req->sync.flags & IORING_FSYNC_DATASYNC);
3322 req_set_fail_links(req);
3323 io_req_complete(req, ret);
3327 static int io_fallocate_prep(struct io_kiocb *req,
3328 const struct io_uring_sqe *sqe)
3330 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3332 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3335 req->sync.off = READ_ONCE(sqe->off);
3336 req->sync.len = READ_ONCE(sqe->addr);
3337 req->sync.mode = READ_ONCE(sqe->len);
3338 req->fsize = rlimit(RLIMIT_FSIZE);
3342 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3346 /* fallocate always requiring blocking context */
3350 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3351 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3353 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3355 req_set_fail_links(req);
3356 io_req_complete(req, ret);
3360 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3362 const char __user *fname;
3365 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3367 if (unlikely(sqe->ioprio || sqe->buf_index))
3369 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3372 /* open.how should be already initialised */
3373 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3374 req->open.how.flags |= O_LARGEFILE;
3376 req->open.dfd = READ_ONCE(sqe->fd);
3377 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3378 req->open.filename = getname(fname);
3379 if (IS_ERR(req->open.filename)) {
3380 ret = PTR_ERR(req->open.filename);
3381 req->open.filename = NULL;
3384 req->open.nofile = rlimit(RLIMIT_NOFILE);
3385 req->flags |= REQ_F_NEED_CLEANUP;
3389 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3393 if (req->flags & REQ_F_NEED_CLEANUP)
3395 mode = READ_ONCE(sqe->len);
3396 flags = READ_ONCE(sqe->open_flags);
3397 req->open.how = build_open_how(flags, mode);
3398 return __io_openat_prep(req, sqe);
3401 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3403 struct open_how __user *how;
3407 if (req->flags & REQ_F_NEED_CLEANUP)
3409 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3410 len = READ_ONCE(sqe->len);
3411 if (len < OPEN_HOW_SIZE_VER0)
3414 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3419 return __io_openat_prep(req, sqe);
3422 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3424 struct open_flags op;
3431 ret = build_open_flags(&req->open.how, &op);
3435 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3439 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3442 ret = PTR_ERR(file);
3444 fsnotify_open(file);
3445 fd_install(ret, file);
3448 putname(req->open.filename);
3449 req->flags &= ~REQ_F_NEED_CLEANUP;
3451 req_set_fail_links(req);
3452 io_req_complete(req, ret);
3456 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3458 return io_openat2(req, force_nonblock);
3461 static int io_remove_buffers_prep(struct io_kiocb *req,
3462 const struct io_uring_sqe *sqe)
3464 struct io_provide_buf *p = &req->pbuf;
3467 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3470 tmp = READ_ONCE(sqe->fd);
3471 if (!tmp || tmp > USHRT_MAX)
3474 memset(p, 0, sizeof(*p));
3476 p->bgid = READ_ONCE(sqe->buf_group);
3480 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3481 int bgid, unsigned nbufs)
3485 /* shouldn't happen */
3489 /* the head kbuf is the list itself */
3490 while (!list_empty(&buf->list)) {
3491 struct io_buffer *nxt;
3493 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3494 list_del(&nxt->list);
3501 idr_remove(&ctx->io_buffer_idr, bgid);
3506 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3507 struct io_comp_state *cs)
3509 struct io_provide_buf *p = &req->pbuf;
3510 struct io_ring_ctx *ctx = req->ctx;
3511 struct io_buffer *head;
3514 io_ring_submit_lock(ctx, !force_nonblock);
3516 lockdep_assert_held(&ctx->uring_lock);
3519 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3521 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3523 io_ring_submit_lock(ctx, !force_nonblock);
3525 req_set_fail_links(req);
3526 __io_req_complete(req, ret, 0, cs);
3530 static int io_provide_buffers_prep(struct io_kiocb *req,
3531 const struct io_uring_sqe *sqe)
3533 struct io_provide_buf *p = &req->pbuf;
3536 if (sqe->ioprio || sqe->rw_flags)
3539 tmp = READ_ONCE(sqe->fd);
3540 if (!tmp || tmp > USHRT_MAX)
3543 p->addr = READ_ONCE(sqe->addr);
3544 p->len = READ_ONCE(sqe->len);
3546 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3549 p->bgid = READ_ONCE(sqe->buf_group);
3550 tmp = READ_ONCE(sqe->off);
3551 if (tmp > USHRT_MAX)
3557 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3559 struct io_buffer *buf;
3560 u64 addr = pbuf->addr;
3561 int i, bid = pbuf->bid;
3563 for (i = 0; i < pbuf->nbufs; i++) {
3564 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3569 buf->len = pbuf->len;
3574 INIT_LIST_HEAD(&buf->list);
3577 list_add_tail(&buf->list, &(*head)->list);
3581 return i ? i : -ENOMEM;
3584 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3585 struct io_comp_state *cs)
3587 struct io_provide_buf *p = &req->pbuf;
3588 struct io_ring_ctx *ctx = req->ctx;
3589 struct io_buffer *head, *list;
3592 io_ring_submit_lock(ctx, !force_nonblock);
3594 lockdep_assert_held(&ctx->uring_lock);
3596 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3598 ret = io_add_buffers(p, &head);
3603 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3606 __io_remove_buffers(ctx, head, p->bgid, -1U);
3611 io_ring_submit_unlock(ctx, !force_nonblock);
3613 req_set_fail_links(req);
3614 __io_req_complete(req, ret, 0, cs);
3618 static int io_epoll_ctl_prep(struct io_kiocb *req,
3619 const struct io_uring_sqe *sqe)
3621 #if defined(CONFIG_EPOLL)
3622 if (sqe->ioprio || sqe->buf_index)
3624 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3627 req->epoll.epfd = READ_ONCE(sqe->fd);
3628 req->epoll.op = READ_ONCE(sqe->len);
3629 req->epoll.fd = READ_ONCE(sqe->off);
3631 if (ep_op_has_event(req->epoll.op)) {
3632 struct epoll_event __user *ev;
3634 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3635 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3645 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3646 struct io_comp_state *cs)
3648 #if defined(CONFIG_EPOLL)
3649 struct io_epoll *ie = &req->epoll;
3652 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3653 if (force_nonblock && ret == -EAGAIN)
3657 req_set_fail_links(req);
3658 __io_req_complete(req, ret, 0, cs);
3665 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3667 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3668 if (sqe->ioprio || sqe->buf_index || sqe->off)
3670 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3673 req->madvise.addr = READ_ONCE(sqe->addr);
3674 req->madvise.len = READ_ONCE(sqe->len);
3675 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3682 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3684 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3685 struct io_madvise *ma = &req->madvise;
3691 ret = do_madvise(ma->addr, ma->len, ma->advice);
3693 req_set_fail_links(req);
3694 io_req_complete(req, ret);
3701 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3703 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3705 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3708 req->fadvise.offset = READ_ONCE(sqe->off);
3709 req->fadvise.len = READ_ONCE(sqe->len);
3710 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3714 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3716 struct io_fadvise *fa = &req->fadvise;
3719 if (force_nonblock) {
3720 switch (fa->advice) {
3721 case POSIX_FADV_NORMAL:
3722 case POSIX_FADV_RANDOM:
3723 case POSIX_FADV_SEQUENTIAL:
3730 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3732 req_set_fail_links(req);
3733 io_req_complete(req, ret);
3737 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3741 if (sqe->ioprio || sqe->buf_index)
3743 if (req->flags & REQ_F_FIXED_FILE)
3746 req->statx.dfd = READ_ONCE(sqe->fd);
3747 req->statx.mask = READ_ONCE(sqe->len);
3748 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3749 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3750 req->statx.flags = READ_ONCE(sqe->statx_flags);
3755 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3757 struct io_statx *ctx = &req->statx;
3760 if (force_nonblock) {
3761 /* only need file table for an actual valid fd */
3762 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3763 req->flags |= REQ_F_NO_FILE_TABLE;
3767 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3771 req_set_fail_links(req);
3772 io_req_complete(req, ret);
3776 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 * If we queue this for async, it must not be cancellable. That would
3780 * leave the 'file' in an undeterminate state, and here need to modify
3781 * io_wq_work.flags, so initialize io_wq_work firstly.
3783 io_req_init_async(req);
3784 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3786 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3788 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3789 sqe->rw_flags || sqe->buf_index)
3791 if (req->flags & REQ_F_FIXED_FILE)
3794 req->close.fd = READ_ONCE(sqe->fd);
3795 if ((req->file && req->file->f_op == &io_uring_fops) ||
3796 req->close.fd == req->ctx->ring_fd)
3799 req->close.put_file = NULL;
3803 static int io_close(struct io_kiocb *req, bool force_nonblock,
3804 struct io_comp_state *cs)
3806 struct io_close *close = &req->close;
3809 /* might be already done during nonblock submission */
3810 if (!close->put_file) {
3811 ret = __close_fd_get_file(close->fd, &close->put_file);
3813 return (ret == -ENOENT) ? -EBADF : ret;
3816 /* if the file has a flush method, be safe and punt to async */
3817 if (close->put_file->f_op->flush && force_nonblock) {
3818 /* was never set, but play safe */
3819 req->flags &= ~REQ_F_NOWAIT;
3820 /* avoid grabbing files - we don't need the files */
3821 req->flags |= REQ_F_NO_FILE_TABLE;
3825 /* No ->flush() or already async, safely close from here */
3826 ret = filp_close(close->put_file, req->work.files);
3828 req_set_fail_links(req);
3829 fput(close->put_file);
3830 close->put_file = NULL;
3831 __io_req_complete(req, ret, 0, cs);
3835 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3837 struct io_ring_ctx *ctx = req->ctx;
3842 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3844 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3847 req->sync.off = READ_ONCE(sqe->off);
3848 req->sync.len = READ_ONCE(sqe->len);
3849 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3853 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3857 /* sync_file_range always requires a blocking context */
3861 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3864 req_set_fail_links(req);
3865 io_req_complete(req, ret);
3869 #if defined(CONFIG_NET)
3870 static int io_setup_async_msg(struct io_kiocb *req,
3871 struct io_async_msghdr *kmsg)
3875 if (io_alloc_async_ctx(req)) {
3876 if (kmsg->iov != kmsg->fast_iov)
3880 req->flags |= REQ_F_NEED_CLEANUP;
3881 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3885 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3886 struct io_async_msghdr *iomsg)
3888 iomsg->iov = iomsg->fast_iov;
3889 iomsg->msg.msg_name = &iomsg->addr;
3890 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3891 req->sr_msg.msg_flags, &iomsg->iov);
3894 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3896 struct io_sr_msg *sr = &req->sr_msg;
3897 struct io_async_ctx *io = req->io;
3900 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3903 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3904 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3905 sr->len = READ_ONCE(sqe->len);
3907 #ifdef CONFIG_COMPAT
3908 if (req->ctx->compat)
3909 sr->msg_flags |= MSG_CMSG_COMPAT;
3912 if (!io || req->opcode == IORING_OP_SEND)
3914 /* iovec is already imported */
3915 if (req->flags & REQ_F_NEED_CLEANUP)
3918 ret = io_sendmsg_copy_hdr(req, &io->msg);
3920 req->flags |= REQ_F_NEED_CLEANUP;
3924 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3925 struct io_comp_state *cs)
3927 struct io_async_msghdr *kmsg = NULL;
3928 struct socket *sock;
3931 sock = sock_from_file(req->file, &ret);
3933 struct io_async_msghdr iomsg;
3937 kmsg = &req->io->msg;
3938 kmsg->msg.msg_name = &req->io->msg.addr;
3939 /* if iov is set, it's allocated already */
3941 kmsg->iov = kmsg->fast_iov;
3942 kmsg->msg.msg_iter.iov = kmsg->iov;
3944 ret = io_sendmsg_copy_hdr(req, &iomsg);
3950 flags = req->sr_msg.msg_flags;
3951 if (flags & MSG_DONTWAIT)
3952 req->flags |= REQ_F_NOWAIT;
3953 else if (force_nonblock)
3954 flags |= MSG_DONTWAIT;
3956 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3957 if (force_nonblock && ret == -EAGAIN)
3958 return io_setup_async_msg(req, kmsg);
3959 if (ret == -ERESTARTSYS)
3963 if (kmsg && kmsg->iov != kmsg->fast_iov)
