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>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
107 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
112 u32 head ____cacheline_aligned_in_smp
;
113 u32 tail ____cacheline_aligned_in_smp
;
117 * This data is shared with the application through the mmap at offsets
118 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
120 * The offsets to the member fields are published through struct
121 * io_sqring_offsets when calling io_uring_setup.
125 * Head and tail offsets into the ring; the offsets need to be
126 * masked to get valid indices.
128 * The kernel controls head of the sq ring and the tail of the cq ring,
129 * and the application controls tail of the sq ring and the head of the
132 struct io_uring sq
, cq
;
134 * Bitmasks to apply to head and tail offsets (constant, equals
137 u32 sq_ring_mask
, cq_ring_mask
;
138 /* Ring sizes (constant, power of 2) */
139 u32 sq_ring_entries
, cq_ring_entries
;
141 * Number of invalid entries dropped by the kernel due to
142 * invalid index stored in array
144 * Written by the kernel, shouldn't be modified by the
145 * application (i.e. get number of "new events" by comparing to
148 * After a new SQ head value was read by the application this
149 * counter includes all submissions that were dropped reaching
150 * the new SQ head (and possibly more).
156 * Written by the kernel, shouldn't be modified by the
159 * The application needs a full memory barrier before checking
160 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
166 * Written by the application, shouldn't be modified by the
171 * Number of completion events lost because the queue was full;
172 * this should be avoided by the application by making sure
173 * there are not more requests pending than there is space in
174 * the completion queue.
176 * Written by the kernel, shouldn't be modified by the
177 * application (i.e. get number of "new events" by comparing to
180 * As completion events come in out of order this counter is not
181 * ordered with any other data.
185 * Ring buffer of completion events.
187 * The kernel writes completion events fresh every time they are
188 * produced, so the application is allowed to modify pending
191 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
194 enum io_uring_cmd_flags
{
195 IO_URING_F_NONBLOCK
= 1,
196 IO_URING_F_COMPLETE_DEFER
= 2,
199 struct io_mapped_ubuf
{
202 struct bio_vec
*bvec
;
203 unsigned int nr_bvecs
;
204 unsigned long acct_pages
;
210 struct list_head list
;
217 struct fixed_rsrc_table
{
221 struct fixed_rsrc_ref_node
{
222 struct percpu_ref refs
;
223 struct list_head node
;
224 struct list_head rsrc_list
;
225 struct fixed_rsrc_data
*rsrc_data
;
226 void (*rsrc_put
)(struct io_ring_ctx
*ctx
,
227 struct io_rsrc_put
*prsrc
);
228 struct llist_node llist
;
232 struct fixed_rsrc_data
{
233 struct fixed_rsrc_table
*table
;
234 struct io_ring_ctx
*ctx
;
236 struct fixed_rsrc_ref_node
*node
;
237 struct percpu_ref refs
;
238 struct completion done
;
243 struct list_head list
;
249 struct io_restriction
{
250 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
251 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
252 u8 sqe_flags_allowed
;
253 u8 sqe_flags_required
;
261 /* ctx's that are using this sqd */
262 struct list_head ctx_list
;
263 struct list_head ctx_new_list
;
264 struct mutex ctx_lock
;
266 struct task_struct
*thread
;
267 struct wait_queue_head wait
;
269 unsigned sq_thread_idle
;
272 #define IO_IOPOLL_BATCH 8
273 #define IO_COMPL_BATCH 32
274 #define IO_REQ_CACHE_SIZE 32
275 #define IO_REQ_ALLOC_BATCH 8
277 struct io_comp_state
{
278 struct io_kiocb
*reqs
[IO_COMPL_BATCH
];
280 unsigned int locked_free_nr
;
281 /* inline/task_work completion list, under ->uring_lock */
282 struct list_head free_list
;
283 /* IRQ completion list, under ->completion_lock */
284 struct list_head locked_free_list
;
287 struct io_submit_link
{
288 struct io_kiocb
*head
;
289 struct io_kiocb
*last
;
292 struct io_submit_state
{
293 struct blk_plug plug
;
294 struct io_submit_link link
;
297 * io_kiocb alloc cache
299 void *reqs
[IO_REQ_CACHE_SIZE
];
300 unsigned int free_reqs
;
305 * Batch completion logic
307 struct io_comp_state comp
;
310 * File reference cache
314 unsigned int file_refs
;
315 unsigned int ios_left
;
320 struct percpu_ref refs
;
321 } ____cacheline_aligned_in_smp
;
325 unsigned int compat
: 1;
326 unsigned int limit_mem
: 1;
327 unsigned int cq_overflow_flushed
: 1;
328 unsigned int drain_next
: 1;
329 unsigned int eventfd_async
: 1;
330 unsigned int restricted
: 1;
331 unsigned int sqo_dead
: 1;
334 * Ring buffer of indices into array of io_uring_sqe, which is
335 * mmapped by the application using the IORING_OFF_SQES offset.
337 * This indirection could e.g. be used to assign fixed
338 * io_uring_sqe entries to operations and only submit them to
339 * the queue when needed.
341 * The kernel modifies neither the indices array nor the entries
345 unsigned cached_sq_head
;
348 unsigned sq_thread_idle
;
349 unsigned cached_sq_dropped
;
350 unsigned cached_cq_overflow
;
351 unsigned long sq_check_overflow
;
353 struct list_head defer_list
;
354 struct list_head timeout_list
;
355 struct list_head cq_overflow_list
;
357 struct io_uring_sqe
*sq_sqes
;
358 } ____cacheline_aligned_in_smp
;
361 struct mutex uring_lock
;
362 wait_queue_head_t wait
;
363 } ____cacheline_aligned_in_smp
;
365 struct io_submit_state submit_state
;
367 struct io_rings
*rings
;
373 * For SQPOLL usage - we hold a reference to the parent task, so we
374 * have access to the ->files
376 struct task_struct
*sqo_task
;
378 /* Only used for accounting purposes */
379 struct mm_struct
*mm_account
;
381 #ifdef CONFIG_BLK_CGROUP
382 struct cgroup_subsys_state
*sqo_blkcg_css
;
385 struct io_sq_data
*sq_data
; /* if using sq thread polling */
387 struct wait_queue_head sqo_sq_wait
;
388 struct list_head sqd_list
;
391 * If used, fixed file set. Writers must ensure that ->refs is dead,
392 * readers must ensure that ->refs is alive as long as the file* is
393 * used. Only updated through io_uring_register(2).
395 struct fixed_rsrc_data
*file_data
;
396 unsigned nr_user_files
;
398 /* if used, fixed mapped user buffers */
399 unsigned nr_user_bufs
;
400 struct io_mapped_ubuf
*user_bufs
;
402 struct user_struct
*user
;
404 const struct cred
*creds
;
408 unsigned int sessionid
;
411 struct completion ref_comp
;
412 struct completion sq_thread_comp
;
414 #if defined(CONFIG_UNIX)
415 struct socket
*ring_sock
;
418 struct idr io_buffer_idr
;
420 struct idr personality_idr
;
423 unsigned cached_cq_tail
;
426 atomic_t cq_timeouts
;
427 unsigned cq_last_tm_flush
;
428 unsigned long cq_check_overflow
;
429 struct wait_queue_head cq_wait
;
430 struct fasync_struct
*cq_fasync
;
431 struct eventfd_ctx
*cq_ev_fd
;
432 } ____cacheline_aligned_in_smp
;
435 spinlock_t completion_lock
;
438 * ->iopoll_list is protected by the ctx->uring_lock for
439 * io_uring instances that don't use IORING_SETUP_SQPOLL.
440 * For SQPOLL, only the single threaded io_sq_thread() will
441 * manipulate the list, hence no extra locking is needed there.
443 struct list_head iopoll_list
;
444 struct hlist_head
*cancel_hash
;
445 unsigned cancel_hash_bits
;
446 bool poll_multi_file
;
448 spinlock_t inflight_lock
;
449 struct list_head inflight_list
;
450 } ____cacheline_aligned_in_smp
;
452 struct delayed_work rsrc_put_work
;
453 struct llist_head rsrc_put_llist
;
454 struct list_head rsrc_ref_list
;
455 spinlock_t rsrc_ref_lock
;
457 struct io_restriction restrictions
;
459 /* Keep this last, we don't need it for the fast path */
460 struct work_struct exit_work
;
464 * First field must be the file pointer in all the
465 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
467 struct io_poll_iocb
{
469 struct wait_queue_head
*head
;
473 struct wait_queue_entry wait
;
476 struct io_poll_remove
{
486 struct io_timeout_data
{
487 struct io_kiocb
*req
;
488 struct hrtimer timer
;
489 struct timespec64 ts
;
490 enum hrtimer_mode mode
;
495 struct sockaddr __user
*addr
;
496 int __user
*addr_len
;
498 unsigned long nofile
;
518 struct list_head list
;
519 /* head of the link, used by linked timeouts only */
520 struct io_kiocb
*head
;
523 struct io_timeout_rem
{
528 struct timespec64 ts
;
533 /* NOTE: kiocb has the file as the first member, so don't do it here */
541 struct sockaddr __user
*addr
;
548 struct user_msghdr __user
*umsg
;
554 struct io_buffer
*kbuf
;
560 struct filename
*filename
;
562 unsigned long nofile
;
565 struct io_rsrc_update
{
591 struct epoll_event event
;
595 struct file
*file_out
;
596 struct file
*file_in
;
603 struct io_provide_buf
{
617 const char __user
*filename
;
618 struct statx __user
*buffer
;
630 struct filename
*oldpath
;
631 struct filename
*newpath
;
639 struct filename
*filename
;
642 struct io_completion
{
644 struct list_head list
;
648 struct io_async_connect
{
649 struct sockaddr_storage address
;
652 struct io_async_msghdr
{
653 struct iovec fast_iov
[UIO_FASTIOV
];
654 /* points to an allocated iov, if NULL we use fast_iov instead */
655 struct iovec
*free_iov
;
656 struct sockaddr __user
*uaddr
;
658 struct sockaddr_storage addr
;
662 struct iovec fast_iov
[UIO_FASTIOV
];
663 const struct iovec
*free_iovec
;
664 struct iov_iter iter
;
666 struct wait_page_queue wpq
;
670 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
671 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
672 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
673 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
674 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
675 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
681 REQ_F_LINK_TIMEOUT_BIT
,
683 REQ_F_NEED_CLEANUP_BIT
,
685 REQ_F_BUFFER_SELECTED_BIT
,
686 REQ_F_NO_FILE_TABLE_BIT
,
687 REQ_F_WORK_INITIALIZED_BIT
,
688 REQ_F_LTIMEOUT_ACTIVE_BIT
,
689 REQ_F_COMPLETE_INLINE_BIT
,
691 /* not a real bit, just to check we're not overflowing the space */
697 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
698 /* drain existing IO first */
699 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
701 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
702 /* doesn't sever on completion < 0 */
703 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
705 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
706 /* IOSQE_BUFFER_SELECT */
707 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
709 /* fail rest of links */
710 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
711 /* on inflight list */
712 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
713 /* read/write uses file position */
714 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
715 /* must not punt to workers */
716 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
717 /* has or had linked timeout */
718 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
720 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
722 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
723 /* already went through poll handler */
724 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
725 /* buffer already selected */
726 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
727 /* doesn't need file table for this request */
728 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
729 /* io_wq_work is initialized */
730 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
731 /* linked timeout is active, i.e. prepared by link's head */
732 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
733 /* completion is deferred through io_comp_state */
734 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
738 struct io_poll_iocb poll
;
739 struct io_poll_iocb
*double_poll
;
742 struct io_task_work
{
743 struct io_wq_work_node node
;
744 task_work_func_t func
;
748 * NOTE! Each of the iocb union members has the file pointer
749 * as the first entry in their struct definition. So you can
750 * access the file pointer through any of the sub-structs,
751 * or directly as just 'ki_filp' in this struct.
757 struct io_poll_iocb poll
;
758 struct io_poll_remove poll_remove
;
759 struct io_accept accept
;
761 struct io_cancel cancel
;
762 struct io_timeout timeout
;
763 struct io_timeout_rem timeout_rem
;
764 struct io_connect connect
;
765 struct io_sr_msg sr_msg
;
767 struct io_close close
;
768 struct io_rsrc_update rsrc_update
;
769 struct io_fadvise fadvise
;
770 struct io_madvise madvise
;
771 struct io_epoll epoll
;
772 struct io_splice splice
;
773 struct io_provide_buf pbuf
;
774 struct io_statx statx
;
775 struct io_shutdown shutdown
;
776 struct io_rename rename
;
777 struct io_unlink unlink
;
778 /* use only after cleaning per-op data, see io_clean_op() */
779 struct io_completion
compl;
782 /* opcode allocated if it needs to store data for async defer */
785 /* polled IO has completed */
791 struct io_ring_ctx
*ctx
;
794 struct task_struct
*task
;
797 struct io_kiocb
*link
;
798 struct percpu_ref
*fixed_rsrc_refs
;
801 * 1. used with ctx->iopoll_list with reads/writes
802 * 2. to track reqs with ->files (see io_op_def::file_table)
804 struct list_head inflight_entry
;
806 struct io_task_work io_task_work
;
807 struct callback_head task_work
;
809 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
810 struct hlist_node hash_node
;
811 struct async_poll
*apoll
;
812 struct io_wq_work work
;
815 struct io_defer_entry
{
816 struct list_head list
;
817 struct io_kiocb
*req
;
822 /* needs req->file assigned */
823 unsigned needs_file
: 1;
824 /* hash wq insertion if file is a regular file */
825 unsigned hash_reg_file
: 1;
826 /* unbound wq insertion if file is a non-regular file */
827 unsigned unbound_nonreg_file
: 1;
828 /* opcode is not supported by this kernel */
829 unsigned not_supported
: 1;
830 /* set if opcode supports polled "wait" */
832 unsigned pollout
: 1;
833 /* op supports buffer selection */
834 unsigned buffer_select
: 1;
835 /* must always have async data allocated */
836 unsigned needs_async_data
: 1;
837 /* should block plug */
839 /* size of async data needed, if any */
840 unsigned short async_size
;
844 static const struct io_op_def io_op_defs
[] = {
845 [IORING_OP_NOP
] = {},
846 [IORING_OP_READV
] = {
848 .unbound_nonreg_file
= 1,
851 .needs_async_data
= 1,
853 .async_size
= sizeof(struct io_async_rw
),
854 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
856 [IORING_OP_WRITEV
] = {
859 .unbound_nonreg_file
= 1,
861 .needs_async_data
= 1,
863 .async_size
= sizeof(struct io_async_rw
),
864 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
867 [IORING_OP_FSYNC
] = {
869 .work_flags
= IO_WQ_WORK_BLKCG
,
871 [IORING_OP_READ_FIXED
] = {
873 .unbound_nonreg_file
= 1,
876 .async_size
= sizeof(struct io_async_rw
),
877 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
879 [IORING_OP_WRITE_FIXED
] = {
882 .unbound_nonreg_file
= 1,
885 .async_size
= sizeof(struct io_async_rw
),
886 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
|
889 [IORING_OP_POLL_ADD
] = {
891 .unbound_nonreg_file
= 1,
893 [IORING_OP_POLL_REMOVE
] = {},
894 [IORING_OP_SYNC_FILE_RANGE
] = {
896 .work_flags
= IO_WQ_WORK_BLKCG
,
898 [IORING_OP_SENDMSG
] = {
900 .unbound_nonreg_file
= 1,
902 .needs_async_data
= 1,
903 .async_size
= sizeof(struct io_async_msghdr
),
904 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
907 [IORING_OP_RECVMSG
] = {
909 .unbound_nonreg_file
= 1,
912 .needs_async_data
= 1,
913 .async_size
= sizeof(struct io_async_msghdr
),
914 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
917 [IORING_OP_TIMEOUT
] = {
918 .needs_async_data
= 1,
919 .async_size
= sizeof(struct io_timeout_data
),
920 .work_flags
= IO_WQ_WORK_MM
,
922 [IORING_OP_TIMEOUT_REMOVE
] = {
923 /* used by timeout updates' prep() */
924 .work_flags
= IO_WQ_WORK_MM
,
926 [IORING_OP_ACCEPT
] = {
928 .unbound_nonreg_file
= 1,
930 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
,
932 [IORING_OP_ASYNC_CANCEL
] = {},
933 [IORING_OP_LINK_TIMEOUT
] = {
934 .needs_async_data
= 1,
935 .async_size
= sizeof(struct io_timeout_data
),
936 .work_flags
= IO_WQ_WORK_MM
,
938 [IORING_OP_CONNECT
] = {
940 .unbound_nonreg_file
= 1,
942 .needs_async_data
= 1,
943 .async_size
= sizeof(struct io_async_connect
),
944 .work_flags
= IO_WQ_WORK_MM
,
946 [IORING_OP_FALLOCATE
] = {
948 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
950 [IORING_OP_OPENAT
] = {
951 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
|
952 IO_WQ_WORK_FS
| IO_WQ_WORK_MM
,
954 [IORING_OP_CLOSE
] = {
955 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
,
957 [IORING_OP_FILES_UPDATE
] = {
958 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
,
960 [IORING_OP_STATX
] = {
961 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
|
962 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
966 .unbound_nonreg_file
= 1,
970 .async_size
= sizeof(struct io_async_rw
),
971 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
973 [IORING_OP_WRITE
] = {
975 .unbound_nonreg_file
= 1,
978 .async_size
= sizeof(struct io_async_rw
),
979 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
982 [IORING_OP_FADVISE
] = {
984 .work_flags
= IO_WQ_WORK_BLKCG
,
986 [IORING_OP_MADVISE
] = {
987 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
991 .unbound_nonreg_file
= 1,
993 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
997 .unbound_nonreg_file
= 1,
1000 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
1002 [IORING_OP_OPENAT2
] = {
1003 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_FS
|
1004 IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
1006 [IORING_OP_EPOLL_CTL
] = {
1007 .unbound_nonreg_file
= 1,
1008 .work_flags
= IO_WQ_WORK_FILES
,
1010 [IORING_OP_SPLICE
] = {
1013 .unbound_nonreg_file
= 1,
1014 .work_flags
= IO_WQ_WORK_BLKCG
,
1016 [IORING_OP_PROVIDE_BUFFERS
] = {},
1017 [IORING_OP_REMOVE_BUFFERS
] = {},
1021 .unbound_nonreg_file
= 1,
1023 [IORING_OP_SHUTDOWN
] = {
1026 [IORING_OP_RENAMEAT
] = {
1027 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
1028 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
1030 [IORING_OP_UNLINKAT
] = {
1031 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
1032 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
1036 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1037 struct task_struct
*task
,
1038 struct files_struct
*files
);
1039 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node
*ref_node
);
1040 static struct fixed_rsrc_ref_node
*alloc_fixed_rsrc_ref_node(
1041 struct io_ring_ctx
*ctx
);
1042 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
);
1044 static bool io_rw_reissue(struct io_kiocb
*req
);
1045 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
1046 static void io_put_req(struct io_kiocb
*req
);
1047 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1048 static void io_double_put_req(struct io_kiocb
*req
);
1049 static void io_dismantle_req(struct io_kiocb
*req
);
1050 static void io_put_task(struct task_struct
*task
, int nr
);
1051 static void io_queue_next(struct io_kiocb
*req
);
1052 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1053 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
1054 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1055 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
1056 struct io_uring_rsrc_update
*ip
,
1058 static void __io_clean_op(struct io_kiocb
*req
);
1059 static struct file
*io_file_get(struct io_submit_state
*state
,
1060 struct io_kiocb
*req
, int fd
, bool fixed
);
1061 static void __io_queue_sqe(struct io_kiocb
*req
);
1062 static void io_rsrc_put_work(struct work_struct
*work
);
1064 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
1065 struct iov_iter
*iter
, bool needs_lock
);
1066 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
1067 const struct iovec
*fast_iov
,
1068 struct iov_iter
*iter
, bool force
);
1069 static void io_req_task_queue(struct io_kiocb
*req
);
1070 static void io_submit_flush_completions(struct io_comp_state
*cs
,
1071 struct io_ring_ctx
*ctx
);
1073 static struct kmem_cache
*req_cachep
;
1075 static const struct file_operations io_uring_fops
;
1077 struct sock
*io_uring_get_socket(struct file
*file
)
1079 #if defined(CONFIG_UNIX)
1080 if (file
->f_op
== &io_uring_fops
) {
1081 struct io_ring_ctx
*ctx
= file
->private_data
;
1083 return ctx
->ring_sock
->sk
;
1088 EXPORT_SYMBOL(io_uring_get_socket
);
1090 #define io_for_each_link(pos, head) \
1091 for (pos = (head); pos; pos = pos->link)
1093 static inline void io_clean_op(struct io_kiocb
*req
)
1095 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
))
1099 static inline void io_set_resource_node(struct io_kiocb
*req
)
1101 struct io_ring_ctx
*ctx
= req
->ctx
;
1103 if (!req
->fixed_rsrc_refs
) {
1104 req
->fixed_rsrc_refs
= &ctx
->file_data
->node
->refs
;
1105 percpu_ref_get(req
->fixed_rsrc_refs
);
1109 static bool io_refs_resurrect(struct percpu_ref
*ref
, struct completion
*compl)
1111 if (!percpu_ref_tryget(ref
)) {
1112 /* already at zero, wait for ->release() */
1113 if (!try_wait_for_completion(compl))
1118 percpu_ref_resurrect(ref
);
1119 reinit_completion(compl);
1120 percpu_ref_put(ref
);
1124 static bool io_match_task(struct io_kiocb
*head
,
1125 struct task_struct
*task
,
1126 struct files_struct
*files
)
1128 struct io_kiocb
*req
;
1130 if (task
&& head
->task
!= task
) {
1131 /* in terms of cancelation, always match if req task is dead */
1132 if (head
->task
->flags
& PF_EXITING
)
1139 io_for_each_link(req
, head
) {
1140 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1142 if (req
->file
&& req
->file
->f_op
== &io_uring_fops
)
1144 if ((req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1145 req
->work
.identity
->files
== files
)
1151 static void io_sq_thread_drop_mm_files(void)
1153 struct files_struct
*files
= current
->files
;
1154 struct mm_struct
*mm
= current
->mm
;
1157 kthread_unuse_mm(mm
);
1162 struct nsproxy
*nsproxy
= current
->nsproxy
;
1165 current
->files
= NULL
;
1166 current
->nsproxy
= NULL
;
1167 task_unlock(current
);
1168 put_files_struct(files
);
1169 put_nsproxy(nsproxy
);
1173 static int __io_sq_thread_acquire_files(struct io_ring_ctx
*ctx
)
1175 if (!current
->files
) {
1176 struct files_struct
*files
;
1177 struct nsproxy
*nsproxy
;
1179 task_lock(ctx
->sqo_task
);
1180 files
= ctx
->sqo_task
->files
;
1182 task_unlock(ctx
->sqo_task
);
1185 atomic_inc(&files
->count
);
1186 get_nsproxy(ctx
->sqo_task
->nsproxy
);
1187 nsproxy
= ctx
->sqo_task
->nsproxy
;
1188 task_unlock(ctx
->sqo_task
);
1191 current
->files
= files
;
1192 current
->nsproxy
= nsproxy
;
1193 task_unlock(current
);
1198 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1200 struct mm_struct
*mm
;
1205 task_lock(ctx
->sqo_task
);
1206 mm
= ctx
->sqo_task
->mm
;
1207 if (unlikely(!mm
|| !mmget_not_zero(mm
)))
1209 task_unlock(ctx
->sqo_task
);
1219 static int __io_sq_thread_acquire_mm_files(struct io_ring_ctx
*ctx
,
1220 struct io_kiocb
*req
)
1222 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1225 if (def
->work_flags
& IO_WQ_WORK_MM
) {
1226 ret
= __io_sq_thread_acquire_mm(ctx
);
1231 if (def
->needs_file
|| (def
->work_flags
& IO_WQ_WORK_FILES
)) {
1232 ret
= __io_sq_thread_acquire_files(ctx
);
1240 static inline int io_sq_thread_acquire_mm_files(struct io_ring_ctx
*ctx
,
1241 struct io_kiocb
*req
)
1243 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
))
1245 return __io_sq_thread_acquire_mm_files(ctx
, req
);
1248 static void io_sq_thread_associate_blkcg(struct io_ring_ctx
*ctx
,
1249 struct cgroup_subsys_state
**cur_css
)
1252 #ifdef CONFIG_BLK_CGROUP
1253 /* puts the old one when swapping */
1254 if (*cur_css
!= ctx
->sqo_blkcg_css
) {
1255 kthread_associate_blkcg(ctx
->sqo_blkcg_css
);
1256 *cur_css
= ctx
->sqo_blkcg_css
;
1261 static void io_sq_thread_unassociate_blkcg(void)
1263 #ifdef CONFIG_BLK_CGROUP
1264 kthread_associate_blkcg(NULL
);
1268 static inline void req_set_fail_links(struct io_kiocb
*req
)
1270 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1271 req
->flags
|= REQ_F_FAIL_LINK
;
1275 * None of these are dereferenced, they are simply used to check if any of
1276 * them have changed. If we're under current and check they are still the
1277 * same, we're fine to grab references to them for actual out-of-line use.
1279 static void io_init_identity(struct io_identity
*id
)
1281 id
->files
= current
->files
;
1282 id
->mm
= current
->mm
;
1283 #ifdef CONFIG_BLK_CGROUP
1285 id
->blkcg_css
= blkcg_css();
1288 id
->creds
= current_cred();
1289 id
->nsproxy
= current
->nsproxy
;
1290 id
->fs
= current
->fs
;
1291 id
->fsize
= rlimit(RLIMIT_FSIZE
);
1293 id
->loginuid
= current
->loginuid
;
1294 id
->sessionid
= current
->sessionid
;
1296 refcount_set(&id
->count
, 1);
1299 static inline void __io_req_init_async(struct io_kiocb
*req
)
1301 memset(&req
->work
, 0, sizeof(req
->work
));
1302 req
->flags
|= REQ_F_WORK_INITIALIZED
;
1306 * Note: must call io_req_init_async() for the first time you
1307 * touch any members of io_wq_work.
1309 static inline void io_req_init_async(struct io_kiocb
*req
)
1311 struct io_uring_task
*tctx
= current
->io_uring
;
1313 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
1316 __io_req_init_async(req
);
1318 /* Grab a ref if this isn't our static identity */
1319 req
->work
.identity
= tctx
->identity
;
1320 if (tctx
->identity
!= &tctx
->__identity
)
1321 refcount_inc(&req
->work
.identity
->count
);
1324 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1326 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1328 complete(&ctx
->ref_comp
);
1331 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1333 return !req
->timeout
.off
;
1336 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1338 struct io_ring_ctx
*ctx
;
1341 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1346 * Use 5 bits less than the max cq entries, that should give us around
1347 * 32 entries per hash list if totally full and uniformly spread.
