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)
108 u32 head ____cacheline_aligned_in_smp
;
109 u32 tail ____cacheline_aligned_in_smp
;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq
, cq
;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask
, cq_ring_mask
;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries
, cq_ring_entries
;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
190 struct io_mapped_ubuf
{
193 struct bio_vec
*bvec
;
194 unsigned int nr_bvecs
;
195 unsigned long acct_pages
;
198 struct fixed_file_table
{
202 struct fixed_file_ref_node
{
203 struct percpu_ref refs
;
204 struct list_head node
;
205 struct list_head file_list
;
206 struct fixed_file_data
*file_data
;
207 struct llist_node llist
;
211 struct fixed_file_data
{
212 struct fixed_file_table
*table
;
213 struct io_ring_ctx
*ctx
;
215 struct fixed_file_ref_node
*node
;
216 struct percpu_ref refs
;
217 struct completion done
;
218 struct list_head ref_list
;
223 struct list_head list
;
229 struct io_restriction
{
230 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
231 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
232 u8 sqe_flags_allowed
;
233 u8 sqe_flags_required
;
241 /* ctx's that are using this sqd */
242 struct list_head ctx_list
;
243 struct list_head ctx_new_list
;
244 struct mutex ctx_lock
;
246 struct task_struct
*thread
;
247 struct wait_queue_head wait
;
249 unsigned sq_thread_idle
;
254 struct percpu_ref refs
;
255 } ____cacheline_aligned_in_smp
;
259 unsigned int compat
: 1;
260 unsigned int limit_mem
: 1;
261 unsigned int cq_overflow_flushed
: 1;
262 unsigned int drain_next
: 1;
263 unsigned int eventfd_async
: 1;
264 unsigned int restricted
: 1;
265 unsigned int sqo_dead
: 1;
268 * Ring buffer of indices into array of io_uring_sqe, which is
269 * mmapped by the application using the IORING_OFF_SQES offset.
271 * This indirection could e.g. be used to assign fixed
272 * io_uring_sqe entries to operations and only submit them to
273 * the queue when needed.
275 * The kernel modifies neither the indices array nor the entries
279 unsigned cached_sq_head
;
282 unsigned sq_thread_idle
;
283 unsigned cached_sq_dropped
;
284 unsigned cached_cq_overflow
;
285 unsigned long sq_check_overflow
;
287 struct list_head defer_list
;
288 struct list_head timeout_list
;
289 struct list_head cq_overflow_list
;
291 struct io_uring_sqe
*sq_sqes
;
292 } ____cacheline_aligned_in_smp
;
294 struct io_rings
*rings
;
300 * For SQPOLL usage - we hold a reference to the parent task, so we
301 * have access to the ->files
303 struct task_struct
*sqo_task
;
305 /* Only used for accounting purposes */
306 struct mm_struct
*mm_account
;
308 #ifdef CONFIG_BLK_CGROUP
309 struct cgroup_subsys_state
*sqo_blkcg_css
;
312 struct io_sq_data
*sq_data
; /* if using sq thread polling */
314 struct wait_queue_head sqo_sq_wait
;
315 struct list_head sqd_list
;
318 * If used, fixed file set. Writers must ensure that ->refs is dead,
319 * readers must ensure that ->refs is alive as long as the file* is
320 * used. Only updated through io_uring_register(2).
322 struct fixed_file_data
*file_data
;
323 unsigned nr_user_files
;
325 /* if used, fixed mapped user buffers */
326 unsigned nr_user_bufs
;
327 struct io_mapped_ubuf
*user_bufs
;
329 struct user_struct
*user
;
331 const struct cred
*creds
;
335 unsigned int sessionid
;
338 struct completion ref_comp
;
339 struct completion sq_thread_comp
;
341 /* if all else fails... */
342 struct io_kiocb
*fallback_req
;
344 #if defined(CONFIG_UNIX)
345 struct socket
*ring_sock
;
348 struct idr io_buffer_idr
;
350 struct idr personality_idr
;
353 unsigned cached_cq_tail
;
356 atomic_t cq_timeouts
;
357 unsigned cq_last_tm_flush
;
358 unsigned long cq_check_overflow
;
359 struct wait_queue_head cq_wait
;
360 struct fasync_struct
*cq_fasync
;
361 struct eventfd_ctx
*cq_ev_fd
;
362 } ____cacheline_aligned_in_smp
;
365 struct mutex uring_lock
;
366 wait_queue_head_t wait
;
367 } ____cacheline_aligned_in_smp
;
370 spinlock_t completion_lock
;
373 * ->iopoll_list is protected by the ctx->uring_lock for
374 * io_uring instances that don't use IORING_SETUP_SQPOLL.
375 * For SQPOLL, only the single threaded io_sq_thread() will
376 * manipulate the list, hence no extra locking is needed there.
378 struct list_head iopoll_list
;
379 struct hlist_head
*cancel_hash
;
380 unsigned cancel_hash_bits
;
381 bool poll_multi_file
;
383 spinlock_t inflight_lock
;
384 struct list_head inflight_list
;
385 } ____cacheline_aligned_in_smp
;
387 struct delayed_work file_put_work
;
388 struct llist_head file_put_llist
;
390 struct work_struct exit_work
;
391 struct io_restriction restrictions
;
395 * First field must be the file pointer in all the
396 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
398 struct io_poll_iocb
{
400 struct wait_queue_head
*head
;
404 struct wait_queue_entry wait
;
407 struct io_poll_remove
{
417 struct io_timeout_data
{
418 struct io_kiocb
*req
;
419 struct hrtimer timer
;
420 struct timespec64 ts
;
421 enum hrtimer_mode mode
;
426 struct sockaddr __user
*addr
;
427 int __user
*addr_len
;
429 unsigned long nofile
;
449 struct list_head list
;
450 /* head of the link, used by linked timeouts only */
451 struct io_kiocb
*head
;
454 struct io_timeout_rem
{
459 struct timespec64 ts
;
464 /* NOTE: kiocb has the file as the first member, so don't do it here */
472 struct sockaddr __user
*addr
;
479 struct user_msghdr __user
*umsg
;
485 struct io_buffer
*kbuf
;
491 bool ignore_nonblock
;
492 struct filename
*filename
;
494 unsigned long nofile
;
497 struct io_files_update
{
523 struct epoll_event event
;
527 struct file
*file_out
;
528 struct file
*file_in
;
535 struct io_provide_buf
{
549 const char __user
*filename
;
550 struct statx __user
*buffer
;
562 struct filename
*oldpath
;
563 struct filename
*newpath
;
571 struct filename
*filename
;
574 struct io_completion
{
576 struct list_head list
;
580 struct io_async_connect
{
581 struct sockaddr_storage address
;
584 struct io_async_msghdr
{
585 struct iovec fast_iov
[UIO_FASTIOV
];
587 struct sockaddr __user
*uaddr
;
589 struct sockaddr_storage addr
;
593 struct iovec fast_iov
[UIO_FASTIOV
];
594 const struct iovec
*free_iovec
;
595 struct iov_iter iter
;
597 struct wait_page_queue wpq
;
601 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
602 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
603 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
604 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
605 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
606 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
612 REQ_F_LINK_TIMEOUT_BIT
,
614 REQ_F_NEED_CLEANUP_BIT
,
616 REQ_F_BUFFER_SELECTED_BIT
,
617 REQ_F_NO_FILE_TABLE_BIT
,
618 REQ_F_WORK_INITIALIZED_BIT
,
619 REQ_F_LTIMEOUT_ACTIVE_BIT
,
621 /* not a real bit, just to check we're not overflowing the space */
627 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
628 /* drain existing IO first */
629 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
631 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
632 /* doesn't sever on completion < 0 */
633 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
635 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
636 /* IOSQE_BUFFER_SELECT */
637 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
639 /* fail rest of links */
640 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
641 /* on inflight list */
642 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
643 /* read/write uses file position */
644 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
645 /* must not punt to workers */
646 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
647 /* has or had linked timeout */
648 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
650 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
652 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
653 /* already went through poll handler */
654 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
655 /* buffer already selected */
656 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
657 /* doesn't need file table for this request */
658 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
659 /* io_wq_work is initialized */
660 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
661 /* linked timeout is active, i.e. prepared by link's head */
662 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
666 struct io_poll_iocb poll
;
667 struct io_poll_iocb
*double_poll
;
671 * NOTE! Each of the iocb union members has the file pointer
672 * as the first entry in their struct definition. So you can
673 * access the file pointer through any of the sub-structs,
674 * or directly as just 'ki_filp' in this struct.
680 struct io_poll_iocb poll
;
681 struct io_poll_remove poll_remove
;
682 struct io_accept accept
;
684 struct io_cancel cancel
;
685 struct io_timeout timeout
;
686 struct io_timeout_rem timeout_rem
;
687 struct io_connect connect
;
688 struct io_sr_msg sr_msg
;
690 struct io_close close
;
691 struct io_files_update files_update
;
692 struct io_fadvise fadvise
;
693 struct io_madvise madvise
;
694 struct io_epoll epoll
;
695 struct io_splice splice
;
696 struct io_provide_buf pbuf
;
697 struct io_statx statx
;
698 struct io_shutdown shutdown
;
699 struct io_rename rename
;
700 struct io_unlink unlink
;
701 /* use only after cleaning per-op data, see io_clean_op() */
702 struct io_completion
compl;
705 /* opcode allocated if it needs to store data for async defer */
708 /* polled IO has completed */
714 struct io_ring_ctx
*ctx
;
717 struct task_struct
*task
;
720 struct io_kiocb
*link
;
721 struct percpu_ref
*fixed_file_refs
;
724 * 1. used with ctx->iopoll_list with reads/writes
725 * 2. to track reqs with ->files (see io_op_def::file_table)
727 struct list_head inflight_entry
;
728 struct callback_head task_work
;
729 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
730 struct hlist_node hash_node
;
731 struct async_poll
*apoll
;
732 struct io_wq_work work
;
735 struct io_defer_entry
{
736 struct list_head list
;
737 struct io_kiocb
*req
;
741 #define IO_IOPOLL_BATCH 8
743 struct io_comp_state
{
745 struct list_head list
;
746 struct io_ring_ctx
*ctx
;
749 struct io_submit_state
{
750 struct blk_plug plug
;
753 * io_kiocb alloc cache
755 void *reqs
[IO_IOPOLL_BATCH
];
756 unsigned int free_reqs
;
761 * Batch completion logic
763 struct io_comp_state comp
;
766 * File reference cache
770 unsigned int file_refs
;
771 unsigned int ios_left
;
775 /* needs req->file assigned */
776 unsigned needs_file
: 1;
777 /* don't fail if file grab fails */
778 unsigned needs_file_no_error
: 1;
779 /* hash wq insertion if file is a regular file */
780 unsigned hash_reg_file
: 1;
781 /* unbound wq insertion if file is a non-regular file */
782 unsigned unbound_nonreg_file
: 1;
783 /* opcode is not supported by this kernel */
784 unsigned not_supported
: 1;
785 /* set if opcode supports polled "wait" */
787 unsigned pollout
: 1;
788 /* op supports buffer selection */
789 unsigned buffer_select
: 1;
790 /* must always have async data allocated */
791 unsigned needs_async_data
: 1;
792 /* should block plug */
794 /* size of async data needed, if any */
795 unsigned short async_size
;
799 static const struct io_op_def io_op_defs
[] = {
800 [IORING_OP_NOP
] = {},
801 [IORING_OP_READV
] = {
803 .unbound_nonreg_file
= 1,
806 .needs_async_data
= 1,
808 .async_size
= sizeof(struct io_async_rw
),
809 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
811 [IORING_OP_WRITEV
] = {
814 .unbound_nonreg_file
= 1,
816 .needs_async_data
= 1,
818 .async_size
= sizeof(struct io_async_rw
),
819 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
822 [IORING_OP_FSYNC
] = {
824 .work_flags
= IO_WQ_WORK_BLKCG
,
826 [IORING_OP_READ_FIXED
] = {
828 .unbound_nonreg_file
= 1,
831 .async_size
= sizeof(struct io_async_rw
),
832 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
834 [IORING_OP_WRITE_FIXED
] = {
837 .unbound_nonreg_file
= 1,
840 .async_size
= sizeof(struct io_async_rw
),
841 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
|
844 [IORING_OP_POLL_ADD
] = {
846 .unbound_nonreg_file
= 1,
848 [IORING_OP_POLL_REMOVE
] = {},
849 [IORING_OP_SYNC_FILE_RANGE
] = {
851 .work_flags
= IO_WQ_WORK_BLKCG
,
853 [IORING_OP_SENDMSG
] = {
855 .unbound_nonreg_file
= 1,
857 .needs_async_data
= 1,
858 .async_size
= sizeof(struct io_async_msghdr
),
859 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
862 [IORING_OP_RECVMSG
] = {
864 .unbound_nonreg_file
= 1,
867 .needs_async_data
= 1,
868 .async_size
= sizeof(struct io_async_msghdr
),
869 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
872 [IORING_OP_TIMEOUT
] = {
873 .needs_async_data
= 1,
874 .async_size
= sizeof(struct io_timeout_data
),
875 .work_flags
= IO_WQ_WORK_MM
,
877 [IORING_OP_TIMEOUT_REMOVE
] = {
878 /* used by timeout updates' prep() */
879 .work_flags
= IO_WQ_WORK_MM
,
881 [IORING_OP_ACCEPT
] = {
883 .unbound_nonreg_file
= 1,
885 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
,
887 [IORING_OP_ASYNC_CANCEL
] = {},
888 [IORING_OP_LINK_TIMEOUT
] = {
889 .needs_async_data
= 1,
890 .async_size
= sizeof(struct io_timeout_data
),
891 .work_flags
= IO_WQ_WORK_MM
,
893 [IORING_OP_CONNECT
] = {
895 .unbound_nonreg_file
= 1,
897 .needs_async_data
= 1,
898 .async_size
= sizeof(struct io_async_connect
),
899 .work_flags
= IO_WQ_WORK_MM
,
901 [IORING_OP_FALLOCATE
] = {
903 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
905 [IORING_OP_OPENAT
] = {
906 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
|
907 IO_WQ_WORK_FS
| IO_WQ_WORK_MM
,
909 [IORING_OP_CLOSE
] = {
910 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
,
912 [IORING_OP_FILES_UPDATE
] = {
913 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
,
915 [IORING_OP_STATX
] = {
916 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
|
917 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
921 .unbound_nonreg_file
= 1,
925 .async_size
= sizeof(struct io_async_rw
),
926 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
928 [IORING_OP_WRITE
] = {
930 .unbound_nonreg_file
= 1,
933 .async_size
= sizeof(struct io_async_rw
),
934 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
937 [IORING_OP_FADVISE
] = {
939 .work_flags
= IO_WQ_WORK_BLKCG
,
941 [IORING_OP_MADVISE
] = {
942 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
946 .unbound_nonreg_file
= 1,
948 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
952 .unbound_nonreg_file
= 1,
955 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
957 [IORING_OP_OPENAT2
] = {
958 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_FS
|
959 IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
961 [IORING_OP_EPOLL_CTL
] = {
962 .unbound_nonreg_file
= 1,
963 .work_flags
= IO_WQ_WORK_FILES
,
965 [IORING_OP_SPLICE
] = {
968 .unbound_nonreg_file
= 1,
969 .work_flags
= IO_WQ_WORK_BLKCG
,
971 [IORING_OP_PROVIDE_BUFFERS
] = {},
972 [IORING_OP_REMOVE_BUFFERS
] = {},
976 .unbound_nonreg_file
= 1,
978 [IORING_OP_SHUTDOWN
] = {
981 [IORING_OP_RENAMEAT
] = {
982 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
983 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
985 [IORING_OP_UNLINKAT
] = {
986 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
987 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
991 enum io_mem_account
{
996 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
997 struct task_struct
*task
,
998 struct files_struct
*files
);
999 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
);
1000 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
1001 struct io_ring_ctx
*ctx
);
1003 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
1004 struct io_comp_state
*cs
);
1005 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
1006 static void io_put_req(struct io_kiocb
*req
);
1007 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1008 static void io_double_put_req(struct io_kiocb
*req
);
1009 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1010 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
1011 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1012 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
1013 struct io_uring_files_update
*ip
,
1015 static void __io_clean_op(struct io_kiocb
*req
);
1016 static struct file
*io_file_get(struct io_submit_state
*state
,
1017 struct io_kiocb
*req
, int fd
, bool fixed
);
1018 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
);
1019 static void io_file_put_work(struct work_struct
*work
);
1021 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
1022 struct iovec
**iovec
, struct iov_iter
*iter
,
1024 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
1025 const struct iovec
*fast_iov
,
1026 struct iov_iter
*iter
, bool force
);
1027 static void io_req_drop_files(struct io_kiocb
*req
);
1028 static void io_req_task_queue(struct io_kiocb
*req
);
1030 static struct kmem_cache
*req_cachep
;
1032 static const struct file_operations io_uring_fops
;
1034 struct sock
*io_uring_get_socket(struct file
*file
)
1036 #if defined(CONFIG_UNIX)
1037 if (file
->f_op
== &io_uring_fops
) {
1038 struct io_ring_ctx
*ctx
= file
->private_data
;
1040 return ctx
->ring_sock
->sk
;
1045 EXPORT_SYMBOL(io_uring_get_socket
);
1047 #define io_for_each_link(pos, head) \
1048 for (pos = (head); pos; pos = pos->link)
1050 static inline void io_clean_op(struct io_kiocb
*req
)
1052 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
))
1056 static inline void io_set_resource_node(struct io_kiocb
*req
)
1058 struct io_ring_ctx
*ctx
= req
->ctx
;
1060 if (!req
->fixed_file_refs
) {
1061 req
->fixed_file_refs
= &ctx
->file_data
->node
->refs
;
1062 percpu_ref_get(req
->fixed_file_refs
);
1066 static bool io_match_task(struct io_kiocb
*head
,
1067 struct task_struct
*task
,
1068 struct files_struct
*files
)
1070 struct io_kiocb
*req
;
1072 if (task
&& head
->task
!= task
) {
1073 /* in terms of cancelation, always match if req task is dead */
1074 if (head
->task
->flags
& PF_EXITING
)
1081 io_for_each_link(req
, head
) {
1082 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1084 if (req
->file
&& req
->file
->f_op
== &io_uring_fops
)
1086 if ((req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1087 req
->work
.identity
->files
== files
)
1093 static void io_sq_thread_drop_mm_files(void)
1095 struct files_struct
*files
= current
->files
;
1096 struct mm_struct
*mm
= current
->mm
;
1099 kthread_unuse_mm(mm
);
1104 struct nsproxy
*nsproxy
= current
->nsproxy
;
1107 current
->files
= NULL
;
1108 current
->nsproxy
= NULL
;
1109 task_unlock(current
);
1110 put_files_struct(files
);
1111 put_nsproxy(nsproxy
);
1115 static int __io_sq_thread_acquire_files(struct io_ring_ctx
*ctx
)
1117 if (current
->flags
& PF_EXITING
)
1120 if (!current
->files
) {
1121 struct files_struct
*files
;
1122 struct nsproxy
*nsproxy
;
1124 task_lock(ctx
->sqo_task
);
1125 files
= ctx
->sqo_task
->files
;
1127 task_unlock(ctx
->sqo_task
);
1130 atomic_inc(&files
->count
);
1131 get_nsproxy(ctx
->sqo_task
->nsproxy
);
1132 nsproxy
= ctx
->sqo_task
->nsproxy
;
1133 task_unlock(ctx
->sqo_task
);
1136 current
->files
= files
;
1137 current
->nsproxy
= nsproxy
;
1138 task_unlock(current
);
1143 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1145 struct mm_struct
*mm
;
1147 if (current
->flags
& PF_EXITING
)
1152 /* Should never happen */
1153 if (unlikely(!(ctx
->flags
& IORING_SETUP_SQPOLL
)))
1156 task_lock(ctx
->sqo_task
);
1157 mm
= ctx
->sqo_task
->mm
;
1158 if (unlikely(!mm
|| !mmget_not_zero(mm
)))
1160 task_unlock(ctx
->sqo_task
);
1170 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx
*ctx
,
1171 struct io_kiocb
*req
)
1173 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1176 if (def
->work_flags
& IO_WQ_WORK_MM
) {
1177 ret
= __io_sq_thread_acquire_mm(ctx
);
1182 if (def
->needs_file
|| (def
->work_flags
& IO_WQ_WORK_FILES
)) {
1183 ret
= __io_sq_thread_acquire_files(ctx
);
1191 static void io_sq_thread_associate_blkcg(struct io_ring_ctx
*ctx
,
1192 struct cgroup_subsys_state
**cur_css
)
1195 #ifdef CONFIG_BLK_CGROUP
1196 /* puts the old one when swapping */
1197 if (*cur_css
!= ctx
->sqo_blkcg_css
) {
1198 kthread_associate_blkcg(ctx
->sqo_blkcg_css
);
1199 *cur_css
= ctx
->sqo_blkcg_css
;
1204 static void io_sq_thread_unassociate_blkcg(void)
1206 #ifdef CONFIG_BLK_CGROUP
1207 kthread_associate_blkcg(NULL
);
1211 static inline void req_set_fail_links(struct io_kiocb
*req
)
1213 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1214 req
->flags
|= REQ_F_FAIL_LINK
;
1218 * None of these are dereferenced, they are simply used to check if any of
1219 * them have changed. If we're under current and check they are still the
1220 * same, we're fine to grab references to them for actual out-of-line use.
1222 static void io_init_identity(struct io_identity
*id
)
1224 id
->files
= current
->files
;
1225 id
->mm
= current
->mm
;
1226 #ifdef CONFIG_BLK_CGROUP
1228 id
->blkcg_css
= blkcg_css();
1231 id
->creds
= current_cred();
1232 id
->nsproxy
= current
->nsproxy
;
1233 id
->fs
= current
->fs
;
1234 id
->fsize
= rlimit(RLIMIT_FSIZE
);
1236 id
->loginuid
= current
->loginuid
;
1237 id
->sessionid
= current
->sessionid
;
1239 refcount_set(&id
->count
, 1);
1242 static inline void __io_req_init_async(struct io_kiocb
*req
)
1244 memset(&req
->work
, 0, sizeof(req
->work
));
1245 req
->flags
|= REQ_F_WORK_INITIALIZED
;
1249 * Note: must call io_req_init_async() for the first time you
1250 * touch any members of io_wq_work.
1252 static inline void io_req_init_async(struct io_kiocb
*req
)
1254 struct io_uring_task
*tctx
= current
->io_uring
;
1256 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
1259 __io_req_init_async(req
);
1261 /* Grab a ref if this isn't our static identity */
1262 req
->work
.identity
= tctx
->identity
;
1263 if (tctx
->identity
!= &tctx
->__identity
)
1264 refcount_inc(&req
->work
.identity
->count
);
1267 static inline bool io_async_submit(struct io_ring_ctx
*ctx
)
1269 return ctx
->flags
& IORING_SETUP_SQPOLL
;
1272 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1274 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1276 complete(&ctx
->ref_comp
);
1279 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1281 return !req
->timeout
.off
;
1284 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1286 struct io_ring_ctx
*ctx
;
1289 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1293 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
1294 if (!ctx
->fallback_req
)
1298 * Use 5 bits less than the max cq entries, that should give us around
1299 * 32 entries per hash list if totally full and uniformly spread.