3965 req->flags &= ~REQ_F_NEED_CLEANUP;
3967 req_set_fail_links(req);
3968 __io_req_complete(req, ret, 0, cs);
3972 static int io_send(struct io_kiocb *req, bool force_nonblock,
3973 struct io_comp_state *cs)
3975 struct socket *sock;
3978 sock = sock_from_file(req->file, &ret);
3980 struct io_sr_msg *sr = &req->sr_msg;
3985 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3990 msg.msg_name = NULL;
3991 msg.msg_control = NULL;
3992 msg.msg_controllen = 0;
3993 msg.msg_namelen = 0;
3995 flags = req->sr_msg.msg_flags;
3996 if (flags & MSG_DONTWAIT)
3997 req->flags |= REQ_F_NOWAIT;
3998 else if (force_nonblock)
3999 flags |= MSG_DONTWAIT;
4001 msg.msg_flags = flags;
4002 ret = sock_sendmsg(sock, &msg);
4003 if (force_nonblock && ret == -EAGAIN)
4005 if (ret == -ERESTARTSYS)
4010 req_set_fail_links(req);
4011 __io_req_complete(req, ret, 0, cs);
4015 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4016 struct io_async_msghdr *iomsg)
4018 struct io_sr_msg *sr = &req->sr_msg;
4019 struct iovec __user *uiov;
4023 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4024 &iomsg->uaddr, &uiov, &iov_len);
4028 if (req->flags & REQ_F_BUFFER_SELECT) {
4031 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4033 sr->len = iomsg->iov[0].iov_len;
4034 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4038 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4039 &iomsg->iov, &iomsg->msg.msg_iter);
4047 #ifdef CONFIG_COMPAT
4048 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4049 struct io_async_msghdr *iomsg)
4051 struct compat_msghdr __user *msg_compat;
4052 struct io_sr_msg *sr = &req->sr_msg;
4053 struct compat_iovec __user *uiov;
4058 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4059 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4064 uiov = compat_ptr(ptr);
4065 if (req->flags & REQ_F_BUFFER_SELECT) {
4066 compat_ssize_t clen;
4070 if (!access_ok(uiov, sizeof(*uiov)))
4072 if (__get_user(clen, &uiov->iov_len))
4076 sr->len = iomsg->iov[0].iov_len;
4079 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4081 &iomsg->msg.msg_iter);
4090 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4091 struct io_async_msghdr *iomsg)
4093 iomsg->msg.msg_name = &iomsg->addr;
4094 iomsg->iov = iomsg->fast_iov;
4096 #ifdef CONFIG_COMPAT
4097 if (req->ctx->compat)
4098 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4101 return __io_recvmsg_copy_hdr(req, iomsg);
4104 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4105 int *cflags, bool needs_lock)
4107 struct io_sr_msg *sr = &req->sr_msg;
4108 struct io_buffer *kbuf;
4110 if (!(req->flags & REQ_F_BUFFER_SELECT))
4113 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4118 req->flags |= REQ_F_BUFFER_SELECTED;
4120 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4121 *cflags |= IORING_CQE_F_BUFFER;
4125 static int io_recvmsg_prep(struct io_kiocb *req,
4126 const struct io_uring_sqe *sqe)
4128 struct io_sr_msg *sr = &req->sr_msg;
4129 struct io_async_ctx *io = req->io;
4132 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4135 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4136 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4137 sr->len = READ_ONCE(sqe->len);
4138 sr->bgid = READ_ONCE(sqe->buf_group);
4140 #ifdef CONFIG_COMPAT
4141 if (req->ctx->compat)
4142 sr->msg_flags |= MSG_CMSG_COMPAT;
4145 if (!io || req->opcode == IORING_OP_RECV)
4147 /* iovec is already imported */
4148 if (req->flags & REQ_F_NEED_CLEANUP)
4151 ret = io_recvmsg_copy_hdr(req, &io->msg);
4153 req->flags |= REQ_F_NEED_CLEANUP;
4157 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4158 struct io_comp_state *cs)
4160 struct io_async_msghdr *kmsg = NULL;
4161 struct socket *sock;
4162 int ret, cflags = 0;
4164 sock = sock_from_file(req->file, &ret);
4166 struct io_buffer *kbuf;
4167 struct io_async_msghdr iomsg;
4171 kmsg = &req->io->msg;
4172 kmsg->msg.msg_name = &req->io->msg.addr;
4173 /* if iov is set, it's allocated already */
4175 kmsg->iov = kmsg->fast_iov;
4176 kmsg->msg.msg_iter.iov = kmsg->iov;
4178 ret = io_recvmsg_copy_hdr(req, &iomsg);
4184 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4186 return PTR_ERR(kbuf);
4188 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4189 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4190 1, req->sr_msg.len);
4193 flags = req->sr_msg.msg_flags;
4194 if (flags & MSG_DONTWAIT)
4195 req->flags |= REQ_F_NOWAIT;
4196 else if (force_nonblock)
4197 flags |= MSG_DONTWAIT;
4199 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4200 kmsg->uaddr, flags);
4201 if (force_nonblock && ret == -EAGAIN) {
4202 ret = io_setup_async_msg(req, kmsg);
4207 if (ret == -ERESTARTSYS)
4213 if (kmsg && kmsg->iov != kmsg->fast_iov)
4215 req->flags &= ~REQ_F_NEED_CLEANUP;
4217 req_set_fail_links(req);
4218 __io_req_complete(req, ret, cflags, cs);
4222 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4223 struct io_comp_state *cs)
4225 struct io_buffer *kbuf = NULL;
4226 struct socket *sock;
4227 int ret, cflags = 0;
4229 sock = sock_from_file(req->file, &ret);
4231 struct io_sr_msg *sr = &req->sr_msg;
4232 void __user *buf = sr->buf;
4237 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4239 return PTR_ERR(kbuf);
4241 buf = u64_to_user_ptr(kbuf->addr);
4243 ret = import_single_range(READ, buf, sr->len, &iov,
4250 req->flags |= REQ_F_NEED_CLEANUP;
4251 msg.msg_name = NULL;
4252 msg.msg_control = NULL;
4253 msg.msg_controllen = 0;
4254 msg.msg_namelen = 0;
4255 msg.msg_iocb = NULL;
4258 flags = req->sr_msg.msg_flags;
4259 if (flags & MSG_DONTWAIT)
4260 req->flags |= REQ_F_NOWAIT;
4261 else if (force_nonblock)
4262 flags |= MSG_DONTWAIT;
4264 ret = sock_recvmsg(sock, &msg, flags);
4265 if (force_nonblock && ret == -EAGAIN)
4267 if (ret == -ERESTARTSYS)
4272 req->flags &= ~REQ_F_NEED_CLEANUP;
4274 req_set_fail_links(req);
4275 __io_req_complete(req, ret, cflags, cs);
4279 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4281 struct io_accept *accept = &req->accept;
4283 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4285 if (sqe->ioprio || sqe->len || sqe->buf_index)
4288 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4289 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4290 accept->flags = READ_ONCE(sqe->accept_flags);
4291 accept->nofile = rlimit(RLIMIT_NOFILE);
4295 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4296 struct io_comp_state *cs)
4298 struct io_accept *accept = &req->accept;
4299 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4302 if (req->file->f_flags & O_NONBLOCK)
4303 req->flags |= REQ_F_NOWAIT;
4305 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4306 accept->addr_len, accept->flags,
4308 if (ret == -EAGAIN && force_nonblock)
4311 if (ret == -ERESTARTSYS)
4313 req_set_fail_links(req);
4315 __io_req_complete(req, ret, 0, cs);
4319 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4321 struct io_connect *conn = &req->connect;
4322 struct io_async_ctx *io = req->io;
4324 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4326 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4329 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4330 conn->addr_len = READ_ONCE(sqe->addr2);
4335 return move_addr_to_kernel(conn->addr, conn->addr_len,
4336 &io->connect.address);
4339 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4340 struct io_comp_state *cs)
4342 struct io_async_ctx __io, *io;
4343 unsigned file_flags;
4349 ret = move_addr_to_kernel(req->connect.addr,
4350 req->connect.addr_len,
4351 &__io.connect.address);
4357 file_flags = force_nonblock ? O_NONBLOCK : 0;
4359 ret = __sys_connect_file(req->file, &io->connect.address,
4360 req->connect.addr_len, file_flags);
4361 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4364 if (io_alloc_async_ctx(req)) {
4368 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4371 if (ret == -ERESTARTSYS)
4375 req_set_fail_links(req);
4376 __io_req_complete(req, ret, 0, cs);
4379 #else /* !CONFIG_NET */
4380 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4385 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4386 struct io_comp_state *cs)
4391 static int io_send(struct io_kiocb *req, bool force_nonblock,
4392 struct io_comp_state *cs)
4397 static int io_recvmsg_prep(struct io_kiocb *req,
4398 const struct io_uring_sqe *sqe)
4403 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4404 struct io_comp_state *cs)
4409 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4410 struct io_comp_state *cs)
4415 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4420 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4421 struct io_comp_state *cs)
4426 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4431 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4432 struct io_comp_state *cs)
4436 #endif /* CONFIG_NET */
4438 struct io_poll_table {
4439 struct poll_table_struct pt;
4440 struct io_kiocb *req;
4444 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4445 __poll_t mask, task_work_func_t func)
4449 /* for instances that support it check for an event match first: */
4450 if (mask && !(mask & poll->events))
4453 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4455 list_del_init(&poll->wait.entry);
4458 init_task_work(&req->task_work, func);
4460 * If this fails, then the task is exiting. When a task exits, the
4461 * work gets canceled, so just cancel this request as well instead
4462 * of executing it. We can't safely execute it anyway, as we may not
4463 * have the needed state needed for it anyway.
4465 ret = io_req_task_work_add(req, &req->task_work);
4466 if (unlikely(ret)) {
4467 struct task_struct *tsk;
4469 WRITE_ONCE(poll->canceled, true);
4470 tsk = io_wq_get_task(req->ctx->io_wq);
4471 task_work_add(tsk, &req->task_work, 0);
4472 wake_up_process(tsk);
4477 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4478 __acquires(&req->ctx->completion_lock)
4480 struct io_ring_ctx *ctx = req->ctx;
4482 if (!req->result && !READ_ONCE(poll->canceled)) {
4483 struct poll_table_struct pt = { ._key = poll->events };
4485 req->result = vfs_poll(req->file, &pt) & poll->events;
4488 spin_lock_irq(&ctx->completion_lock);
4489 if (!req->result && !READ_ONCE(poll->canceled)) {
4490 add_wait_queue(poll->head, &poll->wait);
4497 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4499 struct io_poll_iocb *poll = data;
4501 lockdep_assert_held(&req->ctx->completion_lock);
4503 if (poll && poll->head) {
4504 struct wait_queue_head *head = poll->head;
4506 spin_lock(&head->lock);
4507 list_del_init(&poll->wait.entry);
4508 if (poll->wait.private)
4509 refcount_dec(&req->refs);
4511 spin_unlock(&head->lock);
4515 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4517 struct io_ring_ctx *ctx = req->ctx;
4519 io_poll_remove_double(req, req->io);
4520 req->poll.done = true;
4521 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4522 io_commit_cqring(ctx);
4525 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4527 struct io_ring_ctx *ctx = req->ctx;
4529 if (io_poll_rewait(req, &req->poll)) {
4530 spin_unlock_irq(&ctx->completion_lock);
4534 hash_del(&req->hash_node);
4535 io_poll_complete(req, req->result, 0);
4536 req->flags |= REQ_F_COMP_LOCKED;
4537 *nxt = io_put_req_find_next(req);
4538 spin_unlock_irq(&ctx->completion_lock);
4540 io_cqring_ev_posted(ctx);
4543 static void io_poll_task_func(struct callback_head *cb)
4545 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4546 struct io_kiocb *nxt = NULL;
4548 io_poll_task_handler(req, &nxt);
4550 __io_req_task_submit(nxt);
4553 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4554 int sync, void *key)
4556 struct io_kiocb *req = wait->private;
4557 struct io_poll_iocb *poll = req->apoll->double_poll;
4558 __poll_t mask = key_to_poll(key);
4560 /* for instances that support it check for an event match first: */
4561 if (mask && !(mask & poll->events))
4564 if (poll && poll->head) {
4567 spin_lock(&poll->head->lock);
4568 done = list_empty(&poll->wait.entry);
4570 list_del_init(&poll->wait.entry);
4571 spin_unlock(&poll->head->lock);
4573 __io_async_wake(req, poll, mask, io_poll_task_func);
4575 refcount_dec(&req->refs);
4579 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4580 wait_queue_func_t wake_func)
4584 poll->canceled = false;
4585 poll->events = events;
4586 INIT_LIST_HEAD(&poll->wait.entry);
4587 init_waitqueue_func_entry(&poll->wait, wake_func);
4590 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4591 struct wait_queue_head *head,
4592 struct io_poll_iocb **poll_ptr)
4594 struct io_kiocb *req = pt->req;
4597 * If poll->head is already set, it's because the file being polled
4598 * uses multiple waitqueues for poll handling (eg one for read, one
4599 * for write). Setup a separate io_poll_iocb if this happens.