1349 hash_bits
= ilog2(p
->cq_entries
);
1353 ctx
->cancel_hash_bits
= hash_bits
;
1354 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1356 if (!ctx
->cancel_hash
)
1358 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1360 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1361 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1364 ctx
->flags
= p
->flags
;
1365 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1366 INIT_LIST_HEAD(&ctx
->sqd_list
);
1367 init_waitqueue_head(&ctx
->cq_wait
);
1368 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1369 init_completion(&ctx
->ref_comp
);
1370 init_completion(&ctx
->sq_thread_comp
);
1371 idr_init(&ctx
->io_buffer_idr
);
1372 idr_init(&ctx
->personality_idr
);
1373 mutex_init(&ctx
->uring_lock
);
1374 init_waitqueue_head(&ctx
->wait
);
1375 spin_lock_init(&ctx
->completion_lock
);
1376 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1377 INIT_LIST_HEAD(&ctx
->defer_list
);
1378 INIT_LIST_HEAD(&ctx
->timeout_list
);
1379 spin_lock_init(&ctx
->inflight_lock
);
1380 INIT_LIST_HEAD(&ctx
->inflight_list
);
1381 spin_lock_init(&ctx
->rsrc_ref_lock
);
1382 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1383 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1384 init_llist_head(&ctx
->rsrc_put_llist
);
1385 INIT_LIST_HEAD(&ctx
->submit_state
.comp
.free_list
);
1386 INIT_LIST_HEAD(&ctx
->submit_state
.comp
.locked_free_list
);
1389 kfree(ctx
->cancel_hash
);
1394 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1396 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1397 struct io_ring_ctx
*ctx
= req
->ctx
;
1399 return seq
!= ctx
->cached_cq_tail
1400 + READ_ONCE(ctx
->cached_cq_overflow
);
1406 static void io_put_identity(struct io_uring_task
*tctx
, struct io_kiocb
*req
)
1408 if (req
->work
.identity
== &tctx
->__identity
)
1410 if (refcount_dec_and_test(&req
->work
.identity
->count
))
1411 kfree(req
->work
.identity
);
1414 static void io_req_clean_work(struct io_kiocb
*req
)
1416 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1419 if (req
->work
.flags
& IO_WQ_WORK_MM
)
1420 mmdrop(req
->work
.identity
->mm
);
1421 #ifdef CONFIG_BLK_CGROUP
1422 if (req
->work
.flags
& IO_WQ_WORK_BLKCG
)
1423 css_put(req
->work
.identity
->blkcg_css
);
1425 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1426 put_cred(req
->work
.identity
->creds
);
1427 if (req
->work
.flags
& IO_WQ_WORK_FS
) {
1428 struct fs_struct
*fs
= req
->work
.identity
->fs
;
1430 spin_lock(&req
->work
.identity
->fs
->lock
);
1433 spin_unlock(&req
->work
.identity
->fs
->lock
);
1437 if (req
->work
.flags
& IO_WQ_WORK_FILES
) {
1438 put_files_struct(req
->work
.identity
->files
);
1439 put_nsproxy(req
->work
.identity
->nsproxy
);
1441 if (req
->flags
& REQ_F_INFLIGHT
) {
1442 struct io_ring_ctx
*ctx
= req
->ctx
;
1443 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1444 unsigned long flags
;
1446 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1447 list_del(&req
->inflight_entry
);
1448 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1449 req
->flags
&= ~REQ_F_INFLIGHT
;
1450 if (atomic_read(&tctx
->in_idle
))
1451 wake_up(&tctx
->wait
);
1454 req
->flags
&= ~REQ_F_WORK_INITIALIZED
;
1455 req
->work
.flags
&= ~(IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FS
|
1456 IO_WQ_WORK_CREDS
| IO_WQ_WORK_FILES
);
1457 io_put_identity(req
->task
->io_uring
, req
);
1461 * Create a private copy of io_identity, since some fields don't match
1462 * the current context.
1464 static bool io_identity_cow(struct io_kiocb
*req
)
1466 struct io_uring_task
*tctx
= current
->io_uring
;
1467 const struct cred
*creds
= NULL
;
1468 struct io_identity
*id
;
1470 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1471 creds
= req
->work
.identity
->creds
;
1473 id
= kmemdup(req
->work
.identity
, sizeof(*id
), GFP_KERNEL
);
1474 if (unlikely(!id
)) {
1475 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1480 * We can safely just re-init the creds we copied Either the field
1481 * matches the current one, or we haven't grabbed it yet. The only
1482 * exception is ->creds, through registered personalities, so handle
1483 * that one separately.
1485 io_init_identity(id
);
1489 /* add one for this request */
1490 refcount_inc(&id
->count
);
1492 /* drop tctx and req identity references, if needed */
1493 if (tctx
->identity
!= &tctx
->__identity
&&
1494 refcount_dec_and_test(&tctx
->identity
->count
))
1495 kfree(tctx
->identity
);
1496 if (req
->work
.identity
!= &tctx
->__identity
&&
1497 refcount_dec_and_test(&req
->work
.identity
->count
))
1498 kfree(req
->work
.identity
);
1500 req
->work
.identity
= id
;
1501 tctx
->identity
= id
;
1505 static void io_req_track_inflight(struct io_kiocb
*req
)
1507 struct io_ring_ctx
*ctx
= req
->ctx
;
1509 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1510 io_req_init_async(req
);
1511 req
->flags
|= REQ_F_INFLIGHT
;
1513 spin_lock_irq(&ctx
->inflight_lock
);
1514 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1515 spin_unlock_irq(&ctx
->inflight_lock
);
1519 static bool io_grab_identity(struct io_kiocb
*req
)
1521 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1522 struct io_identity
*id
= req
->work
.identity
;
1524 if (def
->work_flags
& IO_WQ_WORK_FSIZE
) {
1525 if (id
->fsize
!= rlimit(RLIMIT_FSIZE
))
1527 req
->work
.flags
|= IO_WQ_WORK_FSIZE
;
1529 #ifdef CONFIG_BLK_CGROUP
1530 if (!(req
->work
.flags
& IO_WQ_WORK_BLKCG
) &&
1531 (def
->work_flags
& IO_WQ_WORK_BLKCG
)) {
1533 if (id
->blkcg_css
!= blkcg_css()) {
1538 * This should be rare, either the cgroup is dying or the task
1539 * is moving cgroups. Just punt to root for the handful of ios.
1541 if (css_tryget_online(id
->blkcg_css
))
1542 req
->work
.flags
|= IO_WQ_WORK_BLKCG
;
1546 if (!(req
->work
.flags
& IO_WQ_WORK_CREDS
)) {
1547 if (id
->creds
!= current_cred())
1549 get_cred(id
->creds
);
1550 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
1553 if (!uid_eq(current
->loginuid
, id
->loginuid
) ||
1554 current
->sessionid
!= id
->sessionid
)
1557 if (!(req
->work
.flags
& IO_WQ_WORK_FS
) &&
1558 (def
->work_flags
& IO_WQ_WORK_FS
)) {
1559 if (current
->fs
!= id
->fs
)
1561 spin_lock(&id
->fs
->lock
);
1562 if (!id
->fs
->in_exec
) {
1564 req
->work
.flags
|= IO_WQ_WORK_FS
;
1566 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1568 spin_unlock(¤t
->fs
->lock
);
1570 if (!(req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1571 (def
->work_flags
& IO_WQ_WORK_FILES
) &&
1572 !(req
->flags
& REQ_F_NO_FILE_TABLE
)) {
1573 if (id
->files
!= current
->files
||
1574 id
->nsproxy
!= current
->nsproxy
)
1576 atomic_inc(&id
->files
->count
);
1577 get_nsproxy(id
->nsproxy
);
1578 req
->work
.flags
|= IO_WQ_WORK_FILES
;
1579 io_req_track_inflight(req
);
1581 if (!(req
->work
.flags
& IO_WQ_WORK_MM
) &&
1582 (def
->work_flags
& IO_WQ_WORK_MM
)) {
1583 if (id
->mm
!= current
->mm
)
1586 req
->work
.flags
|= IO_WQ_WORK_MM
;
1592 static void io_prep_async_work(struct io_kiocb
*req
)
1594 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1595 struct io_ring_ctx
*ctx
= req
->ctx
;
1597 io_req_init_async(req
);
1599 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1600 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1602 if (req
->flags
& REQ_F_ISREG
) {
1603 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1604 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1606 if (def
->unbound_nonreg_file
)
1607 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1610 /* if we fail grabbing identity, we must COW, regrab, and retry */
1611 if (io_grab_identity(req
))
1614 if (!io_identity_cow(req
))
1617 /* can't fail at this point */
1618 if (!io_grab_identity(req
))
1622 static void io_prep_async_link(struct io_kiocb
*req
)
1624 struct io_kiocb
*cur
;
1626 io_for_each_link(cur
, req
)
1627 io_prep_async_work(cur
);
1630 static struct io_kiocb
*__io_queue_async_work(struct io_kiocb
*req
)
1632 struct io_ring_ctx
*ctx
= req
->ctx
;
1633 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1635 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1636 &req
->work
, req
->flags
);
1637 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1641 static void io_queue_async_work(struct io_kiocb
*req
)
1643 struct io_kiocb
*link
;
1645 /* init ->work of the whole link before punting */
1646 io_prep_async_link(req
);
1647 link
= __io_queue_async_work(req
);
1650 io_queue_linked_timeout(link
);
1653 static void io_kill_timeout(struct io_kiocb
*req
)
1655 struct io_timeout_data
*io
= req
->async_data
;
1658 ret
= hrtimer_try_to_cancel(&io
->timer
);
1660 atomic_set(&req
->ctx
->cq_timeouts
,
1661 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1662 list_del_init(&req
->timeout
.list
);
1663 io_cqring_fill_event(req
, 0);
1664 io_put_req_deferred(req
, 1);
1669 * Returns true if we found and killed one or more timeouts
1671 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
1672 struct files_struct
*files
)
1674 struct io_kiocb
*req
, *tmp
;
1677 spin_lock_irq(&ctx
->completion_lock
);
1678 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1679 if (io_match_task(req
, tsk
, files
)) {
1680 io_kill_timeout(req
);
1684 spin_unlock_irq(&ctx
->completion_lock
);
1685 return canceled
!= 0;
1688 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1691 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1692 struct io_defer_entry
, list
);
1694 if (req_need_defer(de
->req
, de
->seq
))
1696 list_del_init(&de
->list
);
1697 io_req_task_queue(de
->req
);
1699 } while (!list_empty(&ctx
->defer_list
));
1702 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1706 if (list_empty(&ctx
->timeout_list
))
1709 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1712 u32 events_needed
, events_got
;
1713 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1714 struct io_kiocb
, timeout
.list
);
1716 if (io_is_timeout_noseq(req
))
1720 * Since seq can easily wrap around over time, subtract
1721 * the last seq at which timeouts were flushed before comparing.
1722 * Assuming not more than 2^31-1 events have happened since,
1723 * these subtractions won't have wrapped, so we can check if
1724 * target is in [last_seq, current_seq] by comparing the two.
1726 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1727 events_got
= seq
- ctx
->cq_last_tm_flush
;
1728 if (events_got
< events_needed
)
1731 list_del_init(&req
->timeout
.list
);
1732 io_kill_timeout(req
);
1733 } while (!list_empty(&ctx
->timeout_list
));
1735 ctx
->cq_last_tm_flush
= seq
;
1738 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1740 io_flush_timeouts(ctx
);
1742 /* order cqe stores with ring update */
1743 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1745 if (unlikely(!list_empty(&ctx
->defer_list
)))
1746 __io_queue_deferred(ctx
);
1749 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1751 struct io_rings
*r
= ctx
->rings
;
1753 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1756 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1758 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1761 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1763 struct io_rings
*rings
= ctx
->rings
;
1767 * writes to the cq entry need to come after reading head; the
1768 * control dependency is enough as we're using WRITE_ONCE to
1771 if (__io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1774 tail
= ctx
->cached_cq_tail
++;
1775 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1778 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1782 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1784 if (!ctx
->eventfd_async
)
1786 return io_wq_current_is_worker();
1789 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1791 /* see waitqueue_active() comment */
1794 if (waitqueue_active(&ctx
->wait
))
1795 wake_up(&ctx
->wait
);
1796 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1797 wake_up(&ctx
->sq_data
->wait
);
1798 if (io_should_trigger_evfd(ctx
))
1799 eventfd_signal(ctx
->cq_ev_fd
, 1);
1800 if (waitqueue_active(&ctx
->cq_wait
)) {
1801 wake_up_interruptible(&ctx
->cq_wait
);
1802 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1806 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1808 /* see waitqueue_active() comment */
1811 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1812 if (waitqueue_active(&ctx
->wait
))
1813 wake_up(&ctx
->wait
);
1815 if (io_should_trigger_evfd(ctx
))
1816 eventfd_signal(ctx
->cq_ev_fd
, 1);
1817 if (waitqueue_active(&ctx
->cq_wait
)) {
1818 wake_up_interruptible(&ctx
->cq_wait
);
1819 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1823 /* Returns true if there are no backlogged entries after the flush */
1824 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1825 struct task_struct
*tsk
,
1826 struct files_struct
*files
)
1828 struct io_rings
*rings
= ctx
->rings
;
1829 struct io_kiocb
*req
, *tmp
;
1830 struct io_uring_cqe
*cqe
;
1831 unsigned long flags
;
1832 bool all_flushed
, posted
;
1835 if (!force
&& __io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1839 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1840 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1841 if (!io_match_task(req
, tsk
, files
))
1844 cqe
= io_get_cqring(ctx
);
1848 list_move(&req
->compl.list
, &list
);
1850 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1851 WRITE_ONCE(cqe
->res
, req
->result
);
1852 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1854 ctx
->cached_cq_overflow
++;
1855 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1856 ctx
->cached_cq_overflow
);
1861 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1863 clear_bit(0, &ctx
->sq_check_overflow
);
1864 clear_bit(0, &ctx
->cq_check_overflow
);
1865 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1869 io_commit_cqring(ctx
);
1870 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1872 io_cqring_ev_posted(ctx
);
1874 while (!list_empty(&list
)) {
1875 req
= list_first_entry(&list
, struct io_kiocb
, compl.list
);
1876 list_del(&req
->compl.list
);
1883 static void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1884 struct task_struct
*tsk
,
1885 struct files_struct
*files
)
1887 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1888 /* iopoll syncs against uring_lock, not completion_lock */
1889 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1890 mutex_lock(&ctx
->uring_lock
);
1891 __io_cqring_overflow_flush(ctx
, force
, tsk
, files
);
1892 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1893 mutex_unlock(&ctx
->uring_lock
);
1897 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1899 struct io_ring_ctx
*ctx
= req
->ctx
;
1900 struct io_uring_cqe
*cqe
;
1902 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1905 * If we can't get a cq entry, userspace overflowed the
1906 * submission (by quite a lot). Increment the overflow count in
1909 cqe
= io_get_cqring(ctx
);
1911 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1912 WRITE_ONCE(cqe
->res
, res
);
1913 WRITE_ONCE(cqe
->flags
, cflags
);
1914 } else if (ctx
->cq_overflow_flushed
||
1915 atomic_read(&req
->task
->io_uring
->in_idle
)) {
1917 * If we're in ring overflow flush mode, or in task cancel mode,
1918 * then we cannot store the request for later flushing, we need
1919 * to drop it on the floor.
1921 ctx
->cached_cq_overflow
++;
1922 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1924 if (list_empty(&ctx
->cq_overflow_list
)) {
1925 set_bit(0, &ctx
->sq_check_overflow
);
1926 set_bit(0, &ctx
->cq_check_overflow
);
1927 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1931 req
->compl.cflags
= cflags
;
1932 refcount_inc(&req
->refs
);
1933 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1937 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1939 __io_cqring_fill_event(req
, res
, 0);
1942 static inline void io_req_complete_post(struct io_kiocb
*req
, long res
,
1943 unsigned int cflags
)
1945 struct io_ring_ctx
*ctx
= req
->ctx
;
1946 unsigned long flags
;
1948 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1949 __io_cqring_fill_event(req
, res
, cflags
);
1950 io_commit_cqring(ctx
);
1952 * If we're the last reference to this request, add to our locked
1955 if (refcount_dec_and_test(&req
->refs
)) {
1956 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
1958 io_dismantle_req(req
);
1959 io_put_task(req
->task
, 1);
1960 list_add(&req
->compl.list
, &cs
->locked_free_list
);
1961 cs
->locked_free_nr
++;
1964 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1966 io_cqring_ev_posted(ctx
);
1969 percpu_ref_put(&ctx
->refs
);
1973 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1974 unsigned int cflags
)
1978 req
->compl.cflags
= cflags
;
1979 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1982 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1983 long res
, unsigned cflags
)
1985 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1986 io_req_complete_state(req
, res
, cflags
);
1988 io_req_complete_post(req
, res
, cflags
);
1991 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1993 __io_req_complete(req
, 0, res
, 0);
1996 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1998 struct io_submit_state
*state
= &ctx
->submit_state
;
1999 struct io_comp_state
*cs
= &state
->comp
;
2000 struct io_kiocb
*req
= NULL
;
2003 * If we have more than a batch's worth of requests in our IRQ side
2004 * locked cache, grab the lock and move them over to our submission
2007 if (READ_ONCE(cs
->locked_free_nr
) > IO_COMPL_BATCH
) {
2008 spin_lock_irq(&ctx
->completion_lock
);
2009 list_splice_init(&cs
->locked_free_list
, &cs
->free_list
);
2010 cs
->locked_free_nr
= 0;
2011 spin_unlock_irq(&ctx
->completion_lock
);
2014 while (!list_empty(&cs
->free_list
)) {
2015 req
= list_first_entry(&cs
->free_list
, struct io_kiocb
,
2017 list_del(&req
->compl.list
);
2018 state
->reqs
[state
->free_reqs
++] = req
;
2019 if (state
->free_reqs
== ARRAY_SIZE(state
->reqs
))
2026 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
2028 struct io_submit_state
*state
= &ctx
->submit_state
;
2030 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH
> ARRAY_SIZE(state
->reqs
));
2032 if (!state
->free_reqs
) {
2033 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
2036 if (io_flush_cached_reqs(ctx
))
2039 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
2043 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2044 * retry single alloc to be on the safe side.
2046 if (unlikely(ret
<= 0)) {
2047 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
2048 if (!state
->reqs
[0])
2052 state
->free_reqs
= ret
;
2056 return state
->reqs
[state
->free_reqs
];
2059 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
2066 static void io_dismantle_req(struct io_kiocb
*req
)
2070 if (req
->async_data
)
2071 kfree(req
->async_data
);
2073 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
2074 if (req
->fixed_rsrc_refs
)
2075 percpu_ref_put(req
->fixed_rsrc_refs
);
2076 io_req_clean_work(req
);
2079 static inline void io_put_task(struct task_struct
*task
, int nr
)
2081 struct io_uring_task
*tctx
= task
->io_uring
;
2083 percpu_counter_sub(&tctx
->inflight
, nr
);
2084 if (unlikely(atomic_read(&tctx
->in_idle
)))
2085 wake_up(&tctx
->wait
);
2086 put_task_struct_many(task
, nr
);
2089 static void __io_free_req(struct io_kiocb
*req
)
2091 struct io_ring_ctx
*ctx
= req
->ctx
;
2093 io_dismantle_req(req
);
2094 io_put_task(req
->task
, 1);
2096 kmem_cache_free(req_cachep
, req
);
2097 percpu_ref_put(&ctx
->refs
);
2100 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2102 struct io_kiocb
*nxt
= req
->link
;
2104 req
->link
= nxt
->link
;
2108 static void io_kill_linked_timeout(struct io_kiocb
*req
)
2110 struct io_ring_ctx
*ctx
= req
->ctx
;
2111 struct io_kiocb
*link
;
2112 bool cancelled
= false;
2113 unsigned long flags
;
2115 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2119 * Can happen if a linked timeout fired and link had been like
2120 * req -> link t-out -> link t-out [-> ...]
2122 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
2123 struct io_timeout_data
*io
= link
->async_data
;
2126 io_remove_next_linked(req
);
2127 link
->timeout
.head
= NULL
;
2128 ret
= hrtimer_try_to_cancel(&io
->timer
);
2130 io_cqring_fill_event(link
, -ECANCELED
);
2131 io_commit_cqring(ctx
);
2135 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
2136 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2139 io_cqring_ev_posted(ctx
);
2145 static void io_fail_links(struct io_kiocb
*req
)
2147 struct io_kiocb
*link
, *nxt
;
2148 struct io_ring_ctx
*ctx
= req
->ctx
;
2149 unsigned long flags
;
2151 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2159 trace_io_uring_fail_link(req
, link
);
2160 io_cqring_fill_event(link
, -ECANCELED
);
2163 * It's ok to free under spinlock as they're not linked anymore,
2164 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2167 if (link
->flags
& REQ_F_WORK_INITIALIZED
)
2168 io_put_req_deferred(link
, 2);
2170 io_double_put_req(link
);
2173 io_commit_cqring(ctx
);
2174 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2176 io_cqring_ev_posted(ctx
);
2179 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2181 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
2182 io_kill_linked_timeout(req
);
2185 * If LINK is set, we have dependent requests in this chain. If we
2186 * didn't fail this request, queue the first one up, moving any other
2187 * dependencies to the next request. In case of failure, fail the rest
2190 if (likely(!(req
->flags
& REQ_F_FAIL_LINK
))) {
2191 struct io_kiocb
*nxt
= req
->link
;
2200 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2202 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
2204 return __io_req_find_next(req
);
2207 static bool __tctx_task_work(struct io_uring_task
*tctx
)
2209 struct io_ring_ctx
*ctx
= NULL
;
2210 struct io_wq_work_list list
;
2211 struct io_wq_work_node
*node
;
2213 if (wq_list_empty(&tctx
->task_list
))
2216 spin_lock_irq(&tctx
->task_lock
);
2217 list
= tctx
->task_list
;
2218 INIT_WQ_LIST(&tctx
->task_list
);
2219 spin_unlock_irq(&tctx
->task_lock
);
2223 struct io_wq_work_node
*next
= node
->next
;
2224 struct io_ring_ctx
*this_ctx
;
2225 struct io_kiocb
*req
;
2227 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
2228 this_ctx
= req
->ctx
;
2229 req
->task_work
.func(&req
->task_work
);
2234 } else if (ctx
!= this_ctx
) {
2235 mutex_lock(&ctx
->uring_lock
);
2236 io_submit_flush_completions(&ctx
->submit_state
.comp
, ctx
);
2237 mutex_unlock(&ctx
->uring_lock
);
2242 if (ctx
&& ctx
->submit_state
.comp
.nr
) {
2243 mutex_lock(&ctx
->uring_lock
);
2244 io_submit_flush_completions(&ctx
->submit_state
.comp
, ctx
);
2245 mutex_unlock(&ctx
->uring_lock
);
2248 return list
.first
!= NULL
;
2251 static void tctx_task_work(struct callback_head
*cb
)
2253 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
, task_work
);
2255 while (__tctx_task_work(tctx
))
2258 clear_bit(0, &tctx
->task_state
);
2261 static int io_task_work_add(struct task_struct
*tsk
, struct io_kiocb
*req
,
2262 enum task_work_notify_mode notify
)
2264 struct io_uring_task
*tctx
= tsk
->io_uring
;
2265 struct io_wq_work_node
*node
, *prev
;
2266 unsigned long flags
;
2269 WARN_ON_ONCE(!tctx
);
2271 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2272 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
2273 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2275 /* task_work already pending, we're done */
2276 if (test_bit(0, &tctx
->task_state
) ||
2277 test_and_set_bit(0, &tctx
->task_state
))
2280 if (!task_work_add(tsk
, &tctx
->task_work
, notify
))
2284 * Slow path - we failed, find and delete work. if the work is not
2285 * in the list, it got run and we're fine.
2288 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2289 wq_list_for_each(node
, prev
, &tctx
->task_list
) {
2290 if (&req
->io_task_work
.node
== node
) {
2291 wq_list_del(&tctx
->task_list
, node
, prev
);
2296 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2297 clear_bit(0, &tctx
->task_state
);
2301 static int io_req_task_work_add(struct io_kiocb
*req
)
2303 struct task_struct
*tsk
= req
->task
;
2304 struct io_ring_ctx
*ctx
= req
->ctx
;
2305 enum task_work_notify_mode notify
;
2308 if (tsk
->flags
& PF_EXITING
)
2312 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2313 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2314 * processing task_work. There's no reliable way to tell if TWA_RESUME
2318 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
))
2319 notify
= TWA_SIGNAL
;
2321 ret
= io_task_work_add(tsk
, req
, notify
);
2323 wake_up_process(tsk
);
2328 static void io_req_task_work_add_fallback(struct io_kiocb
*req
,
2329 task_work_func_t cb
)
2331 struct task_struct
*tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2333 init_task_work(&req
->task_work
, cb
);
2334 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2335 wake_up_process(tsk
);
2338 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
2340 struct io_ring_ctx
*ctx
= req
->ctx
;
2342 spin_lock_irq(&ctx
->completion_lock
);
2343 io_cqring_fill_event(req
, error
);
2344 io_commit_cqring(ctx
);
2345 spin_unlock_irq(&ctx
->completion_lock
);
2347 io_cqring_ev_posted(ctx
);
2348 req_set_fail_links(req
);
2349 io_double_put_req(req
);
2352 static void io_req_task_cancel(struct callback_head
*cb
)
2354 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2355 struct io_ring_ctx
*ctx
= req
->ctx
;
2357 mutex_lock(&ctx
->uring_lock
);
2358 __io_req_task_cancel(req
, req
->result
);
2359 mutex_unlock(&ctx
->uring_lock
);
2360 percpu_ref_put(&ctx
->refs
);
2363 static void __io_req_task_submit(struct io_kiocb
*req
)
2365 struct io_ring_ctx
*ctx
= req
->ctx
;
2367 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2368 mutex_lock(&ctx
->uring_lock
);
2369 if (!ctx
->sqo_dead
&& !(current
->flags
& PF_EXITING
) &&
2370 !io_sq_thread_acquire_mm_files(ctx
, req
))
2371 __io_queue_sqe(req
);
2373 __io_req_task_cancel(req
, -EFAULT
);
2374 mutex_unlock(&ctx
->uring_lock
);
2376 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
2377 io_sq_thread_drop_mm_files();
2380 static void io_req_task_submit(struct callback_head
*cb
)
2382 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2384 __io_req_task_submit(req
);
2387 static void io_req_task_queue(struct io_kiocb
*req
)
2391 req
->task_work
.func
= io_req_task_submit
;
2392 ret
= io_req_task_work_add(req
);
2393 if (unlikely(ret
)) {
2394 req
->result
= -ECANCELED
;
2395 percpu_ref_get(&req
->ctx
->refs
);
2396 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2400 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2402 percpu_ref_get(&req
->ctx
->refs
);
2404 req
->task_work
.func
= io_req_task_cancel
;
2406 if (unlikely(io_req_task_work_add(req
)))
2407 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2410 static inline void io_queue_next(struct io_kiocb
*req
)
2412 struct io_kiocb
*nxt
= io_req_find_next(req
);
2415 io_req_task_queue(nxt
);
2418 static void io_free_req(struct io_kiocb
*req
)
2425 struct task_struct
*task
;
2430 static inline void io_init_req_batch(struct req_batch
*rb
)
2437 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2438 struct req_batch
*rb
)
2441 io_put_task(rb
->task
, rb
->task_refs
);
2443 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2446 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2447 struct io_submit_state
*state
)
2451 if (req
->task
!= rb
->task
) {
2453 io_put_task(rb
->task
, rb
->task_refs
);
2454 rb
->task
= req
->task
;
2460 io_dismantle_req(req
);
2461 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2462 state
->reqs
[state
->free_reqs
++] = req
;
2464 list_add(&req
->compl.list
, &state
->comp
.free_list
);
2467 static void io_submit_flush_completions(struct io_comp_state
*cs
,
2468 struct io_ring_ctx
*ctx
)
2471 struct io_kiocb
*req
;
2472 struct req_batch rb
;
2474 io_init_req_batch(&rb
);
2475 spin_lock_irq(&ctx
->completion_lock
);
2476 for (i
= 0; i
< nr
; i
++) {
2478 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
2480 io_commit_cqring(ctx
);
2481 spin_unlock_irq(&ctx
->completion_lock
);
2483 io_cqring_ev_posted(ctx
);
2484 for (i
= 0; i
< nr
; i
++) {
2487 /* submission and completion refs */
2488 if (refcount_sub_and_test(2, &req
->refs
))
2489 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2492 io_req_free_batch_finish(ctx
, &rb
);
2497 * Drop reference to request, return next in chain (if there is one) if this
2498 * was the last reference to this request.