1301 hash_bits
= ilog2(p
->cq_entries
);
1305 ctx
->cancel_hash_bits
= hash_bits
;
1306 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1308 if (!ctx
->cancel_hash
)
1310 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1312 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1313 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1316 ctx
->flags
= p
->flags
;
1317 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1318 INIT_LIST_HEAD(&ctx
->sqd_list
);
1319 init_waitqueue_head(&ctx
->cq_wait
);
1320 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1321 init_completion(&ctx
->ref_comp
);
1322 init_completion(&ctx
->sq_thread_comp
);
1323 idr_init(&ctx
->io_buffer_idr
);
1324 idr_init(&ctx
->personality_idr
);
1325 mutex_init(&ctx
->uring_lock
);
1326 init_waitqueue_head(&ctx
->wait
);
1327 spin_lock_init(&ctx
->completion_lock
);
1328 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1329 INIT_LIST_HEAD(&ctx
->defer_list
);
1330 INIT_LIST_HEAD(&ctx
->timeout_list
);
1331 spin_lock_init(&ctx
->inflight_lock
);
1332 INIT_LIST_HEAD(&ctx
->inflight_list
);
1333 INIT_DELAYED_WORK(&ctx
->file_put_work
, io_file_put_work
);
1334 init_llist_head(&ctx
->file_put_llist
);
1337 if (ctx
->fallback_req
)
1338 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1339 kfree(ctx
->cancel_hash
);
1344 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1346 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1347 struct io_ring_ctx
*ctx
= req
->ctx
;
1349 return seq
!= ctx
->cached_cq_tail
1350 + READ_ONCE(ctx
->cached_cq_overflow
);
1356 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1358 struct io_rings
*rings
= ctx
->rings
;
1360 /* order cqe stores with ring update */
1361 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1364 static void io_put_identity(struct io_uring_task
*tctx
, struct io_kiocb
*req
)
1366 if (req
->work
.identity
== &tctx
->__identity
)
1368 if (refcount_dec_and_test(&req
->work
.identity
->count
))
1369 kfree(req
->work
.identity
);
1372 static void io_req_clean_work(struct io_kiocb
*req
)
1374 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1377 req
->flags
&= ~REQ_F_WORK_INITIALIZED
;
1379 if (req
->work
.flags
& IO_WQ_WORK_MM
) {
1380 mmdrop(req
->work
.identity
->mm
);
1381 req
->work
.flags
&= ~IO_WQ_WORK_MM
;
1383 #ifdef CONFIG_BLK_CGROUP
1384 if (req
->work
.flags
& IO_WQ_WORK_BLKCG
) {
1385 css_put(req
->work
.identity
->blkcg_css
);
1386 req
->work
.flags
&= ~IO_WQ_WORK_BLKCG
;
1389 if (req
->work
.flags
& IO_WQ_WORK_CREDS
) {
1390 put_cred(req
->work
.identity
->creds
);
1391 req
->work
.flags
&= ~IO_WQ_WORK_CREDS
;
1393 if (req
->work
.flags
& IO_WQ_WORK_FS
) {
1394 struct fs_struct
*fs
= req
->work
.identity
->fs
;
1396 spin_lock(&req
->work
.identity
->fs
->lock
);
1399 spin_unlock(&req
->work
.identity
->fs
->lock
);
1402 req
->work
.flags
&= ~IO_WQ_WORK_FS
;
1404 if (req
->flags
& REQ_F_INFLIGHT
)
1405 io_req_drop_files(req
);
1407 io_put_identity(req
->task
->io_uring
, req
);
1411 * Create a private copy of io_identity, since some fields don't match
1412 * the current context.
1414 static bool io_identity_cow(struct io_kiocb
*req
)
1416 struct io_uring_task
*tctx
= current
->io_uring
;
1417 const struct cred
*creds
= NULL
;
1418 struct io_identity
*id
;
1420 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1421 creds
= req
->work
.identity
->creds
;
1423 id
= kmemdup(req
->work
.identity
, sizeof(*id
), GFP_KERNEL
);
1424 if (unlikely(!id
)) {
1425 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1430 * We can safely just re-init the creds we copied Either the field
1431 * matches the current one, or we haven't grabbed it yet. The only
1432 * exception is ->creds, through registered personalities, so handle
1433 * that one separately.
1435 io_init_identity(id
);
1439 /* add one for this request */
1440 refcount_inc(&id
->count
);
1442 /* drop tctx and req identity references, if needed */
1443 if (tctx
->identity
!= &tctx
->__identity
&&
1444 refcount_dec_and_test(&tctx
->identity
->count
))
1445 kfree(tctx
->identity
);
1446 if (req
->work
.identity
!= &tctx
->__identity
&&
1447 refcount_dec_and_test(&req
->work
.identity
->count
))
1448 kfree(req
->work
.identity
);
1450 req
->work
.identity
= id
;
1451 tctx
->identity
= id
;
1455 static bool io_grab_identity(struct io_kiocb
*req
)
1457 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1458 struct io_identity
*id
= req
->work
.identity
;
1459 struct io_ring_ctx
*ctx
= req
->ctx
;
1461 if (def
->work_flags
& IO_WQ_WORK_FSIZE
) {
1462 if (id
->fsize
!= rlimit(RLIMIT_FSIZE
))
1464 req
->work
.flags
|= IO_WQ_WORK_FSIZE
;
1466 #ifdef CONFIG_BLK_CGROUP
1467 if (!(req
->work
.flags
& IO_WQ_WORK_BLKCG
) &&
1468 (def
->work_flags
& IO_WQ_WORK_BLKCG
)) {
1470 if (id
->blkcg_css
!= blkcg_css()) {
1475 * This should be rare, either the cgroup is dying or the task
1476 * is moving cgroups. Just punt to root for the handful of ios.
1478 if (css_tryget_online(id
->blkcg_css
))
1479 req
->work
.flags
|= IO_WQ_WORK_BLKCG
;
1483 if (!(req
->work
.flags
& IO_WQ_WORK_CREDS
)) {
1484 if (id
->creds
!= current_cred())
1486 get_cred(id
->creds
);
1487 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
1490 if (!uid_eq(current
->loginuid
, id
->loginuid
) ||
1491 current
->sessionid
!= id
->sessionid
)
1494 if (!(req
->work
.flags
& IO_WQ_WORK_FS
) &&
1495 (def
->work_flags
& IO_WQ_WORK_FS
)) {
1496 if (current
->fs
!= id
->fs
)
1498 spin_lock(&id
->fs
->lock
);
1499 if (!id
->fs
->in_exec
) {
1501 req
->work
.flags
|= IO_WQ_WORK_FS
;
1503 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1505 spin_unlock(¤t
->fs
->lock
);
1507 if (!(req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1508 (def
->work_flags
& IO_WQ_WORK_FILES
) &&
1509 !(req
->flags
& REQ_F_NO_FILE_TABLE
)) {
1510 if (id
->files
!= current
->files
||
1511 id
->nsproxy
!= current
->nsproxy
)
1513 atomic_inc(&id
->files
->count
);
1514 get_nsproxy(id
->nsproxy
);
1516 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1517 req
->flags
|= REQ_F_INFLIGHT
;
1519 spin_lock_irq(&ctx
->inflight_lock
);
1520 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1521 spin_unlock_irq(&ctx
->inflight_lock
);
1523 req
->work
.flags
|= IO_WQ_WORK_FILES
;
1525 if (!(req
->work
.flags
& IO_WQ_WORK_MM
) &&
1526 (def
->work_flags
& IO_WQ_WORK_MM
)) {
1527 if (id
->mm
!= current
->mm
)
1530 req
->work
.flags
|= IO_WQ_WORK_MM
;
1536 static void io_prep_async_work(struct io_kiocb
*req
)
1538 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1539 struct io_ring_ctx
*ctx
= req
->ctx
;
1541 io_req_init_async(req
);
1543 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1544 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1546 if (req
->flags
& REQ_F_ISREG
) {
1547 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1548 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1550 if (def
->unbound_nonreg_file
)
1551 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1554 /* if we fail grabbing identity, we must COW, regrab, and retry */
1555 if (io_grab_identity(req
))
1558 if (!io_identity_cow(req
))
1561 /* can't fail at this point */
1562 if (!io_grab_identity(req
))
1566 static void io_prep_async_link(struct io_kiocb
*req
)
1568 struct io_kiocb
*cur
;
1570 io_for_each_link(cur
, req
)
1571 io_prep_async_work(cur
);
1574 static struct io_kiocb
*__io_queue_async_work(struct io_kiocb
*req
)
1576 struct io_ring_ctx
*ctx
= req
->ctx
;
1577 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1579 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1580 &req
->work
, req
->flags
);
1581 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1585 static void io_queue_async_work(struct io_kiocb
*req
)
1587 struct io_kiocb
*link
;
1589 /* init ->work of the whole link before punting */
1590 io_prep_async_link(req
);
1591 link
= __io_queue_async_work(req
);
1594 io_queue_linked_timeout(link
);
1597 static void io_kill_timeout(struct io_kiocb
*req
)
1599 struct io_timeout_data
*io
= req
->async_data
;
1602 ret
= hrtimer_try_to_cancel(&io
->timer
);
1604 atomic_set(&req
->ctx
->cq_timeouts
,
1605 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1606 list_del_init(&req
->timeout
.list
);
1607 io_cqring_fill_event(req
, 0);
1608 io_put_req_deferred(req
, 1);
1613 * Returns true if we found and killed one or more timeouts
1615 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
1616 struct files_struct
*files
)
1618 struct io_kiocb
*req
, *tmp
;
1621 spin_lock_irq(&ctx
->completion_lock
);
1622 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1623 if (io_match_task(req
, tsk
, files
)) {
1624 io_kill_timeout(req
);
1628 spin_unlock_irq(&ctx
->completion_lock
);
1629 return canceled
!= 0;
1632 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1635 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1636 struct io_defer_entry
, list
);
1638 if (req_need_defer(de
->req
, de
->seq
))
1640 list_del_init(&de
->list
);
1641 io_req_task_queue(de
->req
);
1643 } while (!list_empty(&ctx
->defer_list
));
1646 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1650 if (list_empty(&ctx
->timeout_list
))
1653 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1656 u32 events_needed
, events_got
;
1657 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1658 struct io_kiocb
, timeout
.list
);
1660 if (io_is_timeout_noseq(req
))
1664 * Since seq can easily wrap around over time, subtract
1665 * the last seq at which timeouts were flushed before comparing.
1666 * Assuming not more than 2^31-1 events have happened since,
1667 * these subtractions won't have wrapped, so we can check if
1668 * target is in [last_seq, current_seq] by comparing the two.
1670 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1671 events_got
= seq
- ctx
->cq_last_tm_flush
;
1672 if (events_got
< events_needed
)
1675 list_del_init(&req
->timeout
.list
);
1676 io_kill_timeout(req
);
1677 } while (!list_empty(&ctx
->timeout_list
));
1679 ctx
->cq_last_tm_flush
= seq
;
1682 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1684 io_flush_timeouts(ctx
);
1685 __io_commit_cqring(ctx
);
1687 if (unlikely(!list_empty(&ctx
->defer_list
)))
1688 __io_queue_deferred(ctx
);
1691 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1693 struct io_rings
*r
= ctx
->rings
;
1695 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1698 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1700 struct io_rings
*rings
= ctx
->rings
;
1703 tail
= ctx
->cached_cq_tail
;
1705 * writes to the cq entry need to come after reading head; the
1706 * control dependency is enough as we're using WRITE_ONCE to
1709 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1712 ctx
->cached_cq_tail
++;
1713 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1716 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1720 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1722 if (!ctx
->eventfd_async
)
1724 return io_wq_current_is_worker();
1727 static inline unsigned __io_cqring_events(struct io_ring_ctx
*ctx
)
1729 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1732 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1734 /* see waitqueue_active() comment */
1737 if (waitqueue_active(&ctx
->wait
))
1738 wake_up(&ctx
->wait
);
1739 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1740 wake_up(&ctx
->sq_data
->wait
);
1741 if (io_should_trigger_evfd(ctx
))
1742 eventfd_signal(ctx
->cq_ev_fd
, 1);
1743 if (waitqueue_active(&ctx
->cq_wait
)) {
1744 wake_up_interruptible(&ctx
->cq_wait
);
1745 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1749 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1751 /* see waitqueue_active() comment */
1754 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1755 if (waitqueue_active(&ctx
->wait
))
1756 wake_up(&ctx
->wait
);
1758 if (io_should_trigger_evfd(ctx
))
1759 eventfd_signal(ctx
->cq_ev_fd
, 1);
1760 if (waitqueue_active(&ctx
->cq_wait
)) {
1761 wake_up_interruptible(&ctx
->cq_wait
);
1762 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1766 /* Returns true if there are no backlogged entries after the flush */
1767 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1768 struct task_struct
*tsk
,
1769 struct files_struct
*files
)
1771 struct io_rings
*rings
= ctx
->rings
;
1772 struct io_kiocb
*req
, *tmp
;
1773 struct io_uring_cqe
*cqe
;
1774 unsigned long flags
;
1775 bool all_flushed
, posted
;
1778 if (!force
&& __io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1782 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1783 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1784 if (!io_match_task(req
, tsk
, files
))
1787 cqe
= io_get_cqring(ctx
);
1791 list_move(&req
->compl.list
, &list
);
1793 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1794 WRITE_ONCE(cqe
->res
, req
->result
);
1795 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1797 ctx
->cached_cq_overflow
++;
1798 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1799 ctx
->cached_cq_overflow
);
1804 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1806 clear_bit(0, &ctx
->sq_check_overflow
);
1807 clear_bit(0, &ctx
->cq_check_overflow
);
1808 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1812 io_commit_cqring(ctx
);
1813 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1815 io_cqring_ev_posted(ctx
);
1817 while (!list_empty(&list
)) {
1818 req
= list_first_entry(&list
, struct io_kiocb
, compl.list
);
1819 list_del(&req
->compl.list
);
1826 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1827 struct task_struct
*tsk
,
1828 struct files_struct
*files
)
1832 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1833 /* iopoll syncs against uring_lock, not completion_lock */
1834 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1835 mutex_lock(&ctx
->uring_lock
);
1836 ret
= __io_cqring_overflow_flush(ctx
, force
, tsk
, files
);
1837 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1838 mutex_unlock(&ctx
->uring_lock
);
1844 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1846 struct io_ring_ctx
*ctx
= req
->ctx
;
1847 struct io_uring_cqe
*cqe
;
1849 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1852 * If we can't get a cq entry, userspace overflowed the
1853 * submission (by quite a lot). Increment the overflow count in
1856 cqe
= io_get_cqring(ctx
);
1858 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1859 WRITE_ONCE(cqe
->res
, res
);
1860 WRITE_ONCE(cqe
->flags
, cflags
);
1861 } else if (ctx
->cq_overflow_flushed
||
1862 atomic_read(&req
->task
->io_uring
->in_idle
)) {
1864 * If we're in ring overflow flush mode, or in task cancel mode,
1865 * then we cannot store the request for later flushing, we need
1866 * to drop it on the floor.
1868 ctx
->cached_cq_overflow
++;
1869 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1871 if (list_empty(&ctx
->cq_overflow_list
)) {
1872 set_bit(0, &ctx
->sq_check_overflow
);
1873 set_bit(0, &ctx
->cq_check_overflow
);
1874 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1878 req
->compl.cflags
= cflags
;
1879 refcount_inc(&req
->refs
);
1880 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1884 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1886 __io_cqring_fill_event(req
, res
, 0);
1889 static void io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1891 struct io_ring_ctx
*ctx
= req
->ctx
;
1892 unsigned long flags
;
1894 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1895 __io_cqring_fill_event(req
, res
, cflags
);
1896 io_commit_cqring(ctx
);
1897 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1899 io_cqring_ev_posted(ctx
);
1902 static void io_submit_flush_completions(struct io_comp_state
*cs
)
1904 struct io_ring_ctx
*ctx
= cs
->ctx
;
1906 spin_lock_irq(&ctx
->completion_lock
);
1907 while (!list_empty(&cs
->list
)) {
1908 struct io_kiocb
*req
;
1910 req
= list_first_entry(&cs
->list
, struct io_kiocb
, compl.list
);
1911 list_del(&req
->compl.list
);
1912 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
1915 * io_free_req() doesn't care about completion_lock unless one
1916 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1917 * because of a potential deadlock with req->work.fs->lock
1919 if (req
->flags
& (REQ_F_FAIL_LINK
|REQ_F_LINK_TIMEOUT
1920 |REQ_F_WORK_INITIALIZED
)) {
1921 spin_unlock_irq(&ctx
->completion_lock
);
1923 spin_lock_irq(&ctx
->completion_lock
);
1928 io_commit_cqring(ctx
);
1929 spin_unlock_irq(&ctx
->completion_lock
);
1931 io_cqring_ev_posted(ctx
);
1935 static void __io_req_complete(struct io_kiocb
*req
, long res
, unsigned cflags
,
1936 struct io_comp_state
*cs
)
1939 io_cqring_add_event(req
, res
, cflags
);
1944 req
->compl.cflags
= cflags
;
1945 list_add_tail(&req
->compl.list
, &cs
->list
);
1947 io_submit_flush_completions(cs
);
1951 static void io_req_complete(struct io_kiocb
*req
, long res
)
1953 __io_req_complete(req
, res
, 0, NULL
);
1956 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1958 return req
== (struct io_kiocb
*)
1959 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1962 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1964 struct io_kiocb
*req
;
1966 req
= ctx
->fallback_req
;
1967 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1973 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1974 struct io_submit_state
*state
)
1976 if (!state
->free_reqs
) {
1977 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1981 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1982 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
1985 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1986 * retry single alloc to be on the safe side.
1988 if (unlikely(ret
<= 0)) {
1989 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1990 if (!state
->reqs
[0])
1994 state
->free_reqs
= ret
;
1998 return state
->reqs
[state
->free_reqs
];
2000 return io_get_fallback_req(ctx
);
2003 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
2010 static void io_dismantle_req(struct io_kiocb
*req
)
2014 if (req
->async_data
)
2015 kfree(req
->async_data
);
2017 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
2018 if (req
->fixed_file_refs
)
2019 percpu_ref_put(req
->fixed_file_refs
);
2020 io_req_clean_work(req
);
2023 static void __io_free_req(struct io_kiocb
*req
)
2025 struct io_uring_task
*tctx
= req
->task
->io_uring
;
2026 struct io_ring_ctx
*ctx
= req
->ctx
;
2028 io_dismantle_req(req
);
2030 percpu_counter_dec(&tctx
->inflight
);
2031 if (atomic_read(&tctx
->in_idle
))
2032 wake_up(&tctx
->wait
);
2033 put_task_struct(req
->task
);
2035 if (likely(!io_is_fallback_req(req
)))
2036 kmem_cache_free(req_cachep
, req
);
2038 clear_bit_unlock(0, (unsigned long *) &ctx
->fallback_req
);
2039 percpu_ref_put(&ctx
->refs
);
2042 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2044 struct io_kiocb
*nxt
= req
->link
;
2046 req
->link
= nxt
->link
;
2050 static void io_kill_linked_timeout(struct io_kiocb
*req
)
2052 struct io_ring_ctx
*ctx
= req
->ctx
;
2053 struct io_kiocb
*link
;
2054 bool cancelled
= false;
2055 unsigned long flags
;
2057 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2061 * Can happen if a linked timeout fired and link had been like
2062 * req -> link t-out -> link t-out [-> ...]
2064 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
2065 struct io_timeout_data
*io
= link
->async_data
;
2068 io_remove_next_linked(req
);
2069 link
->timeout
.head
= NULL
;
2070 ret
= hrtimer_try_to_cancel(&io
->timer
);
2072 io_cqring_fill_event(link
, -ECANCELED
);
2073 io_commit_cqring(ctx
);
2077 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
2078 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2081 io_cqring_ev_posted(ctx
);
2087 static void io_fail_links(struct io_kiocb
*req
)
2089 struct io_kiocb
*link
, *nxt
;
2090 struct io_ring_ctx
*ctx
= req
->ctx
;
2091 unsigned long flags
;
2093 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2101 trace_io_uring_fail_link(req
, link
);
2102 io_cqring_fill_event(link
, -ECANCELED
);
2105 * It's ok to free under spinlock as they're not linked anymore,
2106 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2109 if (link
->flags
& REQ_F_WORK_INITIALIZED
)
2110 io_put_req_deferred(link
, 2);
2112 io_double_put_req(link
);
2115 io_commit_cqring(ctx
);
2116 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2118 io_cqring_ev_posted(ctx
);
2121 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2123 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
2124 io_kill_linked_timeout(req
);
2127 * If LINK is set, we have dependent requests in this chain. If we
2128 * didn't fail this request, queue the first one up, moving any other
2129 * dependencies to the next request. In case of failure, fail the rest
2132 if (likely(!(req
->flags
& REQ_F_FAIL_LINK
))) {
2133 struct io_kiocb
*nxt
= req
->link
;
2142 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2144 if (likely(!(req
->link
) && !(req
->flags
& REQ_F_LINK_TIMEOUT
)))
2146 return __io_req_find_next(req
);
2149 static int io_req_task_work_add(struct io_kiocb
*req
)
2151 struct task_struct
*tsk
= req
->task
;
2152 struct io_ring_ctx
*ctx
= req
->ctx
;
2153 enum task_work_notify_mode notify
;
2156 if (tsk
->flags
& PF_EXITING
)
2160 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2161 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2162 * processing task_work. There's no reliable way to tell if TWA_RESUME
2166 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
))
2167 notify
= TWA_SIGNAL
;
2169 ret
= task_work_add(tsk
, &req
->task_work
, notify
);
2171 wake_up_process(tsk
);
2176 static void io_req_task_work_add_fallback(struct io_kiocb
*req
,
2177 void (*cb
)(struct callback_head
*))
2179 struct task_struct
*tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2181 init_task_work(&req
->task_work
, cb
);
2182 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2183 wake_up_process(tsk
);
2186 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
2188 struct io_ring_ctx
*ctx
= req
->ctx
;
2190 spin_lock_irq(&ctx
->completion_lock
);
2191 io_cqring_fill_event(req
, error
);
2192 io_commit_cqring(ctx
);
2193 spin_unlock_irq(&ctx
->completion_lock
);
2195 io_cqring_ev_posted(ctx
);
2196 req_set_fail_links(req
);
2197 io_double_put_req(req
);
2200 static void io_req_task_cancel(struct callback_head
*cb
)
2202 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2203 struct io_ring_ctx
*ctx
= req
->ctx
;
2205 mutex_lock(&ctx
->uring_lock
);
2206 __io_req_task_cancel(req
, -ECANCELED
);
2207 mutex_unlock(&ctx
->uring_lock
);
2208 percpu_ref_put(&ctx
->refs
);
2211 static void __io_req_task_submit(struct io_kiocb
*req
)
2213 struct io_ring_ctx
*ctx
= req
->ctx
;
2215 mutex_lock(&ctx
->uring_lock
);
2216 if (!ctx
->sqo_dead
&&
2217 !__io_sq_thread_acquire_mm(ctx
) &&
2218 !__io_sq_thread_acquire_files(ctx
))
2219 __io_queue_sqe(req
, NULL
);
2221 __io_req_task_cancel(req
, -EFAULT
);
2222 mutex_unlock(&ctx
->uring_lock
);
2224 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
2225 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
2226 io_sq_thread_drop_mm_files();
2229 static void io_req_task_submit(struct callback_head
*cb
)
2231 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2232 struct io_ring_ctx
*ctx
= req
->ctx
;
2234 __io_req_task_submit(req
);
2235 percpu_ref_put(&ctx
->refs
);
2238 static void io_req_task_queue(struct io_kiocb
*req
)
2242 init_task_work(&req
->task_work
, io_req_task_submit
);
2243 percpu_ref_get(&req
->ctx
->refs
);
2245 ret
= io_req_task_work_add(req
);
2247 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2250 static inline void io_queue_next(struct io_kiocb
*req
)
2252 struct io_kiocb
*nxt
= io_req_find_next(req
);
2255 io_req_task_queue(nxt
);
2258 static void io_free_req(struct io_kiocb
*req
)
2265 void *reqs
[IO_IOPOLL_BATCH
];
2268 struct task_struct
*task
;
2272 static inline void io_init_req_batch(struct req_batch
*rb
)
2279 static void __io_req_free_batch_flush(struct io_ring_ctx
*ctx
,
2280 struct req_batch
*rb
)
2282 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
2283 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
2287 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2288 struct req_batch
*rb
)
2291 __io_req_free_batch_flush(ctx
, rb
);
2293 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2295 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2296 if (atomic_read(&tctx
->in_idle
))
2297 wake_up(&tctx
->wait
);
2298 put_task_struct_many(rb
->task
, rb
->task_refs
);
2303 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
)
2305 if (unlikely(io_is_fallback_req(req
))) {
2311 if (req
->task
!= rb
->task
) {
2313 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2315 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2316 if (atomic_read(&tctx
->in_idle
))
2317 wake_up(&tctx
->wait
);
2318 put_task_struct_many(rb
->task
, rb
->task_refs
);
2320 rb
->task
= req
->task
;
2325 io_dismantle_req(req
);
2326 rb
->reqs
[rb
->to_free
++] = req
;
2327 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
2328 __io_req_free_batch_flush(req
->ctx
, rb
);
2332 * Drop reference to request, return next in chain (if there is one) if this
2333 * was the last reference to this request.