4601 if (unlikely(poll->head)) {
4602 /* already have a 2nd entry, fail a third attempt */
4604 pt->error = -EINVAL;
4607 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4609 pt->error = -ENOMEM;
4612 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4613 refcount_inc(&req->refs);
4614 poll->wait.private = req;
4621 if (poll->events & EPOLLEXCLUSIVE)
4622 add_wait_queue_exclusive(head, &poll->wait);
4624 add_wait_queue(head, &poll->wait);
4627 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4628 struct poll_table_struct *p)
4630 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4631 struct async_poll *apoll = pt->req->apoll;
4633 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4636 static void io_async_task_func(struct callback_head *cb)
4638 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4639 struct async_poll *apoll = req->apoll;
4640 struct io_ring_ctx *ctx = req->ctx;
4642 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4644 if (io_poll_rewait(req, &apoll->poll)) {
4645 spin_unlock_irq(&ctx->completion_lock);
4649 /* If req is still hashed, it cannot have been canceled. Don't check. */
4650 if (hash_hashed(&req->hash_node))
4651 hash_del(&req->hash_node);
4653 io_poll_remove_double(req, apoll->double_poll);
4654 spin_unlock_irq(&ctx->completion_lock);
4656 /* restore ->work in case we need to retry again */
4657 if (req->flags & REQ_F_WORK_INITIALIZED)
4658 memcpy(&req->work, &apoll->work, sizeof(req->work));
4660 if (!READ_ONCE(apoll->poll.canceled))
4661 __io_req_task_submit(req);
4663 __io_req_task_cancel(req, -ECANCELED);
4665 kfree(apoll->double_poll);
4669 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4672 struct io_kiocb *req = wait->private;
4673 struct io_poll_iocb *poll = &req->apoll->poll;
4675 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4678 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4681 static void io_poll_req_insert(struct io_kiocb *req)
4683 struct io_ring_ctx *ctx = req->ctx;
4684 struct hlist_head *list;
4686 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4687 hlist_add_head(&req->hash_node, list);
4690 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4691 struct io_poll_iocb *poll,
4692 struct io_poll_table *ipt, __poll_t mask,
4693 wait_queue_func_t wake_func)
4694 __acquires(&ctx->completion_lock)
4696 struct io_ring_ctx *ctx = req->ctx;
4697 bool cancel = false;
4699 io_init_poll_iocb(poll, mask, wake_func);
4700 poll->file = req->file;
4701 poll->wait.private = req;
4703 ipt->pt._key = mask;
4705 ipt->error = -EINVAL;
4707 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4709 spin_lock_irq(&ctx->completion_lock);
4710 if (likely(poll->head)) {
4711 spin_lock(&poll->head->lock);
4712 if (unlikely(list_empty(&poll->wait.entry))) {
4718 if (mask || ipt->error)
4719 list_del_init(&poll->wait.entry);
4721 WRITE_ONCE(poll->canceled, true);
4722 else if (!poll->done) /* actually waiting for an event */
4723 io_poll_req_insert(req);
4724 spin_unlock(&poll->head->lock);
4730 static bool io_arm_poll_handler(struct io_kiocb *req)
4732 const struct io_op_def *def = &io_op_defs[req->opcode];
4733 struct io_ring_ctx *ctx = req->ctx;
4734 struct async_poll *apoll;
4735 struct io_poll_table ipt;
4738 if (!req->file || !file_can_poll(req->file))
4740 if (req->flags & REQ_F_POLLED)
4742 if (!def->pollin && !def->pollout)
4745 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4746 if (unlikely(!apoll))
4748 apoll->double_poll = NULL;
4750 req->flags |= REQ_F_POLLED;
4751 if (req->flags & REQ_F_WORK_INITIALIZED)
4752 memcpy(&apoll->work, &req->work, sizeof(req->work));
4754 io_get_req_task(req);
4756 INIT_HLIST_NODE(&req->hash_node);
4760 mask |= POLLIN | POLLRDNORM;
4762 mask |= POLLOUT | POLLWRNORM;
4763 mask |= POLLERR | POLLPRI;
4765 ipt.pt._qproc = io_async_queue_proc;
4767 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4770 io_poll_remove_double(req, apoll->double_poll);
4771 spin_unlock_irq(&ctx->completion_lock);
4772 if (req->flags & REQ_F_WORK_INITIALIZED)
4773 memcpy(&req->work, &apoll->work, sizeof(req->work));
4774 kfree(apoll->double_poll);
4778 spin_unlock_irq(&ctx->completion_lock);
4779 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4780 apoll->poll.events);
4784 static bool __io_poll_remove_one(struct io_kiocb *req,
4785 struct io_poll_iocb *poll)
4787 bool do_complete = false;
4789 spin_lock(&poll->head->lock);
4790 WRITE_ONCE(poll->canceled, true);
4791 if (!list_empty(&poll->wait.entry)) {
4792 list_del_init(&poll->wait.entry);
4795 spin_unlock(&poll->head->lock);
4796 hash_del(&req->hash_node);
4800 static bool io_poll_remove_one(struct io_kiocb *req)
4804 if (req->opcode == IORING_OP_POLL_ADD) {
4805 io_poll_remove_double(req, req->io);
4806 do_complete = __io_poll_remove_one(req, &req->poll);
4808 struct async_poll *apoll = req->apoll;
4810 io_poll_remove_double(req, apoll->double_poll);
4812 /* non-poll requests have submit ref still */
4813 do_complete = __io_poll_remove_one(req, &apoll->poll);
4817 * restore ->work because we will call
4818 * io_req_clean_work below when dropping the
4821 if (req->flags & REQ_F_WORK_INITIALIZED)
4822 memcpy(&req->work, &apoll->work,
4824 kfree(apoll->double_poll);
4830 io_cqring_fill_event(req, -ECANCELED);
4831 io_commit_cqring(req->ctx);
4832 req->flags |= REQ_F_COMP_LOCKED;
4839 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4841 struct hlist_node *tmp;
4842 struct io_kiocb *req;
4845 spin_lock_irq(&ctx->completion_lock);
4846 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4847 struct hlist_head *list;
4849 list = &ctx->cancel_hash[i];
4850 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4851 posted += io_poll_remove_one(req);
4853 spin_unlock_irq(&ctx->completion_lock);
4856 io_cqring_ev_posted(ctx);
4859 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4861 struct hlist_head *list;
4862 struct io_kiocb *req;
4864 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4865 hlist_for_each_entry(req, list, hash_node) {
4866 if (sqe_addr != req->user_data)
4868 if (io_poll_remove_one(req))
4876 static int io_poll_remove_prep(struct io_kiocb *req,
4877 const struct io_uring_sqe *sqe)
4879 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4881 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4885 req->poll.addr = READ_ONCE(sqe->addr);
4890 * Find a running poll command that matches one specified in sqe->addr,
4891 * and remove it if found.
4893 static int io_poll_remove(struct io_kiocb *req)
4895 struct io_ring_ctx *ctx = req->ctx;
4899 addr = req->poll.addr;
4900 spin_lock_irq(&ctx->completion_lock);
4901 ret = io_poll_cancel(ctx, addr);
4902 spin_unlock_irq(&ctx->completion_lock);
4905 req_set_fail_links(req);
4906 io_req_complete(req, ret);
4910 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4913 struct io_kiocb *req = wait->private;
4914 struct io_poll_iocb *poll = &req->poll;
4916 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4919 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4920 struct poll_table_struct *p)
4922 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4924 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4927 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4929 struct io_poll_iocb *poll = &req->poll;
4932 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4934 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4939 events = READ_ONCE(sqe->poll32_events);
4941 events = swahw32(events);
4943 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4944 (events & EPOLLEXCLUSIVE);
4946 io_get_req_task(req);
4950 static int io_poll_add(struct io_kiocb *req)
4952 struct io_poll_iocb *poll = &req->poll;
4953 struct io_ring_ctx *ctx = req->ctx;
4954 struct io_poll_table ipt;
4957 /* ->work is in union with hash_node and others */
4958 io_req_clean_work(req);
4959 req->flags &= ~REQ_F_WORK_INITIALIZED;
4961 INIT_HLIST_NODE(&req->hash_node);
4962 ipt.pt._qproc = io_poll_queue_proc;
4964 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4967 if (mask) { /* no async, we'd stolen it */
4969 io_poll_complete(req, mask, 0);
4971 spin_unlock_irq(&ctx->completion_lock);
4974 io_cqring_ev_posted(ctx);
4980 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4982 struct io_timeout_data *data = container_of(timer,
4983 struct io_timeout_data, timer);
4984 struct io_kiocb *req = data->req;
4985 struct io_ring_ctx *ctx = req->ctx;
4986 unsigned long flags;
4988 atomic_inc(&ctx->cq_timeouts);
4990 spin_lock_irqsave(&ctx->completion_lock, flags);
4992 * We could be racing with timeout deletion. If the list is empty,
4993 * then timeout lookup already found it and will be handling it.
4995 if (!list_empty(&req->timeout.list))
4996 list_del_init(&req->timeout.list);
4998 io_cqring_fill_event(req, -ETIME);
4999 io_commit_cqring(ctx);
5000 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5002 io_cqring_ev_posted(ctx);
5003 req_set_fail_links(req);
5005 return HRTIMER_NORESTART;
5008 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5010 struct io_kiocb *req;
5013 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5014 if (user_data == req->user_data) {
5015 list_del_init(&req->timeout.list);
5024 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5028 req_set_fail_links(req);
5029 io_cqring_fill_event(req, -ECANCELED);
5034 static int io_timeout_remove_prep(struct io_kiocb *req,
5035 const struct io_uring_sqe *sqe)
5037 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5039 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5041 if (sqe->ioprio || sqe->buf_index || sqe->len)
5044 req->timeout.addr = READ_ONCE(sqe->addr);
5045 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5046 if (req->timeout.flags)
5053 * Remove or update an existing timeout command
5055 static int io_timeout_remove(struct io_kiocb *req)
5057 struct io_ring_ctx *ctx = req->ctx;
5060 spin_lock_irq(&ctx->completion_lock);
5061 ret = io_timeout_cancel(ctx, req->timeout.addr);
5063 io_cqring_fill_event(req, ret);
5064 io_commit_cqring(ctx);
5065 spin_unlock_irq(&ctx->completion_lock);
5066 io_cqring_ev_posted(ctx);
5068 req_set_fail_links(req);
5073 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5074 bool is_timeout_link)
5076 struct io_timeout_data *data;
5078 u32 off = READ_ONCE(sqe->off);
5080 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5082 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5084 if (off && is_timeout_link)
5086 flags = READ_ONCE(sqe->timeout_flags);
5087 if (flags & ~IORING_TIMEOUT_ABS)
5090 req->timeout.off = off;
5092 if (!req->io && io_alloc_async_ctx(req))
5095 data = &req->io->timeout;
5098 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5101 if (flags & IORING_TIMEOUT_ABS)
5102 data->mode = HRTIMER_MODE_ABS;
5104 data->mode = HRTIMER_MODE_REL;
5106 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5110 static int io_timeout(struct io_kiocb *req)
5112 struct io_ring_ctx *ctx = req->ctx;
5113 struct io_timeout_data *data = &req->io->timeout;
5114 struct list_head *entry;
5115 u32 tail, off = req->timeout.off;
5117 spin_lock_irq(&ctx->completion_lock);
5120 * sqe->off holds how many events that need to occur for this
5121 * timeout event to be satisfied. If it isn't set, then this is
5122 * a pure timeout request, sequence isn't used.
5124 if (io_is_timeout_noseq(req)) {
5125 entry = ctx->timeout_list.prev;
5129 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5130 req->timeout.target_seq = tail + off;
5133 * Insertion sort, ensuring the first entry in the list is always
5134 * the one we need first.