2500 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2502 struct io_kiocb
*nxt
= NULL
;
2504 if (refcount_dec_and_test(&req
->refs
)) {
2505 nxt
= io_req_find_next(req
);
2511 static void io_put_req(struct io_kiocb
*req
)
2513 if (refcount_dec_and_test(&req
->refs
))
2517 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2519 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2524 static void io_free_req_deferred(struct io_kiocb
*req
)
2528 req
->task_work
.func
= io_put_req_deferred_cb
;
2529 ret
= io_req_task_work_add(req
);
2531 io_req_task_work_add_fallback(req
, io_put_req_deferred_cb
);
2534 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2536 if (refcount_sub_and_test(refs
, &req
->refs
))
2537 io_free_req_deferred(req
);
2540 static void io_double_put_req(struct io_kiocb
*req
)
2542 /* drop both submit and complete references */
2543 if (refcount_sub_and_test(2, &req
->refs
))
2547 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2549 /* See comment at the top of this file */
2551 return __io_cqring_events(ctx
);
2554 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2556 struct io_rings
*rings
= ctx
->rings
;
2558 /* make sure SQ entry isn't read before tail */
2559 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2562 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2564 unsigned int cflags
;
2566 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2567 cflags
|= IORING_CQE_F_BUFFER
;
2568 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2573 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2575 struct io_buffer
*kbuf
;
2577 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2578 return io_put_kbuf(req
, kbuf
);
2581 static inline bool io_run_task_work(void)
2584 * Not safe to run on exiting task, and the task_work handling will
2585 * not add work to such a task.
2587 if (unlikely(current
->flags
& PF_EXITING
))
2589 if (current
->task_works
) {
2590 __set_current_state(TASK_RUNNING
);
2599 * Find and free completed poll iocbs
2601 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2602 struct list_head
*done
)
2604 struct req_batch rb
;
2605 struct io_kiocb
*req
;
2607 /* order with ->result store in io_complete_rw_iopoll() */
2610 io_init_req_batch(&rb
);
2611 while (!list_empty(done
)) {
2614 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2615 list_del(&req
->inflight_entry
);
2617 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2618 req
->iopoll_completed
= 0;
2619 if (io_rw_reissue(req
))
2623 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2624 cflags
= io_put_rw_kbuf(req
);
2626 __io_cqring_fill_event(req
, req
->result
, cflags
);
2629 if (refcount_dec_and_test(&req
->refs
))
2630 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2633 io_commit_cqring(ctx
);
2634 io_cqring_ev_posted_iopoll(ctx
);
2635 io_req_free_batch_finish(ctx
, &rb
);
2638 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2641 struct io_kiocb
*req
, *tmp
;
2647 * Only spin for completions if we don't have multiple devices hanging
2648 * off our complete list, and we're under the requested amount.
2650 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2653 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2654 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2657 * Move completed and retryable entries to our local lists.
2658 * If we find a request that requires polling, break out
2659 * and complete those lists first, if we have entries there.
2661 if (READ_ONCE(req
->iopoll_completed
)) {
2662 list_move_tail(&req
->inflight_entry
, &done
);
2665 if (!list_empty(&done
))
2668 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2672 /* iopoll may have completed current req */
2673 if (READ_ONCE(req
->iopoll_completed
))
2674 list_move_tail(&req
->inflight_entry
, &done
);
2681 if (!list_empty(&done
))
2682 io_iopoll_complete(ctx
, nr_events
, &done
);
2688 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2689 * non-spinning poll check - we'll still enter the driver poll loop, but only
2690 * as a non-spinning completion check.
2692 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2695 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2698 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2701 if (*nr_events
>= min
)
2709 * We can't just wait for polled events to come to us, we have to actively
2710 * find and complete them.
2712 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2714 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2717 mutex_lock(&ctx
->uring_lock
);
2718 while (!list_empty(&ctx
->iopoll_list
)) {
2719 unsigned int nr_events
= 0;
2721 io_do_iopoll(ctx
, &nr_events
, 0);
2723 /* let it sleep and repeat later if can't complete a request */
2727 * Ensure we allow local-to-the-cpu processing to take place,
2728 * in this case we need to ensure that we reap all events.
2729 * Also let task_work, etc. to progress by releasing the mutex
2731 if (need_resched()) {
2732 mutex_unlock(&ctx
->uring_lock
);
2734 mutex_lock(&ctx
->uring_lock
);
2737 mutex_unlock(&ctx
->uring_lock
);
2740 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2742 unsigned int nr_events
= 0;
2743 int iters
= 0, ret
= 0;
2746 * We disallow the app entering submit/complete with polling, but we
2747 * still need to lock the ring to prevent racing with polled issue
2748 * that got punted to a workqueue.
2750 mutex_lock(&ctx
->uring_lock
);
2753 * Don't enter poll loop if we already have events pending.
2754 * If we do, we can potentially be spinning for commands that
2755 * already triggered a CQE (eg in error).
2757 if (test_bit(0, &ctx
->cq_check_overflow
))
2758 __io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2759 if (io_cqring_events(ctx
))
2763 * If a submit got punted to a workqueue, we can have the
2764 * application entering polling for a command before it gets
2765 * issued. That app will hold the uring_lock for the duration
2766 * of the poll right here, so we need to take a breather every
2767 * now and then to ensure that the issue has a chance to add
2768 * the poll to the issued list. Otherwise we can spin here
2769 * forever, while the workqueue is stuck trying to acquire the
2772 if (!(++iters
& 7)) {
2773 mutex_unlock(&ctx
->uring_lock
);
2775 mutex_lock(&ctx
->uring_lock
);
2778 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2782 } while (min
&& !nr_events
&& !need_resched());
2784 mutex_unlock(&ctx
->uring_lock
);
2788 static void kiocb_end_write(struct io_kiocb
*req
)
2791 * Tell lockdep we inherited freeze protection from submission
2794 if (req
->flags
& REQ_F_ISREG
) {
2795 struct inode
*inode
= file_inode(req
->file
);
2797 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2799 file_end_write(req
->file
);
2803 static bool io_resubmit_prep(struct io_kiocb
*req
)
2805 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2807 struct iov_iter iter
;
2809 /* already prepared */
2810 if (req
->async_data
)
2813 switch (req
->opcode
) {
2814 case IORING_OP_READV
:
2815 case IORING_OP_READ_FIXED
:
2816 case IORING_OP_READ
:
2819 case IORING_OP_WRITEV
:
2820 case IORING_OP_WRITE_FIXED
:
2821 case IORING_OP_WRITE
:
2825 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2830 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2833 return !io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2837 static bool io_rw_reissue(struct io_kiocb
*req
)
2840 umode_t mode
= file_inode(req
->file
)->i_mode
;
2843 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2845 if ((req
->flags
& REQ_F_NOWAIT
) || io_wq_current_is_worker())
2848 lockdep_assert_held(&req
->ctx
->uring_lock
);
2850 ret
= io_sq_thread_acquire_mm_files(req
->ctx
, req
);
2852 if (!ret
&& io_resubmit_prep(req
)) {
2853 refcount_inc(&req
->refs
);
2854 io_queue_async_work(req
);
2857 req_set_fail_links(req
);
2862 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2863 unsigned int issue_flags
)
2867 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) && io_rw_reissue(req
))
2869 if (res
!= req
->result
)
2870 req_set_fail_links(req
);
2872 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
)
2873 kiocb_end_write(req
);
2874 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2875 cflags
= io_put_rw_kbuf(req
);
2876 __io_req_complete(req
, issue_flags
, res
, cflags
);
2879 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2881 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2883 __io_complete_rw(req
, res
, res2
, 0);
2886 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2888 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2890 if (kiocb
->ki_flags
& IOCB_WRITE
)
2891 kiocb_end_write(req
);
2893 if (res
!= -EAGAIN
&& res
!= req
->result
)
2894 req_set_fail_links(req
);
2896 WRITE_ONCE(req
->result
, res
);
2897 /* order with io_poll_complete() checking ->result */
2899 WRITE_ONCE(req
->iopoll_completed
, 1);
2903 * After the iocb has been issued, it's safe to be found on the poll list.
2904 * Adding the kiocb to the list AFTER submission ensures that we don't
2905 * find it from a io_iopoll_getevents() thread before the issuer is done
2906 * accessing the kiocb cookie.
2908 static void io_iopoll_req_issued(struct io_kiocb
*req
, bool in_async
)
2910 struct io_ring_ctx
*ctx
= req
->ctx
;
2913 * Track whether we have multiple files in our lists. This will impact
2914 * how we do polling eventually, not spinning if we're on potentially
2915 * different devices.
2917 if (list_empty(&ctx
->iopoll_list
)) {
2918 ctx
->poll_multi_file
= false;
2919 } else if (!ctx
->poll_multi_file
) {
2920 struct io_kiocb
*list_req
;
2922 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2924 if (list_req
->file
!= req
->file
)
2925 ctx
->poll_multi_file
= true;
2929 * For fast devices, IO may have already completed. If it has, add
2930 * it to the front so we find it first.
2932 if (READ_ONCE(req
->iopoll_completed
))
2933 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2935 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2938 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2939 * task context or in io worker task context. If current task context is
2940 * sq thread, we don't need to check whether should wake up sq thread.
2942 if (in_async
&& (ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2943 wq_has_sleeper(&ctx
->sq_data
->wait
))
2944 wake_up(&ctx
->sq_data
->wait
);
2947 static inline void io_state_file_put(struct io_submit_state
*state
)
2949 if (state
->file_refs
) {
2950 fput_many(state
->file
, state
->file_refs
);
2951 state
->file_refs
= 0;
2956 * Get as many references to a file as we have IOs left in this submission,
2957 * assuming most submissions are for one file, or at least that each file
2958 * has more than one submission.
2960 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2965 if (state
->file_refs
) {
2966 if (state
->fd
== fd
) {
2970 io_state_file_put(state
);
2972 state
->file
= fget_many(fd
, state
->ios_left
);
2973 if (unlikely(!state
->file
))
2977 state
->file_refs
= state
->ios_left
- 1;
2981 static bool io_bdev_nowait(struct block_device
*bdev
)
2983 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2987 * If we tracked the file through the SCM inflight mechanism, we could support
2988 * any file. For now, just ensure that anything potentially problematic is done
2991 static bool io_file_supports_async(struct file
*file
, int rw
)
2993 umode_t mode
= file_inode(file
)->i_mode
;
2995 if (S_ISBLK(mode
)) {
2996 if (IS_ENABLED(CONFIG_BLOCK
) &&
2997 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
3001 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
3003 if (S_ISREG(mode
)) {
3004 if (IS_ENABLED(CONFIG_BLOCK
) &&
3005 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
3006 file
->f_op
!= &io_uring_fops
)
3011 /* any ->read/write should understand O_NONBLOCK */
3012 if (file
->f_flags
& O_NONBLOCK
)
3015 if (!(file
->f_mode
& FMODE_NOWAIT
))
3019 return file
->f_op
->read_iter
!= NULL
;
3021 return file
->f_op
->write_iter
!= NULL
;
3024 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3026 struct io_ring_ctx
*ctx
= req
->ctx
;
3027 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3028 struct file
*file
= req
->file
;
3032 if (S_ISREG(file_inode(file
)->i_mode
))
3033 req
->flags
|= REQ_F_ISREG
;
3035 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
3036 if (kiocb
->ki_pos
== -1 && !(file
->f_mode
& FMODE_STREAM
)) {
3037 req
->flags
|= REQ_F_CUR_POS
;
3038 kiocb
->ki_pos
= file
->f_pos
;
3040 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
3041 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
3042 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
3046 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
3047 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) || (file
->f_flags
& O_NONBLOCK
))
3048 req
->flags
|= REQ_F_NOWAIT
;
3050 ioprio
= READ_ONCE(sqe
->ioprio
);
3052 ret
= ioprio_check_cap(ioprio
);
3056 kiocb
->ki_ioprio
= ioprio
;
3058 kiocb
->ki_ioprio
= get_current_ioprio();
3060 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
3061 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
3062 !kiocb
->ki_filp
->f_op
->iopoll
)
3065 kiocb
->ki_flags
|= IOCB_HIPRI
;
3066 kiocb
->ki_complete
= io_complete_rw_iopoll
;
3067 req
->iopoll_completed
= 0;
3069 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3071 kiocb
->ki_complete
= io_complete_rw
;
3074 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
3075 req
->rw
.len
= READ_ONCE(sqe
->len
);
3076 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
3080 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
3086 case -ERESTARTNOINTR
:
3087 case -ERESTARTNOHAND
:
3088 case -ERESTART_RESTARTBLOCK
:
3090 * We can't just restart the syscall, since previously
3091 * submitted sqes may already be in progress. Just fail this
3097 kiocb
->ki_complete(kiocb
, ret
, 0);
3101 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
3102 unsigned int issue_flags
)
3104 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
3105 struct io_async_rw
*io
= req
->async_data
;
3107 /* add previously done IO, if any */
3108 if (io
&& io
->bytes_done
> 0) {
3110 ret
= io
->bytes_done
;
3112 ret
+= io
->bytes_done
;
3115 if (req
->flags
& REQ_F_CUR_POS
)
3116 req
->file
->f_pos
= kiocb
->ki_pos
;
3117 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
3118 __io_complete_rw(req
, ret
, 0, issue_flags
);
3120 io_rw_done(kiocb
, ret
);
3123 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
3125 struct io_ring_ctx
*ctx
= req
->ctx
;
3126 size_t len
= req
->rw
.len
;
3127 struct io_mapped_ubuf
*imu
;
3128 u16 index
, buf_index
= req
->buf_index
;
3132 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3134 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3135 imu
= &ctx
->user_bufs
[index
];
3136 buf_addr
= req
->rw
.addr
;
3139 if (buf_addr
+ len
< buf_addr
)
3141 /* not inside the mapped region */
3142 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
3146 * May not be a start of buffer, set size appropriately
3147 * and advance us to the beginning.
3149 offset
= buf_addr
- imu
->ubuf
;
3150 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3154 * Don't use iov_iter_advance() here, as it's really slow for
3155 * using the latter parts of a big fixed buffer - it iterates
3156 * over each segment manually. We can cheat a bit here, because
3159 * 1) it's a BVEC iter, we set it up
3160 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3161 * first and last bvec
3163 * So just find our index, and adjust the iterator afterwards.
3164 * If the offset is within the first bvec (or the whole first
3165 * bvec, just use iov_iter_advance(). This makes it easier
3166 * since we can just skip the first segment, which may not
3167 * be PAGE_SIZE aligned.
3169 const struct bio_vec
*bvec
= imu
->bvec
;
3171 if (offset
<= bvec
->bv_len
) {
3172 iov_iter_advance(iter
, offset
);
3174 unsigned long seg_skip
;
3176 /* skip first vec */
3177 offset
-= bvec
->bv_len
;
3178 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3180 iter
->bvec
= bvec
+ seg_skip
;
3181 iter
->nr_segs
-= seg_skip
;
3182 iter
->count
-= bvec
->bv_len
+ offset
;
3183 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3190 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3193 mutex_unlock(&ctx
->uring_lock
);
3196 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3199 * "Normal" inline submissions always hold the uring_lock, since we
3200 * grab it from the system call. Same is true for the SQPOLL offload.
3201 * The only exception is when we've detached the request and issue it
3202 * from an async worker thread, grab the lock for that case.
3205 mutex_lock(&ctx
->uring_lock
);
3208 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3209 int bgid
, struct io_buffer
*kbuf
,
3212 struct io_buffer
*head
;
3214 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3217 io_ring_submit_lock(req
->ctx
, needs_lock
);
3219 lockdep_assert_held(&req
->ctx
->uring_lock
);
3221 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
3223 if (!list_empty(&head
->list
)) {
3224 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3226 list_del(&kbuf
->list
);
3229 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
3231 if (*len
> kbuf
->len
)
3234 kbuf
= ERR_PTR(-ENOBUFS
);
3237 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3242 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3245 struct io_buffer
*kbuf
;
3248 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3249 bgid
= req
->buf_index
;
3250 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3253 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3254 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3255 return u64_to_user_ptr(kbuf
->addr
);
3258 #ifdef CONFIG_COMPAT
3259 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3262 struct compat_iovec __user
*uiov
;
3263 compat_ssize_t clen
;
3267 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3268 if (!access_ok(uiov
, sizeof(*uiov
)))
3270 if (__get_user(clen
, &uiov
->iov_len
))
3276 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3278 return PTR_ERR(buf
);
3279 iov
[0].iov_base
= buf
;
3280 iov
[0].iov_len
= (compat_size_t
) len
;
3285 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3288 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3292 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3295 len
= iov
[0].iov_len
;
3298 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3300 return PTR_ERR(buf
);
3301 iov
[0].iov_base
= buf
;
3302 iov
[0].iov_len
= len
;
3306 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3309 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3310 struct io_buffer
*kbuf
;
3312 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3313 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3314 iov
[0].iov_len
= kbuf
->len
;
3317 if (req
->rw
.len
!= 1)
3320 #ifdef CONFIG_COMPAT
3321 if (req
->ctx
->compat
)
3322 return io_compat_import(req
, iov
, needs_lock
);
3325 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3328 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
3329 struct iov_iter
*iter
, bool needs_lock
)
3331 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3332 size_t sqe_len
= req
->rw
.len
;
3333 u8 opcode
= req
->opcode
;
3336 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3338 return io_import_fixed(req
, rw
, iter
);
3341 /* buffer index only valid with fixed read/write, or buffer select */
3342 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3345 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3346 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3347 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3349 return PTR_ERR(buf
);
3350 req
->rw
.len
= sqe_len
;
3353 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3358 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3359 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3361 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3366 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3370 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3372 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3376 * For files that don't have ->read_iter() and ->write_iter(), handle them
3377 * by looping over ->read() or ->write() manually.
3379 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3381 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3382 struct file
*file
= req
->file
;
3386 * Don't support polled IO through this interface, and we can't
3387 * support non-blocking either. For the latter, this just causes
3388 * the kiocb to be handled from an async context.
3390 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3392 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3395 while (iov_iter_count(iter
)) {
3399 if (!iov_iter_is_bvec(iter
)) {
3400 iovec
= iov_iter_iovec(iter
);
3402 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3403 iovec
.iov_len
= req
->rw
.len
;
3407 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3408 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3410 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3411 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3420 if (nr
!= iovec
.iov_len
)
3424 iov_iter_advance(iter
, nr
);
3430 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3431 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3433 struct io_async_rw
*rw
= req
->async_data
;
3435 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3436 rw
->free_iovec
= iovec
;
3438 /* can only be fixed buffers, no need to do anything */
3439 if (iov_iter_is_bvec(iter
))
3442 unsigned iov_off
= 0;
3444 rw
->iter
.iov
= rw
->fast_iov
;
3445 if (iter
->iov
!= fast_iov
) {
3446 iov_off
= iter
->iov
- fast_iov
;
3447 rw
->iter
.iov
+= iov_off
;
3449 if (rw
->fast_iov
!= fast_iov
)
3450 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3451 sizeof(struct iovec
) * iter
->nr_segs
);
3453 req
->flags
|= REQ_F_NEED_CLEANUP
;
3457 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3459 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3460 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3461 return req
->async_data
== NULL
;
3464 static int io_alloc_async_data(struct io_kiocb
*req
)
3466 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3469 return __io_alloc_async_data(req
);
3472 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3473 const struct iovec
*fast_iov
,
3474 struct iov_iter
*iter
, bool force
)
3476 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3478 if (!req
->async_data
) {
3479 if (__io_alloc_async_data(req
)) {
3484 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3489 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3491 struct io_async_rw
*iorw
= req
->async_data
;
3492 struct iovec
*iov
= iorw
->fast_iov
;
3495 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3496 if (unlikely(ret
< 0))
3499 iorw
->bytes_done
= 0;
3500 iorw
->free_iovec
= iov
;
3502 req
->flags
|= REQ_F_NEED_CLEANUP
;
3506 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3508 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3510 return io_prep_rw(req
, sqe
);
3514 * This is our waitqueue callback handler, registered through lock_page_async()
3515 * when we initially tried to do the IO with the iocb armed our waitqueue.
3516 * This gets called when the page is unlocked, and we generally expect that to
3517 * happen when the page IO is completed and the page is now uptodate. This will
3518 * queue a task_work based retry of the operation, attempting to copy the data
3519 * again. If the latter fails because the page was NOT uptodate, then we will
3520 * do a thread based blocking retry of the operation. That's the unexpected
3523 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3524 int sync
, void *arg
)
3526 struct wait_page_queue
*wpq
;
3527 struct io_kiocb
*req
= wait
->private;
3528 struct wait_page_key
*key
= arg
;
3530 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3532 if (!wake_page_match(wpq
, key
))
3535 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3536 list_del_init(&wait
->entry
);
3538 /* submit ref gets dropped, acquire a new one */
3539 refcount_inc(&req
->refs
);
3540 io_req_task_queue(req
);
3545 * This controls whether a given IO request should be armed for async page
3546 * based retry. If we return false here, the request is handed to the async
3547 * worker threads for retry. If we're doing buffered reads on a regular file,
3548 * we prepare a private wait_page_queue entry and retry the operation. This
3549 * will either succeed because the page is now uptodate and unlocked, or it
3550 * will register a callback when the page is unlocked at IO completion. Through
3551 * that callback, io_uring uses task_work to setup a retry of the operation.
3552 * That retry will attempt the buffered read again. The retry will generally
3553 * succeed, or in rare cases where it fails, we then fall back to using the
3554 * async worker threads for a blocking retry.
3556 static bool io_rw_should_retry(struct io_kiocb
*req
)
3558 struct io_async_rw
*rw
= req
->async_data
;
3559 struct wait_page_queue
*wait
= &rw
->wpq
;
3560 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3562 /* never retry for NOWAIT, we just complete with -EAGAIN */
3563 if (req
->flags
& REQ_F_NOWAIT
)
3566 /* Only for buffered IO */
3567 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3571 * just use poll if we can, and don't attempt if the fs doesn't
3572 * support callback based unlocks
3574 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3577 wait
->wait
.func
= io_async_buf_func
;
3578 wait
->wait
.private = req
;
3579 wait
->wait
.flags
= 0;
3580 INIT_LIST_HEAD(&wait
->wait
.entry
);
3581 kiocb
->ki_flags
|= IOCB_WAITQ
;
3582 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3583 kiocb
->ki_waitq
= wait
;
3587 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3589 if (req
->file
->f_op
->read_iter
)
3590 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3591 else if (req
->file
->f_op
->read
)
3592 return loop_rw_iter(READ
, req
, iter
);
3597 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3599 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3600 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3601 struct iov_iter __iter
, *iter
= &__iter
;
3602 struct io_async_rw
*rw
= req
->async_data
;
3603 ssize_t io_size
, ret
, ret2
;
3604 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3610 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3614 io_size
= iov_iter_count(iter
);
3615 req
->result
= io_size
;
3617 /* Ensure we clear previously set non-block flag */
3618 if (!force_nonblock
)
3619 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3621 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3623 /* If the file doesn't support async, just async punt */
3624 if (force_nonblock
&& !io_file_supports_async(req
->file
, READ
)) {
3625 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3626 return ret
?: -EAGAIN
;
3629 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3630 if (unlikely(ret
)) {
3635 ret
= io_iter_do_read(req
, iter
);
3637 if (ret
== -EIOCBQUEUED
) {
3639 } else if (ret
== -EAGAIN
) {
3640 /* IOPOLL retry should happen for io-wq threads */
3641 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3643 /* no retry on NONBLOCK nor RWF_NOWAIT */
3644 if (req
->flags
& REQ_F_NOWAIT
)
3646 /* some cases will consume bytes even on error returns */
3647 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3649 } else if (ret
<= 0 || ret
== io_size
|| !force_nonblock
||
3650 (req
->flags
& REQ_F_NOWAIT
) || !(req
->flags
& REQ_F_ISREG
)) {
3651 /* read all, failed, already did sync or don't want to retry */
3655 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3660 rw
= req
->async_data
;
3661 /* now use our persistent iterator, if we aren't already */
3666 rw
->bytes_done
+= ret
;
3667 /* if we can retry, do so with the callbacks armed */
3668 if (!io_rw_should_retry(req
)) {
3669 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3674 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3675 * we get -EIOCBQUEUED, then we'll get a notification when the
3676 * desired page gets unlocked. We can also get a partial read
3677 * here, and if we do, then just retry at the new offset.
3679 ret
= io_iter_do_read(req
, iter
);
3680 if (ret
== -EIOCBQUEUED
)
3682 /* we got some bytes, but not all. retry. */
3683 } while (ret
> 0 && ret
< io_size
);
3685 kiocb_done(kiocb
, ret
, issue_flags
);
3687 /* it's faster to check here then delegate to kfree */
3693 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3695 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3697 return io_prep_rw(req
, sqe
);
3700 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3702 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3703 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3704 struct iov_iter __iter
, *iter
= &__iter
;
3705 struct io_async_rw
*rw
= req
->async_data
;
3706 ssize_t ret
, ret2
, io_size
;
3707 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3713 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3717 io_size
= iov_iter_count(iter
);
3718 req
->result
= io_size
;
3720 /* Ensure we clear previously set non-block flag */
3721 if (!force_nonblock
)
3722 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3724 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3726 /* If the file doesn't support async, just async punt */
3727 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3730 /* file path doesn't support NOWAIT for non-direct_IO */
3731 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3732 (req
->flags
& REQ_F_ISREG
))
3735 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3740 * Open-code file_start_write here to grab freeze protection,
3741 * which will be released by another thread in
3742 * io_complete_rw(). Fool lockdep by telling it the lock got
3743 * released so that it doesn't complain about the held lock when
3744 * we return to userspace.
3746 if (req
->flags
& REQ_F_ISREG
) {
3747 sb_start_write(file_inode(req
->file
)->i_sb
);
3748 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3751 kiocb
->ki_flags
|= IOCB_WRITE
;
3753 if (req
->file
->f_op
->write_iter
)
3754 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3755 else if (req
->file
->f_op
->write
)
3756 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3761 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3762 * retry them without IOCB_NOWAIT.