2335 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2337 struct io_kiocb
*nxt
= NULL
;
2339 if (refcount_dec_and_test(&req
->refs
)) {
2340 nxt
= io_req_find_next(req
);
2346 static void io_put_req(struct io_kiocb
*req
)
2348 if (refcount_dec_and_test(&req
->refs
))
2352 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2354 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2359 static void io_free_req_deferred(struct io_kiocb
*req
)
2363 init_task_work(&req
->task_work
, io_put_req_deferred_cb
);
2364 ret
= io_req_task_work_add(req
);
2366 io_req_task_work_add_fallback(req
, io_put_req_deferred_cb
);
2369 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2371 if (refcount_sub_and_test(refs
, &req
->refs
))
2372 io_free_req_deferred(req
);
2375 static void io_double_put_req(struct io_kiocb
*req
)
2377 /* drop both submit and complete references */
2378 if (refcount_sub_and_test(2, &req
->refs
))
2382 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2384 /* See comment at the top of this file */
2386 return __io_cqring_events(ctx
);
2389 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2391 struct io_rings
*rings
= ctx
->rings
;
2393 /* make sure SQ entry isn't read before tail */
2394 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2397 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2399 unsigned int cflags
;
2401 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2402 cflags
|= IORING_CQE_F_BUFFER
;
2403 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2408 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2410 struct io_buffer
*kbuf
;
2412 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2413 return io_put_kbuf(req
, kbuf
);
2416 static inline bool io_run_task_work(void)
2419 * Not safe to run on exiting task, and the task_work handling will
2420 * not add work to such a task.
2422 if (unlikely(current
->flags
& PF_EXITING
))
2424 if (current
->task_works
) {
2425 __set_current_state(TASK_RUNNING
);
2433 static void io_iopoll_queue(struct list_head
*again
)
2435 struct io_kiocb
*req
;
2438 req
= list_first_entry(again
, struct io_kiocb
, inflight_entry
);
2439 list_del(&req
->inflight_entry
);
2440 __io_complete_rw(req
, -EAGAIN
, 0, NULL
);
2441 } while (!list_empty(again
));
2445 * Find and free completed poll iocbs
2447 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2448 struct list_head
*done
)
2450 struct req_batch rb
;
2451 struct io_kiocb
*req
;
2454 /* order with ->result store in io_complete_rw_iopoll() */
2457 io_init_req_batch(&rb
);
2458 while (!list_empty(done
)) {
2461 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2462 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2464 req
->iopoll_completed
= 0;
2465 list_move_tail(&req
->inflight_entry
, &again
);
2468 list_del(&req
->inflight_entry
);
2470 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2471 cflags
= io_put_rw_kbuf(req
);
2473 __io_cqring_fill_event(req
, req
->result
, cflags
);
2476 if (refcount_dec_and_test(&req
->refs
))
2477 io_req_free_batch(&rb
, req
);
2480 io_commit_cqring(ctx
);
2481 io_cqring_ev_posted_iopoll(ctx
);
2482 io_req_free_batch_finish(ctx
, &rb
);
2484 if (!list_empty(&again
))
2485 io_iopoll_queue(&again
);
2488 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2491 struct io_kiocb
*req
, *tmp
;
2497 * Only spin for completions if we don't have multiple devices hanging
2498 * off our complete list, and we're under the requested amount.
2500 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2503 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2504 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2507 * Move completed and retryable entries to our local lists.
2508 * If we find a request that requires polling, break out
2509 * and complete those lists first, if we have entries there.
2511 if (READ_ONCE(req
->iopoll_completed
)) {
2512 list_move_tail(&req
->inflight_entry
, &done
);
2515 if (!list_empty(&done
))
2518 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2522 /* iopoll may have completed current req */
2523 if (READ_ONCE(req
->iopoll_completed
))
2524 list_move_tail(&req
->inflight_entry
, &done
);
2531 if (!list_empty(&done
))
2532 io_iopoll_complete(ctx
, nr_events
, &done
);
2538 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2539 * non-spinning poll check - we'll still enter the driver poll loop, but only
2540 * as a non-spinning completion check.
2542 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2545 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2548 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2551 if (*nr_events
>= min
)
2559 * We can't just wait for polled events to come to us, we have to actively
2560 * find and complete them.
2562 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2564 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2567 mutex_lock(&ctx
->uring_lock
);
2568 while (!list_empty(&ctx
->iopoll_list
)) {
2569 unsigned int nr_events
= 0;
2571 io_do_iopoll(ctx
, &nr_events
, 0);
2573 /* let it sleep and repeat later if can't complete a request */
2577 * Ensure we allow local-to-the-cpu processing to take place,
2578 * in this case we need to ensure that we reap all events.
2579 * Also let task_work, etc. to progress by releasing the mutex
2581 if (need_resched()) {
2582 mutex_unlock(&ctx
->uring_lock
);
2584 mutex_lock(&ctx
->uring_lock
);
2587 mutex_unlock(&ctx
->uring_lock
);
2590 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2592 unsigned int nr_events
= 0;
2593 int iters
= 0, ret
= 0;
2596 * We disallow the app entering submit/complete with polling, but we
2597 * still need to lock the ring to prevent racing with polled issue
2598 * that got punted to a workqueue.
2600 mutex_lock(&ctx
->uring_lock
);
2603 * Don't enter poll loop if we already have events pending.
2604 * If we do, we can potentially be spinning for commands that
2605 * already triggered a CQE (eg in error).
2607 if (test_bit(0, &ctx
->cq_check_overflow
))
2608 __io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2609 if (io_cqring_events(ctx
))
2613 * If a submit got punted to a workqueue, we can have the
2614 * application entering polling for a command before it gets
2615 * issued. That app will hold the uring_lock for the duration
2616 * of the poll right here, so we need to take a breather every
2617 * now and then to ensure that the issue has a chance to add
2618 * the poll to the issued list. Otherwise we can spin here
2619 * forever, while the workqueue is stuck trying to acquire the
2622 if (!(++iters
& 7)) {
2623 mutex_unlock(&ctx
->uring_lock
);
2625 mutex_lock(&ctx
->uring_lock
);
2628 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2632 } while (min
&& !nr_events
&& !need_resched());
2634 mutex_unlock(&ctx
->uring_lock
);
2638 static void kiocb_end_write(struct io_kiocb
*req
)
2641 * Tell lockdep we inherited freeze protection from submission
2644 if (req
->flags
& REQ_F_ISREG
) {
2645 struct inode
*inode
= file_inode(req
->file
);
2647 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2649 file_end_write(req
->file
);
2652 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
,
2653 struct io_comp_state
*cs
)
2655 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2658 if (kiocb
->ki_flags
& IOCB_WRITE
)
2659 kiocb_end_write(req
);
2661 if (res
!= req
->result
)
2662 req_set_fail_links(req
);
2663 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2664 cflags
= io_put_rw_kbuf(req
);
2665 __io_req_complete(req
, res
, cflags
, cs
);
2669 static bool io_resubmit_prep(struct io_kiocb
*req
, int error
)
2671 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2672 ssize_t ret
= -ECANCELED
;
2673 struct iov_iter iter
;
2681 switch (req
->opcode
) {
2682 case IORING_OP_READV
:
2683 case IORING_OP_READ_FIXED
:
2684 case IORING_OP_READ
:
2687 case IORING_OP_WRITEV
:
2688 case IORING_OP_WRITE_FIXED
:
2689 case IORING_OP_WRITE
:
2693 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2698 if (!req
->async_data
) {
2699 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2702 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2710 req_set_fail_links(req
);
2715 static bool io_rw_reissue(struct io_kiocb
*req
, long res
)
2718 umode_t mode
= file_inode(req
->file
)->i_mode
;
2721 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2723 if ((res
!= -EAGAIN
&& res
!= -EOPNOTSUPP
) || io_wq_current_is_worker())
2726 * If ref is dying, we might be running poll reap from the exit work.
2727 * Don't attempt to reissue from that path, just let it fail with
2730 if (percpu_ref_is_dying(&req
->ctx
->refs
))
2733 lockdep_assert_held(&req
->ctx
->uring_lock
);
2735 ret
= io_sq_thread_acquire_mm_files(req
->ctx
, req
);
2737 if (io_resubmit_prep(req
, ret
)) {
2738 refcount_inc(&req
->refs
);
2739 io_queue_async_work(req
);
2747 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2748 struct io_comp_state
*cs
)
2750 if (!io_rw_reissue(req
, res
))
2751 io_complete_rw_common(&req
->rw
.kiocb
, res
, cs
);
2754 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2756 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2758 __io_complete_rw(req
, res
, res2
, NULL
);
2761 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2763 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2765 if (kiocb
->ki_flags
& IOCB_WRITE
)
2766 kiocb_end_write(req
);
2768 if (res
!= -EAGAIN
&& res
!= req
->result
)
2769 req_set_fail_links(req
);
2771 WRITE_ONCE(req
->result
, res
);
2772 /* order with io_poll_complete() checking ->result */
2774 WRITE_ONCE(req
->iopoll_completed
, 1);
2778 * After the iocb has been issued, it's safe to be found on the poll list.
2779 * Adding the kiocb to the list AFTER submission ensures that we don't
2780 * find it from a io_iopoll_getevents() thread before the issuer is done
2781 * accessing the kiocb cookie.
2783 static void io_iopoll_req_issued(struct io_kiocb
*req
, bool in_async
)
2785 struct io_ring_ctx
*ctx
= req
->ctx
;
2788 * Track whether we have multiple files in our lists. This will impact
2789 * how we do polling eventually, not spinning if we're on potentially
2790 * different devices.
2792 if (list_empty(&ctx
->iopoll_list
)) {
2793 ctx
->poll_multi_file
= false;
2794 } else if (!ctx
->poll_multi_file
) {
2795 struct io_kiocb
*list_req
;
2797 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2799 if (list_req
->file
!= req
->file
)
2800 ctx
->poll_multi_file
= true;
2804 * For fast devices, IO may have already completed. If it has, add
2805 * it to the front so we find it first.
2807 if (READ_ONCE(req
->iopoll_completed
))
2808 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2810 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2813 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2814 * task context or in io worker task context. If current task context is
2815 * sq thread, we don't need to check whether should wake up sq thread.
2817 if (in_async
&& (ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2818 wq_has_sleeper(&ctx
->sq_data
->wait
))
2819 wake_up(&ctx
->sq_data
->wait
);
2822 static inline void __io_state_file_put(struct io_submit_state
*state
)
2824 fput_many(state
->file
, state
->file_refs
);
2825 state
->file_refs
= 0;
2828 static inline void io_state_file_put(struct io_submit_state
*state
)
2830 if (state
->file_refs
)
2831 __io_state_file_put(state
);
2835 * Get as many references to a file as we have IOs left in this submission,
2836 * assuming most submissions are for one file, or at least that each file
2837 * has more than one submission.
2839 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2844 if (state
->file_refs
) {
2845 if (state
->fd
== fd
) {
2849 __io_state_file_put(state
);
2851 state
->file
= fget_many(fd
, state
->ios_left
);
2852 if (unlikely(!state
->file
))
2856 state
->file_refs
= state
->ios_left
- 1;
2860 static bool io_bdev_nowait(struct block_device
*bdev
)
2862 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2866 * If we tracked the file through the SCM inflight mechanism, we could support
2867 * any file. For now, just ensure that anything potentially problematic is done
2870 static bool io_file_supports_async(struct file
*file
, int rw
)
2872 umode_t mode
= file_inode(file
)->i_mode
;
2874 if (S_ISBLK(mode
)) {
2875 if (IS_ENABLED(CONFIG_BLOCK
) &&
2876 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2880 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2882 if (S_ISREG(mode
)) {
2883 if (IS_ENABLED(CONFIG_BLOCK
) &&
2884 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2885 file
->f_op
!= &io_uring_fops
)
2890 /* any ->read/write should understand O_NONBLOCK */
2891 if (file
->f_flags
& O_NONBLOCK
)
2894 if (!(file
->f_mode
& FMODE_NOWAIT
))
2898 return file
->f_op
->read_iter
!= NULL
;
2900 return file
->f_op
->write_iter
!= NULL
;
2903 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2905 struct io_ring_ctx
*ctx
= req
->ctx
;
2906 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2910 if (S_ISREG(file_inode(req
->file
)->i_mode
))
2911 req
->flags
|= REQ_F_ISREG
;
2913 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2914 if (kiocb
->ki_pos
== -1 && !(req
->file
->f_mode
& FMODE_STREAM
)) {
2915 req
->flags
|= REQ_F_CUR_POS
;
2916 kiocb
->ki_pos
= req
->file
->f_pos
;
2918 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2919 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2920 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2924 ioprio
= READ_ONCE(sqe
->ioprio
);
2926 ret
= ioprio_check_cap(ioprio
);
2930 kiocb
->ki_ioprio
= ioprio
;
2932 kiocb
->ki_ioprio
= get_current_ioprio();
2934 /* don't allow async punt if RWF_NOWAIT was requested */
2935 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2936 req
->flags
|= REQ_F_NOWAIT
;
2938 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2939 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2940 !kiocb
->ki_filp
->f_op
->iopoll
)
2943 kiocb
->ki_flags
|= IOCB_HIPRI
;
2944 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2945 req
->iopoll_completed
= 0;
2947 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2949 kiocb
->ki_complete
= io_complete_rw
;
2952 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2953 req
->rw
.len
= READ_ONCE(sqe
->len
);
2954 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2958 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2964 case -ERESTARTNOINTR
:
2965 case -ERESTARTNOHAND
:
2966 case -ERESTART_RESTARTBLOCK
:
2968 * We can't just restart the syscall, since previously
2969 * submitted sqes may already be in progress. Just fail this
2975 kiocb
->ki_complete(kiocb
, ret
, 0);
2979 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2980 struct io_comp_state
*cs
)
2982 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2983 struct io_async_rw
*io
= req
->async_data
;
2985 /* add previously done IO, if any */
2986 if (io
&& io
->bytes_done
> 0) {
2988 ret
= io
->bytes_done
;
2990 ret
+= io
->bytes_done
;
2993 if (req
->flags
& REQ_F_CUR_POS
)
2994 req
->file
->f_pos
= kiocb
->ki_pos
;
2995 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2996 __io_complete_rw(req
, ret
, 0, cs
);
2998 io_rw_done(kiocb
, ret
);
3001 static ssize_t
io_import_fixed(struct io_kiocb
*req
, int rw
,
3002 struct iov_iter
*iter
)
3004 struct io_ring_ctx
*ctx
= req
->ctx
;
3005 size_t len
= req
->rw
.len
;
3006 struct io_mapped_ubuf
*imu
;
3007 u16 index
, buf_index
= req
->buf_index
;
3011 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3013 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3014 imu
= &ctx
->user_bufs
[index
];
3015 buf_addr
= req
->rw
.addr
;
3018 if (buf_addr
+ len
< buf_addr
)
3020 /* not inside the mapped region */
3021 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
3025 * May not be a start of buffer, set size appropriately
3026 * and advance us to the beginning.
3028 offset
= buf_addr
- imu
->ubuf
;
3029 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3033 * Don't use iov_iter_advance() here, as it's really slow for
3034 * using the latter parts of a big fixed buffer - it iterates
3035 * over each segment manually. We can cheat a bit here, because
3038 * 1) it's a BVEC iter, we set it up
3039 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3040 * first and last bvec
3042 * So just find our index, and adjust the iterator afterwards.
3043 * If the offset is within the first bvec (or the whole first
3044 * bvec, just use iov_iter_advance(). This makes it easier
3045 * since we can just skip the first segment, which may not
3046 * be PAGE_SIZE aligned.
3048 const struct bio_vec
*bvec
= imu
->bvec
;
3050 if (offset
<= bvec
->bv_len
) {
3051 iov_iter_advance(iter
, offset
);
3053 unsigned long seg_skip
;
3055 /* skip first vec */
3056 offset
-= bvec
->bv_len
;
3057 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3059 iter
->bvec
= bvec
+ seg_skip
;
3060 iter
->nr_segs
-= seg_skip
;
3061 iter
->count
-= bvec
->bv_len
+ offset
;
3062 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3069 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3072 mutex_unlock(&ctx
->uring_lock
);
3075 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3078 * "Normal" inline submissions always hold the uring_lock, since we
3079 * grab it from the system call. Same is true for the SQPOLL offload.
3080 * The only exception is when we've detached the request and issue it
3081 * from an async worker thread, grab the lock for that case.
3084 mutex_lock(&ctx
->uring_lock
);
3087 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3088 int bgid
, struct io_buffer
*kbuf
,
3091 struct io_buffer
*head
;
3093 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3096 io_ring_submit_lock(req
->ctx
, needs_lock
);
3098 lockdep_assert_held(&req
->ctx
->uring_lock
);
3100 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
3102 if (!list_empty(&head
->list
)) {
3103 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3105 list_del(&kbuf
->list
);
3108 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
3110 if (*len
> kbuf
->len
)
3113 kbuf
= ERR_PTR(-ENOBUFS
);
3116 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3121 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3124 struct io_buffer
*kbuf
;
3127 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3128 bgid
= req
->buf_index
;
3129 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3132 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3133 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3134 return u64_to_user_ptr(kbuf
->addr
);
3137 #ifdef CONFIG_COMPAT
3138 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3141 struct compat_iovec __user
*uiov
;
3142 compat_ssize_t clen
;
3146 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3147 if (!access_ok(uiov
, sizeof(*uiov
)))
3149 if (__get_user(clen
, &uiov
->iov_len
))
3155 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3157 return PTR_ERR(buf
);
3158 iov
[0].iov_base
= buf
;
3159 iov
[0].iov_len
= (compat_size_t
) len
;
3164 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3167 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3171 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3174 len
= iov
[0].iov_len
;
3177 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3179 return PTR_ERR(buf
);
3180 iov
[0].iov_base
= buf
;
3181 iov
[0].iov_len
= len
;
3185 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3188 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3189 struct io_buffer
*kbuf
;
3191 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3192 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3193 iov
[0].iov_len
= kbuf
->len
;
3196 if (req
->rw
.len
!= 1)
3199 #ifdef CONFIG_COMPAT
3200 if (req
->ctx
->compat
)
3201 return io_compat_import(req
, iov
, needs_lock
);
3204 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3207 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
3208 struct iovec
**iovec
, struct iov_iter
*iter
,
3211 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3212 size_t sqe_len
= req
->rw
.len
;
3216 opcode
= req
->opcode
;
3217 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3219 return io_import_fixed(req
, rw
, iter
);
3222 /* buffer index only valid with fixed read/write, or buffer select */
3223 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3226 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3227 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3228 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3230 return PTR_ERR(buf
);
3231 req
->rw
.len
= sqe_len
;
3234 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3239 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3240 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3242 ret
= (*iovec
)->iov_len
;
3243 iov_iter_init(iter
, rw
, *iovec
, 1, ret
);
3249 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3253 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3255 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3259 * For files that don't have ->read_iter() and ->write_iter(), handle them
3260 * by looping over ->read() or ->write() manually.
3262 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3264 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3265 struct file
*file
= req
->file
;
3269 * Don't support polled IO through this interface, and we can't
3270 * support non-blocking either. For the latter, this just causes
3271 * the kiocb to be handled from an async context.
3273 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3275 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3278 while (iov_iter_count(iter
)) {
3282 if (!iov_iter_is_bvec(iter
)) {
3283 iovec
= iov_iter_iovec(iter
);
3285 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3286 iovec
.iov_len
= req
->rw
.len
;
3290 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3291 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3293 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3294 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3303 if (nr
!= iovec
.iov_len
)
3307 iov_iter_advance(iter
, nr
);
3313 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3314 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3316 struct io_async_rw
*rw
= req
->async_data
;
3318 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3319 rw
->free_iovec
= iovec
;
3321 /* can only be fixed buffers, no need to do anything */
3322 if (iov_iter_is_bvec(iter
))
3325 unsigned iov_off
= 0;
3327 rw
->iter
.iov
= rw
->fast_iov
;
3328 if (iter
->iov
!= fast_iov
) {
3329 iov_off
= iter
->iov
- fast_iov
;
3330 rw
->iter
.iov
+= iov_off
;
3332 if (rw
->fast_iov
!= fast_iov
)
3333 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3334 sizeof(struct iovec
) * iter
->nr_segs
);
3336 req
->flags
|= REQ_F_NEED_CLEANUP
;
3340 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3342 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3343 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3344 return req
->async_data
== NULL
;
3347 static int io_alloc_async_data(struct io_kiocb
*req
)
3349 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3352 return __io_alloc_async_data(req
);
3355 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3356 const struct iovec
*fast_iov
,
3357 struct iov_iter
*iter
, bool force
)
3359 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3361 if (!req
->async_data
) {
3362 if (__io_alloc_async_data(req
))
3365 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3370 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3372 struct io_async_rw
*iorw
= req
->async_data
;
3373 struct iovec
*iov
= iorw
->fast_iov
;
3376 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3377 if (unlikely(ret
< 0))
3380 iorw
->bytes_done
= 0;
3381 iorw
->free_iovec
= iov
;
3383 req
->flags
|= REQ_F_NEED_CLEANUP
;
3387 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3391 ret
= io_prep_rw(req
, sqe
);
3395 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3398 /* either don't need iovec imported or already have it */
3399 if (!req
->async_data
)
3401 return io_rw_prep_async(req
, READ
);
3405 * This is our waitqueue callback handler, registered through lock_page_async()
3406 * when we initially tried to do the IO with the iocb armed our waitqueue.