5136 list_for_each_prev(entry, &ctx->timeout_list) {
5137 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5140 if (io_is_timeout_noseq(nxt))
5142 /* nxt.seq is behind @tail, otherwise would've been completed */
5143 if (off >= nxt->timeout.target_seq - tail)
5147 list_add(&req->timeout.list, entry);
5148 data->timer.function = io_timeout_fn;
5149 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5150 spin_unlock_irq(&ctx->completion_lock);
5154 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5156 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5158 return req->user_data == (unsigned long) data;
5161 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5163 enum io_wq_cancel cancel_ret;
5166 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5167 switch (cancel_ret) {
5168 case IO_WQ_CANCEL_OK:
5171 case IO_WQ_CANCEL_RUNNING:
5174 case IO_WQ_CANCEL_NOTFOUND:
5182 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5183 struct io_kiocb *req, __u64 sqe_addr,
5186 unsigned long flags;
5189 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5190 if (ret != -ENOENT) {
5191 spin_lock_irqsave(&ctx->completion_lock, flags);
5195 spin_lock_irqsave(&ctx->completion_lock, flags);
5196 ret = io_timeout_cancel(ctx, sqe_addr);
5199 ret = io_poll_cancel(ctx, sqe_addr);
5203 io_cqring_fill_event(req, ret);
5204 io_commit_cqring(ctx);
5205 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5206 io_cqring_ev_posted(ctx);
5209 req_set_fail_links(req);
5213 static int io_async_cancel_prep(struct io_kiocb *req,
5214 const struct io_uring_sqe *sqe)
5216 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5218 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5220 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5223 req->cancel.addr = READ_ONCE(sqe->addr);
5227 static int io_async_cancel(struct io_kiocb *req)
5229 struct io_ring_ctx *ctx = req->ctx;
5231 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5235 static int io_files_update_prep(struct io_kiocb *req,
5236 const struct io_uring_sqe *sqe)
5238 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5240 if (sqe->ioprio || sqe->rw_flags)
5243 req->files_update.offset = READ_ONCE(sqe->off);
5244 req->files_update.nr_args = READ_ONCE(sqe->len);
5245 if (!req->files_update.nr_args)
5247 req->files_update.arg = READ_ONCE(sqe->addr);
5251 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5252 struct io_comp_state *cs)
5254 struct io_ring_ctx *ctx = req->ctx;
5255 struct io_uring_files_update up;
5261 up.offset = req->files_update.offset;
5262 up.fds = req->files_update.arg;
5264 mutex_lock(&ctx->uring_lock);
5265 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5266 mutex_unlock(&ctx->uring_lock);
5269 req_set_fail_links(req);
5270 __io_req_complete(req, ret, 0, cs);
5274 static int io_req_defer_prep(struct io_kiocb *req,
5275 const struct io_uring_sqe *sqe)
5282 if (io_alloc_async_ctx(req))
5285 if (io_op_defs[req->opcode].file_table) {
5286 io_req_init_async(req);
5287 ret = io_grab_files(req);
5292 switch (req->opcode) {
5295 case IORING_OP_READV:
5296 case IORING_OP_READ_FIXED:
5297 case IORING_OP_READ:
5298 ret = io_read_prep(req, sqe, true);
5300 case IORING_OP_WRITEV:
5301 case IORING_OP_WRITE_FIXED:
5302 case IORING_OP_WRITE:
5303 ret = io_write_prep(req, sqe, true);
5305 case IORING_OP_POLL_ADD:
5306 ret = io_poll_add_prep(req, sqe);
5308 case IORING_OP_POLL_REMOVE:
5309 ret = io_poll_remove_prep(req, sqe);
5311 case IORING_OP_FSYNC:
5312 ret = io_prep_fsync(req, sqe);
5314 case IORING_OP_SYNC_FILE_RANGE:
5315 ret = io_prep_sfr(req, sqe);
5317 case IORING_OP_SENDMSG:
5318 case IORING_OP_SEND:
5319 ret = io_sendmsg_prep(req, sqe);
5321 case IORING_OP_RECVMSG:
5322 case IORING_OP_RECV:
5323 ret = io_recvmsg_prep(req, sqe);
5325 case IORING_OP_CONNECT:
5326 ret = io_connect_prep(req, sqe);
5328 case IORING_OP_TIMEOUT:
5329 ret = io_timeout_prep(req, sqe, false);
5331 case IORING_OP_TIMEOUT_REMOVE:
5332 ret = io_timeout_remove_prep(req, sqe);
5334 case IORING_OP_ASYNC_CANCEL:
5335 ret = io_async_cancel_prep(req, sqe);
5337 case IORING_OP_LINK_TIMEOUT:
5338 ret = io_timeout_prep(req, sqe, true);
5340 case IORING_OP_ACCEPT:
5341 ret = io_accept_prep(req, sqe);
5343 case IORING_OP_FALLOCATE:
5344 ret = io_fallocate_prep(req, sqe);
5346 case IORING_OP_OPENAT:
5347 ret = io_openat_prep(req, sqe);
5349 case IORING_OP_CLOSE:
5350 ret = io_close_prep(req, sqe);
5352 case IORING_OP_FILES_UPDATE:
5353 ret = io_files_update_prep(req, sqe);
5355 case IORING_OP_STATX:
5356 ret = io_statx_prep(req, sqe);
5358 case IORING_OP_FADVISE:
5359 ret = io_fadvise_prep(req, sqe);
5361 case IORING_OP_MADVISE:
5362 ret = io_madvise_prep(req, sqe);
5364 case IORING_OP_OPENAT2:
5365 ret = io_openat2_prep(req, sqe);
5367 case IORING_OP_EPOLL_CTL:
5368 ret = io_epoll_ctl_prep(req, sqe);
5370 case IORING_OP_SPLICE:
5371 ret = io_splice_prep(req, sqe);
5373 case IORING_OP_PROVIDE_BUFFERS:
5374 ret = io_provide_buffers_prep(req, sqe);
5376 case IORING_OP_REMOVE_BUFFERS:
5377 ret = io_remove_buffers_prep(req, sqe);
5380 ret = io_tee_prep(req, sqe);
5383 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5392 static u32 io_get_sequence(struct io_kiocb *req)
5394 struct io_kiocb *pos;
5395 struct io_ring_ctx *ctx = req->ctx;
5396 u32 total_submitted, nr_reqs = 1;
5398 if (req->flags & REQ_F_LINK_HEAD)
5399 list_for_each_entry(pos, &req->link_list, link_list)
5402 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5403 return total_submitted - nr_reqs;
5406 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5408 struct io_ring_ctx *ctx = req->ctx;
5409 struct io_defer_entry *de;
5413 /* Still need defer if there is pending req in defer list. */
5414 if (likely(list_empty_careful(&ctx->defer_list) &&
5415 !(req->flags & REQ_F_IO_DRAIN)))
5418 seq = io_get_sequence(req);
5419 /* Still a chance to pass the sequence check */
5420 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5424 ret = io_req_defer_prep(req, sqe);
5428 io_prep_async_link(req);
5429 de = kmalloc(sizeof(*de), GFP_KERNEL);
5433 spin_lock_irq(&ctx->completion_lock);
5434 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5435 spin_unlock_irq(&ctx->completion_lock);
5440 trace_io_uring_defer(ctx, req, req->user_data);
5443 list_add_tail(&de->list, &ctx->defer_list);
5444 spin_unlock_irq(&ctx->completion_lock);
5445 return -EIOCBQUEUED;
5448 static void __io_clean_op(struct io_kiocb *req)
5450 struct io_async_ctx *io = req->io;
5452 switch (req->opcode) {
5453 case IORING_OP_READV:
5454 case IORING_OP_READ_FIXED:
5455 case IORING_OP_READ:
5456 if (req->flags & REQ_F_BUFFER_SELECTED)
5457 kfree((void *)(unsigned long)req->rw.addr);
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 if (req->flags & REQ_F_BUFFER_SELECTED)
5467 kfree(req->sr_msg.kbuf);
5469 case IORING_OP_SENDMSG:
5470 if (io->msg.iov != io->msg.fast_iov)
5473 case IORING_OP_RECV:
5474 if (req->flags & REQ_F_BUFFER_SELECTED)
5475 kfree(req->sr_msg.kbuf);
5477 case IORING_OP_SPLICE:
5479 io_put_file(req, req->splice.file_in,
5480 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5484 req->flags &= ~REQ_F_NEED_CLEANUP;
5487 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5488 bool force_nonblock, struct io_comp_state *cs)
5490 struct io_ring_ctx *ctx = req->ctx;
5493 switch (req->opcode) {
5495 ret = io_nop(req, cs);
5497 case IORING_OP_READV:
5498 case IORING_OP_READ_FIXED:
5499 case IORING_OP_READ:
5501 ret = io_read_prep(req, sqe, force_nonblock);
5505 ret = io_read(req, force_nonblock, cs);
5507 case IORING_OP_WRITEV:
5508 case IORING_OP_WRITE_FIXED:
5509 case IORING_OP_WRITE:
5511 ret = io_write_prep(req, sqe, force_nonblock);
5515 ret = io_write(req, force_nonblock, cs);
5517 case IORING_OP_FSYNC:
5519 ret = io_prep_fsync(req, sqe);
5523 ret = io_fsync(req, force_nonblock);
5525 case IORING_OP_POLL_ADD:
5527 ret = io_poll_add_prep(req, sqe);
5531 ret = io_poll_add(req);
5533 case IORING_OP_POLL_REMOVE:
5535 ret = io_poll_remove_prep(req, sqe);
5539 ret = io_poll_remove(req);
5541 case IORING_OP_SYNC_FILE_RANGE:
5543 ret = io_prep_sfr(req, sqe);
5547 ret = io_sync_file_range(req, force_nonblock);
5549 case IORING_OP_SENDMSG:
5550 case IORING_OP_SEND:
5552 ret = io_sendmsg_prep(req, sqe);
5556 if (req->opcode == IORING_OP_SENDMSG)
5557 ret = io_sendmsg(req, force_nonblock, cs);
5559 ret = io_send(req, force_nonblock, cs);
5561 case IORING_OP_RECVMSG:
5562 case IORING_OP_RECV:
5564 ret = io_recvmsg_prep(req, sqe);
5568 if (req->opcode == IORING_OP_RECVMSG)
5569 ret = io_recvmsg(req, force_nonblock, cs);
5571 ret = io_recv(req, force_nonblock, cs);
5573 case IORING_OP_TIMEOUT:
5575 ret = io_timeout_prep(req, sqe, false);
5579 ret = io_timeout(req);
5581 case IORING_OP_TIMEOUT_REMOVE:
5583 ret = io_timeout_remove_prep(req, sqe);
5587 ret = io_timeout_remove(req);
5589 case IORING_OP_ACCEPT:
5591 ret = io_accept_prep(req, sqe);
5595 ret = io_accept(req, force_nonblock, cs);
5597 case IORING_OP_CONNECT:
5599 ret = io_connect_prep(req, sqe);
5603 ret = io_connect(req, force_nonblock, cs);
5605 case IORING_OP_ASYNC_CANCEL:
5607 ret = io_async_cancel_prep(req, sqe);
5611 ret = io_async_cancel(req);
5613 case IORING_OP_FALLOCATE:
5615 ret = io_fallocate_prep(req, sqe);
5619 ret = io_fallocate(req, force_nonblock);
5621 case IORING_OP_OPENAT:
5623 ret = io_openat_prep(req, sqe);
5627 ret = io_openat(req, force_nonblock);
5629 case IORING_OP_CLOSE:
5631 ret = io_close_prep(req, sqe);
5635 ret = io_close(req, force_nonblock, cs);
5637 case IORING_OP_FILES_UPDATE:
5639 ret = io_files_update_prep(req, sqe);
5643 ret = io_files_update(req, force_nonblock, cs);
5645 case IORING_OP_STATX:
5647 ret = io_statx_prep(req, sqe);
5651 ret = io_statx(req, force_nonblock);
5653 case IORING_OP_FADVISE:
5655 ret = io_fadvise_prep(req, sqe);
5659 ret = io_fadvise(req, force_nonblock);
5661 case IORING_OP_MADVISE:
5663 ret = io_madvise_prep(req, sqe);
5667 ret = io_madvise(req, force_nonblock);
5669 case IORING_OP_OPENAT2:
5671 ret = io_openat2_prep(req, sqe);
5675 ret = io_openat2(req, force_nonblock);
5677 case IORING_OP_EPOLL_CTL:
5679 ret = io_epoll_ctl_prep(req, sqe);
5683 ret = io_epoll_ctl(req, force_nonblock, cs);
5685 case IORING_OP_SPLICE:
5687 ret = io_splice_prep(req, sqe);
5691 ret = io_splice(req, force_nonblock);
5693 case IORING_OP_PROVIDE_BUFFERS:
5695 ret = io_provide_buffers_prep(req, sqe);
5699 ret = io_provide_buffers(req, force_nonblock, cs);
5701 case IORING_OP_REMOVE_BUFFERS:
5703 ret = io_remove_buffers_prep(req, sqe);
5707 ret = io_remove_buffers(req, force_nonblock, cs);
5711 ret = io_tee_prep(req, sqe);
5715 ret = io_tee(req, force_nonblock);
5725 /* If the op doesn't have a file, we're not polling for it */
5726 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5727 const bool in_async = io_wq_current_is_worker();
5729 /* workqueue context doesn't hold uring_lock, grab it now */
5731 mutex_lock(&ctx->uring_lock);
5733 io_iopoll_req_issued(req);
5736 mutex_unlock(&ctx->uring_lock);
5742 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5744 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5745 struct io_kiocb *timeout;
5748 timeout = io_prep_linked_timeout(req);
5750 io_queue_linked_timeout(timeout);
5752 /* if NO_CANCEL is set, we must still run the work */
5753 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5754 IO_WQ_WORK_CANCEL) {
5760 ret = io_issue_sqe(req, NULL, false, NULL);
5762 * We can get EAGAIN for polled IO even though we're
5763 * forcing a sync submission from here, since we can't
5764 * wait for request slots on the block side.
5773 req_set_fail_links(req);
5774 io_req_complete(req, ret);
5777 return io_steal_work(req);
5780 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5783 struct fixed_file_table *table;
5785 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5786 return table->files[index & IORING_FILE_TABLE_MASK];
5789 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5790 int fd, struct file **out_file, bool fixed)
5792 struct io_ring_ctx *ctx = req->ctx;
5796 if (unlikely(!ctx->file_data ||
5797 (unsigned) fd >= ctx->nr_user_files))
5799 fd = array_index_nospec(fd, ctx->nr_user_files);
5800 file = io_file_from_index(ctx, fd);
5802 req->fixed_file_refs = ctx->file_data->cur_refs;
5803 percpu_ref_get(req->fixed_file_refs);
5806 trace_io_uring_file_get(ctx, fd);
5807 file = __io_file_get(state, fd);
5810 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5817 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5822 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5823 if (unlikely(!fixed && io_async_submit(req->ctx)))
5826 return io_file_get(state, req, fd, &req->file, fixed);
5829 static int io_grab_files(struct io_kiocb *req)
5832 struct io_ring_ctx *ctx = req->ctx;
5834 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5836 if (!ctx->ring_file)
5840 spin_lock_irq(&ctx->inflight_lock);
5842 * We use the f_ops->flush() handler to ensure that we can flush
5843 * out work accessing these files if the fd is closed. Check if
5844 * the fd has changed since we started down this path, and disallow
5845 * this operation if it has.
5847 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5848 list_add(&req->inflight_entry, &ctx->inflight_list);
5849 req->flags |= REQ_F_INFLIGHT;
5850 req->work.files = current->files;
5853 spin_unlock_irq(&ctx->inflight_lock);
5859 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5861 struct io_timeout_data *data = container_of(timer,
5862 struct io_timeout_data, timer);
5863 struct io_kiocb *req = data->req;
5864 struct io_ring_ctx *ctx = req->ctx;
5865 struct io_kiocb *prev = NULL;
5866 unsigned long flags;
5868 spin_lock_irqsave(&ctx->completion_lock, flags);
5871 * We don't expect the list to be empty, that will only happen if we
5872 * race with the completion of the linked work.
5874 if (!list_empty(&req->link_list)) {
5875 prev = list_entry(req->link_list.prev, struct io_kiocb,
5877 if (refcount_inc_not_zero(&prev->refs)) {
5878 list_del_init(&req->link_list);
5879 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5884 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5887 req_set_fail_links(prev);
5888 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5891 io_req_complete(req, -ETIME);
5893 return HRTIMER_NORESTART;
5896 static void io_queue_linked_timeout(struct io_kiocb *req)
5898 struct io_ring_ctx *ctx = req->ctx;
5901 * If the list is now empty, then our linked request finished before
5902 * we got a chance to setup the timer
5904 spin_lock_irq(&ctx->completion_lock);
5905 if (!list_empty(&req->link_list)) {
5906 struct io_timeout_data *data = &req->io->timeout;
5908 data->timer.function = io_link_timeout_fn;
5909 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5912 spin_unlock_irq(&ctx->completion_lock);
5914 /* drop submission reference */
5918 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5920 struct io_kiocb *nxt;
5922 if (!(req->flags & REQ_F_LINK_HEAD))
5924 if (req->flags & REQ_F_LINK_TIMEOUT)
5927 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5929 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5932 req->flags |= REQ_F_LINK_TIMEOUT;
5936 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5937 struct io_comp_state *cs)
5939 struct io_kiocb *linked_timeout;
5940 struct io_kiocb *nxt;
5941 const struct cred *old_creds = NULL;
5945 linked_timeout = io_prep_linked_timeout(req);
5947 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5948 req->work.creds != current_cred()) {
5950 revert_creds(old_creds);
5951 if (old_creds == req->work.creds)
5952 old_creds = NULL; /* restored original creds */
5954 old_creds = override_creds(req->work.creds);
5957 ret = io_issue_sqe(req, sqe, true, cs);
5960 * We async punt it if the file wasn't marked NOWAIT, or if the file
5961 * doesn't support non-blocking read/write attempts
5963 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5964 if (io_arm_poll_handler(req)) {
5966 io_queue_linked_timeout(linked_timeout);
5970 io_req_init_async(req);
5972 if (io_op_defs[req->opcode].file_table) {
5973 ret = io_grab_files(req);
5979 * Queued up for async execution, worker will release
5980 * submit reference when the iocb is actually submitted.