3764 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3766 /* no retry on NONBLOCK nor RWF_NOWAIT */
3767 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3769 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3770 /* IOPOLL retry should happen for io-wq threads */
3771 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3774 kiocb_done(kiocb
, ret2
, issue_flags
);
3777 /* some cases will consume bytes even on error returns */
3778 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3779 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3780 return ret
?: -EAGAIN
;
3783 /* it's reportedly faster than delegating the null check to kfree() */
3789 static int io_renameat_prep(struct io_kiocb
*req
,
3790 const struct io_uring_sqe
*sqe
)
3792 struct io_rename
*ren
= &req
->rename
;
3793 const char __user
*oldf
, *newf
;
3795 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3798 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3799 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3800 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3801 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3802 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3804 ren
->oldpath
= getname(oldf
);
3805 if (IS_ERR(ren
->oldpath
))
3806 return PTR_ERR(ren
->oldpath
);
3808 ren
->newpath
= getname(newf
);
3809 if (IS_ERR(ren
->newpath
)) {
3810 putname(ren
->oldpath
);
3811 return PTR_ERR(ren
->newpath
);
3814 req
->flags
|= REQ_F_NEED_CLEANUP
;
3818 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3820 struct io_rename
*ren
= &req
->rename
;
3823 if (issue_flags
& IO_URING_F_NONBLOCK
)
3826 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3827 ren
->newpath
, ren
->flags
);
3829 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3831 req_set_fail_links(req
);
3832 io_req_complete(req
, ret
);
3836 static int io_unlinkat_prep(struct io_kiocb
*req
,
3837 const struct io_uring_sqe
*sqe
)
3839 struct io_unlink
*un
= &req
->unlink
;
3840 const char __user
*fname
;
3842 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3845 un
->dfd
= READ_ONCE(sqe
->fd
);
3847 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3848 if (un
->flags
& ~AT_REMOVEDIR
)
3851 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3852 un
->filename
= getname(fname
);
3853 if (IS_ERR(un
->filename
))
3854 return PTR_ERR(un
->filename
);
3856 req
->flags
|= REQ_F_NEED_CLEANUP
;
3860 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3862 struct io_unlink
*un
= &req
->unlink
;
3865 if (issue_flags
& IO_URING_F_NONBLOCK
)
3868 if (un
->flags
& AT_REMOVEDIR
)
3869 ret
= do_rmdir(un
->dfd
, un
->filename
);
3871 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3873 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3875 req_set_fail_links(req
);
3876 io_req_complete(req
, ret
);
3880 static int io_shutdown_prep(struct io_kiocb
*req
,
3881 const struct io_uring_sqe
*sqe
)
3883 #if defined(CONFIG_NET)
3884 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3886 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3890 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3897 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3899 #if defined(CONFIG_NET)
3900 struct socket
*sock
;
3903 if (issue_flags
& IO_URING_F_NONBLOCK
)
3906 sock
= sock_from_file(req
->file
);
3907 if (unlikely(!sock
))
3910 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3912 req_set_fail_links(req
);
3913 io_req_complete(req
, ret
);
3920 static int __io_splice_prep(struct io_kiocb
*req
,
3921 const struct io_uring_sqe
*sqe
)
3923 struct io_splice
* sp
= &req
->splice
;
3924 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3926 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3930 sp
->len
= READ_ONCE(sqe
->len
);
3931 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3933 if (unlikely(sp
->flags
& ~valid_flags
))
3936 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3937 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3940 req
->flags
|= REQ_F_NEED_CLEANUP
;
3942 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3944 * Splice operation will be punted aync, and here need to
3945 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3947 io_req_init_async(req
);
3948 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3954 static int io_tee_prep(struct io_kiocb
*req
,
3955 const struct io_uring_sqe
*sqe
)
3957 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3959 return __io_splice_prep(req
, sqe
);
3962 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
3964 struct io_splice
*sp
= &req
->splice
;
3965 struct file
*in
= sp
->file_in
;
3966 struct file
*out
= sp
->file_out
;
3967 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3970 if (issue_flags
& IO_URING_F_NONBLOCK
)
3973 ret
= do_tee(in
, out
, sp
->len
, flags
);
3975 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3976 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3979 req_set_fail_links(req
);
3980 io_req_complete(req
, ret
);
3984 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3986 struct io_splice
* sp
= &req
->splice
;
3988 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3989 sp
->off_out
= READ_ONCE(sqe
->off
);
3990 return __io_splice_prep(req
, sqe
);
3993 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
3995 struct io_splice
*sp
= &req
->splice
;
3996 struct file
*in
= sp
->file_in
;
3997 struct file
*out
= sp
->file_out
;
3998 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3999 loff_t
*poff_in
, *poff_out
;
4002 if (issue_flags
& IO_URING_F_NONBLOCK
)
4005 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
4006 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
4009 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
4011 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4012 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4015 req_set_fail_links(req
);
4016 io_req_complete(req
, ret
);
4021 * IORING_OP_NOP just posts a completion event, nothing else.
4023 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
4025 struct io_ring_ctx
*ctx
= req
->ctx
;
4027 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4030 __io_req_complete(req
, issue_flags
, 0, 0);
4034 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4036 struct io_ring_ctx
*ctx
= req
->ctx
;
4041 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4043 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4046 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
4047 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
4050 req
->sync
.off
= READ_ONCE(sqe
->off
);
4051 req
->sync
.len
= READ_ONCE(sqe
->len
);
4055 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
4057 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
4060 /* fsync always requires a blocking context */
4061 if (issue_flags
& IO_URING_F_NONBLOCK
)
4064 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
4065 end
> 0 ? end
: LLONG_MAX
,
4066 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
4068 req_set_fail_links(req
);
4069 io_req_complete(req
, ret
);
4073 static int io_fallocate_prep(struct io_kiocb
*req
,
4074 const struct io_uring_sqe
*sqe
)
4076 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
4078 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4081 req
->sync
.off
= READ_ONCE(sqe
->off
);
4082 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4083 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4087 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
4091 /* fallocate always requiring blocking context */
4092 if (issue_flags
& IO_URING_F_NONBLOCK
)
4094 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4097 req_set_fail_links(req
);
4098 io_req_complete(req
, ret
);
4102 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4104 const char __user
*fname
;
4107 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4109 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4112 /* open.how should be already initialised */
4113 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4114 req
->open
.how
.flags
|= O_LARGEFILE
;
4116 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4117 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4118 req
->open
.filename
= getname(fname
);
4119 if (IS_ERR(req
->open
.filename
)) {
4120 ret
= PTR_ERR(req
->open
.filename
);
4121 req
->open
.filename
= NULL
;
4124 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4125 req
->flags
|= REQ_F_NEED_CLEANUP
;
4129 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4133 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4135 mode
= READ_ONCE(sqe
->len
);
4136 flags
= READ_ONCE(sqe
->open_flags
);
4137 req
->open
.how
= build_open_how(flags
, mode
);
4138 return __io_openat_prep(req
, sqe
);
4141 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4143 struct open_how __user
*how
;
4147 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4149 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4150 len
= READ_ONCE(sqe
->len
);
4151 if (len
< OPEN_HOW_SIZE_VER0
)
4154 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4159 return __io_openat_prep(req
, sqe
);
4162 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
4164 struct open_flags op
;
4167 bool resolve_nonblock
;
4170 ret
= build_open_flags(&req
->open
.how
, &op
);
4173 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
4174 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
4175 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4177 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4178 * it'll always -EAGAIN
4180 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
4182 op
.lookup_flags
|= LOOKUP_CACHED
;
4183 op
.open_flag
|= O_NONBLOCK
;
4186 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4190 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4191 /* only retry if RESOLVE_CACHED wasn't already set by application */
4192 if ((!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)) &&
4193 file
== ERR_PTR(-EAGAIN
)) {
4195 * We could hang on to this 'fd', but seems like marginal
4196 * gain for something that is now known to be a slower path.
4197 * So just put it, and we'll get a new one when we retry.
4205 ret
= PTR_ERR(file
);
4207 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
4208 file
->f_flags
&= ~O_NONBLOCK
;
4209 fsnotify_open(file
);
4210 fd_install(ret
, file
);
4213 putname(req
->open
.filename
);
4214 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4216 req_set_fail_links(req
);
4217 io_req_complete(req
, ret
);
4221 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
4223 return io_openat2(req
, issue_flags
& IO_URING_F_NONBLOCK
);
4226 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4227 const struct io_uring_sqe
*sqe
)
4229 struct io_provide_buf
*p
= &req
->pbuf
;
4232 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
4235 tmp
= READ_ONCE(sqe
->fd
);
4236 if (!tmp
|| tmp
> USHRT_MAX
)
4239 memset(p
, 0, sizeof(*p
));
4241 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4245 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4246 int bgid
, unsigned nbufs
)
4250 /* shouldn't happen */
4254 /* the head kbuf is the list itself */
4255 while (!list_empty(&buf
->list
)) {
4256 struct io_buffer
*nxt
;
4258 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4259 list_del(&nxt
->list
);
4266 idr_remove(&ctx
->io_buffer_idr
, bgid
);
4271 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4273 struct io_provide_buf
*p
= &req
->pbuf
;
4274 struct io_ring_ctx
*ctx
= req
->ctx
;
4275 struct io_buffer
*head
;
4277 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4279 io_ring_submit_lock(ctx
, !force_nonblock
);
4281 lockdep_assert_held(&ctx
->uring_lock
);
4284 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4286 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4288 req_set_fail_links(req
);
4290 /* need to hold the lock to complete IOPOLL requests */
4291 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4292 __io_req_complete(req
, issue_flags
, ret
, 0);
4293 io_ring_submit_unlock(ctx
, !force_nonblock
);
4295 io_ring_submit_unlock(ctx
, !force_nonblock
);
4296 __io_req_complete(req
, issue_flags
, ret
, 0);
4301 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4302 const struct io_uring_sqe
*sqe
)
4304 struct io_provide_buf
*p
= &req
->pbuf
;
4307 if (sqe
->ioprio
|| sqe
->rw_flags
)
4310 tmp
= READ_ONCE(sqe
->fd
);
4311 if (!tmp
|| tmp
> USHRT_MAX
)
4314 p
->addr
= READ_ONCE(sqe
->addr
);
4315 p
->len
= READ_ONCE(sqe
->len
);
4317 if (!access_ok(u64_to_user_ptr(p
->addr
), (p
->len
* p
->nbufs
)))
4320 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4321 tmp
= READ_ONCE(sqe
->off
);
4322 if (tmp
> USHRT_MAX
)
4328 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4330 struct io_buffer
*buf
;
4331 u64 addr
= pbuf
->addr
;
4332 int i
, bid
= pbuf
->bid
;
4334 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4335 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
4340 buf
->len
= pbuf
->len
;
4345 INIT_LIST_HEAD(&buf
->list
);
4348 list_add_tail(&buf
->list
, &(*head
)->list
);
4352 return i
? i
: -ENOMEM
;
4355 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4357 struct io_provide_buf
*p
= &req
->pbuf
;
4358 struct io_ring_ctx
*ctx
= req
->ctx
;
4359 struct io_buffer
*head
, *list
;
4361 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4363 io_ring_submit_lock(ctx
, !force_nonblock
);
4365 lockdep_assert_held(&ctx
->uring_lock
);
4367 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4369 ret
= io_add_buffers(p
, &head
);
4374 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
4377 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4383 req_set_fail_links(req
);
4385 /* need to hold the lock to complete IOPOLL requests */
4386 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4387 __io_req_complete(req
, issue_flags
, ret
, 0);
4388 io_ring_submit_unlock(ctx
, !force_nonblock
);
4390 io_ring_submit_unlock(ctx
, !force_nonblock
);
4391 __io_req_complete(req
, issue_flags
, ret
, 0);
4396 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4397 const struct io_uring_sqe
*sqe
)
4399 #if defined(CONFIG_EPOLL)
4400 if (sqe
->ioprio
|| sqe
->buf_index
)
4402 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4405 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4406 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4407 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4409 if (ep_op_has_event(req
->epoll
.op
)) {
4410 struct epoll_event __user
*ev
;
4412 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4413 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4423 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4425 #if defined(CONFIG_EPOLL)
4426 struct io_epoll
*ie
= &req
->epoll
;
4428 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4430 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4431 if (force_nonblock
&& ret
== -EAGAIN
)
4435 req_set_fail_links(req
);
4436 __io_req_complete(req
, issue_flags
, ret
, 0);
4443 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4445 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4446 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4448 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4451 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4452 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4453 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4460 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4462 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4463 struct io_madvise
*ma
= &req
->madvise
;
4466 if (issue_flags
& IO_URING_F_NONBLOCK
)
4469 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4471 req_set_fail_links(req
);
4472 io_req_complete(req
, ret
);
4479 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4481 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4483 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4486 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4487 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4488 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4492 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4494 struct io_fadvise
*fa
= &req
->fadvise
;
4497 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4498 switch (fa
->advice
) {
4499 case POSIX_FADV_NORMAL
:
4500 case POSIX_FADV_RANDOM
:
4501 case POSIX_FADV_SEQUENTIAL
:
4508 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4510 req_set_fail_links(req
);
4511 io_req_complete(req
, ret
);
4515 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4517 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4519 if (sqe
->ioprio
|| sqe
->buf_index
)
4521 if (req
->flags
& REQ_F_FIXED_FILE
)
4524 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4525 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4526 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4527 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4528 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4533 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4535 struct io_statx
*ctx
= &req
->statx
;
4538 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4539 /* only need file table for an actual valid fd */
4540 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4541 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4545 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4549 req_set_fail_links(req
);
4550 io_req_complete(req
, ret
);
4554 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4556 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4558 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4559 sqe
->rw_flags
|| sqe
->buf_index
)
4561 if (req
->flags
& REQ_F_FIXED_FILE
)
4564 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4568 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4570 struct files_struct
*files
= current
->files
;
4571 struct io_close
*close
= &req
->close
;
4572 struct fdtable
*fdt
;
4578 spin_lock(&files
->file_lock
);
4579 fdt
= files_fdtable(files
);
4580 if (close
->fd
>= fdt
->max_fds
) {
4581 spin_unlock(&files
->file_lock
);
4584 file
= fdt
->fd
[close
->fd
];
4586 spin_unlock(&files
->file_lock
);
4590 if (file
->f_op
== &io_uring_fops
) {
4591 spin_unlock(&files
->file_lock
);
4596 /* if the file has a flush method, be safe and punt to async */
4597 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4598 spin_unlock(&files
->file_lock
);
4602 ret
= __close_fd_get_file(close
->fd
, &file
);
4603 spin_unlock(&files
->file_lock
);
4610 /* No ->flush() or already async, safely close from here */
4611 ret
= filp_close(file
, current
->files
);
4614 req_set_fail_links(req
);
4617 __io_req_complete(req
, issue_flags
, ret
, 0);
4621 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4623 struct io_ring_ctx
*ctx
= req
->ctx
;
4625 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4627 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4630 req
->sync
.off
= READ_ONCE(sqe
->off
);
4631 req
->sync
.len
= READ_ONCE(sqe
->len
);
4632 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4636 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4640 /* sync_file_range always requires a blocking context */
4641 if (issue_flags
& IO_URING_F_NONBLOCK
)
4644 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4647 req_set_fail_links(req
);
4648 io_req_complete(req
, ret
);
4652 #if defined(CONFIG_NET)
4653 static int io_setup_async_msg(struct io_kiocb
*req
,
4654 struct io_async_msghdr
*kmsg
)
4656 struct io_async_msghdr
*async_msg
= req
->async_data
;
4660 if (io_alloc_async_data(req
)) {
4661 kfree(kmsg
->free_iov
);
4664 async_msg
= req
->async_data
;
4665 req
->flags
|= REQ_F_NEED_CLEANUP
;
4666 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4667 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4668 /* if were using fast_iov, set it to the new one */
4669 if (!async_msg
->free_iov
)
4670 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4675 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4676 struct io_async_msghdr
*iomsg
)
4678 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4679 iomsg
->free_iov
= iomsg
->fast_iov
;
4680 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4681 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4684 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4688 if (!io_op_defs
[req
->opcode
].needs_async_data
)
4690 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4692 req
->flags
|= REQ_F_NEED_CLEANUP
;
4696 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4698 struct io_sr_msg
*sr
= &req
->sr_msg
;
4700 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4703 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4704 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4705 sr
->len
= READ_ONCE(sqe
->len
);
4707 #ifdef CONFIG_COMPAT
4708 if (req
->ctx
->compat
)
4709 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4714 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4716 struct io_async_msghdr iomsg
, *kmsg
;
4717 struct socket
*sock
;
4721 sock
= sock_from_file(req
->file
);
4722 if (unlikely(!sock
))
4725 kmsg
= req
->async_data
;
4727 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4733 flags
= req
->sr_msg
.msg_flags
;
4734 if (flags
& MSG_DONTWAIT
)
4735 req
->flags
|= REQ_F_NOWAIT
;
4736 else if (issue_flags
& IO_URING_F_NONBLOCK
)
4737 flags
|= MSG_DONTWAIT
;
4739 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4740 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4741 return io_setup_async_msg(req
, kmsg
);
4742 if (ret
== -ERESTARTSYS
)
4745 /* fast path, check for non-NULL to avoid function call */
4747 kfree(kmsg
->free_iov
);
4748 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4750 req_set_fail_links(req
);
4751 __io_req_complete(req
, issue_flags
, ret
, 0);
4755 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4757 struct io_sr_msg
*sr
= &req
->sr_msg
;
4760 struct socket
*sock
;
4764 sock
= sock_from_file(req
->file
);
4765 if (unlikely(!sock
))
4768 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4772 msg
.msg_name
= NULL
;
4773 msg
.msg_control
= NULL
;
4774 msg
.msg_controllen
= 0;
4775 msg
.msg_namelen
= 0;
4777 flags
= req
->sr_msg
.msg_flags
;
4778 if (flags
& MSG_DONTWAIT
)
4779 req
->flags
|= REQ_F_NOWAIT
;
4780 else if (issue_flags
& IO_URING_F_NONBLOCK
)
4781 flags
|= MSG_DONTWAIT
;
4783 msg
.msg_flags
= flags
;
4784 ret
= sock_sendmsg(sock
, &msg
);
4785 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4787 if (ret
== -ERESTARTSYS
)
4791 req_set_fail_links(req
);
4792 __io_req_complete(req
, issue_flags
, ret
, 0);
4796 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4797 struct io_async_msghdr
*iomsg
)
4799 struct io_sr_msg
*sr
= &req
->sr_msg
;
4800 struct iovec __user
*uiov
;
4804 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4805 &iomsg
->uaddr
, &uiov
, &iov_len
);
4809 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4812 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4814 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4815 iomsg
->free_iov
= NULL
;
4817 iomsg
->free_iov
= iomsg
->fast_iov
;
4818 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4819 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4828 #ifdef CONFIG_COMPAT
4829 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4830 struct io_async_msghdr
*iomsg
)
4832 struct compat_msghdr __user
*msg_compat
;
4833 struct io_sr_msg
*sr
= &req
->sr_msg
;
4834 struct compat_iovec __user
*uiov
;
4839 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4840 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4845 uiov
= compat_ptr(ptr
);
4846 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4847 compat_ssize_t clen
;
4851 if (!access_ok(uiov
, sizeof(*uiov
)))
4853 if (__get_user(clen
, &uiov
->iov_len
))
4858 iomsg
->free_iov
= NULL
;
4860 iomsg
->free_iov
= iomsg
->fast_iov
;
4861 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4862 UIO_FASTIOV
, &iomsg
->free_iov
,
4863 &iomsg
->msg
.msg_iter
, true);
4872 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4873 struct io_async_msghdr
*iomsg
)
4875 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4877 #ifdef CONFIG_COMPAT
4878 if (req
->ctx
->compat
)
4879 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4882 return __io_recvmsg_copy_hdr(req
, iomsg
);
4885 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4888 struct io_sr_msg
*sr
= &req
->sr_msg
;
4889 struct io_buffer
*kbuf
;
4891 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4896 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4900 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4902 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4905 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
4909 if (!io_op_defs
[req
->opcode
].needs_async_data
)
4911 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
4913 req
->flags
|= REQ_F_NEED_CLEANUP
;
4917 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4919 struct io_sr_msg
*sr
= &req
->sr_msg
;
4921 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4924 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4925 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4926 sr
->len
= READ_ONCE(sqe
->len
);
4927 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4929 #ifdef CONFIG_COMPAT
4930 if (req
->ctx
->compat
)
4931 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4936 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4938 struct io_async_msghdr iomsg
, *kmsg
;
4939 struct socket
*sock
;
4940 struct io_buffer
*kbuf
;
4942 int ret
, cflags
= 0;
4943 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4945 sock
= sock_from_file(req
->file
);
4946 if (unlikely(!sock
))
4949 kmsg
= req
->async_data
;
4951 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4957 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4958 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4960 return PTR_ERR(kbuf
);
4961 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4962 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
4963 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
4964 1, req
->sr_msg
.len
);
4967 flags
= req
->sr_msg
.msg_flags
;
4968 if (flags
& MSG_DONTWAIT
)
4969 req
->flags
|= REQ_F_NOWAIT
;
4970 else if (force_nonblock
)
4971 flags
|= MSG_DONTWAIT
;
4973 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4974 kmsg
->uaddr
, flags
);
4975 if (force_nonblock
&& ret
== -EAGAIN
)
4976 return io_setup_async_msg(req
, kmsg
);
4977 if (ret
== -ERESTARTSYS
)
4980 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4981 cflags
= io_put_recv_kbuf(req
);
4982 /* fast path, check for non-NULL to avoid function call */
4984 kfree(kmsg
->free_iov
);
4985 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4987 req_set_fail_links(req
);
4988 __io_req_complete(req
, issue_flags
, ret
, cflags
);
4992 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
4994 struct io_buffer
*kbuf
;
4995 struct io_sr_msg
*sr
= &req
->sr_msg
;
4997 void __user
*buf
= sr
->buf
;
4998 struct socket
*sock
;
5001 int ret
, cflags
= 0;
5002 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5004 sock
= sock_from_file(req
->file
);
5005 if (unlikely(!sock
))
5008 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5009 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5011 return PTR_ERR(kbuf
);
5012 buf
= u64_to_user_ptr(kbuf
->addr
);
5015 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
5019 msg
.msg_name
= NULL
;
5020 msg
.msg_control
= NULL
;
5021 msg
.msg_controllen
= 0;
5022 msg
.msg_namelen
= 0;
5023 msg
.msg_iocb
= NULL
;
5026 flags
= req
->sr_msg
.msg_flags
;
5027 if (flags
& MSG_DONTWAIT
)
5028 req
->flags
|= REQ_F_NOWAIT
;
5029 else if (force_nonblock
)
5030 flags
|= MSG_DONTWAIT
;
5032 ret
= sock_recvmsg(sock
, &msg
, flags
);
5033 if (force_nonblock
&& ret
== -EAGAIN
)
5035 if (ret
== -ERESTARTSYS
)
5038 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5039 cflags
= io_put_recv_kbuf(req
);
5041 req_set_fail_links(req
);
5042 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5046 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5048 struct io_accept
*accept
= &req
->accept
;
5050 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5052 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
5055 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5056 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5057 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
5058 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
5062 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
5064 struct io_accept
*accept
= &req
->accept
;
5065 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5066 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5069 if (req
->file
->f_flags
& O_NONBLOCK
)
5070 req
->flags
|= REQ_F_NOWAIT
;
5072 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
5073 accept
->addr_len
, accept
->flags
,
5075 if (ret
== -EAGAIN
&& force_nonblock
)
5078 if (ret
== -ERESTARTSYS
)
5080 req_set_fail_links(req
);
5082 __io_req_complete(req
, issue_flags
, ret
, 0);
5086 static int io_connect_prep_async(struct io_kiocb
*req
)
5088 struct io_async_connect
*io
= req
->async_data
;
5089 struct io_connect
*conn
= &req
->connect
;
5091 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
5094 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5096 struct io_connect
*conn
= &req
->connect
;
5098 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5100 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
5103 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5104 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5108 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
5110 struct io_async_connect __io
, *io
;
5111 unsigned file_flags
;
5113 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5115 if (req
->async_data
) {
5116 io
= req
->async_data
;
5118 ret
= move_addr_to_kernel(req
->connect
.addr
,
5119 req
->connect
.addr_len
,
5126 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5128 ret
= __sys_connect_file(req
->file
, &io
->address
,
5129 req
->connect
.addr_len
, file_flags
);
5130 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5131 if (req
->async_data
)
5133 if (io_alloc_async_data(req
)) {
5137 io
= req
->async_data
;
5138 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5141 if (ret
== -ERESTARTSYS
)
5145 req_set_fail_links(req
);
5146 __io_req_complete(req
, issue_flags
, ret
, 0);
5149 #else /* !CONFIG_NET */
5150 #define IO_NETOP_FN(op) \
5151 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5153 return -EOPNOTSUPP; \
5156 #define IO_NETOP_PREP(op) \
5158 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5160 return -EOPNOTSUPP; \
5163 #define IO_NETOP_PREP_ASYNC(op) \
5165 static int io_##op##_prep_async(struct io_kiocb *req) \
5167 return -EOPNOTSUPP; \
5170 IO_NETOP_PREP_ASYNC(sendmsg
);
5171 IO_NETOP_PREP_ASYNC(recvmsg
);
5172 IO_NETOP_PREP_ASYNC(connect
);
5173 IO_NETOP_PREP(accept
);
5176 #endif /* CONFIG_NET */
5178 struct io_poll_table
{
5179 struct poll_table_struct pt
;
5180 struct io_kiocb
*req
;
5184 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5185 __poll_t mask
, task_work_func_t func
)
5189 /* for instances that support it check for an event match first: */
5190 if (mask
&& !(mask
& poll
->events
))
5193 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5195 list_del_init(&poll
->wait
.entry
);
5198 req
->task_work
.func
= func
;
5199 percpu_ref_get(&req
->ctx
->refs
);
5202 * If this fails, then the task is exiting. When a task exits, the
5203 * work gets canceled, so just cancel this request as well instead
5204 * of executing it. We can't safely execute it anyway, as we may not
5205 * have the needed state needed for it anyway.