3407 * This gets called when the page is unlocked, and we generally expect that to
3408 * happen when the page IO is completed and the page is now uptodate. This will
3409 * queue a task_work based retry of the operation, attempting to copy the data
3410 * again. If the latter fails because the page was NOT uptodate, then we will
3411 * do a thread based blocking retry of the operation. That's the unexpected
3414 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3415 int sync
, void *arg
)
3417 struct wait_page_queue
*wpq
;
3418 struct io_kiocb
*req
= wait
->private;
3419 struct wait_page_key
*key
= arg
;
3422 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3424 if (!wake_page_match(wpq
, key
))
3427 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3428 list_del_init(&wait
->entry
);
3430 init_task_work(&req
->task_work
, io_req_task_submit
);
3431 percpu_ref_get(&req
->ctx
->refs
);
3433 /* submit ref gets dropped, acquire a new one */
3434 refcount_inc(&req
->refs
);
3435 ret
= io_req_task_work_add(req
);
3437 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
3442 * This controls whether a given IO request should be armed for async page
3443 * based retry. If we return false here, the request is handed to the async
3444 * worker threads for retry. If we're doing buffered reads on a regular file,
3445 * we prepare a private wait_page_queue entry and retry the operation. This
3446 * will either succeed because the page is now uptodate and unlocked, or it
3447 * will register a callback when the page is unlocked at IO completion. Through
3448 * that callback, io_uring uses task_work to setup a retry of the operation.
3449 * That retry will attempt the buffered read again. The retry will generally
3450 * succeed, or in rare cases where it fails, we then fall back to using the
3451 * async worker threads for a blocking retry.
3453 static bool io_rw_should_retry(struct io_kiocb
*req
)
3455 struct io_async_rw
*rw
= req
->async_data
;
3456 struct wait_page_queue
*wait
= &rw
->wpq
;
3457 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3459 /* never retry for NOWAIT, we just complete with -EAGAIN */
3460 if (req
->flags
& REQ_F_NOWAIT
)
3463 /* Only for buffered IO */
3464 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3468 * just use poll if we can, and don't attempt if the fs doesn't
3469 * support callback based unlocks
3471 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3474 wait
->wait
.func
= io_async_buf_func
;
3475 wait
->wait
.private = req
;
3476 wait
->wait
.flags
= 0;
3477 INIT_LIST_HEAD(&wait
->wait
.entry
);
3478 kiocb
->ki_flags
|= IOCB_WAITQ
;
3479 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3480 kiocb
->ki_waitq
= wait
;
3484 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3486 if (req
->file
->f_op
->read_iter
)
3487 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3488 else if (req
->file
->f_op
->read
)
3489 return loop_rw_iter(READ
, req
, iter
);
3494 static int io_read(struct io_kiocb
*req
, bool force_nonblock
,
3495 struct io_comp_state
*cs
)
3497 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3498 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3499 struct iov_iter __iter
, *iter
= &__iter
;
3500 struct io_async_rw
*rw
= req
->async_data
;
3501 ssize_t io_size
, ret
, ret2
;
3508 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3512 io_size
= iov_iter_count(iter
);
3513 req
->result
= io_size
;
3516 /* Ensure we clear previously set non-block flag */
3517 if (!force_nonblock
)
3518 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3520 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3522 /* If the file doesn't support async, just async punt */
3523 no_async
= force_nonblock
&& !io_file_supports_async(req
->file
, READ
);
3527 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3531 ret
= io_iter_do_read(req
, iter
);
3533 if (ret
== -EIOCBQUEUED
) {
3536 } else if (ret
== -EAGAIN
) {
3537 /* IOPOLL retry should happen for io-wq threads */
3538 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3540 /* no retry on NONBLOCK marked file */
3541 if (req
->file
->f_flags
& O_NONBLOCK
)
3543 /* some cases will consume bytes even on error returns */
3544 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3547 } else if (ret
<= 0) {
3548 /* make sure -ERESTARTSYS -> -EINTR is done */
3552 /* read it all, or we did blocking attempt. no retry. */
3553 if (!iov_iter_count(iter
) || !force_nonblock
||
3554 (req
->file
->f_flags
& O_NONBLOCK
) || !(req
->flags
& REQ_F_ISREG
))
3559 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3566 rw
= req
->async_data
;
3567 /* it's copied and will be cleaned with ->io */
3569 /* now use our persistent iterator, if we aren't already */
3572 rw
->bytes_done
+= ret
;
3573 /* if we can retry, do so with the callbacks armed */
3574 if (!io_rw_should_retry(req
)) {
3575 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3580 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3581 * get -EIOCBQUEUED, then we'll get a notification when the desired
3582 * page gets unlocked. We can also get a partial read here, and if we
3583 * do, then just retry at the new offset.
3585 ret
= io_iter_do_read(req
, iter
);
3586 if (ret
== -EIOCBQUEUED
) {
3589 } else if (ret
> 0 && ret
< io_size
) {
3590 /* we got some bytes, but not all. retry. */
3591 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3595 kiocb_done(kiocb
, ret
, cs
);
3598 /* it's reportedly faster than delegating the null check to kfree() */
3604 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3608 ret
= io_prep_rw(req
, sqe
);
3612 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3615 /* either don't need iovec imported or already have it */
3616 if (!req
->async_data
)
3618 return io_rw_prep_async(req
, WRITE
);
3621 static int io_write(struct io_kiocb
*req
, bool force_nonblock
,
3622 struct io_comp_state
*cs
)
3624 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3625 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3626 struct iov_iter __iter
, *iter
= &__iter
;
3627 struct io_async_rw
*rw
= req
->async_data
;
3628 ssize_t ret
, ret2
, io_size
;
3634 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3638 io_size
= iov_iter_count(iter
);
3639 req
->result
= io_size
;
3641 /* Ensure we clear previously set non-block flag */
3642 if (!force_nonblock
)
3643 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3645 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3647 /* If the file doesn't support async, just async punt */
3648 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3651 /* file path doesn't support NOWAIT for non-direct_IO */
3652 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3653 (req
->flags
& REQ_F_ISREG
))
3656 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3661 * Open-code file_start_write here to grab freeze protection,
3662 * which will be released by another thread in
3663 * io_complete_rw(). Fool lockdep by telling it the lock got
3664 * released so that it doesn't complain about the held lock when
3665 * we return to userspace.
3667 if (req
->flags
& REQ_F_ISREG
) {
3668 sb_start_write(file_inode(req
->file
)->i_sb
);
3669 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3672 kiocb
->ki_flags
|= IOCB_WRITE
;
3674 if (req
->file
->f_op
->write_iter
)
3675 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3676 else if (req
->file
->f_op
->write
)
3677 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3682 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3683 * retry them without IOCB_NOWAIT.
3685 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3687 /* no retry on NONBLOCK marked file */
3688 if (ret2
== -EAGAIN
&& (req
->file
->f_flags
& O_NONBLOCK
))
3690 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3691 /* IOPOLL retry should happen for io-wq threads */
3692 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3695 kiocb_done(kiocb
, ret2
, cs
);
3698 /* some cases will consume bytes even on error returns */
3699 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3700 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3705 /* it's reportedly faster than delegating the null check to kfree() */
3711 static int io_renameat_prep(struct io_kiocb
*req
,
3712 const struct io_uring_sqe
*sqe
)
3714 struct io_rename
*ren
= &req
->rename
;
3715 const char __user
*oldf
, *newf
;
3717 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3720 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3721 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3722 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3723 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3724 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3726 ren
->oldpath
= getname(oldf
);
3727 if (IS_ERR(ren
->oldpath
))
3728 return PTR_ERR(ren
->oldpath
);
3730 ren
->newpath
= getname(newf
);
3731 if (IS_ERR(ren
->newpath
)) {
3732 putname(ren
->oldpath
);
3733 return PTR_ERR(ren
->newpath
);
3736 req
->flags
|= REQ_F_NEED_CLEANUP
;
3740 static int io_renameat(struct io_kiocb
*req
, bool force_nonblock
)
3742 struct io_rename
*ren
= &req
->rename
;
3748 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3749 ren
->newpath
, ren
->flags
);
3751 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3753 req_set_fail_links(req
);
3754 io_req_complete(req
, ret
);
3758 static int io_unlinkat_prep(struct io_kiocb
*req
,
3759 const struct io_uring_sqe
*sqe
)
3761 struct io_unlink
*un
= &req
->unlink
;
3762 const char __user
*fname
;
3764 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3767 un
->dfd
= READ_ONCE(sqe
->fd
);
3769 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3770 if (un
->flags
& ~AT_REMOVEDIR
)
3773 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3774 un
->filename
= getname(fname
);
3775 if (IS_ERR(un
->filename
))
3776 return PTR_ERR(un
->filename
);
3778 req
->flags
|= REQ_F_NEED_CLEANUP
;
3782 static int io_unlinkat(struct io_kiocb
*req
, bool force_nonblock
)
3784 struct io_unlink
*un
= &req
->unlink
;
3790 if (un
->flags
& AT_REMOVEDIR
)
3791 ret
= do_rmdir(un
->dfd
, un
->filename
);
3793 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3795 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3797 req_set_fail_links(req
);
3798 io_req_complete(req
, ret
);
3802 static int io_shutdown_prep(struct io_kiocb
*req
,
3803 const struct io_uring_sqe
*sqe
)
3805 #if defined(CONFIG_NET)
3806 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3808 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3812 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3819 static int io_shutdown(struct io_kiocb
*req
, bool force_nonblock
)
3821 #if defined(CONFIG_NET)
3822 struct socket
*sock
;
3828 sock
= sock_from_file(req
->file
);
3829 if (unlikely(!sock
))
3832 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3834 req_set_fail_links(req
);
3835 io_req_complete(req
, ret
);
3842 static int __io_splice_prep(struct io_kiocb
*req
,
3843 const struct io_uring_sqe
*sqe
)
3845 struct io_splice
* sp
= &req
->splice
;
3846 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3848 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3852 sp
->len
= READ_ONCE(sqe
->len
);
3853 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3855 if (unlikely(sp
->flags
& ~valid_flags
))
3858 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3859 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3862 req
->flags
|= REQ_F_NEED_CLEANUP
;
3864 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3866 * Splice operation will be punted aync, and here need to
3867 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3869 io_req_init_async(req
);
3870 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3876 static int io_tee_prep(struct io_kiocb
*req
,
3877 const struct io_uring_sqe
*sqe
)
3879 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3881 return __io_splice_prep(req
, sqe
);
3884 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3886 struct io_splice
*sp
= &req
->splice
;
3887 struct file
*in
= sp
->file_in
;
3888 struct file
*out
= sp
->file_out
;
3889 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3895 ret
= do_tee(in
, out
, sp
->len
, flags
);
3897 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3898 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3901 req_set_fail_links(req
);
3902 io_req_complete(req
, ret
);
3906 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3908 struct io_splice
* sp
= &req
->splice
;
3910 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3911 sp
->off_out
= READ_ONCE(sqe
->off
);
3912 return __io_splice_prep(req
, sqe
);
3915 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3917 struct io_splice
*sp
= &req
->splice
;
3918 struct file
*in
= sp
->file_in
;
3919 struct file
*out
= sp
->file_out
;
3920 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3921 loff_t
*poff_in
, *poff_out
;
3927 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3928 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3931 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3933 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3934 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3937 req_set_fail_links(req
);
3938 io_req_complete(req
, ret
);
3943 * IORING_OP_NOP just posts a completion event, nothing else.
3945 static int io_nop(struct io_kiocb
*req
, struct io_comp_state
*cs
)
3947 struct io_ring_ctx
*ctx
= req
->ctx
;
3949 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3952 __io_req_complete(req
, 0, 0, cs
);
3956 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3958 struct io_ring_ctx
*ctx
= req
->ctx
;
3963 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3965 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3968 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3969 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3972 req
->sync
.off
= READ_ONCE(sqe
->off
);
3973 req
->sync
.len
= READ_ONCE(sqe
->len
);
3977 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3979 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3982 /* fsync always requires a blocking context */
3986 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3987 end
> 0 ? end
: LLONG_MAX
,
3988 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3990 req_set_fail_links(req
);
3991 io_req_complete(req
, ret
);
3995 static int io_fallocate_prep(struct io_kiocb
*req
,
3996 const struct io_uring_sqe
*sqe
)
3998 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
4000 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4003 req
->sync
.off
= READ_ONCE(sqe
->off
);
4004 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4005 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4009 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
4013 /* fallocate always requiring blocking context */
4016 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4019 req_set_fail_links(req
);
4020 io_req_complete(req
, ret
);
4024 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4026 const char __user
*fname
;
4029 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4031 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4034 /* open.how should be already initialised */
4035 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4036 req
->open
.how
.flags
|= O_LARGEFILE
;
4038 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4039 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4040 req
->open
.filename
= getname(fname
);
4041 if (IS_ERR(req
->open
.filename
)) {
4042 ret
= PTR_ERR(req
->open
.filename
);
4043 req
->open
.filename
= NULL
;
4046 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4047 req
->open
.ignore_nonblock
= false;
4048 req
->flags
|= REQ_F_NEED_CLEANUP
;
4052 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4056 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4058 mode
= READ_ONCE(sqe
->len
);
4059 flags
= READ_ONCE(sqe
->open_flags
);
4060 req
->open
.how
= build_open_how(flags
, mode
);
4061 return __io_openat_prep(req
, sqe
);
4064 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4066 struct open_how __user
*how
;
4070 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4072 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4073 len
= READ_ONCE(sqe
->len
);
4074 if (len
< OPEN_HOW_SIZE_VER0
)
4077 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4082 return __io_openat_prep(req
, sqe
);
4085 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
4087 struct open_flags op
;
4091 if (force_nonblock
&& !req
->open
.ignore_nonblock
)
4094 ret
= build_open_flags(&req
->open
.how
, &op
);
4098 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4102 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4105 ret
= PTR_ERR(file
);
4107 * A work-around to ensure that /proc/self works that way
4108 * that it should - if we get -EOPNOTSUPP back, then assume
4109 * that proc_self_get_link() failed us because we're in async
4110 * context. We should be safe to retry this from the task
4111 * itself with force_nonblock == false set, as it should not
4112 * block on lookup. Would be nice to know this upfront and
4113 * avoid the async dance, but doesn't seem feasible.
4115 if (ret
== -EOPNOTSUPP
&& io_wq_current_is_worker()) {
4116 req
->open
.ignore_nonblock
= true;
4117 refcount_inc(&req
->refs
);
4118 io_req_task_queue(req
);
4122 fsnotify_open(file
);
4123 fd_install(ret
, file
);
4126 putname(req
->open
.filename
);
4127 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4129 req_set_fail_links(req
);
4130 io_req_complete(req
, ret
);
4134 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
4136 return io_openat2(req
, force_nonblock
);
4139 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4140 const struct io_uring_sqe
*sqe
)
4142 struct io_provide_buf
*p
= &req
->pbuf
;
4145 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
4148 tmp
= READ_ONCE(sqe
->fd
);
4149 if (!tmp
|| tmp
> USHRT_MAX
)
4152 memset(p
, 0, sizeof(*p
));
4154 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4158 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4159 int bgid
, unsigned nbufs
)
4163 /* shouldn't happen */
4167 /* the head kbuf is the list itself */
4168 while (!list_empty(&buf
->list
)) {
4169 struct io_buffer
*nxt
;
4171 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4172 list_del(&nxt
->list
);
4179 idr_remove(&ctx
->io_buffer_idr
, bgid
);
4184 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4185 struct io_comp_state
*cs
)
4187 struct io_provide_buf
*p
= &req
->pbuf
;
4188 struct io_ring_ctx
*ctx
= req
->ctx
;
4189 struct io_buffer
*head
;
4192 io_ring_submit_lock(ctx
, !force_nonblock
);
4194 lockdep_assert_held(&ctx
->uring_lock
);
4197 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4199 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4201 req_set_fail_links(req
);
4203 /* need to hold the lock to complete IOPOLL requests */
4204 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4205 __io_req_complete(req
, ret
, 0, cs
);
4206 io_ring_submit_unlock(ctx
, !force_nonblock
);
4208 io_ring_submit_unlock(ctx
, !force_nonblock
);
4209 __io_req_complete(req
, ret
, 0, cs
);
4214 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4215 const struct io_uring_sqe
*sqe
)
4218 struct io_provide_buf
*p
= &req
->pbuf
;
4221 if (sqe
->ioprio
|| sqe
->rw_flags
)
4224 tmp
= READ_ONCE(sqe
->fd
);
4225 if (!tmp
|| tmp
> USHRT_MAX
)
4228 p
->addr
= READ_ONCE(sqe
->addr
);
4229 p
->len
= READ_ONCE(sqe
->len
);
4231 size
= (unsigned long)p
->len
* p
->nbufs
;
4232 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4235 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4236 tmp
= READ_ONCE(sqe
->off
);
4237 if (tmp
> USHRT_MAX
)
4243 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4245 struct io_buffer
*buf
;
4246 u64 addr
= pbuf
->addr
;
4247 int i
, bid
= pbuf
->bid
;
4249 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4250 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
4255 buf
->len
= pbuf
->len
;
4260 INIT_LIST_HEAD(&buf
->list
);
4263 list_add_tail(&buf
->list
, &(*head
)->list
);
4267 return i
? i
: -ENOMEM
;
4270 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4271 struct io_comp_state
*cs
)
4273 struct io_provide_buf
*p
= &req
->pbuf
;
4274 struct io_ring_ctx
*ctx
= req
->ctx
;
4275 struct io_buffer
*head
, *list
;
4278 io_ring_submit_lock(ctx
, !force_nonblock
);
4280 lockdep_assert_held(&ctx
->uring_lock
);
4282 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4284 ret
= io_add_buffers(p
, &head
);
4289 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
4292 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4298 req_set_fail_links(req
);
4300 /* need to hold the lock to complete IOPOLL requests */
4301 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4302 __io_req_complete(req
, ret
, 0, cs
);
4303 io_ring_submit_unlock(ctx
, !force_nonblock
);
4305 io_ring_submit_unlock(ctx
, !force_nonblock
);
4306 __io_req_complete(req
, ret
, 0, cs
);
4311 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4312 const struct io_uring_sqe
*sqe
)
4314 #if defined(CONFIG_EPOLL)
4315 if (sqe
->ioprio
|| sqe
->buf_index
)
4317 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4320 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4321 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4322 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4324 if (ep_op_has_event(req
->epoll
.op
)) {
4325 struct epoll_event __user
*ev
;
4327 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4328 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4338 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
,
4339 struct io_comp_state
*cs
)
4341 #if defined(CONFIG_EPOLL)
4342 struct io_epoll
*ie
= &req
->epoll
;
4345 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4346 if (force_nonblock
&& ret
== -EAGAIN
)
4350 req_set_fail_links(req
);
4351 __io_req_complete(req
, ret
, 0, cs
);
4358 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4360 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4361 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4363 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4366 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4367 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4368 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4375 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
4377 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4378 struct io_madvise
*ma
= &req
->madvise
;
4384 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4386 req_set_fail_links(req
);
4387 io_req_complete(req
, ret
);
4394 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4396 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4398 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4401 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4402 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4403 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4407 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
4409 struct io_fadvise
*fa
= &req
->fadvise
;
4412 if (force_nonblock
) {
4413 switch (fa
->advice
) {
4414 case POSIX_FADV_NORMAL
:
4415 case POSIX_FADV_RANDOM
:
4416 case POSIX_FADV_SEQUENTIAL
:
4423 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4425 req_set_fail_links(req
);
4426 io_req_complete(req
, ret
);
4430 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4432 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4434 if (sqe
->ioprio
|| sqe
->buf_index
)
4436 if (req
->flags
& REQ_F_FIXED_FILE
)
4439 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4440 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4441 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4442 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4443 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4448 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
4450 struct io_statx
*ctx
= &req
->statx
;
4453 if (force_nonblock
) {
4454 /* only need file table for an actual valid fd */
4455 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4456 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4460 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4464 req_set_fail_links(req
);
4465 io_req_complete(req
, ret
);
4469 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4471 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4473 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4474 sqe
->rw_flags
|| sqe
->buf_index
)
4476 if (req
->flags
& REQ_F_FIXED_FILE
)
4479 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4483 static int io_close(struct io_kiocb
*req
, bool force_nonblock
,
4484 struct io_comp_state
*cs
)
4486 struct files_struct
*files
= current
->files
;
4487 struct io_close
*close
= &req
->close
;
4488 struct fdtable
*fdt
;
4494 spin_lock(&files
->file_lock
);
4495 fdt
= files_fdtable(files
);
4496 if (close
->fd
>= fdt
->max_fds
) {
4497 spin_unlock(&files
->file_lock
);
4500 file
= fdt
->fd
[close
->fd
];
4502 spin_unlock(&files
->file_lock
);
4506 if (file
->f_op
== &io_uring_fops
) {
4507 spin_unlock(&files
->file_lock
);
4512 /* if the file has a flush method, be safe and punt to async */
4513 if (file
->f_op
->flush
&& force_nonblock
) {
4514 spin_unlock(&files
->file_lock
);
4518 ret
= __close_fd_get_file(close
->fd
, &file
);
4519 spin_unlock(&files
->file_lock
);
4526 /* No ->flush() or already async, safely close from here */
4527 ret
= filp_close(file
, current
->files
);
4530 req_set_fail_links(req
);
4533 __io_req_complete(req
, ret
, 0, cs
);
4537 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4539 struct io_ring_ctx
*ctx
= req
->ctx
;
4544 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4546 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4549 req
->sync
.off
= READ_ONCE(sqe
->off
);
4550 req
->sync
.len
= READ_ONCE(sqe
->len
);
4551 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4555 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
4559 /* sync_file_range always requires a blocking context */
4563 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4566 req_set_fail_links(req
);
4567 io_req_complete(req
, ret
);
4571 #if defined(CONFIG_NET)
4572 static int io_setup_async_msg(struct io_kiocb
*req
,
4573 struct io_async_msghdr
*kmsg
)
4575 struct io_async_msghdr
*async_msg
= req
->async_data
;
4579 if (io_alloc_async_data(req
)) {
4580 if (kmsg
->iov
!= kmsg
->fast_iov
)
4584 async_msg
= req
->async_data
;
4585 req
->flags
|= REQ_F_NEED_CLEANUP
;
4586 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4590 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4591 struct io_async_msghdr
*iomsg
)
4593 iomsg
->iov
= iomsg
->fast_iov
;
4594 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4595 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4596 req
->sr_msg
.msg_flags
, &iomsg
->iov
);
4599 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4601 struct io_async_msghdr
*async_msg
= req
->async_data
;
4602 struct io_sr_msg
*sr
= &req
->sr_msg
;
4605 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4608 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4609 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4610 sr
->len
= READ_ONCE(sqe
->len
);
4612 #ifdef CONFIG_COMPAT
4613 if (req
->ctx
->compat
)
4614 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4617 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4619 ret
= io_sendmsg_copy_hdr(req
, async_msg
);
4621 req
->flags
|= REQ_F_NEED_CLEANUP
;
4625 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4626 struct io_comp_state
*cs
)
4628 struct io_async_msghdr iomsg
, *kmsg
;
4629 struct socket
*sock
;
4634 sock
= sock_from_file(req
->file
);
4635 if (unlikely(!sock
))
4638 if (req
->async_data
) {
4639 kmsg
= req
->async_data
;