5982 io_queue_async_work(req);
5986 if (unlikely(ret)) {
5988 /* un-prep timeout, so it'll be killed as any other linked */
5989 req->flags &= ~REQ_F_LINK_TIMEOUT;
5990 req_set_fail_links(req);
5992 io_req_complete(req, ret);
5996 /* drop submission reference */
5997 nxt = io_put_req_find_next(req);
5999 io_queue_linked_timeout(linked_timeout);
6004 if (req->flags & REQ_F_FORCE_ASYNC)
6010 revert_creds(old_creds);
6013 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6014 struct io_comp_state *cs)
6018 ret = io_req_defer(req, sqe);
6020 if (ret != -EIOCBQUEUED) {
6022 req_set_fail_links(req);
6024 io_req_complete(req, ret);
6026 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6028 ret = io_req_defer_prep(req, sqe);
6034 * Never try inline submit of IOSQE_ASYNC is set, go straight
6035 * to async execution.
6037 io_req_init_async(req);
6038 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6039 io_queue_async_work(req);
6041 __io_queue_sqe(req, sqe, cs);
6045 static inline void io_queue_link_head(struct io_kiocb *req,
6046 struct io_comp_state *cs)
6048 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6050 io_req_complete(req, -ECANCELED);
6052 io_queue_sqe(req, NULL, cs);
6055 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6056 struct io_kiocb **link, struct io_comp_state *cs)
6058 struct io_ring_ctx *ctx = req->ctx;
6062 * If we already have a head request, queue this one for async
6063 * submittal once the head completes. If we don't have a head but
6064 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6065 * submitted sync once the chain is complete. If none of those
6066 * conditions are true (normal request), then just queue it.
6069 struct io_kiocb *head = *link;
6072 * Taking sequential execution of a link, draining both sides
6073 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6074 * requests in the link. So, it drains the head and the
6075 * next after the link request. The last one is done via
6076 * drain_next flag to persist the effect across calls.
6078 if (req->flags & REQ_F_IO_DRAIN) {
6079 head->flags |= REQ_F_IO_DRAIN;
6080 ctx->drain_next = 1;
6082 ret = io_req_defer_prep(req, sqe);
6083 if (unlikely(ret)) {
6084 /* fail even hard links since we don't submit */
6085 head->flags |= REQ_F_FAIL_LINK;
6088 trace_io_uring_link(ctx, req, head);
6089 io_get_req_task(req);
6090 list_add_tail(&req->link_list, &head->link_list);
6092 /* last request of a link, enqueue the link */
6093 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6094 io_queue_link_head(head, cs);
6098 if (unlikely(ctx->drain_next)) {
6099 req->flags |= REQ_F_IO_DRAIN;
6100 ctx->drain_next = 0;
6102 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6103 req->flags |= REQ_F_LINK_HEAD;
6104 INIT_LIST_HEAD(&req->link_list);
6106 ret = io_req_defer_prep(req, sqe);
6108 req->flags |= REQ_F_FAIL_LINK;
6111 io_queue_sqe(req, sqe, cs);
6119 * Batched submission is done, ensure local IO is flushed out.
6121 static void io_submit_state_end(struct io_submit_state *state)
6123 if (!list_empty(&state->comp.list))
6124 io_submit_flush_completions(&state->comp);
6125 blk_finish_plug(&state->plug);
6126 io_state_file_put(state);
6127 if (state->free_reqs)
6128 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6132 * Start submission side cache.
6134 static void io_submit_state_start(struct io_submit_state *state,
6135 struct io_ring_ctx *ctx, unsigned int max_ios)
6137 blk_start_plug(&state->plug);
6139 state->plug.nowait = true;
6142 INIT_LIST_HEAD(&state->comp.list);
6143 state->comp.ctx = ctx;
6144 state->free_reqs = 0;
6146 state->ios_left = max_ios;
6149 static void io_commit_sqring(struct io_ring_ctx *ctx)
6151 struct io_rings *rings = ctx->rings;
6154 * Ensure any loads from the SQEs are done at this point,
6155 * since once we write the new head, the application could
6156 * write new data to them.
6158 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6162 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6163 * that is mapped by userspace. This means that care needs to be taken to
6164 * ensure that reads are stable, as we cannot rely on userspace always
6165 * being a good citizen. If members of the sqe are validated and then later
6166 * used, it's important that those reads are done through READ_ONCE() to
6167 * prevent a re-load down the line.
6169 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6171 u32 *sq_array = ctx->sq_array;
6175 * The cached sq head (or cq tail) serves two purposes:
6177 * 1) allows us to batch the cost of updating the user visible
6179 * 2) allows the kernel side to track the head on its own, even
6180 * though the application is the one updating it.
6182 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6183 if (likely(head < ctx->sq_entries))
6184 return &ctx->sq_sqes[head];
6186 /* drop invalid entries */
6187 ctx->cached_sq_dropped++;
6188 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6192 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6194 ctx->cached_sq_head++;
6197 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6198 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6199 IOSQE_BUFFER_SELECT)
6201 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6202 const struct io_uring_sqe *sqe,
6203 struct io_submit_state *state)
6205 unsigned int sqe_flags;
6208 req->opcode = READ_ONCE(sqe->opcode);
6209 req->user_data = READ_ONCE(sqe->user_data);
6214 /* one is dropped after submission, the other at completion */
6215 refcount_set(&req->refs, 2);
6216 req->task = current;
6219 if (unlikely(req->opcode >= IORING_OP_LAST))
6222 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6225 sqe_flags = READ_ONCE(sqe->flags);
6226 /* enforce forwards compatibility on users */
6227 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6230 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6231 !io_op_defs[req->opcode].buffer_select)
6234 id = READ_ONCE(sqe->personality);
6236 io_req_init_async(req);
6237 req->work.creds = idr_find(&ctx->personality_idr, id);
6238 if (unlikely(!req->work.creds))
6240 get_cred(req->work.creds);
6243 /* same numerical values with corresponding REQ_F_*, safe to copy */
6244 req->flags |= sqe_flags;
6246 if (!io_op_defs[req->opcode].needs_file)
6249 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6252 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6253 struct file *ring_file, int ring_fd)
6255 struct io_submit_state state;
6256 struct io_kiocb *link = NULL;
6257 int i, submitted = 0;
6259 /* if we have a backlog and couldn't flush it all, return BUSY */
6260 if (test_bit(0, &ctx->sq_check_overflow)) {
6261 if (!list_empty(&ctx->cq_overflow_list) &&
6262 !io_cqring_overflow_flush(ctx, false))
6266 /* make sure SQ entry isn't read before tail */
6267 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6269 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6272 io_submit_state_start(&state, ctx, nr);
6274 ctx->ring_fd = ring_fd;
6275 ctx->ring_file = ring_file;
6277 for (i = 0; i < nr; i++) {
6278 const struct io_uring_sqe *sqe;
6279 struct io_kiocb *req;
6282 sqe = io_get_sqe(ctx);
6283 if (unlikely(!sqe)) {
6284 io_consume_sqe(ctx);
6287 req = io_alloc_req(ctx, &state);
6288 if (unlikely(!req)) {
6290 submitted = -EAGAIN;
6294 err = io_init_req(ctx, req, sqe, &state);
6295 io_consume_sqe(ctx);
6296 /* will complete beyond this point, count as submitted */
6299 if (unlikely(err)) {
6302 io_req_complete(req, err);
6306 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6307 true, io_async_submit(ctx));
6308 err = io_submit_sqe(req, sqe, &link, &state.comp);
6313 if (unlikely(submitted != nr)) {
6314 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6316 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6319 io_queue_link_head(link, &state.comp);
6320 io_submit_state_end(&state);
6322 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6323 io_commit_sqring(ctx);
6328 static int io_sq_thread(void *data)
6330 struct io_ring_ctx *ctx = data;
6331 const struct cred *old_cred;
6333 unsigned long timeout;
6336 complete(&ctx->sq_thread_comp);
6338 old_cred = override_creds(ctx->creds);
6340 timeout = jiffies + ctx->sq_thread_idle;
6341 while (!kthread_should_park()) {
6342 unsigned int to_submit;
6344 if (!list_empty(&ctx->iopoll_list)) {
6345 unsigned nr_events = 0;
6347 mutex_lock(&ctx->uring_lock);
6348 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6349 io_do_iopoll(ctx, &nr_events, 0);
6351 timeout = jiffies + ctx->sq_thread_idle;
6352 mutex_unlock(&ctx->uring_lock);
6355 to_submit = io_sqring_entries(ctx);
6358 * If submit got -EBUSY, flag us as needing the application
6359 * to enter the kernel to reap and flush events.
6361 if (!to_submit || ret == -EBUSY || need_resched()) {
6363 * Drop cur_mm before scheduling, we can't hold it for
6364 * long periods (or over schedule()). Do this before
6365 * adding ourselves to the waitqueue, as the unuse/drop
6368 io_sq_thread_drop_mm();
6371 * We're polling. If we're within the defined idle
6372 * period, then let us spin without work before going
6373 * to sleep. The exception is if we got EBUSY doing
6374 * more IO, we should wait for the application to
6375 * reap events and wake us up.
6377 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6378 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6379 !percpu_ref_is_dying(&ctx->refs))) {
6385 prepare_to_wait(&ctx->sqo_wait, &wait,
6386 TASK_INTERRUPTIBLE);
6389 * While doing polled IO, before going to sleep, we need
6390 * to check if there are new reqs added to iopoll_list,
6391 * it is because reqs may have been punted to io worker
6392 * and will be added to iopoll_list later, hence check
6393 * the iopoll_list again.
6395 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6396 !list_empty_careful(&ctx->iopoll_list)) {
6397 finish_wait(&ctx->sqo_wait, &wait);
6401 /* Tell userspace we may need a wakeup call */
6402 spin_lock_irq(&ctx->completion_lock);
6403 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6404 spin_unlock_irq(&ctx->completion_lock);
6406 to_submit = io_sqring_entries(ctx);
6407 if (!to_submit || ret == -EBUSY) {
6408 if (kthread_should_park()) {
6409 finish_wait(&ctx->sqo_wait, &wait);
6412 if (io_run_task_work()) {
6413 finish_wait(&ctx->sqo_wait, &wait);
6416 if (signal_pending(current))
6417 flush_signals(current);
6419 finish_wait(&ctx->sqo_wait, &wait);
6421 spin_lock_irq(&ctx->completion_lock);
6422 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6423 spin_unlock_irq(&ctx->completion_lock);
6427 finish_wait(&ctx->sqo_wait, &wait);
6429 spin_lock_irq(&ctx->completion_lock);
6430 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6431 spin_unlock_irq(&ctx->completion_lock);
6434 mutex_lock(&ctx->uring_lock);
6435 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6436 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6437 mutex_unlock(&ctx->uring_lock);
6438 timeout = jiffies + ctx->sq_thread_idle;
6443 io_sq_thread_drop_mm();
6444 revert_creds(old_cred);
6451 struct io_wait_queue {
6452 struct wait_queue_entry wq;
6453 struct io_ring_ctx *ctx;
6455 unsigned nr_timeouts;
6458 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6460 struct io_ring_ctx *ctx = iowq->ctx;
6463 * Wake up if we have enough events, or if a timeout occurred since we
6464 * started waiting. For timeouts, we always want to return to userspace,
6465 * regardless of event count.
6467 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6468 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6471 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6472 int wake_flags, void *key)
6474 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6477 /* use noflush == true, as we can't safely rely on locking context */
6478 if (!io_should_wake(iowq, true))
6481 return autoremove_wake_function(curr, mode, wake_flags, key);
6485 * Wait until events become available, if we don't already have some. The
6486 * application must reap them itself, as they reside on the shared cq ring.