5207 ret
= io_req_task_work_add(req
);
5208 if (unlikely(ret
)) {
5209 WRITE_ONCE(poll
->canceled
, true);
5210 io_req_task_work_add_fallback(req
, func
);
5215 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5216 __acquires(&req
->ctx
->completion_lock
)
5218 struct io_ring_ctx
*ctx
= req
->ctx
;
5220 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5221 struct poll_table_struct pt
= { ._key
= poll
->events
};
5223 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5226 spin_lock_irq(&ctx
->completion_lock
);
5227 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5228 add_wait_queue(poll
->head
, &poll
->wait
);
5235 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5237 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5238 if (req
->opcode
== IORING_OP_POLL_ADD
)
5239 return req
->async_data
;
5240 return req
->apoll
->double_poll
;
5243 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5245 if (req
->opcode
== IORING_OP_POLL_ADD
)
5247 return &req
->apoll
->poll
;
5250 static void io_poll_remove_double(struct io_kiocb
*req
)
5252 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5254 lockdep_assert_held(&req
->ctx
->completion_lock
);
5256 if (poll
&& poll
->head
) {
5257 struct wait_queue_head
*head
= poll
->head
;
5259 spin_lock(&head
->lock
);
5260 list_del_init(&poll
->wait
.entry
);
5261 if (poll
->wait
.private)
5262 refcount_dec(&req
->refs
);
5264 spin_unlock(&head
->lock
);
5268 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
5270 struct io_ring_ctx
*ctx
= req
->ctx
;
5272 io_poll_remove_double(req
);
5273 req
->poll
.done
= true;
5274 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
5275 io_commit_cqring(ctx
);
5278 static void io_poll_task_func(struct callback_head
*cb
)
5280 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5281 struct io_ring_ctx
*ctx
= req
->ctx
;
5282 struct io_kiocb
*nxt
;
5284 if (io_poll_rewait(req
, &req
->poll
)) {
5285 spin_unlock_irq(&ctx
->completion_lock
);
5287 hash_del(&req
->hash_node
);
5288 io_poll_complete(req
, req
->result
, 0);
5289 spin_unlock_irq(&ctx
->completion_lock
);
5291 nxt
= io_put_req_find_next(req
);
5292 io_cqring_ev_posted(ctx
);
5294 __io_req_task_submit(nxt
);
5297 percpu_ref_put(&ctx
->refs
);
5300 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5301 int sync
, void *key
)
5303 struct io_kiocb
*req
= wait
->private;
5304 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5305 __poll_t mask
= key_to_poll(key
);
5307 /* for instances that support it check for an event match first: */
5308 if (mask
&& !(mask
& poll
->events
))
5311 list_del_init(&wait
->entry
);
5313 if (poll
&& poll
->head
) {
5316 spin_lock(&poll
->head
->lock
);
5317 done
= list_empty(&poll
->wait
.entry
);
5319 list_del_init(&poll
->wait
.entry
);
5320 /* make sure double remove sees this as being gone */
5321 wait
->private = NULL
;
5322 spin_unlock(&poll
->head
->lock
);
5324 /* use wait func handler, so it matches the rq type */
5325 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5328 refcount_dec(&req
->refs
);
5332 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5333 wait_queue_func_t wake_func
)
5337 poll
->canceled
= false;
5338 poll
->events
= events
;
5339 INIT_LIST_HEAD(&poll
->wait
.entry
);
5340 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5343 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5344 struct wait_queue_head
*head
,
5345 struct io_poll_iocb
**poll_ptr
)
5347 struct io_kiocb
*req
= pt
->req
;
5350 * If poll->head is already set, it's because the file being polled
5351 * uses multiple waitqueues for poll handling (eg one for read, one
5352 * for write). Setup a separate io_poll_iocb if this happens.
5354 if (unlikely(poll
->head
)) {
5355 struct io_poll_iocb
*poll_one
= poll
;
5357 /* already have a 2nd entry, fail a third attempt */
5359 pt
->error
= -EINVAL
;
5362 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5364 pt
->error
= -ENOMEM
;
5367 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5368 refcount_inc(&req
->refs
);
5369 poll
->wait
.private = req
;
5376 if (poll
->events
& EPOLLEXCLUSIVE
)
5377 add_wait_queue_exclusive(head
, &poll
->wait
);
5379 add_wait_queue(head
, &poll
->wait
);
5382 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5383 struct poll_table_struct
*p
)
5385 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5386 struct async_poll
*apoll
= pt
->req
->apoll
;
5388 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5391 static void io_async_task_func(struct callback_head
*cb
)
5393 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5394 struct async_poll
*apoll
= req
->apoll
;
5395 struct io_ring_ctx
*ctx
= req
->ctx
;
5397 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5399 if (io_poll_rewait(req
, &apoll
->poll
)) {
5400 spin_unlock_irq(&ctx
->completion_lock
);
5401 percpu_ref_put(&ctx
->refs
);
5405 /* If req is still hashed, it cannot have been canceled. Don't check. */
5406 if (hash_hashed(&req
->hash_node
))
5407 hash_del(&req
->hash_node
);
5409 io_poll_remove_double(req
);
5410 spin_unlock_irq(&ctx
->completion_lock
);
5412 if (!READ_ONCE(apoll
->poll
.canceled
))
5413 __io_req_task_submit(req
);
5415 __io_req_task_cancel(req
, -ECANCELED
);
5417 percpu_ref_put(&ctx
->refs
);
5418 kfree(apoll
->double_poll
);
5422 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5425 struct io_kiocb
*req
= wait
->private;
5426 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5428 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5431 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5434 static void io_poll_req_insert(struct io_kiocb
*req
)
5436 struct io_ring_ctx
*ctx
= req
->ctx
;
5437 struct hlist_head
*list
;
5439 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5440 hlist_add_head(&req
->hash_node
, list
);
5443 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5444 struct io_poll_iocb
*poll
,
5445 struct io_poll_table
*ipt
, __poll_t mask
,
5446 wait_queue_func_t wake_func
)
5447 __acquires(&ctx
->completion_lock
)
5449 struct io_ring_ctx
*ctx
= req
->ctx
;
5450 bool cancel
= false;
5452 INIT_HLIST_NODE(&req
->hash_node
);
5453 io_init_poll_iocb(poll
, mask
, wake_func
);
5454 poll
->file
= req
->file
;
5455 poll
->wait
.private = req
;
5457 ipt
->pt
._key
= mask
;
5459 ipt
->error
= -EINVAL
;
5461 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5463 spin_lock_irq(&ctx
->completion_lock
);
5464 if (likely(poll
->head
)) {
5465 spin_lock(&poll
->head
->lock
);
5466 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5472 if (mask
|| ipt
->error
)
5473 list_del_init(&poll
->wait
.entry
);
5475 WRITE_ONCE(poll
->canceled
, true);
5476 else if (!poll
->done
) /* actually waiting for an event */
5477 io_poll_req_insert(req
);
5478 spin_unlock(&poll
->head
->lock
);
5484 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5486 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5487 struct io_ring_ctx
*ctx
= req
->ctx
;
5488 struct async_poll
*apoll
;
5489 struct io_poll_table ipt
;
5493 if (!req
->file
|| !file_can_poll(req
->file
))
5495 if (req
->flags
& REQ_F_POLLED
)
5499 else if (def
->pollout
)
5503 /* if we can't nonblock try, then no point in arming a poll handler */
5504 if (!io_file_supports_async(req
->file
, rw
))
5507 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5508 if (unlikely(!apoll
))
5510 apoll
->double_poll
= NULL
;
5512 req
->flags
|= REQ_F_POLLED
;
5517 mask
|= POLLIN
| POLLRDNORM
;
5519 mask
|= POLLOUT
| POLLWRNORM
;
5521 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5522 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5523 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5526 mask
|= POLLERR
| POLLPRI
;
5528 ipt
.pt
._qproc
= io_async_queue_proc
;
5530 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5532 if (ret
|| ipt
.error
) {
5533 io_poll_remove_double(req
);
5534 spin_unlock_irq(&ctx
->completion_lock
);
5535 kfree(apoll
->double_poll
);
5539 spin_unlock_irq(&ctx
->completion_lock
);
5540 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5541 apoll
->poll
.events
);
5545 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5546 struct io_poll_iocb
*poll
)
5548 bool do_complete
= false;
5550 spin_lock(&poll
->head
->lock
);
5551 WRITE_ONCE(poll
->canceled
, true);
5552 if (!list_empty(&poll
->wait
.entry
)) {
5553 list_del_init(&poll
->wait
.entry
);
5556 spin_unlock(&poll
->head
->lock
);
5557 hash_del(&req
->hash_node
);
5561 static bool io_poll_remove_one(struct io_kiocb
*req
)
5565 io_poll_remove_double(req
);
5567 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5568 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5570 struct async_poll
*apoll
= req
->apoll
;
5572 /* non-poll requests have submit ref still */
5573 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5576 kfree(apoll
->double_poll
);
5582 io_cqring_fill_event(req
, -ECANCELED
);
5583 io_commit_cqring(req
->ctx
);
5584 req_set_fail_links(req
);
5585 io_put_req_deferred(req
, 1);
5592 * Returns true if we found and killed one or more poll requests
5594 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5595 struct files_struct
*files
)
5597 struct hlist_node
*tmp
;
5598 struct io_kiocb
*req
;
5601 spin_lock_irq(&ctx
->completion_lock
);
5602 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5603 struct hlist_head
*list
;
5605 list
= &ctx
->cancel_hash
[i
];
5606 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5607 if (io_match_task(req
, tsk
, files
))
5608 posted
+= io_poll_remove_one(req
);
5611 spin_unlock_irq(&ctx
->completion_lock
);
5614 io_cqring_ev_posted(ctx
);
5619 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5621 struct hlist_head
*list
;
5622 struct io_kiocb
*req
;
5624 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5625 hlist_for_each_entry(req
, list
, hash_node
) {
5626 if (sqe_addr
!= req
->user_data
)
5628 if (io_poll_remove_one(req
))
5636 static int io_poll_remove_prep(struct io_kiocb
*req
,
5637 const struct io_uring_sqe
*sqe
)
5639 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5641 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5645 req
->poll_remove
.addr
= READ_ONCE(sqe
->addr
);
5650 * Find a running poll command that matches one specified in sqe->addr,
5651 * and remove it if found.
5653 static int io_poll_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
5655 struct io_ring_ctx
*ctx
= req
->ctx
;
5658 spin_lock_irq(&ctx
->completion_lock
);
5659 ret
= io_poll_cancel(ctx
, req
->poll_remove
.addr
);
5660 spin_unlock_irq(&ctx
->completion_lock
);
5663 req_set_fail_links(req
);
5664 io_req_complete(req
, ret
);
5668 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5671 struct io_kiocb
*req
= wait
->private;
5672 struct io_poll_iocb
*poll
= &req
->poll
;
5674 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5677 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5678 struct poll_table_struct
*p
)
5680 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5682 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5685 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5687 struct io_poll_iocb
*poll
= &req
->poll
;
5690 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5692 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5695 events
= READ_ONCE(sqe
->poll32_events
);
5697 events
= swahw32(events
);
5699 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5700 (events
& EPOLLEXCLUSIVE
);
5704 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5706 struct io_poll_iocb
*poll
= &req
->poll
;
5707 struct io_ring_ctx
*ctx
= req
->ctx
;
5708 struct io_poll_table ipt
;
5711 ipt
.pt
._qproc
= io_poll_queue_proc
;
5713 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5716 if (mask
) { /* no async, we'd stolen it */
5718 io_poll_complete(req
, mask
, 0);
5720 spin_unlock_irq(&ctx
->completion_lock
);
5723 io_cqring_ev_posted(ctx
);
5729 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5731 struct io_timeout_data
*data
= container_of(timer
,
5732 struct io_timeout_data
, timer
);
5733 struct io_kiocb
*req
= data
->req
;
5734 struct io_ring_ctx
*ctx
= req
->ctx
;
5735 unsigned long flags
;
5737 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5738 list_del_init(&req
->timeout
.list
);
5739 atomic_set(&req
->ctx
->cq_timeouts
,
5740 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5742 io_cqring_fill_event(req
, -ETIME
);
5743 io_commit_cqring(ctx
);
5744 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5746 io_cqring_ev_posted(ctx
);
5747 req_set_fail_links(req
);
5749 return HRTIMER_NORESTART
;
5752 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5755 struct io_timeout_data
*io
;
5756 struct io_kiocb
*req
;
5759 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5760 if (user_data
== req
->user_data
) {
5767 return ERR_PTR(ret
);
5769 io
= req
->async_data
;
5770 ret
= hrtimer_try_to_cancel(&io
->timer
);
5772 return ERR_PTR(-EALREADY
);
5773 list_del_init(&req
->timeout
.list
);
5777 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5779 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5782 return PTR_ERR(req
);
5784 req_set_fail_links(req
);
5785 io_cqring_fill_event(req
, -ECANCELED
);
5786 io_put_req_deferred(req
, 1);
5790 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5791 struct timespec64
*ts
, enum hrtimer_mode mode
)
5793 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5794 struct io_timeout_data
*data
;
5797 return PTR_ERR(req
);
5799 req
->timeout
.off
= 0; /* noseq */
5800 data
= req
->async_data
;
5801 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5802 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5803 data
->timer
.function
= io_timeout_fn
;
5804 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5808 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5809 const struct io_uring_sqe
*sqe
)
5811 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5813 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5815 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5817 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
5820 tr
->addr
= READ_ONCE(sqe
->addr
);
5821 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5822 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5823 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5825 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5827 } else if (tr
->flags
) {
5828 /* timeout removal doesn't support flags */
5835 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
5837 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
5842 * Remove or update an existing timeout command
5844 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
5846 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5847 struct io_ring_ctx
*ctx
= req
->ctx
;
5850 spin_lock_irq(&ctx
->completion_lock
);
5851 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
))
5852 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5854 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
,
5855 io_translate_timeout_mode(tr
->flags
));
5857 io_cqring_fill_event(req
, ret
);
5858 io_commit_cqring(ctx
);
5859 spin_unlock_irq(&ctx
->completion_lock
);
5860 io_cqring_ev_posted(ctx
);
5862 req_set_fail_links(req
);
5867 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5868 bool is_timeout_link
)
5870 struct io_timeout_data
*data
;
5872 u32 off
= READ_ONCE(sqe
->off
);
5874 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5876 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5878 if (off
&& is_timeout_link
)
5880 flags
= READ_ONCE(sqe
->timeout_flags
);
5881 if (flags
& ~IORING_TIMEOUT_ABS
)
5884 req
->timeout
.off
= off
;
5886 if (!req
->async_data
&& io_alloc_async_data(req
))
5889 data
= req
->async_data
;
5892 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5895 data
->mode
= io_translate_timeout_mode(flags
);
5896 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5900 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
5902 struct io_ring_ctx
*ctx
= req
->ctx
;
5903 struct io_timeout_data
*data
= req
->async_data
;
5904 struct list_head
*entry
;
5905 u32 tail
, off
= req
->timeout
.off
;
5907 spin_lock_irq(&ctx
->completion_lock
);
5910 * sqe->off holds how many events that need to occur for this
5911 * timeout event to be satisfied. If it isn't set, then this is
5912 * a pure timeout request, sequence isn't used.
5914 if (io_is_timeout_noseq(req
)) {
5915 entry
= ctx
->timeout_list
.prev
;
5919 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5920 req
->timeout
.target_seq
= tail
+ off
;
5922 /* Update the last seq here in case io_flush_timeouts() hasn't.
5923 * This is safe because ->completion_lock is held, and submissions
5924 * and completions are never mixed in the same ->completion_lock section.
5926 ctx
->cq_last_tm_flush
= tail
;
5929 * Insertion sort, ensuring the first entry in the list is always
5930 * the one we need first.
5932 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5933 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5936 if (io_is_timeout_noseq(nxt
))
5938 /* nxt.seq is behind @tail, otherwise would've been completed */
5939 if (off
>= nxt
->timeout
.target_seq
- tail
)
5943 list_add(&req
->timeout
.list
, entry
);
5944 data
->timer
.function
= io_timeout_fn
;
5945 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5946 spin_unlock_irq(&ctx
->completion_lock
);
5950 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5952 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5954 return req
->user_data
== (unsigned long) data
;
5957 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5959 enum io_wq_cancel cancel_ret
;
5962 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5963 switch (cancel_ret
) {
5964 case IO_WQ_CANCEL_OK
:
5967 case IO_WQ_CANCEL_RUNNING
:
5970 case IO_WQ_CANCEL_NOTFOUND
:
5978 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5979 struct io_kiocb
*req
, __u64 sqe_addr
,
5982 unsigned long flags
;
5985 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5986 if (ret
!= -ENOENT
) {
5987 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5991 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5992 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5995 ret
= io_poll_cancel(ctx
, sqe_addr
);
5999 io_cqring_fill_event(req
, ret
);
6000 io_commit_cqring(ctx
);
6001 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6002 io_cqring_ev_posted(ctx
);
6005 req_set_fail_links(req
);
6009 static int io_async_cancel_prep(struct io_kiocb
*req
,
6010 const struct io_uring_sqe
*sqe
)
6012 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6014 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6016 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
6019 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
6023 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
6025 struct io_ring_ctx
*ctx
= req
->ctx
;
6027 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
6031 static int io_rsrc_update_prep(struct io_kiocb
*req
,
6032 const struct io_uring_sqe
*sqe
)
6034 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
6036 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6038 if (sqe
->ioprio
|| sqe
->rw_flags
)
6041 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
6042 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
6043 if (!req
->rsrc_update
.nr_args
)
6045 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
6049 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
6051 struct io_ring_ctx
*ctx
= req
->ctx
;
6052 struct io_uring_rsrc_update up
;
6055 if (issue_flags
& IO_URING_F_NONBLOCK
)
6058 up
.offset
= req
->rsrc_update
.offset
;
6059 up
.data
= req
->rsrc_update
.arg
;
6061 mutex_lock(&ctx
->uring_lock
);
6062 ret
= __io_sqe_files_update(ctx
, &up
, req
->rsrc_update
.nr_args
);
6063 mutex_unlock(&ctx
->uring_lock
);
6066 req_set_fail_links(req
);
6067 __io_req_complete(req
, issue_flags
, ret
, 0);
6071 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6073 switch (req
->opcode
) {
6076 case IORING_OP_READV
:
6077 case IORING_OP_READ_FIXED
:
6078 case IORING_OP_READ
:
6079 return io_read_prep(req
, sqe
);
6080 case IORING_OP_WRITEV
:
6081 case IORING_OP_WRITE_FIXED
:
6082 case IORING_OP_WRITE
:
6083 return io_write_prep(req
, sqe
);
6084 case IORING_OP_POLL_ADD
:
6085 return io_poll_add_prep(req
, sqe
);
6086 case IORING_OP_POLL_REMOVE
:
6087 return io_poll_remove_prep(req
, sqe
);
6088 case IORING_OP_FSYNC
:
6089 return io_fsync_prep(req
, sqe
);
6090 case IORING_OP_SYNC_FILE_RANGE
:
6091 return io_sfr_prep(req
, sqe
);
6092 case IORING_OP_SENDMSG
:
6093 case IORING_OP_SEND
:
6094 return io_sendmsg_prep(req
, sqe
);
6095 case IORING_OP_RECVMSG
:
6096 case IORING_OP_RECV
:
6097 return io_recvmsg_prep(req
, sqe
);
6098 case IORING_OP_CONNECT
:
6099 return io_connect_prep(req
, sqe
);
6100 case IORING_OP_TIMEOUT
:
6101 return io_timeout_prep(req
, sqe
, false);
6102 case IORING_OP_TIMEOUT_REMOVE
:
6103 return io_timeout_remove_prep(req
, sqe
);
6104 case IORING_OP_ASYNC_CANCEL
:
6105 return io_async_cancel_prep(req
, sqe
);
6106 case IORING_OP_LINK_TIMEOUT
:
6107 return io_timeout_prep(req
, sqe
, true);
6108 case IORING_OP_ACCEPT
:
6109 return io_accept_prep(req
, sqe
);
6110 case IORING_OP_FALLOCATE
:
6111 return io_fallocate_prep(req
, sqe
);
6112 case IORING_OP_OPENAT
:
6113 return io_openat_prep(req
, sqe
);
6114 case IORING_OP_CLOSE
:
6115 return io_close_prep(req
, sqe
);
6116 case IORING_OP_FILES_UPDATE
:
6117 return io_rsrc_update_prep(req
, sqe
);
6118 case IORING_OP_STATX
:
6119 return io_statx_prep(req
, sqe
);
6120 case IORING_OP_FADVISE
:
6121 return io_fadvise_prep(req
, sqe
);
6122 case IORING_OP_MADVISE
:
6123 return io_madvise_prep(req
, sqe
);
6124 case IORING_OP_OPENAT2
:
6125 return io_openat2_prep(req
, sqe
);
6126 case IORING_OP_EPOLL_CTL
:
6127 return io_epoll_ctl_prep(req
, sqe
);
6128 case IORING_OP_SPLICE
:
6129 return io_splice_prep(req
, sqe
);
6130 case IORING_OP_PROVIDE_BUFFERS
:
6131 return io_provide_buffers_prep(req
, sqe
);
6132 case IORING_OP_REMOVE_BUFFERS
:
6133 return io_remove_buffers_prep(req
, sqe
);
6135 return io_tee_prep(req
, sqe
);
6136 case IORING_OP_SHUTDOWN
:
6137 return io_shutdown_prep(req
, sqe
);
6138 case IORING_OP_RENAMEAT
:
6139 return io_renameat_prep(req
, sqe
);
6140 case IORING_OP_UNLINKAT
:
6141 return io_unlinkat_prep(req
, sqe
);
6144 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6149 static int io_req_prep_async(struct io_kiocb
*req
)
6151 switch (req
->opcode
) {
6152 case IORING_OP_READV
:
6153 case IORING_OP_READ_FIXED
:
6154 case IORING_OP_READ
:
6155 return io_rw_prep_async(req
, READ
);
6156 case IORING_OP_WRITEV
:
6157 case IORING_OP_WRITE_FIXED
:
6158 case IORING_OP_WRITE
:
6159 return io_rw_prep_async(req
, WRITE
);
6160 case IORING_OP_SENDMSG
:
6161 case IORING_OP_SEND
:
6162 return io_sendmsg_prep_async(req
);
6163 case IORING_OP_RECVMSG
:
6164 case IORING_OP_RECV
:
6165 return io_recvmsg_prep_async(req
);
6166 case IORING_OP_CONNECT
:
6167 return io_connect_prep_async(req
);
6172 static int io_req_defer_prep(struct io_kiocb
*req
)
6174 if (!io_op_defs
[req
->opcode
].needs_async_data
)
6176 /* some opcodes init it during the inital prep */
6177 if (req
->async_data
)
6179 if (__io_alloc_async_data(req
))
6181 return io_req_prep_async(req
);
6184 static u32
io_get_sequence(struct io_kiocb
*req
)
6186 struct io_kiocb
*pos
;
6187 struct io_ring_ctx
*ctx
= req
->ctx
;
6188 u32 total_submitted
, nr_reqs
= 0;
6190 io_for_each_link(pos
, req
)
6193 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
6194 return total_submitted
- nr_reqs
;
6197 static int io_req_defer(struct io_kiocb
*req
)
6199 struct io_ring_ctx
*ctx
= req
->ctx
;
6200 struct io_defer_entry
*de
;
6204 /* Still need defer if there is pending req in defer list. */
6205 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6206 !(req
->flags
& REQ_F_IO_DRAIN
)))
6209 seq
= io_get_sequence(req
);
6210 /* Still a chance to pass the sequence check */
6211 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6214 ret
= io_req_defer_prep(req
);
6217 io_prep_async_link(req
);
6218 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6222 spin_lock_irq(&ctx
->completion_lock
);
6223 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6224 spin_unlock_irq(&ctx
->completion_lock
);
6226 io_queue_async_work(req
);
6227 return -EIOCBQUEUED
;
6230 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6233 list_add_tail(&de
->list
, &ctx
->defer_list
);
6234 spin_unlock_irq(&ctx
->completion_lock
);
6235 return -EIOCBQUEUED
;
6238 static void __io_clean_op(struct io_kiocb
*req
)
6240 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6241 switch (req
->opcode
) {
6242 case IORING_OP_READV
:
6243 case IORING_OP_READ_FIXED
:
6244 case IORING_OP_READ
:
6245 kfree((void *)(unsigned long)req
->rw
.addr
);
6247 case IORING_OP_RECVMSG
:
6248 case IORING_OP_RECV
:
6249 kfree(req
->sr_msg
.kbuf
);
6252 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
6255 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6256 switch (req
->opcode
) {
6257 case IORING_OP_READV
:
6258 case IORING_OP_READ_FIXED
:
6259 case IORING_OP_READ
:
6260 case IORING_OP_WRITEV
:
6261 case IORING_OP_WRITE_FIXED
:
6262 case IORING_OP_WRITE
: {
6263 struct io_async_rw
*io
= req
->async_data
;
6265 kfree(io
->free_iovec
);
6268 case IORING_OP_RECVMSG
:
6269 case IORING_OP_SENDMSG
: {
6270 struct io_async_msghdr
*io
= req
->async_data
;
6272 kfree(io
->free_iov
);
6275 case IORING_OP_SPLICE
:
6277 io_put_file(req
, req
->splice
.file_in
,
6278 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
6280 case IORING_OP_OPENAT
:
6281 case IORING_OP_OPENAT2
:
6282 if (req
->open
.filename
)
6283 putname(req
->open
.filename
);
6285 case IORING_OP_RENAMEAT
:
6286 putname(req
->rename
.oldpath
);
6287 putname(req
->rename
.newpath
);
6289 case IORING_OP_UNLINKAT
:
6290 putname(req
->unlink
.filename
);
6293 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6297 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6299 struct io_ring_ctx
*ctx
= req
->ctx
;
6302 switch (req
->opcode
) {
6304 ret
= io_nop(req
, issue_flags
);
6306 case IORING_OP_READV
:
6307 case IORING_OP_READ_FIXED
:
6308 case IORING_OP_READ
:
6309 ret
= io_read(req
, issue_flags
);
6311 case IORING_OP_WRITEV
:
6312 case IORING_OP_WRITE_FIXED
:
6313 case IORING_OP_WRITE
:
6314 ret
= io_write(req
, issue_flags
);
6316 case IORING_OP_FSYNC
:
6317 ret
= io_fsync(req
, issue_flags
);
6319 case IORING_OP_POLL_ADD
:
6320 ret
= io_poll_add(req
, issue_flags
);
6322 case IORING_OP_POLL_REMOVE
:
6323 ret
= io_poll_remove(req
, issue_flags
);
6325 case IORING_OP_SYNC_FILE_RANGE
:
6326 ret
= io_sync_file_range(req
, issue_flags
);
6328 case IORING_OP_SENDMSG
:
6329 ret
= io_sendmsg(req
, issue_flags
);
6331 case IORING_OP_SEND
:
6332 ret
= io_send(req
, issue_flags
);
6334 case IORING_OP_RECVMSG
:
6335 ret
= io_recvmsg(req
, issue_flags
);
6337 case IORING_OP_RECV
:
6338 ret
= io_recv(req
, issue_flags
);
6340 case IORING_OP_TIMEOUT
:
6341 ret
= io_timeout(req
, issue_flags
);
6343 case IORING_OP_TIMEOUT_REMOVE
:
6344 ret
= io_timeout_remove(req
, issue_flags
);
6346 case IORING_OP_ACCEPT
:
6347 ret
= io_accept(req
, issue_flags
);
6349 case IORING_OP_CONNECT
:
6350 ret
= io_connect(req
, issue_flags
);
6352 case IORING_OP_ASYNC_CANCEL
:
6353 ret
= io_async_cancel(req
, issue_flags
);
6355 case IORING_OP_FALLOCATE
:
6356 ret
= io_fallocate(req
, issue_flags
);
6358 case IORING_OP_OPENAT
:
6359 ret
= io_openat(req
, issue_flags
);
6361 case IORING_OP_CLOSE
:
6362 ret
= io_close(req
, issue_flags
);
6364 case IORING_OP_FILES_UPDATE
:
6365 ret
= io_files_update(req
, issue_flags
);
6367 case IORING_OP_STATX
:
6368 ret
= io_statx(req
, issue_flags
);
6370 case IORING_OP_FADVISE
:
6371 ret
= io_fadvise(req
, issue_flags
);
6373 case IORING_OP_MADVISE
:
6374 ret
= io_madvise(req
, issue_flags
);
6376 case IORING_OP_OPENAT2
:
6377 ret
= io_openat2(req
, issue_flags
);
6379 case IORING_OP_EPOLL_CTL
:
6380 ret
= io_epoll_ctl(req
, issue_flags
);
6382 case IORING_OP_SPLICE
:
6383 ret
= io_splice(req
, issue_flags
);
6385 case IORING_OP_PROVIDE_BUFFERS
:
6386 ret
= io_provide_buffers(req
, issue_flags
);
6388 case IORING_OP_REMOVE_BUFFERS
:
6389 ret
= io_remove_buffers(req
, issue_flags
);
6392 ret
= io_tee(req
, issue_flags
);
6394 case IORING_OP_SHUTDOWN
:
6395 ret
= io_shutdown(req
, issue_flags
);
6397 case IORING_OP_RENAMEAT
:
6398 ret
= io_renameat(req
, issue_flags
);
6400 case IORING_OP_UNLINKAT
:
6401 ret
= io_unlinkat(req
, issue_flags
);
6411 /* If the op doesn't have a file, we're not polling for it */
6412 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6413 const bool in_async
= io_wq_current_is_worker();
6415 /* workqueue context doesn't hold uring_lock, grab it now */
6417 mutex_lock(&ctx
->uring_lock
);
6419 io_iopoll_req_issued(req
, in_async
);
6422 mutex_unlock(&ctx
->uring_lock
);
6428 static void io_wq_submit_work(struct io_wq_work
*work
)
6430 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6431 struct io_kiocb
*timeout
;
6434 timeout
= io_prep_linked_timeout(req
);
6436 io_queue_linked_timeout(timeout
);
6438 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6443 ret
= io_issue_sqe(req
, 0);
6445 * We can get EAGAIN for polled IO even though we're
6446 * forcing a sync submission from here, since we can't
6447 * wait for request slots on the block side.