4640 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4641 /* if iov is set, it's allocated already */
4643 kmsg
->iov
= kmsg
->fast_iov
;
4644 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4646 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4652 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4653 if (flags
& MSG_DONTWAIT
)
4654 req
->flags
|= REQ_F_NOWAIT
;
4655 else if (force_nonblock
)
4656 flags
|= MSG_DONTWAIT
;
4658 if (flags
& MSG_WAITALL
)
4659 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4661 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4662 if (force_nonblock
&& ret
== -EAGAIN
)
4663 return io_setup_async_msg(req
, kmsg
);
4664 if (ret
== -ERESTARTSYS
)
4667 if (kmsg
->iov
!= kmsg
->fast_iov
)
4669 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4671 req_set_fail_links(req
);
4672 __io_req_complete(req
, ret
, 0, cs
);
4676 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4677 struct io_comp_state
*cs
)
4679 struct io_sr_msg
*sr
= &req
->sr_msg
;
4682 struct socket
*sock
;
4687 sock
= sock_from_file(req
->file
);
4688 if (unlikely(!sock
))
4691 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4695 msg
.msg_name
= NULL
;
4696 msg
.msg_control
= NULL
;
4697 msg
.msg_controllen
= 0;
4698 msg
.msg_namelen
= 0;
4700 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4701 if (flags
& MSG_DONTWAIT
)
4702 req
->flags
|= REQ_F_NOWAIT
;
4703 else if (force_nonblock
)
4704 flags
|= MSG_DONTWAIT
;
4706 if (flags
& MSG_WAITALL
)
4707 min_ret
= iov_iter_count(&msg
.msg_iter
);
4709 msg
.msg_flags
= flags
;
4710 ret
= sock_sendmsg(sock
, &msg
);
4711 if (force_nonblock
&& ret
== -EAGAIN
)
4713 if (ret
== -ERESTARTSYS
)
4717 req_set_fail_links(req
);
4718 __io_req_complete(req
, ret
, 0, cs
);
4722 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4723 struct io_async_msghdr
*iomsg
)
4725 struct io_sr_msg
*sr
= &req
->sr_msg
;
4726 struct iovec __user
*uiov
;
4730 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4731 &iomsg
->uaddr
, &uiov
, &iov_len
);
4735 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4738 if (copy_from_user(iomsg
->iov
, uiov
, sizeof(*uiov
)))
4740 sr
->len
= iomsg
->iov
[0].iov_len
;
4741 iov_iter_init(&iomsg
->msg
.msg_iter
, READ
, iomsg
->iov
, 1,
4745 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4746 &iomsg
->iov
, &iomsg
->msg
.msg_iter
,
4755 #ifdef CONFIG_COMPAT
4756 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4757 struct io_async_msghdr
*iomsg
)
4759 struct compat_msghdr __user
*msg_compat
;
4760 struct io_sr_msg
*sr
= &req
->sr_msg
;
4761 struct compat_iovec __user
*uiov
;
4766 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4767 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4772 uiov
= compat_ptr(ptr
);
4773 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4774 compat_ssize_t clen
;
4778 if (!access_ok(uiov
, sizeof(*uiov
)))
4780 if (__get_user(clen
, &uiov
->iov_len
))
4785 iomsg
->iov
[0].iov_len
= clen
;
4788 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4789 UIO_FASTIOV
, &iomsg
->iov
,
4790 &iomsg
->msg
.msg_iter
, true);
4799 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4800 struct io_async_msghdr
*iomsg
)
4802 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4803 iomsg
->iov
= iomsg
->fast_iov
;
4805 #ifdef CONFIG_COMPAT
4806 if (req
->ctx
->compat
)
4807 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4810 return __io_recvmsg_copy_hdr(req
, iomsg
);
4813 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4816 struct io_sr_msg
*sr
= &req
->sr_msg
;
4817 struct io_buffer
*kbuf
;
4819 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4824 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4828 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4830 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4833 static int io_recvmsg_prep(struct io_kiocb
*req
,
4834 const struct io_uring_sqe
*sqe
)
4836 struct io_async_msghdr
*async_msg
= req
->async_data
;
4837 struct io_sr_msg
*sr
= &req
->sr_msg
;
4840 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4843 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4844 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4845 sr
->len
= READ_ONCE(sqe
->len
);
4846 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4848 #ifdef CONFIG_COMPAT
4849 if (req
->ctx
->compat
)
4850 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4853 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4855 ret
= io_recvmsg_copy_hdr(req
, async_msg
);
4857 req
->flags
|= REQ_F_NEED_CLEANUP
;
4861 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
4862 struct io_comp_state
*cs
)
4864 struct io_async_msghdr iomsg
, *kmsg
;
4865 struct socket
*sock
;
4866 struct io_buffer
*kbuf
;
4869 int ret
, cflags
= 0;
4871 sock
= sock_from_file(req
->file
);
4872 if (unlikely(!sock
))
4875 if (req
->async_data
) {
4876 kmsg
= req
->async_data
;
4877 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4878 /* if iov is set, it's allocated already */
4880 kmsg
->iov
= kmsg
->fast_iov
;
4881 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4883 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4889 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4890 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4892 return PTR_ERR(kbuf
);
4893 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4894 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4895 1, req
->sr_msg
.len
);
4898 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4899 if (flags
& MSG_DONTWAIT
)
4900 req
->flags
|= REQ_F_NOWAIT
;
4901 else if (force_nonblock
)
4902 flags
|= MSG_DONTWAIT
;
4904 if (flags
& MSG_WAITALL
)
4905 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4907 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4908 kmsg
->uaddr
, flags
);
4909 if (force_nonblock
&& ret
== -EAGAIN
)
4910 return io_setup_async_msg(req
, kmsg
);
4911 if (ret
== -ERESTARTSYS
)
4914 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4915 cflags
= io_put_recv_kbuf(req
);
4916 if (kmsg
->iov
!= kmsg
->fast_iov
)
4918 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4919 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4920 req_set_fail_links(req
);
4921 __io_req_complete(req
, ret
, cflags
, cs
);
4925 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
4926 struct io_comp_state
*cs
)
4928 struct io_buffer
*kbuf
;
4929 struct io_sr_msg
*sr
= &req
->sr_msg
;
4931 void __user
*buf
= sr
->buf
;
4932 struct socket
*sock
;
4936 int ret
, cflags
= 0;
4938 sock
= sock_from_file(req
->file
);
4939 if (unlikely(!sock
))
4942 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4943 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4945 return PTR_ERR(kbuf
);
4946 buf
= u64_to_user_ptr(kbuf
->addr
);
4949 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4953 msg
.msg_name
= NULL
;
4954 msg
.msg_control
= NULL
;
4955 msg
.msg_controllen
= 0;
4956 msg
.msg_namelen
= 0;
4957 msg
.msg_iocb
= NULL
;
4960 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4961 if (flags
& MSG_DONTWAIT
)
4962 req
->flags
|= REQ_F_NOWAIT
;
4963 else if (force_nonblock
)
4964 flags
|= MSG_DONTWAIT
;
4966 if (flags
& MSG_WAITALL
)
4967 min_ret
= iov_iter_count(&msg
.msg_iter
);
4969 ret
= sock_recvmsg(sock
, &msg
, flags
);
4970 if (force_nonblock
&& ret
== -EAGAIN
)
4972 if (ret
== -ERESTARTSYS
)
4975 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4976 cflags
= io_put_recv_kbuf(req
);
4977 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4978 req_set_fail_links(req
);
4979 __io_req_complete(req
, ret
, cflags
, cs
);
4983 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4985 struct io_accept
*accept
= &req
->accept
;
4987 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4989 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4992 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4993 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4994 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4995 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4999 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
5000 struct io_comp_state
*cs
)
5002 struct io_accept
*accept
= &req
->accept
;
5003 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5006 if (req
->file
->f_flags
& O_NONBLOCK
)
5007 req
->flags
|= REQ_F_NOWAIT
;
5009 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
5010 accept
->addr_len
, accept
->flags
,
5012 if (ret
== -EAGAIN
&& force_nonblock
)
5015 if (ret
== -ERESTARTSYS
)
5017 req_set_fail_links(req
);
5019 __io_req_complete(req
, ret
, 0, cs
);
5023 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5025 struct io_connect
*conn
= &req
->connect
;
5026 struct io_async_connect
*io
= req
->async_data
;
5028 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5030 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
5033 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5034 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5039 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
5043 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5044 struct io_comp_state
*cs
)
5046 struct io_async_connect __io
, *io
;
5047 unsigned file_flags
;
5050 if (req
->async_data
) {
5051 io
= req
->async_data
;
5053 ret
= move_addr_to_kernel(req
->connect
.addr
,
5054 req
->connect
.addr_len
,
5061 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5063 ret
= __sys_connect_file(req
->file
, &io
->address
,
5064 req
->connect
.addr_len
, file_flags
);
5065 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5066 if (req
->async_data
)
5068 if (io_alloc_async_data(req
)) {
5072 io
= req
->async_data
;
5073 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5076 if (ret
== -ERESTARTSYS
)
5080 req_set_fail_links(req
);
5081 __io_req_complete(req
, ret
, 0, cs
);
5084 #else /* !CONFIG_NET */
5085 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5090 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
5091 struct io_comp_state
*cs
)
5096 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
5097 struct io_comp_state
*cs
)
5102 static int io_recvmsg_prep(struct io_kiocb
*req
,
5103 const struct io_uring_sqe
*sqe
)
5108 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
5109 struct io_comp_state
*cs
)
5114 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
5115 struct io_comp_state
*cs
)
5120 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5125 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
5126 struct io_comp_state
*cs
)
5131 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5136 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5137 struct io_comp_state
*cs
)
5141 #endif /* CONFIG_NET */
5143 struct io_poll_table
{
5144 struct poll_table_struct pt
;
5145 struct io_kiocb
*req
;
5149 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5150 __poll_t mask
, task_work_func_t func
)
5154 /* for instances that support it check for an event match first: */
5155 if (mask
&& !(mask
& poll
->events
))
5158 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5160 list_del_init(&poll
->wait
.entry
);
5163 init_task_work(&req
->task_work
, func
);
5164 percpu_ref_get(&req
->ctx
->refs
);
5167 * If this fails, then the task is exiting. When a task exits, the
5168 * work gets canceled, so just cancel this request as well instead
5169 * of executing it. We can't safely execute it anyway, as we may not
5170 * have the needed state needed for it anyway.
5172 ret
= io_req_task_work_add(req
);
5173 if (unlikely(ret
)) {
5174 WRITE_ONCE(poll
->canceled
, true);
5175 io_req_task_work_add_fallback(req
, func
);
5180 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5181 __acquires(&req
->ctx
->completion_lock
)
5183 struct io_ring_ctx
*ctx
= req
->ctx
;
5185 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5186 struct poll_table_struct pt
= { ._key
= poll
->events
};
5188 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5191 spin_lock_irq(&ctx
->completion_lock
);
5192 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5193 add_wait_queue(poll
->head
, &poll
->wait
);
5200 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5202 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5203 if (req
->opcode
== IORING_OP_POLL_ADD
)
5204 return req
->async_data
;
5205 return req
->apoll
->double_poll
;
5208 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5210 if (req
->opcode
== IORING_OP_POLL_ADD
)
5212 return &req
->apoll
->poll
;
5215 static void io_poll_remove_double(struct io_kiocb
*req
)
5217 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5219 lockdep_assert_held(&req
->ctx
->completion_lock
);
5221 if (poll
&& poll
->head
) {
5222 struct wait_queue_head
*head
= poll
->head
;
5224 spin_lock(&head
->lock
);
5225 list_del_init(&poll
->wait
.entry
);
5226 if (poll
->wait
.private)
5227 refcount_dec(&req
->refs
);
5229 spin_unlock(&head
->lock
);
5233 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
5235 struct io_ring_ctx
*ctx
= req
->ctx
;
5237 io_poll_remove_double(req
);
5238 req
->poll
.done
= true;
5239 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
5240 io_commit_cqring(ctx
);
5243 static void io_poll_task_func(struct callback_head
*cb
)
5245 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5246 struct io_ring_ctx
*ctx
= req
->ctx
;
5247 struct io_kiocb
*nxt
;
5249 if (io_poll_rewait(req
, &req
->poll
)) {
5250 spin_unlock_irq(&ctx
->completion_lock
);
5252 hash_del(&req
->hash_node
);
5253 io_poll_complete(req
, req
->result
, 0);
5254 spin_unlock_irq(&ctx
->completion_lock
);
5256 nxt
= io_put_req_find_next(req
);
5257 io_cqring_ev_posted(ctx
);
5259 __io_req_task_submit(nxt
);
5262 percpu_ref_put(&ctx
->refs
);
5265 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5266 int sync
, void *key
)
5268 struct io_kiocb
*req
= wait
->private;
5269 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5270 __poll_t mask
= key_to_poll(key
);
5272 /* for instances that support it check for an event match first: */
5273 if (mask
&& !(mask
& poll
->events
))
5276 list_del_init(&wait
->entry
);
5278 if (poll
&& poll
->head
) {
5281 spin_lock(&poll
->head
->lock
);
5282 done
= list_empty(&poll
->wait
.entry
);
5284 list_del_init(&poll
->wait
.entry
);
5285 /* make sure double remove sees this as being gone */
5286 wait
->private = NULL
;
5287 spin_unlock(&poll
->head
->lock
);
5289 /* use wait func handler, so it matches the rq type */
5290 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5293 refcount_dec(&req
->refs
);
5297 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5298 wait_queue_func_t wake_func
)
5302 poll
->canceled
= false;
5303 poll
->events
= events
;
5304 INIT_LIST_HEAD(&poll
->wait
.entry
);
5305 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5308 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5309 struct wait_queue_head
*head
,
5310 struct io_poll_iocb
**poll_ptr
)
5312 struct io_kiocb
*req
= pt
->req
;
5315 * If poll->head is already set, it's because the file being polled
5316 * uses multiple waitqueues for poll handling (eg one for read, one
5317 * for write). Setup a separate io_poll_iocb if this happens.
5319 if (unlikely(poll
->head
)) {
5320 struct io_poll_iocb
*poll_one
= poll
;
5322 /* already have a 2nd entry, fail a third attempt */
5324 pt
->error
= -EINVAL
;
5327 /* double add on the same waitqueue head, ignore */
5328 if (poll
->head
== head
)
5330 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5332 pt
->error
= -ENOMEM
;
5335 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5336 refcount_inc(&req
->refs
);
5337 poll
->wait
.private = req
;
5344 if (poll
->events
& EPOLLEXCLUSIVE
)
5345 add_wait_queue_exclusive(head
, &poll
->wait
);
5347 add_wait_queue(head
, &poll
->wait
);
5350 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5351 struct poll_table_struct
*p
)
5353 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5354 struct async_poll
*apoll
= pt
->req
->apoll
;
5356 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5359 static void io_async_task_func(struct callback_head
*cb
)
5361 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5362 struct async_poll
*apoll
= req
->apoll
;
5363 struct io_ring_ctx
*ctx
= req
->ctx
;
5365 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5367 if (io_poll_rewait(req
, &apoll
->poll
)) {
5368 spin_unlock_irq(&ctx
->completion_lock
);
5369 percpu_ref_put(&ctx
->refs
);
5373 /* If req is still hashed, it cannot have been canceled. Don't check. */
5374 if (hash_hashed(&req
->hash_node
))
5375 hash_del(&req
->hash_node
);
5377 io_poll_remove_double(req
);
5378 spin_unlock_irq(&ctx
->completion_lock
);
5380 if (!READ_ONCE(apoll
->poll
.canceled
))
5381 __io_req_task_submit(req
);
5383 __io_req_task_cancel(req
, -ECANCELED
);
5385 percpu_ref_put(&ctx
->refs
);
5386 kfree(apoll
->double_poll
);
5390 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5393 struct io_kiocb
*req
= wait
->private;
5394 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5396 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5399 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5402 static void io_poll_req_insert(struct io_kiocb
*req
)
5404 struct io_ring_ctx
*ctx
= req
->ctx
;
5405 struct hlist_head
*list
;
5407 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5408 hlist_add_head(&req
->hash_node
, list
);
5411 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5412 struct io_poll_iocb
*poll
,
5413 struct io_poll_table
*ipt
, __poll_t mask
,
5414 wait_queue_func_t wake_func
)
5415 __acquires(&ctx
->completion_lock
)
5417 struct io_ring_ctx
*ctx
= req
->ctx
;
5418 bool cancel
= false;
5420 INIT_HLIST_NODE(&req
->hash_node
);
5421 io_init_poll_iocb(poll
, mask
, wake_func
);
5422 poll
->file
= req
->file
;
5423 poll
->wait
.private = req
;
5425 ipt
->pt
._key
= mask
;
5427 ipt
->error
= -EINVAL
;
5429 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5431 spin_lock_irq(&ctx
->completion_lock
);
5432 if (likely(poll
->head
)) {
5433 spin_lock(&poll
->head
->lock
);
5434 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5440 if (mask
|| ipt
->error
)
5441 list_del_init(&poll
->wait
.entry
);
5443 WRITE_ONCE(poll
->canceled
, true);
5444 else if (!poll
->done
) /* actually waiting for an event */
5445 io_poll_req_insert(req
);
5446 spin_unlock(&poll
->head
->lock
);
5452 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5454 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5455 struct io_ring_ctx
*ctx
= req
->ctx
;
5456 struct async_poll
*apoll
;
5457 struct io_poll_table ipt
;
5461 if (!req
->file
|| !file_can_poll(req
->file
))
5463 if (req
->flags
& REQ_F_POLLED
)
5467 else if (def
->pollout
)
5471 /* if we can't nonblock try, then no point in arming a poll handler */
5472 if (!io_file_supports_async(req
->file
, rw
))
5475 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5476 if (unlikely(!apoll
))
5478 apoll
->double_poll
= NULL
;
5480 req
->flags
|= REQ_F_POLLED
;
5485 mask
|= POLLIN
| POLLRDNORM
;
5487 mask
|= POLLOUT
| POLLWRNORM
;
5489 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5490 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5491 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5494 mask
|= POLLERR
| POLLPRI
;
5496 ipt
.pt
._qproc
= io_async_queue_proc
;
5498 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5500 if (ret
|| ipt
.error
) {
5501 io_poll_remove_double(req
);
5502 spin_unlock_irq(&ctx
->completion_lock
);
5503 kfree(apoll
->double_poll
);
5507 spin_unlock_irq(&ctx
->completion_lock
);
5508 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5509 apoll
->poll
.events
);
5513 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5514 struct io_poll_iocb
*poll
)
5516 bool do_complete
= false;
5518 spin_lock(&poll
->head
->lock
);
5519 WRITE_ONCE(poll
->canceled
, true);
5520 if (!list_empty(&poll
->wait
.entry
)) {
5521 list_del_init(&poll
->wait
.entry
);
5524 spin_unlock(&poll
->head
->lock
);
5525 hash_del(&req
->hash_node
);
5529 static bool io_poll_remove_one(struct io_kiocb
*req
)
5533 io_poll_remove_double(req
);
5535 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5536 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5538 struct async_poll
*apoll
= req
->apoll
;
5540 /* non-poll requests have submit ref still */
5541 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5544 kfree(apoll
->double_poll
);
5550 io_cqring_fill_event(req
, -ECANCELED
);
5551 io_commit_cqring(req
->ctx
);
5552 req_set_fail_links(req
);
5553 io_put_req_deferred(req
, 1);
5560 * Returns true if we found and killed one or more poll requests
5562 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5563 struct files_struct
*files
)
5565 struct hlist_node
*tmp
;
5566 struct io_kiocb
*req
;
5569 spin_lock_irq(&ctx
->completion_lock
);
5570 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5571 struct hlist_head
*list
;
5573 list
= &ctx
->cancel_hash
[i
];
5574 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5575 if (io_match_task(req
, tsk
, files
))
5576 posted
+= io_poll_remove_one(req
);
5579 spin_unlock_irq(&ctx
->completion_lock
);
5582 io_cqring_ev_posted(ctx
);
5587 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5589 struct hlist_head
*list
;
5590 struct io_kiocb
*req
;
5592 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5593 hlist_for_each_entry(req
, list
, hash_node
) {
5594 if (sqe_addr
!= req
->user_data
)
5596 if (io_poll_remove_one(req
))
5604 static int io_poll_remove_prep(struct io_kiocb
*req
,
5605 const struct io_uring_sqe
*sqe
)
5607 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5609 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5613 req
->poll_remove
.addr
= READ_ONCE(sqe
->addr
);
5618 * Find a running poll command that matches one specified in sqe->addr,
5619 * and remove it if found.
5621 static int io_poll_remove(struct io_kiocb
*req
)
5623 struct io_ring_ctx
*ctx
= req
->ctx
;
5626 spin_lock_irq(&ctx
->completion_lock
);
5627 ret
= io_poll_cancel(ctx
, req
->poll_remove
.addr
);
5628 spin_unlock_irq(&ctx
->completion_lock
);
5631 req_set_fail_links(req
);
5632 io_req_complete(req
, ret
);
5636 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5639 struct io_kiocb
*req
= wait
->private;
5640 struct io_poll_iocb
*poll
= &req
->poll
;
5642 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5645 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5646 struct poll_table_struct
*p
)
5648 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5650 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5653 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5655 struct io_poll_iocb
*poll
= &req
->poll
;
5658 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5660 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5663 events
= READ_ONCE(sqe
->poll32_events
);
5665 events
= swahw32(events
);
5667 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5668 (events
& EPOLLEXCLUSIVE
);
5672 static int io_poll_add(struct io_kiocb
*req
)
5674 struct io_poll_iocb
*poll
= &req
->poll
;
5675 struct io_ring_ctx
*ctx
= req
->ctx
;
5676 struct io_poll_table ipt
;
5679 ipt
.pt
._qproc
= io_poll_queue_proc
;
5681 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5684 if (mask
) { /* no async, we'd stolen it */
5686 io_poll_complete(req
, mask
, 0);
5688 spin_unlock_irq(&ctx
->completion_lock
);
5691 io_cqring_ev_posted(ctx
);
5697 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5699 struct io_timeout_data
*data
= container_of(timer
,
5700 struct io_timeout_data
, timer
);
5701 struct io_kiocb
*req
= data
->req
;
5702 struct io_ring_ctx
*ctx
= req
->ctx
;
5703 unsigned long flags
;
5705 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5706 list_del_init(&req
->timeout
.list
);
5707 atomic_set(&req
->ctx
->cq_timeouts
,
5708 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5710 io_cqring_fill_event(req
, -ETIME
);
5711 io_commit_cqring(ctx
);
5712 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5714 io_cqring_ev_posted(ctx
);
5715 req_set_fail_links(req
);
5717 return HRTIMER_NORESTART
;
5720 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5723 struct io_timeout_data
*io
;
5724 struct io_kiocb
*req
;
5727 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5728 if (user_data
== req
->user_data
) {
5735 return ERR_PTR(ret
);
5737 io
= req
->async_data
;
5738 ret
= hrtimer_try_to_cancel(&io
->timer
);
5740 return ERR_PTR(-EALREADY
);
5741 list_del_init(&req
->timeout
.list
);
5745 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5747 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5750 return PTR_ERR(req
);
5752 req_set_fail_links(req
);
5753 io_cqring_fill_event(req
, -ECANCELED
);
5754 io_put_req_deferred(req
, 1);
5758 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5759 struct timespec64
*ts
, enum hrtimer_mode mode
)
5761 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5762 struct io_timeout_data
*data
;
5765 return PTR_ERR(req
);
5767 req
->timeout
.off
= 0; /* noseq */
5768 data
= req
->async_data
;
5769 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5770 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5771 data
->timer
.function
= io_timeout_fn
;
5772 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5776 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5777 const struct io_uring_sqe
*sqe
)
5779 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5781 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5783 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5785 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
5788 tr
->addr
= READ_ONCE(sqe
->addr
);
5789 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5790 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5791 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5793 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5795 } else if (tr
->flags
) {
5796 /* timeout removal doesn't support flags */
5804 * Remove or update an existing timeout command
5806 static int io_timeout_remove(struct io_kiocb
*req
)
5808 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5809 struct io_ring_ctx
*ctx
= req
->ctx
;
5812 spin_lock_irq(&ctx
->completion_lock
);
5813 if (req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
) {
5814 enum hrtimer_mode mode
= (tr
->flags
& IORING_TIMEOUT_ABS
)
5815 ? HRTIMER_MODE_ABS
: HRTIMER_MODE_REL
;
5817 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
5819 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5822 io_cqring_fill_event(req
, ret
);
5823 io_commit_cqring(ctx
);
5824 spin_unlock_irq(&ctx
->completion_lock
);
5825 io_cqring_ev_posted(ctx
);
5827 req_set_fail_links(req
);
5832 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5833 bool is_timeout_link
)
5835 struct io_timeout_data
*data
;
5837 u32 off
= READ_ONCE(sqe
->off
);
5839 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5841 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5843 if (off
&& is_timeout_link
)
5845 flags
= READ_ONCE(sqe
->timeout_flags
);
5846 if (flags
& ~IORING_TIMEOUT_ABS
)
5849 req
->timeout
.off
= off
;
5851 if (!req
->async_data
&& io_alloc_async_data(req
))
5854 data
= req
->async_data
;
5857 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5860 if (flags
& IORING_TIMEOUT_ABS
)
5861 data
->mode
= HRTIMER_MODE_ABS
;
5863 data
->mode
= HRTIMER_MODE_REL
;
5865 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5869 static int io_timeout(struct io_kiocb
*req
)
5871 struct io_ring_ctx
*ctx
= req
->ctx
;
5872 struct io_timeout_data
*data
= req
->async_data
;
5873 struct list_head
*entry
;
5874 u32 tail
, off
= req
->timeout
.off
;
5876 spin_lock_irq(&ctx
->completion_lock
);
5879 * sqe->off holds how many events that need to occur for this
5880 * timeout event to be satisfied. If it isn't set, then this is
5881 * a pure timeout request, sequence isn't used.