6488 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6489 const sigset_t __user *sig, size_t sigsz)
6491 struct io_wait_queue iowq = {
6494 .func = io_wake_function,
6495 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6498 .to_wait = min_events,
6500 struct io_rings *rings = ctx->rings;
6504 if (io_cqring_events(ctx, false) >= min_events)
6506 if (!io_run_task_work())
6511 #ifdef CONFIG_COMPAT
6512 if (in_compat_syscall())
6513 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6517 ret = set_user_sigmask(sig, sigsz);
6523 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6524 trace_io_uring_cqring_wait(ctx, min_events);
6526 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6527 TASK_INTERRUPTIBLE);
6528 /* make sure we run task_work before checking for signals */
6529 if (io_run_task_work())
6531 if (signal_pending(current)) {
6532 if (current->jobctl & JOBCTL_TASK_WORK) {
6533 spin_lock_irq(¤t->sighand->siglock);
6534 current->jobctl &= ~JOBCTL_TASK_WORK;
6535 recalc_sigpending();
6536 spin_unlock_irq(¤t->sighand->siglock);
6542 if (io_should_wake(&iowq, false))
6546 finish_wait(&ctx->wait, &iowq.wq);
6548 restore_saved_sigmask_unless(ret == -EINTR);
6550 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6553 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6555 #if defined(CONFIG_UNIX)
6556 if (ctx->ring_sock) {
6557 struct sock *sock = ctx->ring_sock->sk;
6558 struct sk_buff *skb;
6560 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6566 for (i = 0; i < ctx->nr_user_files; i++) {
6569 file = io_file_from_index(ctx, i);
6576 static void io_file_ref_kill(struct percpu_ref *ref)
6578 struct fixed_file_data *data;
6580 data = container_of(ref, struct fixed_file_data, refs);
6581 complete(&data->done);
6584 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6586 struct fixed_file_data *data = ctx->file_data;
6587 struct fixed_file_ref_node *ref_node = NULL;
6588 unsigned nr_tables, i;
6593 spin_lock(&data->lock);
6594 if (!list_empty(&data->ref_list))
6595 ref_node = list_first_entry(&data->ref_list,
6596 struct fixed_file_ref_node, node);
6597 spin_unlock(&data->lock);
6599 percpu_ref_kill(&ref_node->refs);
6601 percpu_ref_kill(&data->refs);
6603 /* wait for all refs nodes to complete */
6604 flush_delayed_work(&ctx->file_put_work);
6605 wait_for_completion(&data->done);
6607 __io_sqe_files_unregister(ctx);
6608 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6609 for (i = 0; i < nr_tables; i++)
6610 kfree(data->table[i].files);
6612 percpu_ref_exit(&data->refs);
6614 ctx->file_data = NULL;
6615 ctx->nr_user_files = 0;
6619 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6621 if (ctx->sqo_thread) {
6622 wait_for_completion(&ctx->sq_thread_comp);
6624 * The park is a bit of a work-around, without it we get
6625 * warning spews on shutdown with SQPOLL set and affinity
6626 * set to a single CPU.
6628 kthread_park(ctx->sqo_thread);
6629 kthread_stop(ctx->sqo_thread);
6630 ctx->sqo_thread = NULL;
6634 static void io_finish_async(struct io_ring_ctx *ctx)
6636 io_sq_thread_stop(ctx);
6639 io_wq_destroy(ctx->io_wq);
6644 #if defined(CONFIG_UNIX)
6646 * Ensure the UNIX gc is aware of our file set, so we are certain that
6647 * the io_uring can be safely unregistered on process exit, even if we have
6648 * loops in the file referencing.
6650 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6652 struct sock *sk = ctx->ring_sock->sk;
6653 struct scm_fp_list *fpl;
6654 struct sk_buff *skb;
6657 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6661 skb = alloc_skb(0, GFP_KERNEL);
6670 fpl->user = get_uid(ctx->user);
6671 for (i = 0; i < nr; i++) {
6672 struct file *file = io_file_from_index(ctx, i + offset);
6676 fpl->fp[nr_files] = get_file(file);
6677 unix_inflight(fpl->user, fpl->fp[nr_files]);
6682 fpl->max = SCM_MAX_FD;
6683 fpl->count = nr_files;
6684 UNIXCB(skb).fp = fpl;
6685 skb->destructor = unix_destruct_scm;
6686 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6687 skb_queue_head(&sk->sk_receive_queue, skb);
6689 for (i = 0; i < nr_files; i++)
6700 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6701 * causes regular reference counting to break down. We rely on the UNIX
6702 * garbage collection to take care of this problem for us.
6704 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6706 unsigned left, total;
6710 left = ctx->nr_user_files;
6712 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6714 ret = __io_sqe_files_scm(ctx, this_files, total);
6718 total += this_files;
6724 while (total < ctx->nr_user_files) {
6725 struct file *file = io_file_from_index(ctx, total);
6735 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6741 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6746 for (i = 0; i < nr_tables; i++) {
6747 struct fixed_file_table *table = &ctx->file_data->table[i];
6748 unsigned this_files;
6750 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6751 table->files = kcalloc(this_files, sizeof(struct file *),
6755 nr_files -= this_files;
6761 for (i = 0; i < nr_tables; i++) {
6762 struct fixed_file_table *table = &ctx->file_data->table[i];
6763 kfree(table->files);
6768 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6770 #if defined(CONFIG_UNIX)
6771 struct sock *sock = ctx->ring_sock->sk;
6772 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6773 struct sk_buff *skb;
6776 __skb_queue_head_init(&list);
6779 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6780 * remove this entry and rearrange the file array.
6782 skb = skb_dequeue(head);
6784 struct scm_fp_list *fp;
6786 fp = UNIXCB(skb).fp;
6787 for (i = 0; i < fp->count; i++) {
6790 if (fp->fp[i] != file)
6793 unix_notinflight(fp->user, fp->fp[i]);
6794 left = fp->count - 1 - i;
6796 memmove(&fp->fp[i], &fp->fp[i + 1],
6797 left * sizeof(struct file *));
6804 __skb_queue_tail(&list, skb);
6814 __skb_queue_tail(&list, skb);
6816 skb = skb_dequeue(head);
6819 if (skb_peek(&list)) {
6820 spin_lock_irq(&head->lock);
6821 while ((skb = __skb_dequeue(&list)) != NULL)
6822 __skb_queue_tail(head, skb);
6823 spin_unlock_irq(&head->lock);
6830 struct io_file_put {
6831 struct list_head list;
6835 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6837 struct fixed_file_data *file_data = ref_node->file_data;
6838 struct io_ring_ctx *ctx = file_data->ctx;
6839 struct io_file_put *pfile, *tmp;
6841 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6842 list_del(&pfile->list);
6843 io_ring_file_put(ctx, pfile->file);
6847 spin_lock(&file_data->lock);
6848 list_del(&ref_node->node);
6849 spin_unlock(&file_data->lock);
6851 percpu_ref_exit(&ref_node->refs);
6853 percpu_ref_put(&file_data->refs);
6856 static void io_file_put_work(struct work_struct *work)
6858 struct io_ring_ctx *ctx;
6859 struct llist_node *node;
6861 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6862 node = llist_del_all(&ctx->file_put_llist);
6865 struct fixed_file_ref_node *ref_node;
6866 struct llist_node *next = node->next;
6868 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6869 __io_file_put_work(ref_node);
6874 static void io_file_data_ref_zero(struct percpu_ref *ref)
6876 struct fixed_file_ref_node *ref_node;
6877 struct io_ring_ctx *ctx;
6881 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6882 ctx = ref_node->file_data->ctx;
6884 if (percpu_ref_is_dying(&ctx->file_data->refs))
6887 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6889 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6891 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6894 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6895 struct io_ring_ctx *ctx)
6897 struct fixed_file_ref_node *ref_node;
6899 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6901 return ERR_PTR(-ENOMEM);
6903 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6906 return ERR_PTR(-ENOMEM);
6908 INIT_LIST_HEAD(&ref_node->node);
6909 INIT_LIST_HEAD(&ref_node->file_list);
6910 ref_node->file_data = ctx->file_data;
6914 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6916 percpu_ref_exit(&ref_node->refs);
6920 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6923 __s32 __user *fds = (__s32 __user *) arg;
6928 struct fixed_file_ref_node *ref_node;
6934 if (nr_args > IORING_MAX_FIXED_FILES)
6937 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6938 if (!ctx->file_data)
6940 ctx->file_data->ctx = ctx;
6941 init_completion(&ctx->file_data->done);
6942 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6943 spin_lock_init(&ctx->file_data->lock);
6945 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6946 ctx->file_data->table = kcalloc(nr_tables,
6947 sizeof(struct fixed_file_table),
6949 if (!ctx->file_data->table) {
6950 kfree(ctx->file_data);
6951 ctx->file_data = NULL;
6955 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6956 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6957 kfree(ctx->file_data->table);
6958 kfree(ctx->file_data);
6959 ctx->file_data = NULL;
6963 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6964 percpu_ref_exit(&ctx->file_data->refs);
6965 kfree(ctx->file_data->table);
6966 kfree(ctx->file_data);
6967 ctx->file_data = NULL;
6971 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6972 struct fixed_file_table *table;
6976 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6978 /* allow sparse sets */
6984 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6985 index = i & IORING_FILE_TABLE_MASK;
6993 * Don't allow io_uring instances to be registered. If UNIX
6994 * isn't enabled, then this causes a reference cycle and this
6995 * instance can never get freed. If UNIX is enabled we'll
6996 * handle it just fine, but there's still no point in allowing
6997 * a ring fd as it doesn't support regular read/write anyway.
6999 if (file->f_op == &io_uring_fops) {
7004 table->files[index] = file;
7008 for (i = 0; i < ctx->nr_user_files; i++) {
7009 file = io_file_from_index(ctx, i);
7013 for (i = 0; i < nr_tables; i++)
7014 kfree(ctx->file_data->table[i].files);
7016 percpu_ref_exit(&ctx->file_data->refs);
7017 kfree(ctx->file_data->table);
7018 kfree(ctx->file_data);
7019 ctx->file_data = NULL;
7020 ctx->nr_user_files = 0;
7024 ret = io_sqe_files_scm(ctx);
7026 io_sqe_files_unregister(ctx);
7030 ref_node = alloc_fixed_file_ref_node(ctx);
7031 if (IS_ERR(ref_node)) {
7032 io_sqe_files_unregister(ctx);
7033 return PTR_ERR(ref_node);
7036 ctx->file_data->cur_refs = &ref_node->refs;
7037 spin_lock(&ctx->file_data->lock);
7038 list_add(&ref_node->node, &ctx->file_data->ref_list);
7039 spin_unlock(&ctx->file_data->lock);
7040 percpu_ref_get(&ctx->file_data->refs);
7044 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7047 #if defined(CONFIG_UNIX)
7048 struct sock *sock = ctx->ring_sock->sk;
7049 struct sk_buff_head *head = &sock->sk_receive_queue;
7050 struct sk_buff *skb;
7053 * See if we can merge this file into an existing skb SCM_RIGHTS
7054 * file set. If there's no room, fall back to allocating a new skb
7055 * and filling it in.
7057 spin_lock_irq(&head->lock);
7058 skb = skb_peek(head);
7060 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7062 if (fpl->count < SCM_MAX_FD) {
7063 __skb_unlink(skb, head);
7064 spin_unlock_irq(&head->lock);
7065 fpl->fp[fpl->count] = get_file(file);
7066 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7068 spin_lock_irq(&head->lock);
7069 __skb_queue_head(head, skb);
7074 spin_unlock_irq(&head->lock);
7081 return __io_sqe_files_scm(ctx, 1, index);
7087 static int io_queue_file_removal(struct fixed_file_data *data,
7090 struct io_file_put *pfile;
7091 struct percpu_ref *refs = data->cur_refs;
7092 struct fixed_file_ref_node *ref_node;
7094 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7098 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7100 list_add(&pfile->list, &ref_node->file_list);
7105 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7106 struct io_uring_files_update *up,
7109 struct fixed_file_data *data = ctx->file_data;
7110 struct fixed_file_ref_node *ref_node;
7115 bool needs_switch = false;
7117 if (check_add_overflow(up->offset, nr_args, &done))
7119 if (done > ctx->nr_user_files)
7122 ref_node = alloc_fixed_file_ref_node(ctx);
7123 if (IS_ERR(ref_node))
7124 return PTR_ERR(ref_node);
7127 fds = u64_to_user_ptr(up->fds);
7129 struct fixed_file_table *table;
7133 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7137 i = array_index_nospec(up->offset, ctx->nr_user_files);
7138 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7139 index = i & IORING_FILE_TABLE_MASK;
7140 if (table->files[index]) {
7141 file = io_file_from_index(ctx, index);
7142 err = io_queue_file_removal(data, file);
7145 table->files[index] = NULL;
7146 needs_switch = true;
7155 * Don't allow io_uring instances to be registered. If
7156 * UNIX isn't enabled, then this causes a reference
7157 * cycle and this instance can never get freed. If UNIX
7158 * is enabled we'll handle it just fine, but there's
7159 * still no point in allowing a ring fd as it doesn't
7160 * support regular read/write anyway.