6455 /* avoid locking problems by failing it from a clean context */
6457 /* io-wq is going to take one down */
6458 refcount_inc(&req
->refs
);
6459 io_req_task_queue_fail(req
, ret
);
6463 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6466 struct fixed_rsrc_table
*table
;
6468 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6469 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6472 static struct file
*io_file_get(struct io_submit_state
*state
,
6473 struct io_kiocb
*req
, int fd
, bool fixed
)
6475 struct io_ring_ctx
*ctx
= req
->ctx
;
6479 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6481 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6482 file
= io_file_from_index(ctx
, fd
);
6483 io_set_resource_node(req
);
6485 trace_io_uring_file_get(ctx
, fd
);
6486 file
= __io_file_get(state
, fd
);
6489 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6490 io_req_track_inflight(req
);
6494 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6496 struct io_timeout_data
*data
= container_of(timer
,
6497 struct io_timeout_data
, timer
);
6498 struct io_kiocb
*prev
, *req
= data
->req
;
6499 struct io_ring_ctx
*ctx
= req
->ctx
;
6500 unsigned long flags
;
6502 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6503 prev
= req
->timeout
.head
;
6504 req
->timeout
.head
= NULL
;
6507 * We don't expect the list to be empty, that will only happen if we
6508 * race with the completion of the linked work.
6510 if (prev
&& refcount_inc_not_zero(&prev
->refs
))
6511 io_remove_next_linked(prev
);
6514 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6517 req_set_fail_links(prev
);
6518 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6519 io_put_req_deferred(prev
, 1);
6521 io_req_complete_post(req
, -ETIME
, 0);
6522 io_put_req_deferred(req
, 1);
6524 return HRTIMER_NORESTART
;
6527 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6530 * If the back reference is NULL, then our linked request finished
6531 * before we got a chance to setup the timer
6533 if (req
->timeout
.head
) {
6534 struct io_timeout_data
*data
= req
->async_data
;
6536 data
->timer
.function
= io_link_timeout_fn
;
6537 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6542 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6544 struct io_ring_ctx
*ctx
= req
->ctx
;
6546 spin_lock_irq(&ctx
->completion_lock
);
6547 __io_queue_linked_timeout(req
);
6548 spin_unlock_irq(&ctx
->completion_lock
);
6550 /* drop submission reference */
6554 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6556 struct io_kiocb
*nxt
= req
->link
;
6558 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6559 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6562 nxt
->timeout
.head
= req
;
6563 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6564 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6568 static void __io_queue_sqe(struct io_kiocb
*req
)
6570 struct io_kiocb
*linked_timeout
= io_prep_linked_timeout(req
);
6571 const struct cred
*old_creds
= NULL
;
6574 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
6575 (req
->work
.flags
& IO_WQ_WORK_CREDS
) &&
6576 req
->work
.identity
->creds
!= current_cred())
6577 old_creds
= override_creds(req
->work
.identity
->creds
);
6579 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
6582 revert_creds(old_creds
);
6585 * We async punt it if the file wasn't marked NOWAIT, or if the file
6586 * doesn't support non-blocking read/write attempts
6588 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6589 if (!io_arm_poll_handler(req
)) {
6591 * Queued up for async execution, worker will release
6592 * submit reference when the iocb is actually submitted.
6594 io_queue_async_work(req
);
6596 } else if (likely(!ret
)) {
6597 /* drop submission reference */
6598 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
6599 struct io_ring_ctx
*ctx
= req
->ctx
;
6600 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
6602 cs
->reqs
[cs
->nr
++] = req
;
6603 if (cs
->nr
== ARRAY_SIZE(cs
->reqs
))
6604 io_submit_flush_completions(cs
, ctx
);
6609 req_set_fail_links(req
);
6611 io_req_complete(req
, ret
);
6614 io_queue_linked_timeout(linked_timeout
);
6617 static void io_queue_sqe(struct io_kiocb
*req
)
6621 ret
= io_req_defer(req
);
6623 if (ret
!= -EIOCBQUEUED
) {
6625 req_set_fail_links(req
);
6627 io_req_complete(req
, ret
);
6629 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6630 ret
= io_req_defer_prep(req
);
6633 io_queue_async_work(req
);
6635 __io_queue_sqe(req
);
6640 * Check SQE restrictions (opcode and flags).
6642 * Returns 'true' if SQE is allowed, 'false' otherwise.
6644 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6645 struct io_kiocb
*req
,
6646 unsigned int sqe_flags
)
6648 if (!ctx
->restricted
)
6651 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6654 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6655 ctx
->restrictions
.sqe_flags_required
)
6658 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6659 ctx
->restrictions
.sqe_flags_required
))
6665 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6666 const struct io_uring_sqe
*sqe
)
6668 struct io_submit_state
*state
;
6669 unsigned int sqe_flags
;
6672 req
->opcode
= READ_ONCE(sqe
->opcode
);
6673 /* same numerical values with corresponding REQ_F_*, safe to copy */
6674 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
6675 req
->user_data
= READ_ONCE(sqe
->user_data
);
6676 req
->async_data
= NULL
;
6680 req
->fixed_rsrc_refs
= NULL
;
6681 /* one is dropped after submission, the other at completion */
6682 refcount_set(&req
->refs
, 2);
6683 req
->task
= current
;
6686 /* enforce forwards compatibility on users */
6687 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
)) {
6692 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6695 if (unlikely(io_sq_thread_acquire_mm_files(ctx
, req
)))
6698 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6701 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6702 !io_op_defs
[req
->opcode
].buffer_select
)
6705 id
= READ_ONCE(sqe
->personality
);
6707 struct io_identity
*iod
;
6709 iod
= idr_find(&ctx
->personality_idr
, id
);
6712 refcount_inc(&iod
->count
);
6714 __io_req_init_async(req
);
6715 get_cred(iod
->creds
);
6716 req
->work
.identity
= iod
;
6717 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
6720 state
= &ctx
->submit_state
;
6723 * Plug now if we have more than 1 IO left after this, and the target
6724 * is potentially a read/write to block based storage.
6726 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6727 io_op_defs
[req
->opcode
].plug
) {
6728 blk_start_plug(&state
->plug
);
6729 state
->plug_started
= true;
6732 if (io_op_defs
[req
->opcode
].needs_file
) {
6733 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6735 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6736 if (unlikely(!req
->file
))
6744 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6745 const struct io_uring_sqe
*sqe
)
6747 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
6750 ret
= io_init_req(ctx
, req
, sqe
);
6751 if (unlikely(ret
)) {
6754 io_req_complete(req
, ret
);
6756 /* fail even hard links since we don't submit */
6757 link
->head
->flags
|= REQ_F_FAIL_LINK
;
6758 io_put_req(link
->head
);
6759 io_req_complete(link
->head
, -ECANCELED
);
6764 ret
= io_req_prep(req
, sqe
);
6768 /* don't need @sqe from now on */
6769 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6770 true, ctx
->flags
& IORING_SETUP_SQPOLL
);
6773 * If we already have a head request, queue this one for async
6774 * submittal once the head completes. If we don't have a head but
6775 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6776 * submitted sync once the chain is complete. If none of those
6777 * conditions are true (normal request), then just queue it.
6780 struct io_kiocb
*head
= link
->head
;
6783 * Taking sequential execution of a link, draining both sides
6784 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6785 * requests in the link. So, it drains the head and the
6786 * next after the link request. The last one is done via
6787 * drain_next flag to persist the effect across calls.
6789 if (req
->flags
& REQ_F_IO_DRAIN
) {
6790 head
->flags
|= REQ_F_IO_DRAIN
;
6791 ctx
->drain_next
= 1;
6793 ret
= io_req_defer_prep(req
);
6796 trace_io_uring_link(ctx
, req
, head
);
6797 link
->last
->link
= req
;
6800 /* last request of a link, enqueue the link */
6801 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6806 if (unlikely(ctx
->drain_next
)) {
6807 req
->flags
|= REQ_F_IO_DRAIN
;
6808 ctx
->drain_next
= 0;
6810 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6822 * Batched submission is done, ensure local IO is flushed out.
6824 static void io_submit_state_end(struct io_submit_state
*state
,
6825 struct io_ring_ctx
*ctx
)
6827 if (state
->link
.head
)
6828 io_queue_sqe(state
->link
.head
);
6830 io_submit_flush_completions(&state
->comp
, ctx
);
6831 if (state
->plug_started
)
6832 blk_finish_plug(&state
->plug
);
6833 io_state_file_put(state
);
6837 * Start submission side cache.
6839 static void io_submit_state_start(struct io_submit_state
*state
,
6840 unsigned int max_ios
)
6842 state
->plug_started
= false;
6843 state
->ios_left
= max_ios
;
6844 /* set only head, no need to init link_last in advance */
6845 state
->link
.head
= NULL
;
6848 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6850 struct io_rings
*rings
= ctx
->rings
;
6853 * Ensure any loads from the SQEs are done at this point,
6854 * since once we write the new head, the application could
6855 * write new data to them.
6857 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6861 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6862 * that is mapped by userspace. This means that care needs to be taken to
6863 * ensure that reads are stable, as we cannot rely on userspace always
6864 * being a good citizen. If members of the sqe are validated and then later
6865 * used, it's important that those reads are done through READ_ONCE() to
6866 * prevent a re-load down the line.
6868 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6870 u32
*sq_array
= ctx
->sq_array
;
6874 * The cached sq head (or cq tail) serves two purposes:
6876 * 1) allows us to batch the cost of updating the user visible
6878 * 2) allows the kernel side to track the head on its own, even
6879 * though the application is the one updating it.
6881 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
++ & ctx
->sq_mask
]);
6882 if (likely(head
< ctx
->sq_entries
))
6883 return &ctx
->sq_sqes
[head
];
6885 /* drop invalid entries */
6886 ctx
->cached_sq_dropped
++;
6887 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6891 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6895 /* if we have a backlog and couldn't flush it all, return BUSY */
6896 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6897 if (!__io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6901 /* make sure SQ entry isn't read before tail */
6902 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6904 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6907 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6908 refcount_add(nr
, ¤t
->usage
);
6909 io_submit_state_start(&ctx
->submit_state
, nr
);
6911 while (submitted
< nr
) {
6912 const struct io_uring_sqe
*sqe
;
6913 struct io_kiocb
*req
;
6915 req
= io_alloc_req(ctx
);
6916 if (unlikely(!req
)) {
6918 submitted
= -EAGAIN
;
6921 sqe
= io_get_sqe(ctx
);
6922 if (unlikely(!sqe
)) {
6923 kmem_cache_free(req_cachep
, req
);
6926 /* will complete beyond this point, count as submitted */
6928 if (io_submit_sqe(ctx
, req
, sqe
))
6932 if (unlikely(submitted
!= nr
)) {
6933 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6934 struct io_uring_task
*tctx
= current
->io_uring
;
6935 int unused
= nr
- ref_used
;
6937 percpu_ref_put_many(&ctx
->refs
, unused
);
6938 percpu_counter_sub(&tctx
->inflight
, unused
);
6939 put_task_struct_many(current
, unused
);
6942 io_submit_state_end(&ctx
->submit_state
, ctx
);
6943 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6944 io_commit_sqring(ctx
);
6949 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6951 /* Tell userspace we may need a wakeup call */
6952 spin_lock_irq(&ctx
->completion_lock
);
6953 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6954 spin_unlock_irq(&ctx
->completion_lock
);
6957 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6959 spin_lock_irq(&ctx
->completion_lock
);
6960 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6961 spin_unlock_irq(&ctx
->completion_lock
);
6964 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
6966 unsigned int to_submit
;
6969 to_submit
= io_sqring_entries(ctx
);
6970 /* if we're handling multiple rings, cap submit size for fairness */
6971 if (cap_entries
&& to_submit
> 8)
6974 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
6975 unsigned nr_events
= 0;
6977 mutex_lock(&ctx
->uring_lock
);
6978 if (!list_empty(&ctx
->iopoll_list
))
6979 io_do_iopoll(ctx
, &nr_events
, 0);
6981 if (to_submit
&& !ctx
->sqo_dead
&&
6982 likely(!percpu_ref_is_dying(&ctx
->refs
)))
6983 ret
= io_submit_sqes(ctx
, to_submit
);
6984 mutex_unlock(&ctx
->uring_lock
);
6987 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
6988 wake_up(&ctx
->sqo_sq_wait
);
6993 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
6995 struct io_ring_ctx
*ctx
;
6996 unsigned sq_thread_idle
= 0;
6998 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6999 if (sq_thread_idle
< ctx
->sq_thread_idle
)
7000 sq_thread_idle
= ctx
->sq_thread_idle
;
7003 sqd
->sq_thread_idle
= sq_thread_idle
;
7006 static void io_sqd_init_new(struct io_sq_data
*sqd
)
7008 struct io_ring_ctx
*ctx
;
7010 while (!list_empty(&sqd
->ctx_new_list
)) {
7011 ctx
= list_first_entry(&sqd
->ctx_new_list
, struct io_ring_ctx
, sqd_list
);
7012 list_move_tail(&ctx
->sqd_list
, &sqd
->ctx_list
);
7013 complete(&ctx
->sq_thread_comp
);
7016 io_sqd_update_thread_idle(sqd
);
7019 static int io_sq_thread(void *data
)
7021 struct cgroup_subsys_state
*cur_css
= NULL
;
7022 struct files_struct
*old_files
= current
->files
;
7023 struct nsproxy
*old_nsproxy
= current
->nsproxy
;
7024 const struct cred
*old_cred
= NULL
;
7025 struct io_sq_data
*sqd
= data
;
7026 struct io_ring_ctx
*ctx
;
7027 unsigned long timeout
= 0;
7031 current
->files
= NULL
;
7032 current
->nsproxy
= NULL
;
7033 task_unlock(current
);
7035 while (!kthread_should_stop()) {
7037 bool cap_entries
, sqt_spin
, needs_sched
;
7040 * Any changes to the sqd lists are synchronized through the
7041 * kthread parking. This synchronizes the thread vs users,
7042 * the users are synchronized on the sqd->ctx_lock.
7044 if (kthread_should_park()) {
7047 * When sq thread is unparked, in case the previous park operation
7048 * comes from io_put_sq_data(), which means that sq thread is going
7049 * to be stopped, so here needs to have a check.
7051 if (kthread_should_stop())
7055 if (unlikely(!list_empty(&sqd
->ctx_new_list
))) {
7056 io_sqd_init_new(sqd
);
7057 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7061 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7062 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7063 if (current
->cred
!= ctx
->creds
) {
7065 revert_creds(old_cred
);
7066 old_cred
= override_creds(ctx
->creds
);
7068 io_sq_thread_associate_blkcg(ctx
, &cur_css
);
7070 current
->loginuid
= ctx
->loginuid
;
7071 current
->sessionid
= ctx
->sessionid
;
7074 ret
= __io_sq_thread(ctx
, cap_entries
);
7075 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7078 io_sq_thread_drop_mm_files();
7081 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7083 io_sq_thread_drop_mm_files();
7086 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7091 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7092 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7093 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7094 !list_empty_careful(&ctx
->iopoll_list
)) {
7095 needs_sched
= false;
7098 if (io_sqring_entries(ctx
)) {
7099 needs_sched
= false;
7104 if (needs_sched
&& !kthread_should_park()) {
7105 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7106 io_ring_set_wakeup_flag(ctx
);
7109 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7110 io_ring_clear_wakeup_flag(ctx
);
7113 finish_wait(&sqd
->wait
, &wait
);
7114 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7118 io_sq_thread_drop_mm_files();
7121 io_sq_thread_unassociate_blkcg();
7123 revert_creds(old_cred
);
7126 current
->files
= old_files
;
7127 current
->nsproxy
= old_nsproxy
;
7128 task_unlock(current
);
7135 struct io_wait_queue
{
7136 struct wait_queue_entry wq
;
7137 struct io_ring_ctx
*ctx
;
7139 unsigned nr_timeouts
;
7142 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7144 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7147 * Wake up if we have enough events, or if a timeout occurred since we
7148 * started waiting. For timeouts, we always want to return to userspace,
7149 * regardless of event count.
7151 return io_cqring_events(ctx
) >= iowq
->to_wait
||
7152 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7155 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7156 int wake_flags
, void *key
)
7158 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7162 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7163 * the task, and the next invocation will do it.
7165 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->cq_check_overflow
))
7166 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7170 static int io_run_task_work_sig(void)
7172 if (io_run_task_work())
7174 if (!signal_pending(current
))
7176 if (test_tsk_thread_flag(current
, TIF_NOTIFY_SIGNAL
))
7177 return -ERESTARTSYS
;
7181 /* when returns >0, the caller should retry */
7182 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7183 struct io_wait_queue
*iowq
,
7184 signed long *timeout
)
7188 /* make sure we run task_work before checking for signals */
7189 ret
= io_run_task_work_sig();
7190 if (ret
|| io_should_wake(iowq
))
7192 /* let the caller flush overflows, retry */
7193 if (test_bit(0, &ctx
->cq_check_overflow
))
7196 *timeout
= schedule_timeout(*timeout
);
7197 return !*timeout
? -ETIME
: 1;
7201 * Wait until events become available, if we don't already have some. The
7202 * application must reap them itself, as they reside on the shared cq ring.
7204 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7205 const sigset_t __user
*sig
, size_t sigsz
,
7206 struct __kernel_timespec __user
*uts
)
7208 struct io_wait_queue iowq
= {
7211 .func
= io_wake_function
,
7212 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
7215 .to_wait
= min_events
,
7217 struct io_rings
*rings
= ctx
->rings
;
7218 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
7222 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7223 if (io_cqring_events(ctx
) >= min_events
)
7225 if (!io_run_task_work())
7230 #ifdef CONFIG_COMPAT
7231 if (in_compat_syscall())
7232 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7236 ret
= set_user_sigmask(sig
, sigsz
);
7243 struct timespec64 ts
;
7245 if (get_timespec64(&ts
, uts
))
7247 timeout
= timespec64_to_jiffies(&ts
);
7250 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7251 trace_io_uring_cqring_wait(ctx
, min_events
);
7253 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7254 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
7255 TASK_INTERRUPTIBLE
);
7256 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
7257 finish_wait(&ctx
->wait
, &iowq
.wq
);
7260 restore_saved_sigmask_unless(ret
== -EINTR
);
7262 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7265 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7267 #if defined(CONFIG_UNIX)
7268 if (ctx
->ring_sock
) {
7269 struct sock
*sock
= ctx
->ring_sock
->sk
;
7270 struct sk_buff
*skb
;
7272 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7278 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7281 file
= io_file_from_index(ctx
, i
);
7288 static void io_rsrc_data_ref_zero(struct percpu_ref
*ref
)
7290 struct fixed_rsrc_data
*data
;
7292 data
= container_of(ref
, struct fixed_rsrc_data
, refs
);
7293 complete(&data
->done
);
7296 static inline void io_rsrc_ref_lock(struct io_ring_ctx
*ctx
)
7298 spin_lock_bh(&ctx
->rsrc_ref_lock
);
7301 static inline void io_rsrc_ref_unlock(struct io_ring_ctx
*ctx
)
7303 spin_unlock_bh(&ctx
->rsrc_ref_lock
);
7306 static void io_sqe_rsrc_set_node(struct io_ring_ctx
*ctx
,
7307 struct fixed_rsrc_data
*rsrc_data
,
7308 struct fixed_rsrc_ref_node
*ref_node
)
7310 io_rsrc_ref_lock(ctx
);
7311 rsrc_data
->node
= ref_node
;
7312 list_add_tail(&ref_node
->node
, &ctx
->rsrc_ref_list
);
7313 io_rsrc_ref_unlock(ctx
);
7314 percpu_ref_get(&rsrc_data
->refs
);
7317 static void io_sqe_rsrc_kill_node(struct io_ring_ctx
*ctx
, struct fixed_rsrc_data
*data
)
7319 struct fixed_rsrc_ref_node
*ref_node
= NULL
;
7321 io_rsrc_ref_lock(ctx
);
7322 ref_node
= data
->node
;
7324 io_rsrc_ref_unlock(ctx
);
7326 percpu_ref_kill(&ref_node
->refs
);
7329 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data
*data
,
7330 struct io_ring_ctx
*ctx
,
7331 void (*rsrc_put
)(struct io_ring_ctx
*ctx
,
7332 struct io_rsrc_put
*prsrc
))
7334 struct fixed_rsrc_ref_node
*backup_node
;
7340 data
->quiesce
= true;
7343 backup_node
= alloc_fixed_rsrc_ref_node(ctx
);
7346 backup_node
->rsrc_data
= data
;
7347 backup_node
->rsrc_put
= rsrc_put
;
7349 io_sqe_rsrc_kill_node(ctx
, data
);
7350 percpu_ref_kill(&data
->refs
);
7351 flush_delayed_work(&ctx
->rsrc_put_work
);
7353 ret
= wait_for_completion_interruptible(&data
->done
);
7354 if (!ret
|| !io_refs_resurrect(&data
->refs
, &data
->done
))
7357 io_sqe_rsrc_set_node(ctx
, data
, backup_node
);
7359 mutex_unlock(&ctx
->uring_lock
);
7360 ret
= io_run_task_work_sig();
7361 mutex_lock(&ctx
->uring_lock
);
7363 data
->quiesce
= false;
7366 destroy_fixed_rsrc_ref_node(backup_node
);
7370 static struct fixed_rsrc_data
*alloc_fixed_rsrc_data(struct io_ring_ctx
*ctx
)
7372 struct fixed_rsrc_data
*data
;
7374 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7378 if (percpu_ref_init(&data
->refs
, io_rsrc_data_ref_zero
,
7379 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
)) {
7384 init_completion(&data
->done
);
7388 static void free_fixed_rsrc_data(struct fixed_rsrc_data
*data
)
7390 percpu_ref_exit(&data
->refs
);
7395 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7397 struct fixed_rsrc_data
*data
= ctx
->file_data
;
7398 unsigned nr_tables
, i
;
7402 * percpu_ref_is_dying() is to stop parallel files unregister
7403 * Since we possibly drop uring lock later in this function to
7406 if (!data
|| percpu_ref_is_dying(&data
->refs
))
7408 ret
= io_rsrc_ref_quiesce(data
, ctx
, io_ring_file_put
);
7412 __io_sqe_files_unregister(ctx
);
7413 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
7414 for (i
= 0; i
< nr_tables
; i
++)
7415 kfree(data
->table
[i
].files
);
7416 free_fixed_rsrc_data(data
);
7417 ctx
->file_data
= NULL
;
7418 ctx
->nr_user_files
= 0;
7422 static void io_put_sq_data(struct io_sq_data
*sqd
)
7424 if (refcount_dec_and_test(&sqd
->refs
)) {
7426 * The park is a bit of a work-around, without it we get
7427 * warning spews on shutdown with SQPOLL set and affinity
7428 * set to a single CPU.
7431 kthread_park(sqd
->thread
);
7432 kthread_stop(sqd
->thread
);
7439 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7441 struct io_ring_ctx
*ctx_attach
;
7442 struct io_sq_data
*sqd
;
7445 f
= fdget(p
->wq_fd
);
7447 return ERR_PTR(-ENXIO
);
7448 if (f
.file
->f_op
!= &io_uring_fops
) {
7450 return ERR_PTR(-EINVAL
);
7453 ctx_attach
= f
.file
->private_data
;
7454 sqd
= ctx_attach
->sq_data
;
7457 return ERR_PTR(-EINVAL
);
7460 refcount_inc(&sqd
->refs
);
7465 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
)
7467 struct io_sq_data
*sqd
;
7469 if (p
->flags
& IORING_SETUP_ATTACH_WQ
)
7470 return io_attach_sq_data(p
);
7472 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7474 return ERR_PTR(-ENOMEM
);
7476 refcount_set(&sqd
->refs
, 1);
7477 INIT_LIST_HEAD(&sqd
->ctx_list
);
7478 INIT_LIST_HEAD(&sqd
->ctx_new_list
);
7479 mutex_init(&sqd
->ctx_lock
);
7480 mutex_init(&sqd
->lock
);
7481 init_waitqueue_head(&sqd
->wait
);
7485 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7486 __releases(&sqd
->lock
)
7490 kthread_unpark(sqd
->thread
);
7491 mutex_unlock(&sqd
->lock
);
7494 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7495 __acquires(&sqd
->lock
)
7499 mutex_lock(&sqd
->lock
);
7500 kthread_park(sqd
->thread
);
7503 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
7505 struct io_sq_data
*sqd
= ctx
->sq_data
;
7510 * We may arrive here from the error branch in
7511 * io_sq_offload_create() where the kthread is created
7512 * without being waked up, thus wake it up now to make
7513 * sure the wait will complete.
7515 wake_up_process(sqd
->thread
);
7516 wait_for_completion(&ctx
->sq_thread_comp
);
7518 io_sq_thread_park(sqd
);
7521 mutex_lock(&sqd
->ctx_lock
);
7522 list_del(&ctx
->sqd_list
);
7523 io_sqd_update_thread_idle(sqd
);
7524 mutex_unlock(&sqd
->ctx_lock
);
7527 io_sq_thread_unpark(sqd
);
7529 io_put_sq_data(sqd
);
7530 ctx
->sq_data
= NULL
;
7534 static void io_finish_async(struct io_ring_ctx
*ctx
)
7536 io_sq_thread_stop(ctx
);
7539 io_wq_destroy(ctx
->io_wq
);
7544 #if defined(CONFIG_UNIX)
7546 * Ensure the UNIX gc is aware of our file set, so we are certain that
7547 * the io_uring can be safely unregistered on process exit, even if we have
7548 * loops in the file referencing.