5883 if (io_is_timeout_noseq(req
)) {
5884 entry
= ctx
->timeout_list
.prev
;
5888 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5889 req
->timeout
.target_seq
= tail
+ off
;
5891 /* Update the last seq here in case io_flush_timeouts() hasn't.
5892 * This is safe because ->completion_lock is held, and submissions
5893 * and completions are never mixed in the same ->completion_lock section.
5895 ctx
->cq_last_tm_flush
= tail
;
5898 * Insertion sort, ensuring the first entry in the list is always
5899 * the one we need first.
5901 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5902 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5905 if (io_is_timeout_noseq(nxt
))
5907 /* nxt.seq is behind @tail, otherwise would've been completed */
5908 if (off
>= nxt
->timeout
.target_seq
- tail
)
5912 list_add(&req
->timeout
.list
, entry
);
5913 data
->timer
.function
= io_timeout_fn
;
5914 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5915 spin_unlock_irq(&ctx
->completion_lock
);
5919 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5921 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5923 return req
->user_data
== (unsigned long) data
;
5926 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5928 enum io_wq_cancel cancel_ret
;
5931 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5932 switch (cancel_ret
) {
5933 case IO_WQ_CANCEL_OK
:
5936 case IO_WQ_CANCEL_RUNNING
:
5939 case IO_WQ_CANCEL_NOTFOUND
:
5947 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5948 struct io_kiocb
*req
, __u64 sqe_addr
,
5951 unsigned long flags
;
5954 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5955 if (ret
!= -ENOENT
) {
5956 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5960 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5961 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5964 ret
= io_poll_cancel(ctx
, sqe_addr
);
5968 io_cqring_fill_event(req
, ret
);
5969 io_commit_cqring(ctx
);
5970 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5971 io_cqring_ev_posted(ctx
);
5974 req_set_fail_links(req
);
5978 static int io_async_cancel_prep(struct io_kiocb
*req
,
5979 const struct io_uring_sqe
*sqe
)
5981 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5983 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5985 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5988 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5992 static int io_async_cancel(struct io_kiocb
*req
)
5994 struct io_ring_ctx
*ctx
= req
->ctx
;
5996 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
6000 static int io_files_update_prep(struct io_kiocb
*req
,
6001 const struct io_uring_sqe
*sqe
)
6003 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
6005 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6007 if (sqe
->ioprio
|| sqe
->rw_flags
)
6010 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
6011 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
6012 if (!req
->files_update
.nr_args
)
6014 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
6018 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
,
6019 struct io_comp_state
*cs
)
6021 struct io_ring_ctx
*ctx
= req
->ctx
;
6022 struct io_uring_files_update up
;
6028 up
.offset
= req
->files_update
.offset
;
6029 up
.fds
= req
->files_update
.arg
;
6031 mutex_lock(&ctx
->uring_lock
);
6032 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
6033 mutex_unlock(&ctx
->uring_lock
);
6036 req_set_fail_links(req
);
6037 __io_req_complete(req
, ret
, 0, cs
);
6041 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6043 switch (req
->opcode
) {
6046 case IORING_OP_READV
:
6047 case IORING_OP_READ_FIXED
:
6048 case IORING_OP_READ
:
6049 return io_read_prep(req
, sqe
);
6050 case IORING_OP_WRITEV
:
6051 case IORING_OP_WRITE_FIXED
:
6052 case IORING_OP_WRITE
:
6053 return io_write_prep(req
, sqe
);
6054 case IORING_OP_POLL_ADD
:
6055 return io_poll_add_prep(req
, sqe
);
6056 case IORING_OP_POLL_REMOVE
:
6057 return io_poll_remove_prep(req
, sqe
);
6058 case IORING_OP_FSYNC
:
6059 return io_prep_fsync(req
, sqe
);
6060 case IORING_OP_SYNC_FILE_RANGE
:
6061 return io_prep_sfr(req
, sqe
);
6062 case IORING_OP_SENDMSG
:
6063 case IORING_OP_SEND
:
6064 return io_sendmsg_prep(req
, sqe
);
6065 case IORING_OP_RECVMSG
:
6066 case IORING_OP_RECV
:
6067 return io_recvmsg_prep(req
, sqe
);
6068 case IORING_OP_CONNECT
:
6069 return io_connect_prep(req
, sqe
);
6070 case IORING_OP_TIMEOUT
:
6071 return io_timeout_prep(req
, sqe
, false);
6072 case IORING_OP_TIMEOUT_REMOVE
:
6073 return io_timeout_remove_prep(req
, sqe
);
6074 case IORING_OP_ASYNC_CANCEL
:
6075 return io_async_cancel_prep(req
, sqe
);
6076 case IORING_OP_LINK_TIMEOUT
:
6077 return io_timeout_prep(req
, sqe
, true);
6078 case IORING_OP_ACCEPT
:
6079 return io_accept_prep(req
, sqe
);
6080 case IORING_OP_FALLOCATE
:
6081 return io_fallocate_prep(req
, sqe
);
6082 case IORING_OP_OPENAT
:
6083 return io_openat_prep(req
, sqe
);
6084 case IORING_OP_CLOSE
:
6085 return io_close_prep(req
, sqe
);
6086 case IORING_OP_FILES_UPDATE
:
6087 return io_files_update_prep(req
, sqe
);
6088 case IORING_OP_STATX
:
6089 return io_statx_prep(req
, sqe
);
6090 case IORING_OP_FADVISE
:
6091 return io_fadvise_prep(req
, sqe
);
6092 case IORING_OP_MADVISE
:
6093 return io_madvise_prep(req
, sqe
);
6094 case IORING_OP_OPENAT2
:
6095 return io_openat2_prep(req
, sqe
);
6096 case IORING_OP_EPOLL_CTL
:
6097 return io_epoll_ctl_prep(req
, sqe
);
6098 case IORING_OP_SPLICE
:
6099 return io_splice_prep(req
, sqe
);
6100 case IORING_OP_PROVIDE_BUFFERS
:
6101 return io_provide_buffers_prep(req
, sqe
);
6102 case IORING_OP_REMOVE_BUFFERS
:
6103 return io_remove_buffers_prep(req
, sqe
);
6105 return io_tee_prep(req
, sqe
);
6106 case IORING_OP_SHUTDOWN
:
6107 return io_shutdown_prep(req
, sqe
);
6108 case IORING_OP_RENAMEAT
:
6109 return io_renameat_prep(req
, sqe
);
6110 case IORING_OP_UNLINKAT
:
6111 return io_unlinkat_prep(req
, sqe
);
6114 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6119 static int io_req_defer_prep(struct io_kiocb
*req
,
6120 const struct io_uring_sqe
*sqe
)
6124 if (io_alloc_async_data(req
))
6126 return io_req_prep(req
, sqe
);
6129 static u32
io_get_sequence(struct io_kiocb
*req
)
6131 struct io_kiocb
*pos
;
6132 struct io_ring_ctx
*ctx
= req
->ctx
;
6133 u32 total_submitted
, nr_reqs
= 0;
6135 io_for_each_link(pos
, req
)
6138 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
6139 return total_submitted
- nr_reqs
;
6142 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6144 struct io_ring_ctx
*ctx
= req
->ctx
;
6145 struct io_defer_entry
*de
;
6149 /* Still need defer if there is pending req in defer list. */
6150 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6151 !(req
->flags
& REQ_F_IO_DRAIN
)))
6154 seq
= io_get_sequence(req
);
6155 /* Still a chance to pass the sequence check */
6156 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6159 if (!req
->async_data
) {
6160 ret
= io_req_defer_prep(req
, sqe
);
6164 io_prep_async_link(req
);
6165 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6169 spin_lock_irq(&ctx
->completion_lock
);
6170 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6171 spin_unlock_irq(&ctx
->completion_lock
);
6173 io_queue_async_work(req
);
6174 return -EIOCBQUEUED
;
6177 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6180 list_add_tail(&de
->list
, &ctx
->defer_list
);
6181 spin_unlock_irq(&ctx
->completion_lock
);
6182 return -EIOCBQUEUED
;
6185 static void io_req_drop_files(struct io_kiocb
*req
)
6187 struct io_ring_ctx
*ctx
= req
->ctx
;
6188 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6189 unsigned long flags
;
6191 if (req
->work
.flags
& IO_WQ_WORK_FILES
) {
6192 put_files_struct(req
->work
.identity
->files
);
6193 put_nsproxy(req
->work
.identity
->nsproxy
);
6195 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
6196 list_del(&req
->inflight_entry
);
6197 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
6198 req
->flags
&= ~REQ_F_INFLIGHT
;
6199 req
->work
.flags
&= ~IO_WQ_WORK_FILES
;
6200 if (atomic_read(&tctx
->in_idle
))
6201 wake_up(&tctx
->wait
);
6204 static void __io_clean_op(struct io_kiocb
*req
)
6206 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6207 switch (req
->opcode
) {
6208 case IORING_OP_READV
:
6209 case IORING_OP_READ_FIXED
:
6210 case IORING_OP_READ
:
6211 kfree((void *)(unsigned long)req
->rw
.addr
);
6213 case IORING_OP_RECVMSG
:
6214 case IORING_OP_RECV
:
6215 kfree(req
->sr_msg
.kbuf
);
6218 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
6221 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6222 switch (req
->opcode
) {
6223 case IORING_OP_READV
:
6224 case IORING_OP_READ_FIXED
:
6225 case IORING_OP_READ
:
6226 case IORING_OP_WRITEV
:
6227 case IORING_OP_WRITE_FIXED
:
6228 case IORING_OP_WRITE
: {
6229 struct io_async_rw
*io
= req
->async_data
;
6231 kfree(io
->free_iovec
);
6234 case IORING_OP_RECVMSG
:
6235 case IORING_OP_SENDMSG
: {
6236 struct io_async_msghdr
*io
= req
->async_data
;
6237 if (io
->iov
!= io
->fast_iov
)
6241 case IORING_OP_SPLICE
:
6243 io_put_file(req
, req
->splice
.file_in
,
6244 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
6246 case IORING_OP_OPENAT
:
6247 case IORING_OP_OPENAT2
:
6248 if (req
->open
.filename
)
6249 putname(req
->open
.filename
);
6251 case IORING_OP_RENAMEAT
:
6252 putname(req
->rename
.oldpath
);
6253 putname(req
->rename
.newpath
);
6255 case IORING_OP_UNLINKAT
:
6256 putname(req
->unlink
.filename
);
6259 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6263 static int io_issue_sqe(struct io_kiocb
*req
, bool force_nonblock
,
6264 struct io_comp_state
*cs
)
6266 struct io_ring_ctx
*ctx
= req
->ctx
;
6269 switch (req
->opcode
) {
6271 ret
= io_nop(req
, cs
);
6273 case IORING_OP_READV
:
6274 case IORING_OP_READ_FIXED
:
6275 case IORING_OP_READ
:
6276 ret
= io_read(req
, force_nonblock
, cs
);
6278 case IORING_OP_WRITEV
:
6279 case IORING_OP_WRITE_FIXED
:
6280 case IORING_OP_WRITE
:
6281 ret
= io_write(req
, force_nonblock
, cs
);
6283 case IORING_OP_FSYNC
:
6284 ret
= io_fsync(req
, force_nonblock
);
6286 case IORING_OP_POLL_ADD
:
6287 ret
= io_poll_add(req
);
6289 case IORING_OP_POLL_REMOVE
:
6290 ret
= io_poll_remove(req
);
6292 case IORING_OP_SYNC_FILE_RANGE
:
6293 ret
= io_sync_file_range(req
, force_nonblock
);
6295 case IORING_OP_SENDMSG
:
6296 ret
= io_sendmsg(req
, force_nonblock
, cs
);
6298 case IORING_OP_SEND
:
6299 ret
= io_send(req
, force_nonblock
, cs
);
6301 case IORING_OP_RECVMSG
:
6302 ret
= io_recvmsg(req
, force_nonblock
, cs
);
6304 case IORING_OP_RECV
:
6305 ret
= io_recv(req
, force_nonblock
, cs
);
6307 case IORING_OP_TIMEOUT
:
6308 ret
= io_timeout(req
);
6310 case IORING_OP_TIMEOUT_REMOVE
:
6311 ret
= io_timeout_remove(req
);
6313 case IORING_OP_ACCEPT
:
6314 ret
= io_accept(req
, force_nonblock
, cs
);
6316 case IORING_OP_CONNECT
:
6317 ret
= io_connect(req
, force_nonblock
, cs
);
6319 case IORING_OP_ASYNC_CANCEL
:
6320 ret
= io_async_cancel(req
);
6322 case IORING_OP_FALLOCATE
:
6323 ret
= io_fallocate(req
, force_nonblock
);
6325 case IORING_OP_OPENAT
:
6326 ret
= io_openat(req
, force_nonblock
);
6328 case IORING_OP_CLOSE
:
6329 ret
= io_close(req
, force_nonblock
, cs
);
6331 case IORING_OP_FILES_UPDATE
:
6332 ret
= io_files_update(req
, force_nonblock
, cs
);
6334 case IORING_OP_STATX
:
6335 ret
= io_statx(req
, force_nonblock
);
6337 case IORING_OP_FADVISE
:
6338 ret
= io_fadvise(req
, force_nonblock
);
6340 case IORING_OP_MADVISE
:
6341 ret
= io_madvise(req
, force_nonblock
);
6343 case IORING_OP_OPENAT2
:
6344 ret
= io_openat2(req
, force_nonblock
);
6346 case IORING_OP_EPOLL_CTL
:
6347 ret
= io_epoll_ctl(req
, force_nonblock
, cs
);
6349 case IORING_OP_SPLICE
:
6350 ret
= io_splice(req
, force_nonblock
);
6352 case IORING_OP_PROVIDE_BUFFERS
:
6353 ret
= io_provide_buffers(req
, force_nonblock
, cs
);
6355 case IORING_OP_REMOVE_BUFFERS
:
6356 ret
= io_remove_buffers(req
, force_nonblock
, cs
);
6359 ret
= io_tee(req
, force_nonblock
);
6361 case IORING_OP_SHUTDOWN
:
6362 ret
= io_shutdown(req
, force_nonblock
);
6364 case IORING_OP_RENAMEAT
:
6365 ret
= io_renameat(req
, force_nonblock
);
6367 case IORING_OP_UNLINKAT
:
6368 ret
= io_unlinkat(req
, force_nonblock
);
6378 /* If the op doesn't have a file, we're not polling for it */
6379 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6380 const bool in_async
= io_wq_current_is_worker();
6382 /* workqueue context doesn't hold uring_lock, grab it now */
6384 mutex_lock(&ctx
->uring_lock
);
6386 io_iopoll_req_issued(req
, in_async
);
6389 mutex_unlock(&ctx
->uring_lock
);
6395 static void io_wq_submit_work(struct io_wq_work
*work
)
6397 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6398 struct io_kiocb
*timeout
;
6401 timeout
= io_prep_linked_timeout(req
);
6403 io_queue_linked_timeout(timeout
);
6405 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
6406 /* io-wq is going to take down one */
6407 refcount_inc(&req
->refs
);
6408 percpu_ref_get(&req
->ctx
->refs
);
6409 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
6415 ret
= io_issue_sqe(req
, false, NULL
);
6417 * We can get EAGAIN for polled IO even though we're
6418 * forcing a sync submission from here, since we can't
6419 * wait for request slots on the block side.
6428 struct io_ring_ctx
*lock_ctx
= NULL
;
6430 if (req
->ctx
->flags
& IORING_SETUP_IOPOLL
)
6431 lock_ctx
= req
->ctx
;
6434 * io_iopoll_complete() does not hold completion_lock to
6435 * complete polled io, so here for polled io, we can not call
6436 * io_req_complete() directly, otherwise there maybe concurrent
6437 * access to cqring, defer_list, etc, which is not safe. Given
6438 * that io_iopoll_complete() is always called under uring_lock,
6439 * so here for polled io, we also get uring_lock to complete
6443 mutex_lock(&lock_ctx
->uring_lock
);
6445 req_set_fail_links(req
);
6446 io_req_complete(req
, ret
);
6449 mutex_unlock(&lock_ctx
->uring_lock
);
6453 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6456 struct fixed_file_table
*table
;
6458 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6459 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6462 static struct file
*io_file_get(struct io_submit_state
*state
,
6463 struct io_kiocb
*req
, int fd
, bool fixed
)
6465 struct io_ring_ctx
*ctx
= req
->ctx
;
6469 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6471 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6472 file
= io_file_from_index(ctx
, fd
);
6473 io_set_resource_node(req
);
6475 trace_io_uring_file_get(ctx
, fd
);
6476 file
= __io_file_get(state
, fd
);
6479 if (file
&& file
->f_op
== &io_uring_fops
&&
6480 !(req
->flags
& REQ_F_INFLIGHT
)) {
6481 io_req_init_async(req
);
6482 req
->flags
|= REQ_F_INFLIGHT
;
6484 spin_lock_irq(&ctx
->inflight_lock
);
6485 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
6486 spin_unlock_irq(&ctx
->inflight_lock
);
6492 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6494 struct io_timeout_data
*data
= container_of(timer
,
6495 struct io_timeout_data
, timer
);
6496 struct io_kiocb
*prev
, *req
= data
->req
;
6497 struct io_ring_ctx
*ctx
= req
->ctx
;
6498 unsigned long flags
;
6500 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6501 prev
= req
->timeout
.head
;
6502 req
->timeout
.head
= NULL
;
6505 * We don't expect the list to be empty, that will only happen if we
6506 * race with the completion of the linked work.
6508 if (prev
&& refcount_inc_not_zero(&prev
->refs
))
6509 io_remove_next_linked(prev
);
6512 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6515 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6516 io_put_req_deferred(prev
, 1);
6518 io_cqring_add_event(req
, -ETIME
, 0);
6519 io_put_req_deferred(req
, 1);
6521 return HRTIMER_NORESTART
;
6524 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6527 * If the back reference is NULL, then our linked request finished
6528 * before we got a chance to setup the timer
6530 if (req
->timeout
.head
) {
6531 struct io_timeout_data
*data
= req
->async_data
;
6533 data
->timer
.function
= io_link_timeout_fn
;
6534 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6539 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6541 struct io_ring_ctx
*ctx
= req
->ctx
;
6543 spin_lock_irq(&ctx
->completion_lock
);
6544 __io_queue_linked_timeout(req
);
6545 spin_unlock_irq(&ctx
->completion_lock
);
6547 /* drop submission reference */
6551 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6553 struct io_kiocb
*nxt
= req
->link
;
6555 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6556 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6559 nxt
->timeout
.head
= req
;
6560 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6561 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6565 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
)
6567 struct io_kiocb
*linked_timeout
;
6568 const struct cred
*old_creds
= NULL
;
6572 linked_timeout
= io_prep_linked_timeout(req
);
6574 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
6575 (req
->work
.flags
& IO_WQ_WORK_CREDS
) &&
6576 req
->work
.identity
->creds
!= current_cred()) {
6578 revert_creds(old_creds
);
6579 if (old_creds
== req
->work
.identity
->creds
)
6580 old_creds
= NULL
; /* restored original creds */
6582 old_creds
= override_creds(req
->work
.identity
->creds
);
6585 ret
= io_issue_sqe(req
, true, cs
);
6588 * We async punt it if the file wasn't marked NOWAIT, or if the file
6589 * doesn't support non-blocking read/write attempts
6591 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6592 if (!io_arm_poll_handler(req
)) {
6594 * Queued up for async execution, worker will release
6595 * submit reference when the iocb is actually submitted.
6597 io_queue_async_work(req
);
6601 io_queue_linked_timeout(linked_timeout
);
6602 } else if (likely(!ret
)) {
6603 /* drop submission reference */
6604 req
= io_put_req_find_next(req
);
6606 io_queue_linked_timeout(linked_timeout
);
6609 if (!(req
->flags
& REQ_F_FORCE_ASYNC
))
6611 io_queue_async_work(req
);
6614 /* un-prep timeout, so it'll be killed as any other linked */
6615 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
6616 req_set_fail_links(req
);
6618 io_req_complete(req
, ret
);
6622 revert_creds(old_creds
);
6625 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6626 struct io_comp_state
*cs
)
6630 ret
= io_req_defer(req
, sqe
);
6632 if (ret
!= -EIOCBQUEUED
) {
6634 req_set_fail_links(req
);
6636 io_req_complete(req
, ret
);
6638 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6639 if (!req
->async_data
) {
6640 ret
= io_req_defer_prep(req
, sqe
);
6644 io_queue_async_work(req
);
6647 ret
= io_req_prep(req
, sqe
);
6651 __io_queue_sqe(req
, cs
);
6655 static inline void io_queue_link_head(struct io_kiocb
*req
,
6656 struct io_comp_state
*cs
)
6658 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
6660 io_req_complete(req
, -ECANCELED
);
6662 io_queue_sqe(req
, NULL
, cs
);
6665 struct io_submit_link
{
6666 struct io_kiocb
*head
;
6667 struct io_kiocb
*last
;
6670 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6671 struct io_submit_link
*link
, struct io_comp_state
*cs
)
6673 struct io_ring_ctx
*ctx
= req
->ctx
;
6677 * If we already have a head request, queue this one for async
6678 * submittal once the head completes. If we don't have a head but
6679 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6680 * submitted sync once the chain is complete. If none of those
6681 * conditions are true (normal request), then just queue it.