7162 if (file->f_op == &io_uring_fops) {
7167 table->files[index] = file;
7168 err = io_sqe_file_register(ctx, file, i);
7180 percpu_ref_kill(data->cur_refs);
7181 spin_lock(&data->lock);
7182 list_add(&ref_node->node, &data->ref_list);
7183 data->cur_refs = &ref_node->refs;
7184 spin_unlock(&data->lock);
7185 percpu_ref_get(&ctx->file_data->refs);
7187 destroy_fixed_file_ref_node(ref_node);
7189 return done ? done : err;
7192 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7195 struct io_uring_files_update up;
7197 if (!ctx->file_data)
7201 if (copy_from_user(&up, arg, sizeof(up)))
7206 return __io_sqe_files_update(ctx, &up, nr_args);
7209 static void io_free_work(struct io_wq_work *work)
7211 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7213 /* Consider that io_steal_work() relies on this ref */
7217 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7218 struct io_uring_params *p)
7220 struct io_wq_data data;
7222 struct io_ring_ctx *ctx_attach;
7223 unsigned int concurrency;
7226 data.user = ctx->user;
7227 data.free_work = io_free_work;
7228 data.do_work = io_wq_submit_work;
7230 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7231 /* Do QD, or 4 * CPUS, whatever is smallest */
7232 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7234 ctx->io_wq = io_wq_create(concurrency, &data);
7235 if (IS_ERR(ctx->io_wq)) {
7236 ret = PTR_ERR(ctx->io_wq);
7242 f = fdget(p->wq_fd);
7246 if (f.file->f_op != &io_uring_fops) {
7251 ctx_attach = f.file->private_data;
7252 /* @io_wq is protected by holding the fd */
7253 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7258 ctx->io_wq = ctx_attach->io_wq;
7264 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7265 struct io_uring_params *p)
7269 if (ctx->flags & IORING_SETUP_SQPOLL) {
7270 mmgrab(current->mm);
7271 ctx->sqo_mm = current->mm;
7274 if (!capable(CAP_SYS_ADMIN))
7277 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7278 if (!ctx->sq_thread_idle)
7279 ctx->sq_thread_idle = HZ;
7281 if (p->flags & IORING_SETUP_SQ_AFF) {
7282 int cpu = p->sq_thread_cpu;
7285 if (cpu >= nr_cpu_ids)
7287 if (!cpu_online(cpu))
7290 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7294 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7297 if (IS_ERR(ctx->sqo_thread)) {
7298 ret = PTR_ERR(ctx->sqo_thread);
7299 ctx->sqo_thread = NULL;
7302 wake_up_process(ctx->sqo_thread);
7303 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7304 /* Can't have SQ_AFF without SQPOLL */
7309 ret = io_init_wq_offload(ctx, p);
7315 io_finish_async(ctx);
7317 mmdrop(ctx->sqo_mm);
7323 static inline void __io_unaccount_mem(struct user_struct *user,
7324 unsigned long nr_pages)
7326 atomic_long_sub(nr_pages, &user->locked_vm);
7329 static inline int __io_account_mem(struct user_struct *user,
7330 unsigned long nr_pages)
7332 unsigned long page_limit, cur_pages, new_pages;
7334 /* Don't allow more pages than we can safely lock */
7335 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7338 cur_pages = atomic_long_read(&user->locked_vm);
7339 new_pages = cur_pages + nr_pages;
7340 if (new_pages > page_limit)
7342 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7343 new_pages) != cur_pages);
7348 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7349 enum io_mem_account acct)
7352 __io_unaccount_mem(ctx->user, nr_pages);
7355 if (acct == ACCT_LOCKED)
7356 ctx->sqo_mm->locked_vm -= nr_pages;
7357 else if (acct == ACCT_PINNED)
7358 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7362 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7363 enum io_mem_account acct)
7367 if (ctx->limit_mem) {
7368 ret = __io_account_mem(ctx->user, nr_pages);
7374 if (acct == ACCT_LOCKED)
7375 ctx->sqo_mm->locked_vm += nr_pages;
7376 else if (acct == ACCT_PINNED)
7377 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7383 static void io_mem_free(void *ptr)
7390 page = virt_to_head_page(ptr);
7391 if (put_page_testzero(page))
7392 free_compound_page(page);
7395 static void *io_mem_alloc(size_t size)
7397 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7400 return (void *) __get_free_pages(gfp_flags, get_order(size));
7403 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7406 struct io_rings *rings;
7407 size_t off, sq_array_size;
7409 off = struct_size(rings, cqes, cq_entries);
7410 if (off == SIZE_MAX)
7414 off = ALIGN(off, SMP_CACHE_BYTES);
7422 sq_array_size = array_size(sizeof(u32), sq_entries);
7423 if (sq_array_size == SIZE_MAX)
7426 if (check_add_overflow(off, sq_array_size, &off))
7432 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7436 pages = (size_t)1 << get_order(
7437 rings_size(sq_entries, cq_entries, NULL));
7438 pages += (size_t)1 << get_order(
7439 array_size(sizeof(struct io_uring_sqe), sq_entries));
7444 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7448 if (!ctx->user_bufs)
7451 for (i = 0; i < ctx->nr_user_bufs; i++) {
7452 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7454 for (j = 0; j < imu->nr_bvecs; j++)
7455 unpin_user_page(imu->bvec[j].bv_page);
7457 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7462 kfree(ctx->user_bufs);
7463 ctx->user_bufs = NULL;
7464 ctx->nr_user_bufs = 0;
7468 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7469 void __user *arg, unsigned index)
7471 struct iovec __user *src;
7473 #ifdef CONFIG_COMPAT
7475 struct compat_iovec __user *ciovs;
7476 struct compat_iovec ciov;
7478 ciovs = (struct compat_iovec __user *) arg;
7479 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7482 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7483 dst->iov_len = ciov.iov_len;
7487 src = (struct iovec __user *) arg;
7488 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7493 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7496 struct vm_area_struct **vmas = NULL;
7497 struct page **pages = NULL;
7498 int i, j, got_pages = 0;
7503 if (!nr_args || nr_args > UIO_MAXIOV)
7506 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7508 if (!ctx->user_bufs)
7511 for (i = 0; i < nr_args; i++) {
7512 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7513 unsigned long off, start, end, ubuf;
7518 ret = io_copy_iov(ctx, &iov, arg, i);
7523 * Don't impose further limits on the size and buffer
7524 * constraints here, we'll -EINVAL later when IO is
7525 * submitted if they are wrong.
7528 if (!iov.iov_base || !iov.iov_len)
7531 /* arbitrary limit, but we need something */
7532 if (iov.iov_len > SZ_1G)
7535 ubuf = (unsigned long) iov.iov_base;
7536 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7537 start = ubuf >> PAGE_SHIFT;
7538 nr_pages = end - start;
7540 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7545 if (!pages || nr_pages > got_pages) {
7548 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7550 vmas = kvmalloc_array(nr_pages,
7551 sizeof(struct vm_area_struct *),
7553 if (!pages || !vmas) {
7555 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7558 got_pages = nr_pages;
7561 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7565 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7570 mmap_read_lock(current->mm);
7571 pret = pin_user_pages(ubuf, nr_pages,
7572 FOLL_WRITE | FOLL_LONGTERM,
7574 if (pret == nr_pages) {
7575 /* don't support file backed memory */
7576 for (j = 0; j < nr_pages; j++) {
7577 struct vm_area_struct *vma = vmas[j];
7580 !is_file_hugepages(vma->vm_file)) {
7586 ret = pret < 0 ? pret : -EFAULT;
7588 mmap_read_unlock(current->mm);
7591 * if we did partial map, or found file backed vmas,
7592 * release any pages we did get
7595 unpin_user_pages(pages, pret);
7596 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7601 off = ubuf & ~PAGE_MASK;
7603 for (j = 0; j < nr_pages; j++) {
7606 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7607 imu->bvec[j].bv_page = pages[j];
7608 imu->bvec[j].bv_len = vec_len;
7609 imu->bvec[j].bv_offset = off;
7613 /* store original address for later verification */
7615 imu->len = iov.iov_len;
7616 imu->nr_bvecs = nr_pages;
7618 ctx->nr_user_bufs++;
7626 io_sqe_buffer_unregister(ctx);
7630 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7632 __s32 __user *fds = arg;
7638 if (copy_from_user(&fd, fds, sizeof(*fds)))
7641 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7642 if (IS_ERR(ctx->cq_ev_fd)) {
7643 int ret = PTR_ERR(ctx->cq_ev_fd);
7644 ctx->cq_ev_fd = NULL;
7651 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7653 if (ctx->cq_ev_fd) {
7654 eventfd_ctx_put(ctx->cq_ev_fd);
7655 ctx->cq_ev_fd = NULL;
7662 static int __io_destroy_buffers(int id, void *p, void *data)
7664 struct io_ring_ctx *ctx = data;
7665 struct io_buffer *buf = p;
7667 __io_remove_buffers(ctx, buf, id, -1U);
7671 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7673 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7674 idr_destroy(&ctx->io_buffer_idr);
7677 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7679 io_finish_async(ctx);
7681 mmdrop(ctx->sqo_mm);
7685 io_sqe_buffer_unregister(ctx);
7686 io_sqe_files_unregister(ctx);
7687 io_eventfd_unregister(ctx);
7688 io_destroy_buffers(ctx);
7689 idr_destroy(&ctx->personality_idr);
7691 #if defined(CONFIG_UNIX)
7692 if (ctx->ring_sock) {
7693 ctx->ring_sock->file = NULL; /* so that iput() is called */
7694 sock_release(ctx->ring_sock);
7698 io_mem_free(ctx->rings);
7699 io_mem_free(ctx->sq_sqes);
7701 percpu_ref_exit(&ctx->refs);
7702 free_uid(ctx->user);
7703 put_cred(ctx->creds);
7704 kfree(ctx->cancel_hash);
7705 kmem_cache_free(req_cachep, ctx->fallback_req);
7709 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7711 struct io_ring_ctx *ctx = file->private_data;
7714 poll_wait(file, &ctx->cq_wait, wait);
7716 * synchronizes with barrier from wq_has_sleeper call in
7720 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7721 ctx->rings->sq_ring_entries)
7722 mask |= EPOLLOUT | EPOLLWRNORM;
7723 if (io_cqring_events(ctx, false))
7724 mask |= EPOLLIN | EPOLLRDNORM;
7729 static int io_uring_fasync(int fd, struct file *file, int on)
7731 struct io_ring_ctx *ctx = file->private_data;
7733 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7736 static int io_remove_personalities(int id, void *p, void *data)
7738 struct io_ring_ctx *ctx = data;
7739 const struct cred *cred;
7741 cred = idr_remove(&ctx->personality_idr, id);
7747 static void io_ring_exit_work(struct work_struct *work)
7749 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7753 * If we're doing polled IO and end up having requests being
7754 * submitted async (out-of-line), then completions can come in while
7755 * we're waiting for refs to drop. We need to reap these manually,
7756 * as nobody else will be looking for them.
7760 io_cqring_overflow_flush(ctx, true);
7761 io_iopoll_try_reap_events(ctx);
7762 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7763 io_ring_ctx_free(ctx);
7766 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7768 mutex_lock(&ctx->uring_lock);
7769 percpu_ref_kill(&ctx->refs);
7770 mutex_unlock(&ctx->uring_lock);
7772 io_kill_timeouts(ctx);
7773 io_poll_remove_all(ctx);
7776 io_wq_cancel_all(ctx->io_wq);
7778 /* if we failed setting up the ctx, we might not have any rings */
7780 io_cqring_overflow_flush(ctx, true);
7781 io_iopoll_try_reap_events(ctx);
7782 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7785 * Do this upfront, so we won't have a grace period where the ring
7786 * is closed but resources aren't reaped yet. This can cause
7787 * spurious failure in setting up a new ring.
7789 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7792 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7793 queue_work(system_wq, &ctx->exit_work);
7796 static int io_uring_release(struct inode *inode, struct file *file)
7798 struct io_ring_ctx *ctx = file->private_data;
7800 file->private_data = NULL;
7801 io_ring_ctx_wait_and_kill(ctx);
7805 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7807 struct files_struct *files = data;
7809 return work->files == files;
7812 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7813 struct files_struct *files)
7815 if (list_empty_careful(&ctx->inflight_list))
7818 /* cancel all at once, should be faster than doing it one by one*/
7819 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7821 while (!list_empty_careful(&ctx->inflight_list)) {
7822 struct io_kiocb *cancel_req = NULL, *req;
7825 spin_lock_irq(&ctx->inflight_lock);
7826 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7827 if (req->work.files != files)
7829 /* req is being completed, ignore */
7830 if (!refcount_inc_not_zero(&req->refs))
7836 prepare_to_wait(&ctx->inflight_wait, &wait,
7837 TASK_UNINTERRUPTIBLE);
7838 spin_unlock_irq(&ctx->inflight_lock);
7840 /* We need to keep going until we don't find a matching req */
7844 if (cancel_req->flags & REQ_F_OVERFLOW) {
7845 spin_lock_irq(&ctx->completion_lock);
7846 list_del(&cancel_req->compl.list);
7847 cancel_req->flags &= ~REQ_F_OVERFLOW;
7848 if (list_empty(&ctx->cq_overflow_list)) {
7849 clear_bit(0, &ctx->sq_check_overflow);
7850 clear_bit(0, &ctx->cq_check_overflow);
7851 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
7853 spin_unlock_irq(&ctx->completion_lock);
7855 WRITE_ONCE(ctx->rings->cq_overflow,
7856 atomic_inc_return(&ctx->cached_cq_overflow));
7859 * Put inflight ref and overflow ref. If that's
7860 * all we had, then we're done with this request.
7862 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7863 io_free_req(cancel_req);
7864 finish_wait(&ctx->inflight_wait, &wait);
7868 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7869 io_put_req(cancel_req);
7873 finish_wait(&ctx->inflight_wait, &wait);
7877 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7879 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7880 struct task_struct *task = data;
7882 return req->task == task;
7885 static int io_uring_flush(struct file *file, void *data)
7887 struct io_ring_ctx *ctx = file->private_data;
7889 io_uring_cancel_files(ctx, data);
7892 * If the task is going away, cancel work it may have pending
7894 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7895 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7900 static void *io_uring_validate_mmap_request(struct file *file,
7901 loff_t pgoff, size_t sz)
7903 struct io_ring_ctx *ctx = file->private_data;
7904 loff_t offset = pgoff << PAGE_SHIFT;
7909 case IORING_OFF_SQ_RING:
7910 case IORING_OFF_CQ_RING:
7913 case IORING_OFF_SQES:
7917 return ERR_PTR(-EINVAL);
7920 page = virt_to_head_page(ptr);
7921 if (sz > page_size(page))
7922 return ERR_PTR(-EINVAL);
7929 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7931 size_t sz = vma->vm_end - vma->vm_start;
7935 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7937 return PTR_ERR(ptr);
7939 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7940 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7943 #else /* !CONFIG_MMU */
7945 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7947 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7950 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7952 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7955 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7956 unsigned long addr, unsigned long len,
7957 unsigned long pgoff, unsigned long flags)
7961 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7963 return PTR_ERR(ptr);
7965 return (unsigned long) ptr;
7968 #endif /* !CONFIG_MMU */
7970 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7971 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7974 struct io_ring_ctx *ctx;
7981 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7989 if (f.file->f_op != &io_uring_fops)
7993 ctx = f.file->private_data;
7994 if (!percpu_ref_tryget(&ctx->refs))
7998 * For SQ polling, the thread will do all submissions and completions.