7550 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7552 struct sock
*sk
= ctx
->ring_sock
->sk
;
7553 struct scm_fp_list
*fpl
;
7554 struct sk_buff
*skb
;
7557 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7561 skb
= alloc_skb(0, GFP_KERNEL
);
7570 fpl
->user
= get_uid(ctx
->user
);
7571 for (i
= 0; i
< nr
; i
++) {
7572 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7576 fpl
->fp
[nr_files
] = get_file(file
);
7577 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7582 fpl
->max
= SCM_MAX_FD
;
7583 fpl
->count
= nr_files
;
7584 UNIXCB(skb
).fp
= fpl
;
7585 skb
->destructor
= unix_destruct_scm
;
7586 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7587 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7589 for (i
= 0; i
< nr_files
; i
++)
7600 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7601 * causes regular reference counting to break down. We rely on the UNIX
7602 * garbage collection to take care of this problem for us.
7604 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7606 unsigned left
, total
;
7610 left
= ctx
->nr_user_files
;
7612 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7614 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7618 total
+= this_files
;
7624 while (total
< ctx
->nr_user_files
) {
7625 struct file
*file
= io_file_from_index(ctx
, total
);
7635 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7641 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data
*file_data
,
7642 unsigned nr_tables
, unsigned nr_files
)
7646 for (i
= 0; i
< nr_tables
; i
++) {
7647 struct fixed_rsrc_table
*table
= &file_data
->table
[i
];
7648 unsigned this_files
;
7650 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7651 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7655 nr_files
-= this_files
;
7661 for (i
= 0; i
< nr_tables
; i
++) {
7662 struct fixed_rsrc_table
*table
= &file_data
->table
[i
];
7663 kfree(table
->files
);
7668 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
7670 struct file
*file
= prsrc
->file
;
7671 #if defined(CONFIG_UNIX)
7672 struct sock
*sock
= ctx
->ring_sock
->sk
;
7673 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7674 struct sk_buff
*skb
;
7677 __skb_queue_head_init(&list
);
7680 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7681 * remove this entry and rearrange the file array.
7683 skb
= skb_dequeue(head
);
7685 struct scm_fp_list
*fp
;
7687 fp
= UNIXCB(skb
).fp
;
7688 for (i
= 0; i
< fp
->count
; i
++) {
7691 if (fp
->fp
[i
] != file
)
7694 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7695 left
= fp
->count
- 1 - i
;
7697 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7698 left
* sizeof(struct file
*));
7705 __skb_queue_tail(&list
, skb
);
7715 __skb_queue_tail(&list
, skb
);
7717 skb
= skb_dequeue(head
);
7720 if (skb_peek(&list
)) {
7721 spin_lock_irq(&head
->lock
);
7722 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7723 __skb_queue_tail(head
, skb
);
7724 spin_unlock_irq(&head
->lock
);
7731 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node
*ref_node
)
7733 struct fixed_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
7734 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
7735 struct io_rsrc_put
*prsrc
, *tmp
;
7737 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
7738 list_del(&prsrc
->list
);
7739 ref_node
->rsrc_put(ctx
, prsrc
);
7743 percpu_ref_exit(&ref_node
->refs
);
7745 percpu_ref_put(&rsrc_data
->refs
);
7748 static void io_rsrc_put_work(struct work_struct
*work
)
7750 struct io_ring_ctx
*ctx
;
7751 struct llist_node
*node
;
7753 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
7754 node
= llist_del_all(&ctx
->rsrc_put_llist
);
7757 struct fixed_rsrc_ref_node
*ref_node
;
7758 struct llist_node
*next
= node
->next
;
7760 ref_node
= llist_entry(node
, struct fixed_rsrc_ref_node
, llist
);
7761 __io_rsrc_put_work(ref_node
);
7766 static struct file
**io_fixed_file_slot(struct fixed_rsrc_data
*file_data
,
7769 struct fixed_rsrc_table
*table
;
7771 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7772 return &table
->files
[i
& IORING_FILE_TABLE_MASK
];
7775 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7777 struct fixed_rsrc_ref_node
*ref_node
;
7778 struct fixed_rsrc_data
*data
;
7779 struct io_ring_ctx
*ctx
;
7780 bool first_add
= false;
7783 ref_node
= container_of(ref
, struct fixed_rsrc_ref_node
, refs
);
7784 data
= ref_node
->rsrc_data
;
7787 io_rsrc_ref_lock(ctx
);
7788 ref_node
->done
= true;
7790 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7791 ref_node
= list_first_entry(&ctx
->rsrc_ref_list
,
7792 struct fixed_rsrc_ref_node
, node
);
7793 /* recycle ref nodes in order */
7794 if (!ref_node
->done
)
7796 list_del(&ref_node
->node
);
7797 first_add
|= llist_add(&ref_node
->llist
, &ctx
->rsrc_put_llist
);
7799 io_rsrc_ref_unlock(ctx
);
7801 if (percpu_ref_is_dying(&data
->refs
))
7805 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, 0);
7807 queue_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
7810 static struct fixed_rsrc_ref_node
*alloc_fixed_rsrc_ref_node(
7811 struct io_ring_ctx
*ctx
)
7813 struct fixed_rsrc_ref_node
*ref_node
;
7815 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7819 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7824 INIT_LIST_HEAD(&ref_node
->node
);
7825 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7826 ref_node
->done
= false;
7830 static void init_fixed_file_ref_node(struct io_ring_ctx
*ctx
,
7831 struct fixed_rsrc_ref_node
*ref_node
)
7833 ref_node
->rsrc_data
= ctx
->file_data
;
7834 ref_node
->rsrc_put
= io_ring_file_put
;
7837 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node
*ref_node
)
7839 percpu_ref_exit(&ref_node
->refs
);
7844 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7847 __s32 __user
*fds
= (__s32 __user
*) arg
;
7848 unsigned nr_tables
, i
;
7850 int fd
, ret
= -ENOMEM
;
7851 struct fixed_rsrc_ref_node
*ref_node
;
7852 struct fixed_rsrc_data
*file_data
;
7858 if (nr_args
> IORING_MAX_FIXED_FILES
)
7861 file_data
= alloc_fixed_rsrc_data(ctx
);
7864 ctx
->file_data
= file_data
;
7866 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7867 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7869 if (!file_data
->table
)
7872 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7875 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7876 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7880 /* allow sparse sets */
7890 * Don't allow io_uring instances to be registered. If UNIX
7891 * isn't enabled, then this causes a reference cycle and this
7892 * instance can never get freed. If UNIX is enabled we'll
7893 * handle it just fine, but there's still no point in allowing
7894 * a ring fd as it doesn't support regular read/write anyway.
7896 if (file
->f_op
== &io_uring_fops
) {
7900 *io_fixed_file_slot(file_data
, i
) = file
;
7903 ret
= io_sqe_files_scm(ctx
);
7905 io_sqe_files_unregister(ctx
);
7909 ref_node
= alloc_fixed_rsrc_ref_node(ctx
);
7911 io_sqe_files_unregister(ctx
);
7914 init_fixed_file_ref_node(ctx
, ref_node
);
7916 io_sqe_rsrc_set_node(ctx
, file_data
, ref_node
);
7919 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7920 file
= io_file_from_index(ctx
, i
);
7924 for (i
= 0; i
< nr_tables
; i
++)
7925 kfree(file_data
->table
[i
].files
);
7926 ctx
->nr_user_files
= 0;
7928 free_fixed_rsrc_data(ctx
->file_data
);
7929 ctx
->file_data
= NULL
;
7933 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7936 #if defined(CONFIG_UNIX)
7937 struct sock
*sock
= ctx
->ring_sock
->sk
;
7938 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7939 struct sk_buff
*skb
;
7942 * See if we can merge this file into an existing skb SCM_RIGHTS
7943 * file set. If there's no room, fall back to allocating a new skb
7944 * and filling it in.
7946 spin_lock_irq(&head
->lock
);
7947 skb
= skb_peek(head
);
7949 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7951 if (fpl
->count
< SCM_MAX_FD
) {
7952 __skb_unlink(skb
, head
);
7953 spin_unlock_irq(&head
->lock
);
7954 fpl
->fp
[fpl
->count
] = get_file(file
);
7955 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7957 spin_lock_irq(&head
->lock
);
7958 __skb_queue_head(head
, skb
);
7963 spin_unlock_irq(&head
->lock
);
7970 return __io_sqe_files_scm(ctx
, 1, index
);
7976 static int io_queue_rsrc_removal(struct fixed_rsrc_data
*data
, void *rsrc
)
7978 struct io_rsrc_put
*prsrc
;
7979 struct fixed_rsrc_ref_node
*ref_node
= data
->node
;
7981 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
7986 list_add(&prsrc
->list
, &ref_node
->rsrc_list
);
7991 static inline int io_queue_file_removal(struct fixed_rsrc_data
*data
,
7994 return io_queue_rsrc_removal(data
, (void *)file
);
7997 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7998 struct io_uring_rsrc_update
*up
,
8001 struct fixed_rsrc_data
*data
= ctx
->file_data
;
8002 struct fixed_rsrc_ref_node
*ref_node
;
8003 struct file
*file
, **file_slot
;
8007 bool needs_switch
= false;
8009 if (check_add_overflow(up
->offset
, nr_args
, &done
))
8011 if (done
> ctx
->nr_user_files
)
8014 ref_node
= alloc_fixed_rsrc_ref_node(ctx
);
8017 init_fixed_file_ref_node(ctx
, ref_node
);
8019 fds
= u64_to_user_ptr(up
->data
);
8020 for (done
= 0; done
< nr_args
; done
++) {
8022 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8026 if (fd
== IORING_REGISTER_FILES_SKIP
)
8029 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
8030 file_slot
= io_fixed_file_slot(ctx
->file_data
, i
);
8033 err
= io_queue_file_removal(data
, *file_slot
);
8037 needs_switch
= true;
8046 * Don't allow io_uring instances to be registered. If
8047 * UNIX isn't enabled, then this causes a reference
8048 * cycle and this instance can never get freed. If UNIX
8049 * is enabled we'll handle it just fine, but there's
8050 * still no point in allowing a ring fd as it doesn't
8051 * support regular read/write anyway.
8053 if (file
->f_op
== &io_uring_fops
) {
8059 err
= io_sqe_file_register(ctx
, file
, i
);
8069 percpu_ref_kill(&data
->node
->refs
);
8070 io_sqe_rsrc_set_node(ctx
, data
, ref_node
);
8072 destroy_fixed_rsrc_ref_node(ref_node
);
8074 return done
? done
: err
;
8077 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
8080 struct io_uring_rsrc_update up
;
8082 if (!ctx
->file_data
)
8086 if (copy_from_user(&up
, arg
, sizeof(up
)))
8091 return __io_sqe_files_update(ctx
, &up
, nr_args
);
8094 static struct io_wq_work
*io_free_work(struct io_wq_work
*work
)
8096 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8098 req
= io_put_req_find_next(req
);
8099 return req
? &req
->work
: NULL
;
8102 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
8103 struct io_uring_params
*p
)
8105 struct io_wq_data data
;
8107 struct io_ring_ctx
*ctx_attach
;
8108 unsigned int concurrency
;
8111 data
.user
= ctx
->user
;
8112 data
.free_work
= io_free_work
;
8113 data
.do_work
= io_wq_submit_work
;
8115 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
8116 /* Do QD, or 4 * CPUS, whatever is smallest */
8117 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8119 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
8120 if (IS_ERR(ctx
->io_wq
)) {
8121 ret
= PTR_ERR(ctx
->io_wq
);
8127 f
= fdget(p
->wq_fd
);
8131 if (f
.file
->f_op
!= &io_uring_fops
) {
8136 ctx_attach
= f
.file
->private_data
;
8137 /* @io_wq is protected by holding the fd */
8138 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
8143 ctx
->io_wq
= ctx_attach
->io_wq
;
8149 static int io_uring_alloc_task_context(struct task_struct
*task
)
8151 struct io_uring_task
*tctx
;
8154 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
8155 if (unlikely(!tctx
))
8158 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8159 if (unlikely(ret
)) {
8165 init_waitqueue_head(&tctx
->wait
);
8167 atomic_set(&tctx
->in_idle
, 0);
8168 tctx
->sqpoll
= false;
8169 io_init_identity(&tctx
->__identity
);
8170 tctx
->identity
= &tctx
->__identity
;
8171 task
->io_uring
= tctx
;
8172 spin_lock_init(&tctx
->task_lock
);
8173 INIT_WQ_LIST(&tctx
->task_list
);
8174 tctx
->task_state
= 0;
8175 init_task_work(&tctx
->task_work
, tctx_task_work
);
8179 void __io_uring_free(struct task_struct
*tsk
)
8181 struct io_uring_task
*tctx
= tsk
->io_uring
;
8183 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8184 WARN_ON_ONCE(refcount_read(&tctx
->identity
->count
) != 1);
8185 if (tctx
->identity
!= &tctx
->__identity
)
8186 kfree(tctx
->identity
);
8187 percpu_counter_destroy(&tctx
->inflight
);
8189 tsk
->io_uring
= NULL
;
8192 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8193 struct io_uring_params
*p
)
8197 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8198 struct io_sq_data
*sqd
;
8201 if (!capable(CAP_SYS_ADMIN
) && !capable(CAP_SYS_NICE
))
8204 sqd
= io_get_sq_data(p
);
8211 io_sq_thread_park(sqd
);
8212 mutex_lock(&sqd
->ctx_lock
);
8213 list_add(&ctx
->sqd_list
, &sqd
->ctx_new_list
);
8214 mutex_unlock(&sqd
->ctx_lock
);
8215 io_sq_thread_unpark(sqd
);
8217 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8218 if (!ctx
->sq_thread_idle
)
8219 ctx
->sq_thread_idle
= HZ
;
8224 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8225 int cpu
= p
->sq_thread_cpu
;
8228 if (cpu
>= nr_cpu_ids
)
8230 if (!cpu_online(cpu
))
8233 sqd
->thread
= kthread_create_on_cpu(io_sq_thread
, sqd
,
8234 cpu
, "io_uring-sq");
8236 sqd
->thread
= kthread_create(io_sq_thread
, sqd
,
8239 if (IS_ERR(sqd
->thread
)) {
8240 ret
= PTR_ERR(sqd
->thread
);
8244 ret
= io_uring_alloc_task_context(sqd
->thread
);
8247 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8248 /* Can't have SQ_AFF without SQPOLL */
8254 ret
= io_init_wq_offload(ctx
, p
);
8260 io_finish_async(ctx
);
8264 static void io_sq_offload_start(struct io_ring_ctx
*ctx
)
8266 struct io_sq_data
*sqd
= ctx
->sq_data
;
8268 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && sqd
->thread
)
8269 wake_up_process(sqd
->thread
);
8272 static inline void __io_unaccount_mem(struct user_struct
*user
,
8273 unsigned long nr_pages
)
8275 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8278 static inline int __io_account_mem(struct user_struct
*user
,
8279 unsigned long nr_pages
)
8281 unsigned long page_limit
, cur_pages
, new_pages
;
8283 /* Don't allow more pages than we can safely lock */
8284 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8287 cur_pages
= atomic_long_read(&user
->locked_vm
);
8288 new_pages
= cur_pages
+ nr_pages
;
8289 if (new_pages
> page_limit
)
8291 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8292 new_pages
) != cur_pages
);
8297 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8300 __io_unaccount_mem(ctx
->user
, nr_pages
);
8302 if (ctx
->mm_account
)
8303 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8306 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8310 if (ctx
->limit_mem
) {
8311 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8316 if (ctx
->mm_account
)
8317 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8322 static void io_mem_free(void *ptr
)
8329 page
= virt_to_head_page(ptr
);
8330 if (put_page_testzero(page
))
8331 free_compound_page(page
);
8334 static void *io_mem_alloc(size_t size
)
8336 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8337 __GFP_NORETRY
| __GFP_ACCOUNT
;
8339 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8342 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8345 struct io_rings
*rings
;
8346 size_t off
, sq_array_size
;
8348 off
= struct_size(rings
, cqes
, cq_entries
);
8349 if (off
== SIZE_MAX
)
8353 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8361 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8362 if (sq_array_size
== SIZE_MAX
)
8365 if (check_add_overflow(off
, sq_array_size
, &off
))
8371 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8375 if (!ctx
->user_bufs
)
8378 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8379 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8381 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
8382 unpin_user_page(imu
->bvec
[j
].bv_page
);
8384 if (imu
->acct_pages
)
8385 io_unaccount_mem(ctx
, imu
->acct_pages
);
8390 kfree(ctx
->user_bufs
);
8391 ctx
->user_bufs
= NULL
;
8392 ctx
->nr_user_bufs
= 0;
8396 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8397 void __user
*arg
, unsigned index
)
8399 struct iovec __user
*src
;
8401 #ifdef CONFIG_COMPAT
8403 struct compat_iovec __user
*ciovs
;
8404 struct compat_iovec ciov
;
8406 ciovs
= (struct compat_iovec __user
*) arg
;
8407 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8410 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8411 dst
->iov_len
= ciov
.iov_len
;
8415 src
= (struct iovec __user
*) arg
;
8416 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8422 * Not super efficient, but this is just a registration time. And we do cache
8423 * the last compound head, so generally we'll only do a full search if we don't
8426 * We check if the given compound head page has already been accounted, to
8427 * avoid double accounting it. This allows us to account the full size of the
8428 * page, not just the constituent pages of a huge page.
8430 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8431 int nr_pages
, struct page
*hpage
)
8435 /* check current page array */
8436 for (i
= 0; i
< nr_pages
; i
++) {
8437 if (!PageCompound(pages
[i
]))
8439 if (compound_head(pages
[i
]) == hpage
)
8443 /* check previously registered pages */
8444 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8445 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8447 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8448 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8450 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8458 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8459 int nr_pages
, struct io_mapped_ubuf
*imu
,
8460 struct page
**last_hpage
)
8464 for (i
= 0; i
< nr_pages
; i
++) {
8465 if (!PageCompound(pages
[i
])) {
8470 hpage
= compound_head(pages
[i
]);
8471 if (hpage
== *last_hpage
)
8473 *last_hpage
= hpage
;
8474 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8476 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8480 if (!imu
->acct_pages
)
8483 ret
= io_account_mem(ctx
, imu
->acct_pages
);
8485 imu
->acct_pages
= 0;
8489 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
8490 struct io_mapped_ubuf
*imu
,
8491 struct page
**last_hpage
)
8493 struct vm_area_struct
**vmas
= NULL
;
8494 struct page
**pages
= NULL
;
8495 unsigned long off
, start
, end
, ubuf
;
8497 int ret
, pret
, nr_pages
, i
;
8499 ubuf
= (unsigned long) iov
->iov_base
;
8500 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8501 start
= ubuf
>> PAGE_SHIFT
;
8502 nr_pages
= end
- start
;
8506 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
8510 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
8515 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8521 mmap_read_lock(current
->mm
);
8522 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
8524 if (pret
== nr_pages
) {
8525 /* don't support file backed memory */
8526 for (i
= 0; i
< nr_pages
; i
++) {
8527 struct vm_area_struct
*vma
= vmas
[i
];
8530 !is_file_hugepages(vma
->vm_file
)) {
8536 ret
= pret
< 0 ? pret
: -EFAULT
;
8538 mmap_read_unlock(current
->mm
);
8541 * if we did partial map, or found file backed vmas,
8542 * release any pages we did get
8545 unpin_user_pages(pages
, pret
);
8550 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
8552 unpin_user_pages(pages
, pret
);
8557 off
= ubuf
& ~PAGE_MASK
;
8558 size
= iov
->iov_len
;
8559 for (i
= 0; i
< nr_pages
; i
++) {
8562 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8563 imu
->bvec
[i
].bv_page
= pages
[i
];
8564 imu
->bvec
[i
].bv_len
= vec_len
;
8565 imu
->bvec
[i
].bv_offset
= off
;
8569 /* store original address for later verification */
8571 imu
->len
= iov
->iov_len
;
8572 imu
->nr_bvecs
= nr_pages
;
8580 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
8584 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8587 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8589 if (!ctx
->user_bufs
)
8595 static int io_buffer_validate(struct iovec
*iov
)
8598 * Don't impose further limits on the size and buffer
8599 * constraints here, we'll -EINVAL later when IO is
8600 * submitted if they are wrong.
8602 if (!iov
->iov_base
|| !iov
->iov_len
)
8605 /* arbitrary limit, but we need something */
8606 if (iov
->iov_len
> SZ_1G
)
8612 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8613 unsigned int nr_args
)
8617 struct page
*last_hpage
= NULL
;
8619 ret
= io_buffers_map_alloc(ctx
, nr_args
);
8623 for (i
= 0; i
< nr_args
; i
++) {
8624 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8626 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8630 ret
= io_buffer_validate(&iov
);
8634 ret
= io_sqe_buffer_register(ctx
, &iov
, imu
, &last_hpage
);
8638 ctx
->nr_user_bufs
++;
8642 io_sqe_buffers_unregister(ctx
);
8647 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8649 __s32 __user
*fds
= arg
;
8655 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8658 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8659 if (IS_ERR(ctx
->cq_ev_fd
)) {
8660 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8661 ctx
->cq_ev_fd
= NULL
;
8668 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8670 if (ctx
->cq_ev_fd
) {
8671 eventfd_ctx_put(ctx
->cq_ev_fd
);
8672 ctx
->cq_ev_fd
= NULL
;
8679 static int __io_destroy_buffers(int id
, void *p
, void *data
)
8681 struct io_ring_ctx
*ctx
= data
;
8682 struct io_buffer
*buf
= p
;
8684 __io_remove_buffers(ctx
, buf
, id
, -1U);
8688 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8690 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
8691 idr_destroy(&ctx
->io_buffer_idr
);
8694 static void io_req_cache_free(struct list_head
*list
, struct task_struct
*tsk
)
8696 struct io_kiocb
*req
, *nxt
;
8698 list_for_each_entry_safe(req
, nxt
, list
, compl.list
) {
8699 if (tsk
&& req
->task
!= tsk
)
8701 list_del(&req
->compl.list
);
8702 kmem_cache_free(req_cachep
, req
);
8706 static void io_req_caches_free(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
)
8708 struct io_submit_state
*submit_state
= &ctx
->submit_state
;
8710 mutex_lock(&ctx
->uring_lock
);
8712 if (submit_state
->free_reqs
)
8713 kmem_cache_free_bulk(req_cachep
, submit_state
->free_reqs
,
8714 submit_state
->reqs
);
8716 io_req_cache_free(&submit_state
->comp
.free_list
, NULL
);
8718 spin_lock_irq(&ctx
->completion_lock
);
8719 io_req_cache_free(&submit_state
->comp
.locked_free_list
, NULL
);
8720 spin_unlock_irq(&ctx
->completion_lock
);
8722 mutex_unlock(&ctx
->uring_lock
);
8725 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8728 * Some may use context even when all refs and requests have been put,
8729 * and they are free to do so while still holding uring_lock, see
8730 * __io_req_task_submit(). Wait for them to finish.
8732 mutex_lock(&ctx
->uring_lock
);
8733 mutex_unlock(&ctx
->uring_lock
);
8735 io_finish_async(ctx
);
8736 io_sqe_buffers_unregister(ctx
);
8738 if (ctx
->sqo_task
) {
8739 put_task_struct(ctx
->sqo_task
);
8740 ctx
->sqo_task
= NULL
;
8741 mmdrop(ctx
->mm_account
);
8742 ctx
->mm_account
= NULL
;
8745 #ifdef CONFIG_BLK_CGROUP
8746 if (ctx
->sqo_blkcg_css
)
8747 css_put(ctx
->sqo_blkcg_css
);
8750 mutex_lock(&ctx
->uring_lock
);
8751 io_sqe_files_unregister(ctx
);
8752 mutex_unlock(&ctx
->uring_lock
);
8753 io_eventfd_unregister(ctx
);
8754 io_destroy_buffers(ctx
);
8755 idr_destroy(&ctx
->personality_idr
);
8757 #if defined(CONFIG_UNIX)
8758 if (ctx
->ring_sock
) {
8759 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8760 sock_release(ctx
->ring_sock
);
8764 io_mem_free(ctx
->rings
);
8765 io_mem_free(ctx
->sq_sqes
);
8767 percpu_ref_exit(&ctx
->refs
);
8768 free_uid(ctx
->user
);
8769 put_cred(ctx
->creds
);
8770 io_req_caches_free(ctx
, NULL
);
8771 kfree(ctx
->cancel_hash
);
8775 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8777 struct io_ring_ctx
*ctx
= file
->private_data
;
8780 poll_wait(file
, &ctx
->cq_wait
, wait
);
8782 * synchronizes with barrier from wq_has_sleeper call in
8786 if (!io_sqring_full(ctx
))
8787 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8790 * Don't flush cqring overflow list here, just do a simple check.
8791 * Otherwise there could possible be ABBA deadlock:
8794 * lock(&ctx->uring_lock);
8796 * lock(&ctx->uring_lock);
8799 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8800 * pushs them to do the flush.
8802 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->cq_check_overflow
))
8803 mask
|= EPOLLIN
| EPOLLRDNORM
;
8808 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8810 struct io_ring_ctx
*ctx
= file
->private_data
;
8812 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8815 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8817 struct io_identity
*iod
;
8819 iod
= idr_remove(&ctx
->personality_idr
, id
);
8821 put_cred(iod
->creds
);
8822 if (refcount_dec_and_test(&iod
->count
))
8830 static int io_remove_personalities(int id
, void *p
, void *data
)
8832 struct io_ring_ctx
*ctx
= data
;
8834 io_unregister_personality(ctx
, id
);
8838 static void io_ring_exit_work(struct work_struct
*work
)
8840 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
8844 * If we're doing polled IO and end up having requests being
8845 * submitted async (out-of-line), then completions can come in while
8846 * we're waiting for refs to drop. We need to reap these manually,
8847 * as nobody else will be looking for them.
8850 io_uring_try_cancel_requests(ctx
, NULL
, NULL
);
8851 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8852 io_ring_ctx_free(ctx
);
8855 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8857 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8859 return req
->ctx
== data
;
8862 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8864 mutex_lock(&ctx
->uring_lock
);
8865 percpu_ref_kill(&ctx
->refs
);
8867 if (WARN_ON_ONCE((ctx
->flags
& IORING_SETUP_SQPOLL
) && !ctx
->sqo_dead
))
8870 /* if force is set, the ring is going away. always drop after that */
8871 ctx
->cq_overflow_flushed
= 1;
8873 __io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8874 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
8875 mutex_unlock(&ctx
->uring_lock
);
8877 io_kill_timeouts(ctx
, NULL
, NULL
);
8878 io_poll_remove_all(ctx
, NULL
, NULL
);
8881 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8883 /* if we failed setting up the ctx, we might not have any rings */
8884 io_iopoll_try_reap_events(ctx
);
8886 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8888 * Use system_unbound_wq to avoid spawning tons of event kworkers
8889 * if we're exiting a ton of rings at the same time. It just adds
8890 * noise and overhead, there's no discernable change in runtime
8891 * over using system_wq.