6684 struct io_kiocb
*head
= link
->head
;
6687 * Taking sequential execution of a link, draining both sides
6688 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6689 * requests in the link. So, it drains the head and the
6690 * next after the link request. The last one is done via
6691 * drain_next flag to persist the effect across calls.
6693 if (req
->flags
& REQ_F_IO_DRAIN
) {
6694 head
->flags
|= REQ_F_IO_DRAIN
;
6695 ctx
->drain_next
= 1;
6697 ret
= io_req_defer_prep(req
, sqe
);
6698 if (unlikely(ret
)) {
6699 /* fail even hard links since we don't submit */
6700 head
->flags
|= REQ_F_FAIL_LINK
;
6703 trace_io_uring_link(ctx
, req
, head
);
6704 link
->last
->link
= req
;
6707 /* last request of a link, enqueue the link */
6708 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6709 io_queue_link_head(head
, cs
);
6713 if (unlikely(ctx
->drain_next
)) {
6714 req
->flags
|= REQ_F_IO_DRAIN
;
6715 ctx
->drain_next
= 0;
6717 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6718 ret
= io_req_defer_prep(req
, sqe
);
6720 req
->flags
|= REQ_F_FAIL_LINK
;
6724 io_queue_sqe(req
, sqe
, cs
);
6732 * Batched submission is done, ensure local IO is flushed out.
6734 static void io_submit_state_end(struct io_submit_state
*state
)
6736 if (!list_empty(&state
->comp
.list
))
6737 io_submit_flush_completions(&state
->comp
);
6738 if (state
->plug_started
)
6739 blk_finish_plug(&state
->plug
);
6740 io_state_file_put(state
);
6741 if (state
->free_reqs
)
6742 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6746 * Start submission side cache.
6748 static void io_submit_state_start(struct io_submit_state
*state
,
6749 struct io_ring_ctx
*ctx
, unsigned int max_ios
)
6751 state
->plug_started
= false;
6753 INIT_LIST_HEAD(&state
->comp
.list
);
6754 state
->comp
.ctx
= ctx
;
6755 state
->free_reqs
= 0;
6756 state
->file_refs
= 0;
6757 state
->ios_left
= max_ios
;
6760 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6762 struct io_rings
*rings
= ctx
->rings
;
6765 * Ensure any loads from the SQEs are done at this point,
6766 * since once we write the new head, the application could
6767 * write new data to them.
6769 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6773 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6774 * that is mapped by userspace. This means that care needs to be taken to
6775 * ensure that reads are stable, as we cannot rely on userspace always
6776 * being a good citizen. If members of the sqe are validated and then later
6777 * used, it's important that those reads are done through READ_ONCE() to
6778 * prevent a re-load down the line.
6780 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6782 u32
*sq_array
= ctx
->sq_array
;
6786 * The cached sq head (or cq tail) serves two purposes:
6788 * 1) allows us to batch the cost of updating the user visible
6790 * 2) allows the kernel side to track the head on its own, even
6791 * though the application is the one updating it.
6793 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6794 if (likely(head
< ctx
->sq_entries
))
6795 return &ctx
->sq_sqes
[head
];
6797 /* drop invalid entries */
6798 ctx
->cached_sq_dropped
++;
6799 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6803 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6805 ctx
->cached_sq_head
++;
6809 * Check SQE restrictions (opcode and flags).
6811 * Returns 'true' if SQE is allowed, 'false' otherwise.
6813 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6814 struct io_kiocb
*req
,
6815 unsigned int sqe_flags
)
6817 if (!ctx
->restricted
)
6820 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6823 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6824 ctx
->restrictions
.sqe_flags_required
)
6827 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6828 ctx
->restrictions
.sqe_flags_required
))
6834 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6835 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6836 IOSQE_BUFFER_SELECT)
6838 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6839 const struct io_uring_sqe
*sqe
,
6840 struct io_submit_state
*state
)
6842 unsigned int sqe_flags
;
6845 req
->opcode
= READ_ONCE(sqe
->opcode
);
6846 req
->user_data
= READ_ONCE(sqe
->user_data
);
6847 req
->async_data
= NULL
;
6852 req
->fixed_file_refs
= NULL
;
6853 /* one is dropped after submission, the other at completion */
6854 refcount_set(&req
->refs
, 2);
6855 req
->task
= current
;
6858 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6861 if (unlikely(io_sq_thread_acquire_mm_files(ctx
, req
)))
6864 sqe_flags
= READ_ONCE(sqe
->flags
);
6865 /* enforce forwards compatibility on users */
6866 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6869 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6872 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6873 !io_op_defs
[req
->opcode
].buffer_select
)
6876 id
= READ_ONCE(sqe
->personality
);
6878 struct io_identity
*iod
;
6880 iod
= idr_find(&ctx
->personality_idr
, id
);
6883 refcount_inc(&iod
->count
);
6885 __io_req_init_async(req
);
6886 get_cred(iod
->creds
);
6887 req
->work
.identity
= iod
;
6888 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
6891 /* same numerical values with corresponding REQ_F_*, safe to copy */
6892 req
->flags
|= sqe_flags
;
6895 * Plug now if we have more than 1 IO left after this, and the target
6896 * is potentially a read/write to block based storage.
6898 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6899 io_op_defs
[req
->opcode
].plug
) {
6900 blk_start_plug(&state
->plug
);
6901 state
->plug_started
= true;
6905 if (io_op_defs
[req
->opcode
].needs_file
) {
6906 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6908 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6909 if (unlikely(!req
->file
&&
6910 !io_op_defs
[req
->opcode
].needs_file_no_error
))
6918 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6920 struct io_submit_state state
;
6921 struct io_submit_link link
;
6922 int i
, submitted
= 0;
6924 /* if we have a backlog and couldn't flush it all, return BUSY */
6925 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6926 if (!__io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6930 /* make sure SQ entry isn't read before tail */
6931 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6933 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6936 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6937 refcount_add(nr
, ¤t
->usage
);
6939 io_submit_state_start(&state
, ctx
, nr
);
6942 for (i
= 0; i
< nr
; i
++) {
6943 const struct io_uring_sqe
*sqe
;
6944 struct io_kiocb
*req
;
6947 sqe
= io_get_sqe(ctx
);
6948 if (unlikely(!sqe
)) {
6949 io_consume_sqe(ctx
);
6952 req
= io_alloc_req(ctx
, &state
);
6953 if (unlikely(!req
)) {
6955 submitted
= -EAGAIN
;
6958 io_consume_sqe(ctx
);
6959 /* will complete beyond this point, count as submitted */
6962 err
= io_init_req(ctx
, req
, sqe
, &state
);
6963 if (unlikely(err
)) {
6966 io_req_complete(req
, err
);
6970 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6971 true, io_async_submit(ctx
));
6972 err
= io_submit_sqe(req
, sqe
, &link
, &state
.comp
);
6977 if (unlikely(submitted
!= nr
)) {
6978 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6979 struct io_uring_task
*tctx
= current
->io_uring
;
6980 int unused
= nr
- ref_used
;
6982 percpu_ref_put_many(&ctx
->refs
, unused
);
6983 percpu_counter_sub(&tctx
->inflight
, unused
);
6984 put_task_struct_many(current
, unused
);
6987 io_queue_link_head(link
.head
, &state
.comp
);
6988 io_submit_state_end(&state
);
6990 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6991 io_commit_sqring(ctx
);
6996 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6998 /* Tell userspace we may need a wakeup call */
6999 spin_lock_irq(&ctx
->completion_lock
);
7000 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
7001 spin_unlock_irq(&ctx
->completion_lock
);
7004 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
7006 spin_lock_irq(&ctx
->completion_lock
);
7007 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
7008 spin_unlock_irq(&ctx
->completion_lock
);
7011 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
7013 unsigned int to_submit
;
7016 to_submit
= io_sqring_entries(ctx
);
7017 /* if we're handling multiple rings, cap submit size for fairness */
7018 if (cap_entries
&& to_submit
> 8)
7021 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
7022 unsigned nr_events
= 0;
7024 mutex_lock(&ctx
->uring_lock
);
7025 if (!list_empty(&ctx
->iopoll_list
))
7026 io_do_iopoll(ctx
, &nr_events
, 0);
7028 if (to_submit
&& !ctx
->sqo_dead
&&
7029 likely(!percpu_ref_is_dying(&ctx
->refs
)))
7030 ret
= io_submit_sqes(ctx
, to_submit
);
7031 mutex_unlock(&ctx
->uring_lock
);
7034 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
7035 wake_up(&ctx
->sqo_sq_wait
);
7040 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7042 struct io_ring_ctx
*ctx
;
7043 unsigned sq_thread_idle
= 0;
7045 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7046 if (sq_thread_idle
< ctx
->sq_thread_idle
)
7047 sq_thread_idle
= ctx
->sq_thread_idle
;
7050 sqd
->sq_thread_idle
= sq_thread_idle
;
7053 static void io_sqd_init_new(struct io_sq_data
*sqd
)
7055 struct io_ring_ctx
*ctx
;
7057 while (!list_empty(&sqd
->ctx_new_list
)) {
7058 ctx
= list_first_entry(&sqd
->ctx_new_list
, struct io_ring_ctx
, sqd_list
);
7059 list_move_tail(&ctx
->sqd_list
, &sqd
->ctx_list
);
7060 complete(&ctx
->sq_thread_comp
);
7063 io_sqd_update_thread_idle(sqd
);
7066 static int io_sq_thread(void *data
)
7068 struct cgroup_subsys_state
*cur_css
= NULL
;
7069 struct files_struct
*old_files
= current
->files
;
7070 struct nsproxy
*old_nsproxy
= current
->nsproxy
;
7071 const struct cred
*old_cred
= NULL
;
7072 struct io_sq_data
*sqd
= data
;
7073 struct io_ring_ctx
*ctx
;
7074 unsigned long timeout
= 0;
7078 current
->files
= NULL
;
7079 current
->nsproxy
= NULL
;
7080 task_unlock(current
);
7082 while (!kthread_should_stop()) {
7084 bool cap_entries
, sqt_spin
, needs_sched
;
7087 * Any changes to the sqd lists are synchronized through the
7088 * kthread parking. This synchronizes the thread vs users,
7089 * the users are synchronized on the sqd->ctx_lock.
7091 if (kthread_should_park()) {
7094 * When sq thread is unparked, in case the previous park operation
7095 * comes from io_put_sq_data(), which means that sq thread is going
7096 * to be stopped, so here needs to have a check.
7098 if (kthread_should_stop())
7102 if (unlikely(!list_empty(&sqd
->ctx_new_list
))) {
7103 io_sqd_init_new(sqd
);
7104 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7108 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7109 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7110 if (current
->cred
!= ctx
->creds
) {
7112 revert_creds(old_cred
);
7113 old_cred
= override_creds(ctx
->creds
);
7115 io_sq_thread_associate_blkcg(ctx
, &cur_css
);
7117 current
->loginuid
= ctx
->loginuid
;
7118 current
->sessionid
= ctx
->sessionid
;
7121 ret
= __io_sq_thread(ctx
, cap_entries
);
7122 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7125 io_sq_thread_drop_mm_files();
7128 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7130 io_sq_thread_drop_mm_files();
7133 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7137 if (kthread_should_park())
7141 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7142 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7143 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7144 !list_empty_careful(&ctx
->iopoll_list
)) {
7145 needs_sched
= false;
7148 if (io_sqring_entries(ctx
)) {
7149 needs_sched
= false;
7155 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7156 io_ring_set_wakeup_flag(ctx
);
7159 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7160 io_ring_clear_wakeup_flag(ctx
);
7163 finish_wait(&sqd
->wait
, &wait
);
7164 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7168 io_sq_thread_drop_mm_files();
7171 io_sq_thread_unassociate_blkcg();
7173 revert_creds(old_cred
);
7176 current
->files
= old_files
;
7177 current
->nsproxy
= old_nsproxy
;
7178 task_unlock(current
);
7185 struct io_wait_queue
{
7186 struct wait_queue_entry wq
;
7187 struct io_ring_ctx
*ctx
;
7189 unsigned nr_timeouts
;
7192 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7194 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7197 * Wake up if we have enough events, or if a timeout occurred since we
7198 * started waiting. For timeouts, we always want to return to userspace,
7199 * regardless of event count.
7201 return io_cqring_events(ctx
) >= iowq
->to_wait
||
7202 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7205 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7206 int wake_flags
, void *key
)
7208 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7212 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7213 * the task, and the next invocation will do it.
7215 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->cq_check_overflow
))
7216 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7220 static int io_run_task_work_sig(void)
7222 if (io_run_task_work())
7224 if (!signal_pending(current
))
7226 if (test_tsk_thread_flag(current
, TIF_NOTIFY_SIGNAL
))
7227 return -ERESTARTSYS
;
7231 /* when returns >0, the caller should retry */
7232 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7233 struct io_wait_queue
*iowq
,
7234 signed long *timeout
)
7238 /* make sure we run task_work before checking for signals */
7239 ret
= io_run_task_work_sig();
7240 if (ret
|| io_should_wake(iowq
))
7242 /* let the caller flush overflows, retry */
7243 if (test_bit(0, &ctx
->cq_check_overflow
))
7246 *timeout
= schedule_timeout(*timeout
);
7247 return !*timeout
? -ETIME
: 1;
7251 * Wait until events become available, if we don't already have some. The
7252 * application must reap them itself, as they reside on the shared cq ring.
7254 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7255 const sigset_t __user
*sig
, size_t sigsz
,
7256 struct __kernel_timespec __user
*uts
)
7258 struct io_wait_queue iowq
= {
7261 .func
= io_wake_function
,
7262 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
7265 .to_wait
= min_events
,
7267 struct io_rings
*rings
= ctx
->rings
;
7268 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
7272 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7273 if (io_cqring_events(ctx
) >= min_events
)
7275 if (!io_run_task_work())
7280 #ifdef CONFIG_COMPAT
7281 if (in_compat_syscall())
7282 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7286 ret
= set_user_sigmask(sig
, sigsz
);
7293 struct timespec64 ts
;
7295 if (get_timespec64(&ts
, uts
))
7297 timeout
= timespec64_to_jiffies(&ts
);
7300 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7301 trace_io_uring_cqring_wait(ctx
, min_events
);
7303 /* if we can't even flush overflow, don't wait for more */
7304 if (!io_cqring_overflow_flush(ctx
, false, NULL
, NULL
)) {
7308 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
7309 TASK_INTERRUPTIBLE
);
7310 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
7311 finish_wait(&ctx
->wait
, &iowq
.wq
);
7315 restore_saved_sigmask_unless(ret
== -EINTR
);
7317 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7320 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7322 #if defined(CONFIG_UNIX)
7323 if (ctx
->ring_sock
) {
7324 struct sock
*sock
= ctx
->ring_sock
->sk
;
7325 struct sk_buff
*skb
;
7327 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7333 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7336 file
= io_file_from_index(ctx
, i
);
7343 static void io_file_ref_kill(struct percpu_ref
*ref
)
7345 struct fixed_file_data
*data
;
7347 data
= container_of(ref
, struct fixed_file_data
, refs
);
7348 complete(&data
->done
);
7351 static void io_sqe_files_set_node(struct fixed_file_data
*file_data
,
7352 struct fixed_file_ref_node
*ref_node
)
7354 spin_lock_bh(&file_data
->lock
);
7355 file_data
->node
= ref_node
;
7356 list_add_tail(&ref_node
->node
, &file_data
->ref_list
);
7357 spin_unlock_bh(&file_data
->lock
);
7358 percpu_ref_get(&file_data
->refs
);
7361 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7363 struct fixed_file_data
*data
= ctx
->file_data
;
7364 struct fixed_file_ref_node
*backup_node
, *ref_node
= NULL
;
7365 unsigned nr_tables
, i
;
7370 backup_node
= alloc_fixed_file_ref_node(ctx
);
7374 spin_lock_bh(&data
->lock
);
7375 ref_node
= data
->node
;
7376 spin_unlock_bh(&data
->lock
);
7378 percpu_ref_kill(&ref_node
->refs
);
7380 percpu_ref_kill(&data
->refs
);
7382 /* wait for all refs nodes to complete */
7383 flush_delayed_work(&ctx
->file_put_work
);
7385 ret
= wait_for_completion_interruptible(&data
->done
);
7388 ret
= io_run_task_work_sig();
7390 percpu_ref_resurrect(&data
->refs
);
7391 reinit_completion(&data
->done
);
7392 io_sqe_files_set_node(data
, backup_node
);
7397 __io_sqe_files_unregister(ctx
);
7398 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
7399 for (i
= 0; i
< nr_tables
; i
++)
7400 kfree(data
->table
[i
].files
);
7402 percpu_ref_exit(&data
->refs
);
7404 ctx
->file_data
= NULL
;
7405 ctx
->nr_user_files
= 0;
7406 destroy_fixed_file_ref_node(backup_node
);
7410 static void io_put_sq_data(struct io_sq_data
*sqd
)
7412 if (refcount_dec_and_test(&sqd
->refs
)) {
7414 * The park is a bit of a work-around, without it we get
7415 * warning spews on shutdown with SQPOLL set and affinity
7416 * set to a single CPU.
7419 kthread_park(sqd
->thread
);
7420 kthread_stop(sqd
->thread
);
7427 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7429 struct io_ring_ctx
*ctx_attach
;
7430 struct io_sq_data
*sqd
;
7433 f
= fdget(p
->wq_fd
);
7435 return ERR_PTR(-ENXIO
);
7436 if (f
.file
->f_op
!= &io_uring_fops
) {
7438 return ERR_PTR(-EINVAL
);
7441 ctx_attach
= f
.file
->private_data
;
7442 sqd
= ctx_attach
->sq_data
;
7445 return ERR_PTR(-EINVAL
);
7448 refcount_inc(&sqd
->refs
);
7453 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
)
7455 struct io_sq_data
*sqd
;
7457 if (p
->flags
& IORING_SETUP_ATTACH_WQ
)
7458 return io_attach_sq_data(p
);
7460 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7462 return ERR_PTR(-ENOMEM
);
7464 refcount_set(&sqd
->refs
, 1);
7465 INIT_LIST_HEAD(&sqd
->ctx_list
);
7466 INIT_LIST_HEAD(&sqd
->ctx_new_list
);
7467 mutex_init(&sqd
->ctx_lock
);
7468 mutex_init(&sqd
->lock
);
7469 init_waitqueue_head(&sqd
->wait
);
7473 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7474 __releases(&sqd
->lock
)
7478 kthread_unpark(sqd
->thread
);
7479 mutex_unlock(&sqd
->lock
);
7482 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7483 __acquires(&sqd
->lock
)
7487 mutex_lock(&sqd
->lock
);
7488 kthread_park(sqd
->thread
);
7491 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
7493 struct io_sq_data
*sqd
= ctx
->sq_data
;
7498 * We may arrive here from the error branch in
7499 * io_sq_offload_create() where the kthread is created
7500 * without being waked up, thus wake it up now to make
7501 * sure the wait will complete.
7503 wake_up_process(sqd
->thread
);
7504 wait_for_completion(&ctx
->sq_thread_comp
);
7506 io_sq_thread_park(sqd
);
7509 mutex_lock(&sqd
->ctx_lock
);
7510 list_del(&ctx
->sqd_list
);
7511 io_sqd_update_thread_idle(sqd
);
7512 mutex_unlock(&sqd
->ctx_lock
);
7515 io_sq_thread_unpark(sqd
);
7517 io_put_sq_data(sqd
);
7518 ctx
->sq_data
= NULL
;
7522 static void io_finish_async(struct io_ring_ctx
*ctx
)
7524 io_sq_thread_stop(ctx
);
7527 io_wq_destroy(ctx
->io_wq
);
7532 #if defined(CONFIG_UNIX)
7534 * Ensure the UNIX gc is aware of our file set, so we are certain that
7535 * the io_uring can be safely unregistered on process exit, even if we have
7536 * loops in the file referencing.
7538 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7540 struct sock
*sk
= ctx
->ring_sock
->sk
;
7541 struct scm_fp_list
*fpl
;
7542 struct sk_buff
*skb
;
7545 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7549 skb
= alloc_skb(0, GFP_KERNEL
);
7558 fpl
->user
= get_uid(ctx
->user
);
7559 for (i
= 0; i
< nr
; i
++) {
7560 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7564 fpl
->fp
[nr_files
] = get_file(file
);
7565 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7570 fpl
->max
= SCM_MAX_FD
;
7571 fpl
->count
= nr_files
;
7572 UNIXCB(skb
).fp
= fpl
;
7573 skb
->destructor
= unix_destruct_scm
;
7574 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7575 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7577 for (i
= 0; i
< nr_files
; i
++)
7588 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7589 * causes regular reference counting to break down. We rely on the UNIX
7590 * garbage collection to take care of this problem for us.
7592 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7594 unsigned left
, total
;
7598 left
= ctx
->nr_user_files
;
7600 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7602 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7606 total
+= this_files
;
7612 while (total
< ctx
->nr_user_files
) {
7613 struct file
*file
= io_file_from_index(ctx
, total
);
7623 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7629 static int io_sqe_alloc_file_tables(struct fixed_file_data
*file_data
,
7630 unsigned nr_tables
, unsigned nr_files
)
7634 for (i
= 0; i
< nr_tables
; i
++) {
7635 struct fixed_file_table
*table
= &file_data
->table
[i
];
7636 unsigned this_files
;
7638 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7639 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7643 nr_files
-= this_files
;
7649 for (i
= 0; i
< nr_tables
; i
++) {
7650 struct fixed_file_table
*table
= &file_data
->table
[i
];
7651 kfree(table
->files
);
7656 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
7658 #if defined(CONFIG_UNIX)
7659 struct sock
*sock
= ctx
->ring_sock
->sk
;
7660 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7661 struct sk_buff
*skb
;
7664 __skb_queue_head_init(&list
);
7667 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7668 * remove this entry and rearrange the file array.