7999 * Just return the requested submit count, and wake the thread if
8003 if (ctx->flags & IORING_SETUP_SQPOLL) {
8004 if (!list_empty_careful(&ctx->cq_overflow_list))
8005 io_cqring_overflow_flush(ctx, false);
8006 if (flags & IORING_ENTER_SQ_WAKEUP)
8007 wake_up(&ctx->sqo_wait);
8008 submitted = to_submit;
8009 } else if (to_submit) {
8010 mutex_lock(&ctx->uring_lock);
8011 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8012 mutex_unlock(&ctx->uring_lock);
8014 if (submitted != to_submit)
8017 if (flags & IORING_ENTER_GETEVENTS) {
8018 min_complete = min(min_complete, ctx->cq_entries);
8021 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8022 * space applications don't need to do io completion events
8023 * polling again, they can rely on io_sq_thread to do polling
8024 * work, which can reduce cpu usage and uring_lock contention.
8026 if (ctx->flags & IORING_SETUP_IOPOLL &&
8027 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8028 ret = io_iopoll_check(ctx, min_complete);
8030 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8035 percpu_ref_put(&ctx->refs);
8038 return submitted ? submitted : ret;
8041 #ifdef CONFIG_PROC_FS
8042 static int io_uring_show_cred(int id, void *p, void *data)
8044 const struct cred *cred = p;
8045 struct seq_file *m = data;
8046 struct user_namespace *uns = seq_user_ns(m);
8047 struct group_info *gi;
8052 seq_printf(m, "%5d\n", id);
8053 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8054 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8055 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8056 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8057 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8058 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8059 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8060 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8061 seq_puts(m, "\n\tGroups:\t");
8062 gi = cred->group_info;
8063 for (g = 0; g < gi->ngroups; g++) {
8064 seq_put_decimal_ull(m, g ? " " : "",
8065 from_kgid_munged(uns, gi->gid[g]));
8067 seq_puts(m, "\n\tCapEff:\t");
8068 cap = cred->cap_effective;
8069 CAP_FOR_EACH_U32(__capi)
8070 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8075 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8079 mutex_lock(&ctx->uring_lock);
8080 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8081 for (i = 0; i < ctx->nr_user_files; i++) {
8082 struct fixed_file_table *table;
8085 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8086 f = table->files[i & IORING_FILE_TABLE_MASK];
8088 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8090 seq_printf(m, "%5u: <none>\n", i);
8092 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8093 for (i = 0; i < ctx->nr_user_bufs; i++) {
8094 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8096 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8097 (unsigned int) buf->len);
8099 if (!idr_is_empty(&ctx->personality_idr)) {
8100 seq_printf(m, "Personalities:\n");
8101 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8103 seq_printf(m, "PollList:\n");
8104 spin_lock_irq(&ctx->completion_lock);
8105 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8106 struct hlist_head *list = &ctx->cancel_hash[i];
8107 struct io_kiocb *req;
8109 hlist_for_each_entry(req, list, hash_node)
8110 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8111 req->task->task_works != NULL);
8113 spin_unlock_irq(&ctx->completion_lock);
8114 mutex_unlock(&ctx->uring_lock);
8117 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8119 struct io_ring_ctx *ctx = f->private_data;
8121 if (percpu_ref_tryget(&ctx->refs)) {
8122 __io_uring_show_fdinfo(ctx, m);
8123 percpu_ref_put(&ctx->refs);
8128 static const struct file_operations io_uring_fops = {
8129 .release = io_uring_release,
8130 .flush = io_uring_flush,
8131 .mmap = io_uring_mmap,
8133 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8134 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8136 .poll = io_uring_poll,
8137 .fasync = io_uring_fasync,
8138 #ifdef CONFIG_PROC_FS
8139 .show_fdinfo = io_uring_show_fdinfo,
8143 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8144 struct io_uring_params *p)
8146 struct io_rings *rings;
8147 size_t size, sq_array_offset;
8149 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8150 if (size == SIZE_MAX)
8153 rings = io_mem_alloc(size);
8158 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8159 rings->sq_ring_mask = p->sq_entries - 1;
8160 rings->cq_ring_mask = p->cq_entries - 1;
8161 rings->sq_ring_entries = p->sq_entries;
8162 rings->cq_ring_entries = p->cq_entries;
8163 ctx->sq_mask = rings->sq_ring_mask;
8164 ctx->cq_mask = rings->cq_ring_mask;
8165 ctx->sq_entries = rings->sq_ring_entries;
8166 ctx->cq_entries = rings->cq_ring_entries;
8168 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8169 if (size == SIZE_MAX) {
8170 io_mem_free(ctx->rings);
8175 ctx->sq_sqes = io_mem_alloc(size);
8176 if (!ctx->sq_sqes) {
8177 io_mem_free(ctx->rings);
8186 * Allocate an anonymous fd, this is what constitutes the application
8187 * visible backing of an io_uring instance. The application mmaps this
8188 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8189 * we have to tie this fd to a socket for file garbage collection purposes.
8191 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8196 #if defined(CONFIG_UNIX)
8197 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8203 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8207 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8208 O_RDWR | O_CLOEXEC);
8211 ret = PTR_ERR(file);
8215 #if defined(CONFIG_UNIX)
8216 ctx->ring_sock->file = file;
8218 fd_install(ret, file);
8221 #if defined(CONFIG_UNIX)
8222 sock_release(ctx->ring_sock);
8223 ctx->ring_sock = NULL;
8228 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8229 struct io_uring_params __user *params)
8231 struct user_struct *user = NULL;
8232 struct io_ring_ctx *ctx;
8238 if (entries > IORING_MAX_ENTRIES) {
8239 if (!(p->flags & IORING_SETUP_CLAMP))
8241 entries = IORING_MAX_ENTRIES;
8245 * Use twice as many entries for the CQ ring. It's possible for the
8246 * application to drive a higher depth than the size of the SQ ring,
8247 * since the sqes are only used at submission time. This allows for
8248 * some flexibility in overcommitting a bit. If the application has
8249 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8250 * of CQ ring entries manually.
8252 p->sq_entries = roundup_pow_of_two(entries);
8253 if (p->flags & IORING_SETUP_CQSIZE) {
8255 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8256 * to a power-of-two, if it isn't already. We do NOT impose
8257 * any cq vs sq ring sizing.
8259 if (p->cq_entries < p->sq_entries)
8261 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8262 if (!(p->flags & IORING_SETUP_CLAMP))
8264 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8266 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8268 p->cq_entries = 2 * p->sq_entries;
8271 user = get_uid(current_user());
8272 limit_mem = !capable(CAP_IPC_LOCK);
8275 ret = __io_account_mem(user,
8276 ring_pages(p->sq_entries, p->cq_entries));
8283 ctx = io_ring_ctx_alloc(p);
8286 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8291 ctx->compat = in_compat_syscall();
8293 ctx->creds = get_current_cred();
8295 ret = io_allocate_scq_urings(ctx, p);
8299 ret = io_sq_offload_start(ctx, p);
8303 memset(&p->sq_off, 0, sizeof(p->sq_off));
8304 p->sq_off.head = offsetof(struct io_rings, sq.head);
8305 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8306 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8307 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8308 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8309 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8310 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8312 memset(&p->cq_off, 0, sizeof(p->cq_off));
8313 p->cq_off.head = offsetof(struct io_rings, cq.head);
8314 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8315 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8316 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8317 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8318 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8319 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8321 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8322 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8323 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8324 IORING_FEAT_POLL_32BITS;
8326 if (copy_to_user(params, p, sizeof(*p))) {
8331 * Install ring fd as the very last thing, so we don't risk someone
8332 * having closed it before we finish setup
8334 ret = io_uring_get_fd(ctx);
8338 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8339 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8341 ctx->limit_mem = limit_mem;
8344 io_ring_ctx_wait_and_kill(ctx);
8349 * Sets up an aio uring context, and returns the fd. Applications asks for a
8350 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8351 * params structure passed in.
8353 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8355 struct io_uring_params p;
8358 if (copy_from_user(&p, params, sizeof(p)))
8360 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8365 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8366 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8367 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8370 return io_uring_create(entries, &p, params);
8373 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8374 struct io_uring_params __user *, params)
8376 return io_uring_setup(entries, params);
8379 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8381 struct io_uring_probe *p;
8385 size = struct_size(p, ops, nr_args);
8386 if (size == SIZE_MAX)
8388 p = kzalloc(size, GFP_KERNEL);
8393 if (copy_from_user(p, arg, size))
8396 if (memchr_inv(p, 0, size))
8399 p->last_op = IORING_OP_LAST - 1;
8400 if (nr_args > IORING_OP_LAST)
8401 nr_args = IORING_OP_LAST;
8403 for (i = 0; i < nr_args; i++) {
8405 if (!io_op_defs[i].not_supported)
8406 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8411 if (copy_to_user(arg, p, size))
8418 static int io_register_personality(struct io_ring_ctx *ctx)
8420 const struct cred *creds = get_current_cred();
8423 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8424 USHRT_MAX, GFP_KERNEL);
8430 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8432 const struct cred *old_creds;
8434 old_creds = idr_remove(&ctx->personality_idr, id);
8436 put_cred(old_creds);
8443 static bool io_register_op_must_quiesce(int op)
8446 case IORING_UNREGISTER_FILES:
8447 case IORING_REGISTER_FILES_UPDATE:
8448 case IORING_REGISTER_PROBE:
8449 case IORING_REGISTER_PERSONALITY:
8450 case IORING_UNREGISTER_PERSONALITY:
8457 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8458 void __user *arg, unsigned nr_args)
8459 __releases(ctx->uring_lock)
8460 __acquires(ctx->uring_lock)
8465 * We're inside the ring mutex, if the ref is already dying, then
8466 * someone else killed the ctx or is already going through
8467 * io_uring_register().
8469 if (percpu_ref_is_dying(&ctx->refs))
8472 if (io_register_op_must_quiesce(opcode)) {
8473 percpu_ref_kill(&ctx->refs);
8476 * Drop uring mutex before waiting for references to exit. If
8477 * another thread is currently inside io_uring_enter() it might
8478 * need to grab the uring_lock to make progress. If we hold it
8479 * here across the drain wait, then we can deadlock. It's safe
8480 * to drop the mutex here, since no new references will come in
8481 * after we've killed the percpu ref.
8483 mutex_unlock(&ctx->uring_lock);
8484 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8485 mutex_lock(&ctx->uring_lock);
8487 percpu_ref_resurrect(&ctx->refs);
8494 case IORING_REGISTER_BUFFERS:
8495 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8497 case IORING_UNREGISTER_BUFFERS:
8501 ret = io_sqe_buffer_unregister(ctx);
8503 case IORING_REGISTER_FILES:
8504 ret = io_sqe_files_register(ctx, arg, nr_args);
8506 case IORING_UNREGISTER_FILES:
8510 ret = io_sqe_files_unregister(ctx);
8512 case IORING_REGISTER_FILES_UPDATE:
8513 ret = io_sqe_files_update(ctx, arg, nr_args);
8515 case IORING_REGISTER_EVENTFD:
8516 case IORING_REGISTER_EVENTFD_ASYNC:
8520 ret = io_eventfd_register(ctx, arg);
8523 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8524 ctx->eventfd_async = 1;
8526 ctx->eventfd_async = 0;
8528 case IORING_UNREGISTER_EVENTFD:
8532 ret = io_eventfd_unregister(ctx);
8534 case IORING_REGISTER_PROBE:
8536 if (!arg || nr_args > 256)
8538 ret = io_probe(ctx, arg, nr_args);
8540 case IORING_REGISTER_PERSONALITY:
8544 ret = io_register_personality(ctx);
8546 case IORING_UNREGISTER_PERSONALITY:
8550 ret = io_unregister_personality(ctx, nr_args);
8557 if (io_register_op_must_quiesce(opcode)) {
8558 /* bring the ctx back to life */
8559 percpu_ref_reinit(&ctx->refs);
8561 reinit_completion(&ctx->ref_comp);
8566 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8567 void __user *, arg, unsigned int, nr_args)
8569 struct io_ring_ctx *ctx;
8578 if (f.file->f_op != &io_uring_fops)
8581 ctx = f.file->private_data;
8583 mutex_lock(&ctx->uring_lock);
8584 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8585 mutex_unlock(&ctx->uring_lock);
8586 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8587 ctx->cq_ev_fd != NULL, ret);
8593 static int __init io_uring_init(void)
8595 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8596 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8597 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8600 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8601 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8602 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8603 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8604 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8605 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8606 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8607 BUILD_BUG_SQE_ELEM(8, __u64, off);
8608 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8609 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8610 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8611 BUILD_BUG_SQE_ELEM(24, __u32, len);
8612 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8613 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8614 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8615 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8616 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8617 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8618 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8619 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8620 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8621 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8622 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8623 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8624 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8625 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8626 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8627 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8628 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8629 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8630 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8632 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8633 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8634 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8637 __initcall(io_uring_init);