8893 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8896 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8898 struct io_ring_ctx
*ctx
= file
->private_data
;
8900 file
->private_data
= NULL
;
8901 io_ring_ctx_wait_and_kill(ctx
);
8905 struct io_task_cancel
{
8906 struct task_struct
*task
;
8907 struct files_struct
*files
;
8910 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8912 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8913 struct io_task_cancel
*cancel
= data
;
8916 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8917 unsigned long flags
;
8918 struct io_ring_ctx
*ctx
= req
->ctx
;
8920 /* protect against races with linked timeouts */
8921 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8922 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8923 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8925 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8930 static void io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8931 struct task_struct
*task
,
8932 struct files_struct
*files
)
8934 struct io_defer_entry
*de
= NULL
;
8937 spin_lock_irq(&ctx
->completion_lock
);
8938 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8939 if (io_match_task(de
->req
, task
, files
)) {
8940 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8944 spin_unlock_irq(&ctx
->completion_lock
);
8946 while (!list_empty(&list
)) {
8947 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8948 list_del_init(&de
->list
);
8949 req_set_fail_links(de
->req
);
8950 io_put_req(de
->req
);
8951 io_req_complete(de
->req
, -ECANCELED
);
8956 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
8957 struct task_struct
*task
,
8958 struct files_struct
*files
)
8960 struct io_task_cancel cancel
= { .task
= task
, .files
= files
, };
8963 enum io_wq_cancel cret
;
8967 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
,
8969 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8972 /* SQPOLL thread does its own polling */
8973 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) && !files
) {
8974 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8975 io_iopoll_try_reap_events(ctx
);
8980 ret
|= io_poll_remove_all(ctx
, task
, files
);
8981 ret
|= io_kill_timeouts(ctx
, task
, files
);
8982 ret
|= io_run_task_work();
8983 io_cqring_overflow_flush(ctx
, true, task
, files
);
8990 static int io_uring_count_inflight(struct io_ring_ctx
*ctx
,
8991 struct task_struct
*task
,
8992 struct files_struct
*files
)
8994 struct io_kiocb
*req
;
8997 spin_lock_irq(&ctx
->inflight_lock
);
8998 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
)
8999 cnt
+= io_match_task(req
, task
, files
);
9000 spin_unlock_irq(&ctx
->inflight_lock
);
9004 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
9005 struct task_struct
*task
,
9006 struct files_struct
*files
)
9008 while (!list_empty_careful(&ctx
->inflight_list
)) {
9012 inflight
= io_uring_count_inflight(ctx
, task
, files
);
9016 io_uring_try_cancel_requests(ctx
, task
, files
);
9019 io_sq_thread_unpark(ctx
->sq_data
);
9020 prepare_to_wait(&task
->io_uring
->wait
, &wait
,
9021 TASK_UNINTERRUPTIBLE
);
9022 if (inflight
== io_uring_count_inflight(ctx
, task
, files
))
9024 finish_wait(&task
->io_uring
->wait
, &wait
);
9026 io_sq_thread_park(ctx
->sq_data
);
9030 static void io_disable_sqo_submit(struct io_ring_ctx
*ctx
)
9032 mutex_lock(&ctx
->uring_lock
);
9034 mutex_unlock(&ctx
->uring_lock
);
9036 /* make sure callers enter the ring to get error */
9038 io_ring_set_wakeup_flag(ctx
);
9042 * We need to iteratively cancel requests, in case a request has dependent
9043 * hard links. These persist even for failure of cancelations, hence keep
9044 * looping until none are found.
9046 static void io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
9047 struct files_struct
*files
)
9049 struct task_struct
*task
= current
;
9051 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
9052 io_disable_sqo_submit(ctx
);
9053 task
= ctx
->sq_data
->thread
;
9054 atomic_inc(&task
->io_uring
->in_idle
);
9055 io_sq_thread_park(ctx
->sq_data
);
9058 io_cancel_defer_files(ctx
, task
, files
);
9060 io_uring_cancel_files(ctx
, task
, files
);
9062 io_uring_try_cancel_requests(ctx
, task
, NULL
);
9064 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
9065 atomic_dec(&task
->io_uring
->in_idle
);
9066 io_sq_thread_unpark(ctx
->sq_data
);
9071 * Note that this task has used io_uring. We use it for cancelation purposes.
9073 static int io_uring_add_task_file(struct io_ring_ctx
*ctx
, struct file
*file
)
9075 struct io_uring_task
*tctx
= current
->io_uring
;
9078 if (unlikely(!tctx
)) {
9079 ret
= io_uring_alloc_task_context(current
);
9082 tctx
= current
->io_uring
;
9084 if (tctx
->last
!= file
) {
9085 void *old
= xa_load(&tctx
->xa
, (unsigned long)file
);
9089 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)file
,
9096 /* one and only SQPOLL file note, held by sqo_task */
9097 WARN_ON_ONCE((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
9098 current
!= ctx
->sqo_task
);
9104 * This is race safe in that the task itself is doing this, hence it
9105 * cannot be going through the exit/cancel paths at the same time.
9106 * This cannot be modified while exit/cancel is running.
9108 if (!tctx
->sqpoll
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9109 tctx
->sqpoll
= true;
9115 * Remove this io_uring_file -> task mapping.
9117 static void io_uring_del_task_file(struct file
*file
)
9119 struct io_uring_task
*tctx
= current
->io_uring
;
9121 if (tctx
->last
== file
)
9123 file
= xa_erase(&tctx
->xa
, (unsigned long)file
);
9128 static void io_uring_remove_task_files(struct io_uring_task
*tctx
)
9131 unsigned long index
;
9133 xa_for_each(&tctx
->xa
, index
, file
)
9134 io_uring_del_task_file(file
);
9137 void __io_uring_files_cancel(struct files_struct
*files
)
9139 struct io_uring_task
*tctx
= current
->io_uring
;
9141 unsigned long index
;
9143 /* make sure overflow events are dropped */
9144 atomic_inc(&tctx
->in_idle
);
9145 xa_for_each(&tctx
->xa
, index
, file
)
9146 io_uring_cancel_task_requests(file
->private_data
, files
);
9147 atomic_dec(&tctx
->in_idle
);
9150 io_uring_remove_task_files(tctx
);
9153 static s64
tctx_inflight(struct io_uring_task
*tctx
)
9155 return percpu_counter_sum(&tctx
->inflight
);
9158 static void io_uring_cancel_sqpoll(struct io_ring_ctx
*ctx
)
9160 struct io_uring_task
*tctx
;
9166 tctx
= ctx
->sq_data
->thread
->io_uring
;
9167 io_disable_sqo_submit(ctx
);
9169 atomic_inc(&tctx
->in_idle
);
9171 /* read completions before cancelations */
9172 inflight
= tctx_inflight(tctx
);
9175 io_uring_cancel_task_requests(ctx
, NULL
);
9177 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9179 * If we've seen completions, retry without waiting. This
9180 * avoids a race where a completion comes in before we did
9181 * prepare_to_wait().
9183 if (inflight
== tctx_inflight(tctx
))
9185 finish_wait(&tctx
->wait
, &wait
);
9187 atomic_dec(&tctx
->in_idle
);
9191 * Find any io_uring fd that this task has registered or done IO on, and cancel
9194 void __io_uring_task_cancel(void)
9196 struct io_uring_task
*tctx
= current
->io_uring
;
9200 /* make sure overflow events are dropped */
9201 atomic_inc(&tctx
->in_idle
);
9203 /* trigger io_disable_sqo_submit() */
9206 unsigned long index
;
9208 xa_for_each(&tctx
->xa
, index
, file
)
9209 io_uring_cancel_sqpoll(file
->private_data
);
9213 /* read completions before cancelations */
9214 inflight
= tctx_inflight(tctx
);
9217 __io_uring_files_cancel(NULL
);
9219 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9222 * If we've seen completions, retry without waiting. This
9223 * avoids a race where a completion comes in before we did
9224 * prepare_to_wait().
9226 if (inflight
== tctx_inflight(tctx
))
9228 finish_wait(&tctx
->wait
, &wait
);
9231 atomic_dec(&tctx
->in_idle
);
9233 io_uring_remove_task_files(tctx
);
9236 static int io_uring_flush(struct file
*file
, void *data
)
9238 struct io_uring_task
*tctx
= current
->io_uring
;
9239 struct io_ring_ctx
*ctx
= file
->private_data
;
9241 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
)) {
9242 io_uring_cancel_task_requests(ctx
, NULL
);
9243 io_req_caches_free(ctx
, current
);
9249 /* we should have cancelled and erased it before PF_EXITING */
9250 WARN_ON_ONCE((current
->flags
& PF_EXITING
) &&
9251 xa_load(&tctx
->xa
, (unsigned long)file
));
9254 * fput() is pending, will be 2 if the only other ref is our potential
9255 * task file note. If the task is exiting, drop regardless of count.
9257 if (atomic_long_read(&file
->f_count
) != 2)
9260 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9261 /* there is only one file note, which is owned by sqo_task */
9262 WARN_ON_ONCE(ctx
->sqo_task
!= current
&&
9263 xa_load(&tctx
->xa
, (unsigned long)file
));
9264 /* sqo_dead check is for when this happens after cancellation */
9265 WARN_ON_ONCE(ctx
->sqo_task
== current
&& !ctx
->sqo_dead
&&
9266 !xa_load(&tctx
->xa
, (unsigned long)file
));
9268 io_disable_sqo_submit(ctx
);
9271 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) || ctx
->sqo_task
== current
)
9272 io_uring_del_task_file(file
);
9276 static void *io_uring_validate_mmap_request(struct file
*file
,
9277 loff_t pgoff
, size_t sz
)
9279 struct io_ring_ctx
*ctx
= file
->private_data
;
9280 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9285 case IORING_OFF_SQ_RING
:
9286 case IORING_OFF_CQ_RING
:
9289 case IORING_OFF_SQES
:
9293 return ERR_PTR(-EINVAL
);
9296 page
= virt_to_head_page(ptr
);
9297 if (sz
> page_size(page
))
9298 return ERR_PTR(-EINVAL
);
9305 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9307 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9311 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9313 return PTR_ERR(ptr
);
9315 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9316 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9319 #else /* !CONFIG_MMU */
9321 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9323 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9326 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9328 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9331 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9332 unsigned long addr
, unsigned long len
,
9333 unsigned long pgoff
, unsigned long flags
)
9337 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9339 return PTR_ERR(ptr
);
9341 return (unsigned long) ptr
;
9344 #endif /* !CONFIG_MMU */
9346 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9352 if (!io_sqring_full(ctx
))
9355 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9357 if (unlikely(ctx
->sqo_dead
)) {
9362 if (!io_sqring_full(ctx
))
9366 } while (!signal_pending(current
));
9368 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9373 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9374 struct __kernel_timespec __user
**ts
,
9375 const sigset_t __user
**sig
)
9377 struct io_uring_getevents_arg arg
;
9380 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9381 * is just a pointer to the sigset_t.
9383 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9384 *sig
= (const sigset_t __user
*) argp
;
9390 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9391 * timespec and sigset_t pointers if good.
9393 if (*argsz
!= sizeof(arg
))
9395 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9397 *sig
= u64_to_user_ptr(arg
.sigmask
);
9398 *argsz
= arg
.sigmask_sz
;
9399 *ts
= u64_to_user_ptr(arg
.ts
);
9403 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9404 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9407 struct io_ring_ctx
*ctx
;
9414 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9415 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
))
9423 if (f
.file
->f_op
!= &io_uring_fops
)
9427 ctx
= f
.file
->private_data
;
9428 if (!percpu_ref_tryget(&ctx
->refs
))
9432 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9436 * For SQ polling, the thread will do all submissions and completions.
9437 * Just return the requested submit count, and wake the thread if
9441 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9442 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
9445 if (unlikely(ctx
->sqo_dead
))
9447 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9448 wake_up(&ctx
->sq_data
->wait
);
9449 if (flags
& IORING_ENTER_SQ_WAIT
) {
9450 ret
= io_sqpoll_wait_sq(ctx
);
9454 submitted
= to_submit
;
9455 } else if (to_submit
) {
9456 ret
= io_uring_add_task_file(ctx
, f
.file
);
9459 mutex_lock(&ctx
->uring_lock
);
9460 submitted
= io_submit_sqes(ctx
, to_submit
);
9461 mutex_unlock(&ctx
->uring_lock
);
9463 if (submitted
!= to_submit
)
9466 if (flags
& IORING_ENTER_GETEVENTS
) {
9467 const sigset_t __user
*sig
;
9468 struct __kernel_timespec __user
*ts
;
9470 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9474 min_complete
= min(min_complete
, ctx
->cq_entries
);
9477 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9478 * space applications don't need to do io completion events
9479 * polling again, they can rely on io_sq_thread to do polling
9480 * work, which can reduce cpu usage and uring_lock contention.
9482 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9483 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9484 ret
= io_iopoll_check(ctx
, min_complete
);
9486 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9491 percpu_ref_put(&ctx
->refs
);
9494 return submitted
? submitted
: ret
;
9497 #ifdef CONFIG_PROC_FS
9498 static int io_uring_show_cred(int id
, void *p
, void *data
)
9500 struct io_identity
*iod
= p
;
9501 const struct cred
*cred
= iod
->creds
;
9502 struct seq_file
*m
= data
;
9503 struct user_namespace
*uns
= seq_user_ns(m
);
9504 struct group_info
*gi
;
9509 seq_printf(m
, "%5d\n", id
);
9510 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9511 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9512 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9513 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9514 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9515 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9516 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9517 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9518 seq_puts(m
, "\n\tGroups:\t");
9519 gi
= cred
->group_info
;
9520 for (g
= 0; g
< gi
->ngroups
; g
++) {
9521 seq_put_decimal_ull(m
, g
? " " : "",
9522 from_kgid_munged(uns
, gi
->gid
[g
]));
9524 seq_puts(m
, "\n\tCapEff:\t");
9525 cap
= cred
->cap_effective
;
9526 CAP_FOR_EACH_U32(__capi
)
9527 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9532 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9534 struct io_sq_data
*sq
= NULL
;
9539 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9540 * since fdinfo case grabs it in the opposite direction of normal use
9541 * cases. If we fail to get the lock, we just don't iterate any
9542 * structures that could be going away outside the io_uring mutex.
9544 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9546 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9549 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9550 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9551 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9552 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9553 struct file
*f
= *io_fixed_file_slot(ctx
->file_data
, i
);
9556 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9558 seq_printf(m
, "%5u: <none>\n", i
);
9560 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9561 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9562 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9564 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9565 (unsigned int) buf
->len
);
9567 if (has_lock
&& !idr_is_empty(&ctx
->personality_idr
)) {
9568 seq_printf(m
, "Personalities:\n");
9569 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
9571 seq_printf(m
, "PollList:\n");
9572 spin_lock_irq(&ctx
->completion_lock
);
9573 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9574 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9575 struct io_kiocb
*req
;
9577 hlist_for_each_entry(req
, list
, hash_node
)
9578 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9579 req
->task
->task_works
!= NULL
);
9581 spin_unlock_irq(&ctx
->completion_lock
);
9583 mutex_unlock(&ctx
->uring_lock
);
9586 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9588 struct io_ring_ctx
*ctx
= f
->private_data
;
9590 if (percpu_ref_tryget(&ctx
->refs
)) {
9591 __io_uring_show_fdinfo(ctx
, m
);
9592 percpu_ref_put(&ctx
->refs
);
9597 static const struct file_operations io_uring_fops
= {
9598 .release
= io_uring_release
,
9599 .flush
= io_uring_flush
,
9600 .mmap
= io_uring_mmap
,
9602 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9603 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9605 .poll
= io_uring_poll
,
9606 .fasync
= io_uring_fasync
,
9607 #ifdef CONFIG_PROC_FS
9608 .show_fdinfo
= io_uring_show_fdinfo
,
9612 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9613 struct io_uring_params
*p
)
9615 struct io_rings
*rings
;
9616 size_t size
, sq_array_offset
;
9618 /* make sure these are sane, as we already accounted them */
9619 ctx
->sq_entries
= p
->sq_entries
;
9620 ctx
->cq_entries
= p
->cq_entries
;
9622 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9623 if (size
== SIZE_MAX
)
9626 rings
= io_mem_alloc(size
);
9631 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9632 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9633 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9634 rings
->sq_ring_entries
= p
->sq_entries
;
9635 rings
->cq_ring_entries
= p
->cq_entries
;
9636 ctx
->sq_mask
= rings
->sq_ring_mask
;
9637 ctx
->cq_mask
= rings
->cq_ring_mask
;
9639 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9640 if (size
== SIZE_MAX
) {
9641 io_mem_free(ctx
->rings
);
9646 ctx
->sq_sqes
= io_mem_alloc(size
);
9647 if (!ctx
->sq_sqes
) {
9648 io_mem_free(ctx
->rings
);
9656 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9660 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9664 ret
= io_uring_add_task_file(ctx
, file
);
9669 fd_install(fd
, file
);
9674 * Allocate an anonymous fd, this is what constitutes the application
9675 * visible backing of an io_uring instance. The application mmaps this
9676 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9677 * we have to tie this fd to a socket for file garbage collection purposes.
9679 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9682 #if defined(CONFIG_UNIX)
9685 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9688 return ERR_PTR(ret
);
9691 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9692 O_RDWR
| O_CLOEXEC
);
9693 #if defined(CONFIG_UNIX)
9695 sock_release(ctx
->ring_sock
);
9696 ctx
->ring_sock
= NULL
;
9698 ctx
->ring_sock
->file
= file
;
9704 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9705 struct io_uring_params __user
*params
)
9707 struct user_struct
*user
= NULL
;
9708 struct io_ring_ctx
*ctx
;
9714 if (entries
> IORING_MAX_ENTRIES
) {
9715 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9717 entries
= IORING_MAX_ENTRIES
;
9721 * Use twice as many entries for the CQ ring. It's possible for the
9722 * application to drive a higher depth than the size of the SQ ring,
9723 * since the sqes are only used at submission time. This allows for
9724 * some flexibility in overcommitting a bit. If the application has
9725 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9726 * of CQ ring entries manually.
9728 p
->sq_entries
= roundup_pow_of_two(entries
);
9729 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9731 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9732 * to a power-of-two, if it isn't already. We do NOT impose
9733 * any cq vs sq ring sizing.
9737 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9738 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9740 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9742 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9743 if (p
->cq_entries
< p
->sq_entries
)
9746 p
->cq_entries
= 2 * p
->sq_entries
;
9749 user
= get_uid(current_user());
9751 ctx
= io_ring_ctx_alloc(p
);
9756 ctx
->compat
= in_compat_syscall();
9757 ctx
->limit_mem
= !capable(CAP_IPC_LOCK
);
9759 ctx
->creds
= get_current_cred();
9761 ctx
->loginuid
= current
->loginuid
;
9762 ctx
->sessionid
= current
->sessionid
;
9764 ctx
->sqo_task
= get_task_struct(current
);
9767 * This is just grabbed for accounting purposes. When a process exits,
9768 * the mm is exited and dropped before the files, hence we need to hang
9769 * on to this mm purely for the purposes of being able to unaccount
9770 * memory (locked/pinned vm). It's not used for anything else.
9772 mmgrab(current
->mm
);
9773 ctx
->mm_account
= current
->mm
;
9775 #ifdef CONFIG_BLK_CGROUP
9777 * The sq thread will belong to the original cgroup it was inited in.
9778 * If the cgroup goes offline (e.g. disabling the io controller), then
9779 * issued bios will be associated with the closest cgroup later in the
9783 ctx
->sqo_blkcg_css
= blkcg_css();
9784 ret
= css_tryget_online(ctx
->sqo_blkcg_css
);
9787 /* don't init against a dying cgroup, have the user try again */
9788 ctx
->sqo_blkcg_css
= NULL
;
9793 ret
= io_allocate_scq_urings(ctx
, p
);
9797 ret
= io_sq_offload_create(ctx
, p
);
9801 if (!(p
->flags
& IORING_SETUP_R_DISABLED
))
9802 io_sq_offload_start(ctx
);
9804 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9805 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9806 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9807 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9808 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9809 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9810 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9811 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9813 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9814 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9815 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9816 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9817 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9818 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9819 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9820 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9822 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9823 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9824 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9825 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9826 IORING_FEAT_EXT_ARG
;
9828 if (copy_to_user(params
, p
, sizeof(*p
))) {
9833 file
= io_uring_get_file(ctx
);
9835 ret
= PTR_ERR(file
);
9840 * Install ring fd as the very last thing, so we don't risk someone
9841 * having closed it before we finish setup
9843 ret
= io_uring_install_fd(ctx
, file
);
9845 io_disable_sqo_submit(ctx
);
9846 /* fput will clean it up */
9851 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9854 io_disable_sqo_submit(ctx
);
9855 io_ring_ctx_wait_and_kill(ctx
);
9860 * Sets up an aio uring context, and returns the fd. Applications asks for a
9861 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9862 * params structure passed in.
9864 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9866 struct io_uring_params p
;
9869 if (copy_from_user(&p
, params
, sizeof(p
)))
9871 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9876 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9877 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9878 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9879 IORING_SETUP_R_DISABLED
))
9882 return io_uring_create(entries
, &p
, params
);
9885 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9886 struct io_uring_params __user
*, params
)
9888 return io_uring_setup(entries
, params
);
9891 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9893 struct io_uring_probe
*p
;
9897 size
= struct_size(p
, ops
, nr_args
);
9898 if (size
== SIZE_MAX
)
9900 p
= kzalloc(size
, GFP_KERNEL
);
9905 if (copy_from_user(p
, arg
, size
))
9908 if (memchr_inv(p
, 0, size
))
9911 p
->last_op
= IORING_OP_LAST
- 1;
9912 if (nr_args
> IORING_OP_LAST
)
9913 nr_args
= IORING_OP_LAST
;
9915 for (i
= 0; i
< nr_args
; i
++) {
9917 if (!io_op_defs
[i
].not_supported
)
9918 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9923 if (copy_to_user(arg
, p
, size
))
9930 static int io_register_personality(struct io_ring_ctx
*ctx
)
9932 struct io_identity
*id
;
9935 id
= kmalloc(sizeof(*id
), GFP_KERNEL
);
9939 io_init_identity(id
);
9940 id
->creds
= get_current_cred();
9942 ret
= idr_alloc_cyclic(&ctx
->personality_idr
, id
, 1, USHRT_MAX
, GFP_KERNEL
);
9944 put_cred(id
->creds
);
9950 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9951 unsigned int nr_args
)
9953 struct io_uring_restriction
*res
;
9957 /* Restrictions allowed only if rings started disabled */
9958 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9961 /* We allow only a single restrictions registration */
9962 if (ctx
->restrictions
.registered
)
9965 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9968 size
= array_size(nr_args
, sizeof(*res
));
9969 if (size
== SIZE_MAX
)
9972 res
= memdup_user(arg
, size
);
9974 return PTR_ERR(res
);
9978 for (i
= 0; i
< nr_args
; i
++) {
9979 switch (res
[i
].opcode
) {
9980 case IORING_RESTRICTION_REGISTER_OP
:
9981 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9986 __set_bit(res
[i
].register_op
,
9987 ctx
->restrictions
.register_op
);
9989 case IORING_RESTRICTION_SQE_OP
:
9990 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9995 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9997 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9998 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
10000 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
10001 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
10010 /* Reset all restrictions if an error happened */
10012 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
10014 ctx
->restrictions
.registered
= true;
10020 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10022 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10025 if (ctx
->restrictions
.registered
)
10026 ctx
->restricted
= 1;
10028 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
10030 io_sq_offload_start(ctx
);
10035 static bool io_register_op_must_quiesce(int op
)
10038 case IORING_UNREGISTER_FILES
:
10039 case IORING_REGISTER_FILES_UPDATE
:
10040 case IORING_REGISTER_PROBE
:
10041 case IORING_REGISTER_PERSONALITY
:
10042 case IORING_UNREGISTER_PERSONALITY
:
10049 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10050 void __user
*arg
, unsigned nr_args
)
10051 __releases(ctx
->uring_lock
)
10052 __acquires(ctx
->uring_lock
)
10057 * We're inside the ring mutex, if the ref is already dying, then
10058 * someone else killed the ctx or is already going through
10059 * io_uring_register().
10061 if (percpu_ref_is_dying(&ctx
->refs
))
10064 if (io_register_op_must_quiesce(opcode
)) {
10065 percpu_ref_kill(&ctx
->refs
);
10068 * Drop uring mutex before waiting for references to exit. If
10069 * another thread is currently inside io_uring_enter() it might
10070 * need to grab the uring_lock to make progress. If we hold it
10071 * here across the drain wait, then we can deadlock. It's safe
10072 * to drop the mutex here, since no new references will come in
10073 * after we've killed the percpu ref.
10075 mutex_unlock(&ctx
->uring_lock
);
10077 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10080 ret
= io_run_task_work_sig();
10085 mutex_lock(&ctx
->uring_lock
);
10087 if (ret
&& io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
))
10091 if (ctx
->restricted
) {
10092 if (opcode
>= IORING_REGISTER_LAST
) {
10097 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
10104 case IORING_REGISTER_BUFFERS
:
10105 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
);
10107 case IORING_UNREGISTER_BUFFERS
:
10109 if (arg
|| nr_args
)
10111 ret
= io_sqe_buffers_unregister(ctx
);
10113 case IORING_REGISTER_FILES
:
10114 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
10116 case IORING_UNREGISTER_FILES
:
10118 if (arg
|| nr_args
)
10120 ret
= io_sqe_files_unregister(ctx
);
10122 case IORING_REGISTER_FILES_UPDATE
:
10123 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
10125 case IORING_REGISTER_EVENTFD
:
10126 case IORING_REGISTER_EVENTFD_ASYNC
:
10130 ret
= io_eventfd_register(ctx
, arg
);
10133 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10134 ctx
->eventfd_async
= 1;
10136 ctx
->eventfd_async
= 0;
10138 case IORING_UNREGISTER_EVENTFD
:
10140 if (arg
|| nr_args
)
10142 ret
= io_eventfd_unregister(ctx
);
10144 case IORING_REGISTER_PROBE
:
10146 if (!arg
|| nr_args
> 256)
10148 ret
= io_probe(ctx
, arg
, nr_args
);
10150 case IORING_REGISTER_PERSONALITY
:
10152 if (arg
|| nr_args
)
10154 ret
= io_register_personality(ctx
);
10156 case IORING_UNREGISTER_PERSONALITY
:
10160 ret
= io_unregister_personality(ctx
, nr_args
);
10162 case IORING_REGISTER_ENABLE_RINGS
:
10164 if (arg
|| nr_args
)
10166 ret
= io_register_enable_rings(ctx
);
10168 case IORING_REGISTER_RESTRICTIONS
:
10169 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10177 if (io_register_op_must_quiesce(opcode
)) {
10178 /* bring the ctx back to life */
10179 percpu_ref_reinit(&ctx
->refs
);
10180 reinit_completion(&ctx
->ref_comp
);
10185 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10186 void __user
*, arg
, unsigned int, nr_args
)
10188 struct io_ring_ctx
*ctx
;
10197 if (f
.file
->f_op
!= &io_uring_fops
)
10200 ctx
= f
.file
->private_data
;
10202 io_run_task_work();
10204 mutex_lock(&ctx
->uring_lock
);
10205 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10206 mutex_unlock(&ctx
->uring_lock
);
10207 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10208 ctx
->cq_ev_fd
!= NULL
, ret
);
10214 static int __init
io_uring_init(void)
10216 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10217 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10218 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10221 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10222 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10223 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10224 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10225 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10226 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10227 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10228 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10229 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10230 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10231 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10232 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10233 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10234 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10235 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10236 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10237 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10238 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10239 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10240 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10241 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10242 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10243 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10244 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10245 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10246 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10247 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10248 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10249 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10250 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10251 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10253 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10254 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
10255 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
10259 __initcall(io_uring_init
);