7670 skb
= skb_dequeue(head
);
7672 struct scm_fp_list
*fp
;
7674 fp
= UNIXCB(skb
).fp
;
7675 for (i
= 0; i
< fp
->count
; i
++) {
7678 if (fp
->fp
[i
] != file
)
7681 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7682 left
= fp
->count
- 1 - i
;
7684 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7685 left
* sizeof(struct file
*));
7692 __skb_queue_tail(&list
, skb
);
7702 __skb_queue_tail(&list
, skb
);
7704 skb
= skb_dequeue(head
);
7707 if (skb_peek(&list
)) {
7708 spin_lock_irq(&head
->lock
);
7709 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7710 __skb_queue_tail(head
, skb
);
7711 spin_unlock_irq(&head
->lock
);
7718 struct io_file_put
{
7719 struct list_head list
;
7723 static void __io_file_put_work(struct fixed_file_ref_node
*ref_node
)
7725 struct fixed_file_data
*file_data
= ref_node
->file_data
;
7726 struct io_ring_ctx
*ctx
= file_data
->ctx
;
7727 struct io_file_put
*pfile
, *tmp
;
7729 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
7730 list_del(&pfile
->list
);
7731 io_ring_file_put(ctx
, pfile
->file
);
7735 percpu_ref_exit(&ref_node
->refs
);
7737 percpu_ref_put(&file_data
->refs
);
7740 static void io_file_put_work(struct work_struct
*work
)
7742 struct io_ring_ctx
*ctx
;
7743 struct llist_node
*node
;
7745 ctx
= container_of(work
, struct io_ring_ctx
, file_put_work
.work
);
7746 node
= llist_del_all(&ctx
->file_put_llist
);
7749 struct fixed_file_ref_node
*ref_node
;
7750 struct llist_node
*next
= node
->next
;
7752 ref_node
= llist_entry(node
, struct fixed_file_ref_node
, llist
);
7753 __io_file_put_work(ref_node
);
7758 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
7760 struct fixed_file_ref_node
*ref_node
;
7761 struct fixed_file_data
*data
;
7762 struct io_ring_ctx
*ctx
;
7763 bool first_add
= false;
7766 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
7767 data
= ref_node
->file_data
;
7770 spin_lock_bh(&data
->lock
);
7771 ref_node
->done
= true;
7773 while (!list_empty(&data
->ref_list
)) {
7774 ref_node
= list_first_entry(&data
->ref_list
,
7775 struct fixed_file_ref_node
, node
);
7776 /* recycle ref nodes in order */
7777 if (!ref_node
->done
)
7779 list_del(&ref_node
->node
);
7780 first_add
|= llist_add(&ref_node
->llist
, &ctx
->file_put_llist
);
7782 spin_unlock_bh(&data
->lock
);
7784 if (percpu_ref_is_dying(&data
->refs
))
7788 mod_delayed_work(system_wq
, &ctx
->file_put_work
, 0);
7790 queue_delayed_work(system_wq
, &ctx
->file_put_work
, delay
);
7793 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
7794 struct io_ring_ctx
*ctx
)
7796 struct fixed_file_ref_node
*ref_node
;
7798 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7802 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
7807 INIT_LIST_HEAD(&ref_node
->node
);
7808 INIT_LIST_HEAD(&ref_node
->file_list
);
7809 ref_node
->file_data
= ctx
->file_data
;
7810 ref_node
->done
= false;
7814 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
7816 percpu_ref_exit(&ref_node
->refs
);
7820 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7823 __s32 __user
*fds
= (__s32 __user
*) arg
;
7824 unsigned nr_tables
, i
;
7826 int fd
, ret
= -ENOMEM
;
7827 struct fixed_file_ref_node
*ref_node
;
7828 struct fixed_file_data
*file_data
;
7834 if (nr_args
> IORING_MAX_FIXED_FILES
)
7837 file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL
);
7840 file_data
->ctx
= ctx
;
7841 init_completion(&file_data
->done
);
7842 INIT_LIST_HEAD(&file_data
->ref_list
);
7843 spin_lock_init(&file_data
->lock
);
7845 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7846 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7848 if (!file_data
->table
)
7851 if (percpu_ref_init(&file_data
->refs
, io_file_ref_kill
,
7852 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
7855 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7857 ctx
->file_data
= file_data
;
7859 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7860 struct fixed_file_table
*table
;
7863 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7867 /* allow sparse sets */
7877 * Don't allow io_uring instances to be registered. If UNIX
7878 * isn't enabled, then this causes a reference cycle and this
7879 * instance can never get freed. If UNIX is enabled we'll
7880 * handle it just fine, but there's still no point in allowing
7881 * a ring fd as it doesn't support regular read/write anyway.
7883 if (file
->f_op
== &io_uring_fops
) {
7887 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7888 index
= i
& IORING_FILE_TABLE_MASK
;
7889 table
->files
[index
] = file
;
7892 ret
= io_sqe_files_scm(ctx
);
7894 io_sqe_files_unregister(ctx
);
7898 ref_node
= alloc_fixed_file_ref_node(ctx
);
7900 io_sqe_files_unregister(ctx
);
7904 io_sqe_files_set_node(file_data
, ref_node
);
7907 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7908 file
= io_file_from_index(ctx
, i
);
7912 for (i
= 0; i
< nr_tables
; i
++)
7913 kfree(file_data
->table
[i
].files
);
7914 ctx
->nr_user_files
= 0;
7916 percpu_ref_exit(&file_data
->refs
);
7918 kfree(file_data
->table
);
7920 ctx
->file_data
= NULL
;
7924 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7927 #if defined(CONFIG_UNIX)
7928 struct sock
*sock
= ctx
->ring_sock
->sk
;
7929 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7930 struct sk_buff
*skb
;
7933 * See if we can merge this file into an existing skb SCM_RIGHTS
7934 * file set. If there's no room, fall back to allocating a new skb
7935 * and filling it in.
7937 spin_lock_irq(&head
->lock
);
7938 skb
= skb_peek(head
);
7940 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7942 if (fpl
->count
< SCM_MAX_FD
) {
7943 __skb_unlink(skb
, head
);
7944 spin_unlock_irq(&head
->lock
);
7945 fpl
->fp
[fpl
->count
] = get_file(file
);
7946 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7948 spin_lock_irq(&head
->lock
);
7949 __skb_queue_head(head
, skb
);
7954 spin_unlock_irq(&head
->lock
);
7961 return __io_sqe_files_scm(ctx
, 1, index
);
7967 static int io_queue_file_removal(struct fixed_file_data
*data
,
7970 struct io_file_put
*pfile
;
7971 struct fixed_file_ref_node
*ref_node
= data
->node
;
7973 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
7978 list_add(&pfile
->list
, &ref_node
->file_list
);
7983 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7984 struct io_uring_files_update
*up
,
7987 struct fixed_file_data
*data
= ctx
->file_data
;
7988 struct fixed_file_ref_node
*ref_node
;
7993 bool needs_switch
= false;
7995 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7997 if (done
> ctx
->nr_user_files
)
8000 ref_node
= alloc_fixed_file_ref_node(ctx
);
8005 fds
= u64_to_user_ptr(up
->fds
);
8007 struct fixed_file_table
*table
;
8011 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8015 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
8016 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
8017 index
= i
& IORING_FILE_TABLE_MASK
;
8018 if (table
->files
[index
]) {
8019 file
= table
->files
[index
];
8020 err
= io_queue_file_removal(data
, file
);
8023 table
->files
[index
] = NULL
;
8024 needs_switch
= true;
8033 * Don't allow io_uring instances to be registered. If
8034 * UNIX isn't enabled, then this causes a reference
8035 * cycle and this instance can never get freed. If UNIX
8036 * is enabled we'll handle it just fine, but there's
8037 * still no point in allowing a ring fd as it doesn't
8038 * support regular read/write anyway.
8040 if (file
->f_op
== &io_uring_fops
) {
8045 table
->files
[index
] = file
;
8046 err
= io_sqe_file_register(ctx
, file
, i
);
8048 table
->files
[index
] = NULL
;
8059 percpu_ref_kill(&data
->node
->refs
);
8060 io_sqe_files_set_node(data
, ref_node
);
8062 destroy_fixed_file_ref_node(ref_node
);
8064 return done
? done
: err
;
8067 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
8070 struct io_uring_files_update up
;
8072 if (!ctx
->file_data
)
8076 if (copy_from_user(&up
, arg
, sizeof(up
)))
8081 return __io_sqe_files_update(ctx
, &up
, nr_args
);
8084 static struct io_wq_work
*io_free_work(struct io_wq_work
*work
)
8086 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8088 req
= io_put_req_find_next(req
);
8089 return req
? &req
->work
: NULL
;
8092 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
8093 struct io_uring_params
*p
)
8095 struct io_wq_data data
;
8097 struct io_ring_ctx
*ctx_attach
;
8098 unsigned int concurrency
;
8101 data
.user
= ctx
->user
;
8102 data
.free_work
= io_free_work
;
8103 data
.do_work
= io_wq_submit_work
;
8105 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
8106 /* Do QD, or 4 * CPUS, whatever is smallest */
8107 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8109 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
8110 if (IS_ERR(ctx
->io_wq
)) {
8111 ret
= PTR_ERR(ctx
->io_wq
);
8117 f
= fdget(p
->wq_fd
);
8121 if (f
.file
->f_op
!= &io_uring_fops
) {
8126 ctx_attach
= f
.file
->private_data
;
8127 /* @io_wq is protected by holding the fd */
8128 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
8133 ctx
->io_wq
= ctx_attach
->io_wq
;
8139 static int io_uring_alloc_task_context(struct task_struct
*task
)
8141 struct io_uring_task
*tctx
;
8144 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
8145 if (unlikely(!tctx
))
8148 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8149 if (unlikely(ret
)) {
8155 init_waitqueue_head(&tctx
->wait
);
8157 atomic_set(&tctx
->in_idle
, 0);
8158 tctx
->sqpoll
= false;
8159 io_init_identity(&tctx
->__identity
);
8160 tctx
->identity
= &tctx
->__identity
;
8161 task
->io_uring
= tctx
;
8165 void __io_uring_free(struct task_struct
*tsk
)
8167 struct io_uring_task
*tctx
= tsk
->io_uring
;
8169 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8170 WARN_ON_ONCE(refcount_read(&tctx
->identity
->count
) != 1);
8171 if (tctx
->identity
!= &tctx
->__identity
)
8172 kfree(tctx
->identity
);
8173 percpu_counter_destroy(&tctx
->inflight
);
8175 tsk
->io_uring
= NULL
;
8178 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8179 struct io_uring_params
*p
)
8183 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8184 struct io_sq_data
*sqd
;
8187 if (!capable(CAP_SYS_ADMIN
) && !capable(CAP_SYS_NICE
))
8190 sqd
= io_get_sq_data(p
);
8197 io_sq_thread_park(sqd
);
8198 mutex_lock(&sqd
->ctx_lock
);
8199 list_add(&ctx
->sqd_list
, &sqd
->ctx_new_list
);
8200 mutex_unlock(&sqd
->ctx_lock
);
8201 io_sq_thread_unpark(sqd
);
8203 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8204 if (!ctx
->sq_thread_idle
)
8205 ctx
->sq_thread_idle
= HZ
;
8210 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8211 int cpu
= p
->sq_thread_cpu
;
8214 if (cpu
>= nr_cpu_ids
)
8216 if (!cpu_online(cpu
))
8219 sqd
->thread
= kthread_create_on_cpu(io_sq_thread
, sqd
,
8220 cpu
, "io_uring-sq");
8222 sqd
->thread
= kthread_create(io_sq_thread
, sqd
,
8225 if (IS_ERR(sqd
->thread
)) {
8226 ret
= PTR_ERR(sqd
->thread
);
8230 ret
= io_uring_alloc_task_context(sqd
->thread
);
8233 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8234 /* Can't have SQ_AFF without SQPOLL */
8240 ret
= io_init_wq_offload(ctx
, p
);
8246 io_finish_async(ctx
);
8250 static void io_sq_offload_start(struct io_ring_ctx
*ctx
)
8252 struct io_sq_data
*sqd
= ctx
->sq_data
;
8254 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && sqd
->thread
)
8255 wake_up_process(sqd
->thread
);
8258 static inline void __io_unaccount_mem(struct user_struct
*user
,
8259 unsigned long nr_pages
)
8261 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8264 static inline int __io_account_mem(struct user_struct
*user
,
8265 unsigned long nr_pages
)
8267 unsigned long page_limit
, cur_pages
, new_pages
;
8269 /* Don't allow more pages than we can safely lock */
8270 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8273 cur_pages
= atomic_long_read(&user
->locked_vm
);
8274 new_pages
= cur_pages
+ nr_pages
;
8275 if (new_pages
> page_limit
)
8277 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8278 new_pages
) != cur_pages
);
8283 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8284 enum io_mem_account acct
)
8287 __io_unaccount_mem(ctx
->user
, nr_pages
);
8289 if (ctx
->mm_account
) {
8290 if (acct
== ACCT_LOCKED
) {
8291 mmap_write_lock(ctx
->mm_account
);
8292 ctx
->mm_account
->locked_vm
-= nr_pages
;
8293 mmap_write_unlock(ctx
->mm_account
);
8294 }else if (acct
== ACCT_PINNED
) {
8295 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8300 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8301 enum io_mem_account acct
)
8305 if (ctx
->limit_mem
) {
8306 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8311 if (ctx
->mm_account
) {
8312 if (acct
== ACCT_LOCKED
) {
8313 mmap_write_lock(ctx
->mm_account
);
8314 ctx
->mm_account
->locked_vm
+= nr_pages
;
8315 mmap_write_unlock(ctx
->mm_account
);
8316 } else if (acct
== ACCT_PINNED
) {
8317 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8324 static void io_mem_free(void *ptr
)
8331 page
= virt_to_head_page(ptr
);
8332 if (put_page_testzero(page
))
8333 free_compound_page(page
);
8336 static void *io_mem_alloc(size_t size
)
8338 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8341 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8344 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8347 struct io_rings
*rings
;
8348 size_t off
, sq_array_size
;
8350 off
= struct_size(rings
, cqes
, cq_entries
);
8351 if (off
== SIZE_MAX
)
8355 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8363 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8364 if (sq_array_size
== SIZE_MAX
)
8367 if (check_add_overflow(off
, sq_array_size
, &off
))
8373 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
8377 pages
= (size_t)1 << get_order(
8378 rings_size(sq_entries
, cq_entries
, NULL
));
8379 pages
+= (size_t)1 << get_order(
8380 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
8385 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
8389 if (!ctx
->user_bufs
)
8392 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8393 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8395 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
8396 unpin_user_page(imu
->bvec
[j
].bv_page
);
8398 if (imu
->acct_pages
)
8399 io_unaccount_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8404 kfree(ctx
->user_bufs
);
8405 ctx
->user_bufs
= NULL
;
8406 ctx
->nr_user_bufs
= 0;
8410 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8411 void __user
*arg
, unsigned index
)
8413 struct iovec __user
*src
;
8415 #ifdef CONFIG_COMPAT
8417 struct compat_iovec __user
*ciovs
;
8418 struct compat_iovec ciov
;
8420 ciovs
= (struct compat_iovec __user
*) arg
;
8421 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8424 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8425 dst
->iov_len
= ciov
.iov_len
;
8429 src
= (struct iovec __user
*) arg
;
8430 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8436 * Not super efficient, but this is just a registration time. And we do cache
8437 * the last compound head, so generally we'll only do a full search if we don't
8440 * We check if the given compound head page has already been accounted, to
8441 * avoid double accounting it. This allows us to account the full size of the
8442 * page, not just the constituent pages of a huge page.
8444 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8445 int nr_pages
, struct page
*hpage
)
8449 /* check current page array */
8450 for (i
= 0; i
< nr_pages
; i
++) {
8451 if (!PageCompound(pages
[i
]))
8453 if (compound_head(pages
[i
]) == hpage
)
8457 /* check previously registered pages */
8458 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8459 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8461 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8462 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8464 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8472 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8473 int nr_pages
, struct io_mapped_ubuf
*imu
,
8474 struct page
**last_hpage
)
8478 for (i
= 0; i
< nr_pages
; i
++) {
8479 if (!PageCompound(pages
[i
])) {
8484 hpage
= compound_head(pages
[i
]);
8485 if (hpage
== *last_hpage
)
8487 *last_hpage
= hpage
;
8488 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8490 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8494 if (!imu
->acct_pages
)
8497 ret
= io_account_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8499 imu
->acct_pages
= 0;
8503 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8506 struct vm_area_struct
**vmas
= NULL
;
8507 struct page
**pages
= NULL
;
8508 struct page
*last_hpage
= NULL
;
8509 int i
, j
, got_pages
= 0;
8514 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8517 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8519 if (!ctx
->user_bufs
)
8522 for (i
= 0; i
< nr_args
; i
++) {
8523 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8524 unsigned long off
, start
, end
, ubuf
;
8529 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8534 * Don't impose further limits on the size and buffer
8535 * constraints here, we'll -EINVAL later when IO is
8536 * submitted if they are wrong.
8539 if (!iov
.iov_base
|| !iov
.iov_len
)
8542 /* arbitrary limit, but we need something */
8543 if (iov
.iov_len
> SZ_1G
)
8546 ubuf
= (unsigned long) iov
.iov_base
;
8547 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8548 start
= ubuf
>> PAGE_SHIFT
;
8549 nr_pages
= end
- start
;
8552 if (!pages
|| nr_pages
> got_pages
) {
8555 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
8557 vmas
= kvmalloc_array(nr_pages
,
8558 sizeof(struct vm_area_struct
*),
8560 if (!pages
|| !vmas
) {
8564 got_pages
= nr_pages
;
8567 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8574 mmap_read_lock(current
->mm
);
8575 pret
= pin_user_pages(ubuf
, nr_pages
,
8576 FOLL_WRITE
| FOLL_LONGTERM
,
8578 if (pret
== nr_pages
) {
8579 /* don't support file backed memory */
8580 for (j
= 0; j
< nr_pages
; j
++) {
8581 struct vm_area_struct
*vma
= vmas
[j
];
8584 !is_file_hugepages(vma
->vm_file
)) {
8590 ret
= pret
< 0 ? pret
: -EFAULT
;
8592 mmap_read_unlock(current
->mm
);
8595 * if we did partial map, or found file backed vmas,
8596 * release any pages we did get
8599 unpin_user_pages(pages
, pret
);
8604 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, &last_hpage
);
8606 unpin_user_pages(pages
, pret
);
8611 off
= ubuf
& ~PAGE_MASK
;
8613 for (j
= 0; j
< nr_pages
; j
++) {
8616 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8617 imu
->bvec
[j
].bv_page
= pages
[j
];
8618 imu
->bvec
[j
].bv_len
= vec_len
;
8619 imu
->bvec
[j
].bv_offset
= off
;
8623 /* store original address for later verification */
8625 imu
->len
= iov
.iov_len
;
8626 imu
->nr_bvecs
= nr_pages
;
8628 ctx
->nr_user_bufs
++;
8636 io_sqe_buffer_unregister(ctx
);
8640 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8642 __s32 __user
*fds
= arg
;
8648 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8651 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8652 if (IS_ERR(ctx
->cq_ev_fd
)) {
8653 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8654 ctx
->cq_ev_fd
= NULL
;
8661 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8663 if (ctx
->cq_ev_fd
) {
8664 eventfd_ctx_put(ctx
->cq_ev_fd
);
8665 ctx
->cq_ev_fd
= NULL
;
8672 static int __io_destroy_buffers(int id
, void *p
, void *data
)
8674 struct io_ring_ctx
*ctx
= data
;
8675 struct io_buffer
*buf
= p
;
8677 __io_remove_buffers(ctx
, buf
, id
, -1U);
8681 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8683 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
8684 idr_destroy(&ctx
->io_buffer_idr
);
8687 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8689 io_finish_async(ctx
);
8690 io_sqe_buffer_unregister(ctx
);
8692 if (ctx
->sqo_task
) {
8693 put_task_struct(ctx
->sqo_task
);
8694 ctx
->sqo_task
= NULL
;
8695 mmdrop(ctx
->mm_account
);
8696 ctx
->mm_account
= NULL
;
8699 #ifdef CONFIG_BLK_CGROUP
8700 if (ctx
->sqo_blkcg_css
)
8701 css_put(ctx
->sqo_blkcg_css
);
8704 io_sqe_files_unregister(ctx
);
8705 io_eventfd_unregister(ctx
);
8706 io_destroy_buffers(ctx
);
8707 idr_destroy(&ctx
->personality_idr
);
8709 #if defined(CONFIG_UNIX)
8710 if (ctx
->ring_sock
) {
8711 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8712 sock_release(ctx
->ring_sock
);
8716 io_mem_free(ctx
->rings
);
8717 io_mem_free(ctx
->sq_sqes
);
8719 percpu_ref_exit(&ctx
->refs
);
8720 free_uid(ctx
->user
);
8721 put_cred(ctx
->creds
);
8722 kfree(ctx
->cancel_hash
);
8723 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
8727 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8729 struct io_ring_ctx
*ctx
= file
->private_data
;
8732 poll_wait(file
, &ctx
->cq_wait
, wait
);
8734 * synchronizes with barrier from wq_has_sleeper call in
8738 if (!io_sqring_full(ctx
))
8739 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8742 * Don't flush cqring overflow list here, just do a simple check.
8743 * Otherwise there could possible be ABBA deadlock:
8746 * lock(&ctx->uring_lock);
8748 * lock(&ctx->uring_lock);
8751 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8752 * pushs them to do the flush.
8754 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->cq_check_overflow
))
8755 mask
|= EPOLLIN
| EPOLLRDNORM
;
8760 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8762 struct io_ring_ctx
*ctx
= file
->private_data
;
8764 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8767 static int io_remove_personalities(int id
, void *p
, void *data
)
8769 struct io_ring_ctx
*ctx
= data
;
8770 struct io_identity
*iod
;
8772 iod
= idr_remove(&ctx
->personality_idr
, id
);
8774 put_cred(iod
->creds
);
8775 if (refcount_dec_and_test(&iod
->count
))
8781 static void io_ring_exit_work(struct work_struct
*work
)
8783 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
8787 * If we're doing polled IO and end up having requests being
8788 * submitted async (out-of-line), then completions can come in while
8789 * we're waiting for refs to drop. We need to reap these manually,
8790 * as nobody else will be looking for them.
8793 io_uring_try_cancel_requests(ctx
, NULL
, NULL
);
8794 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8795 io_ring_ctx_free(ctx
);
8798 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8800 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8802 return req
->ctx
== data
;
8805 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8807 mutex_lock(&ctx
->uring_lock
);
8808 percpu_ref_kill(&ctx
->refs
);
8810 if (WARN_ON_ONCE((ctx
->flags
& IORING_SETUP_SQPOLL
) && !ctx
->sqo_dead
))
8813 /* if force is set, the ring is going away. always drop after that */
8814 ctx
->cq_overflow_flushed
= 1;
8816 __io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8817 mutex_unlock(&ctx
->uring_lock
);
8819 /* prevent SQPOLL from submitting new requests */
8821 io_sq_thread_park(ctx
->sq_data
);
8822 list_del_init(&ctx
->sqd_list
);
8823 io_sqd_update_thread_idle(ctx
->sq_data
);
8824 io_sq_thread_unpark(ctx
->sq_data
);
8827 io_kill_timeouts(ctx
, NULL
, NULL
);
8828 io_poll_remove_all(ctx
, NULL
, NULL
);
8831 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8833 /* if we failed setting up the ctx, we might not have any rings */
8834 io_iopoll_try_reap_events(ctx
);
8835 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
8838 * Do this upfront, so we won't have a grace period where the ring
8839 * is closed but resources aren't reaped yet. This can cause
8840 * spurious failure in setting up a new ring.
8842 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
8845 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8847 * Use system_unbound_wq to avoid spawning tons of event kworkers
8848 * if we're exiting a ton of rings at the same time. It just adds
8849 * noise and overhead, there's no discernable change in runtime
8850 * over using system_wq.
8852 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8855 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8857 struct io_ring_ctx
*ctx
= file
->private_data
;
8859 file
->private_data
= NULL
;
8860 io_ring_ctx_wait_and_kill(ctx
);
8864 struct io_task_cancel
{
8865 struct task_struct
*task
;
8866 struct files_struct
*files
;
8869 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8871 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8872 struct io_task_cancel
*cancel
= data
;
8875 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8876 unsigned long flags
;
8877 struct io_ring_ctx
*ctx
= req
->ctx
;
8879 /* protect against races with linked timeouts */
8880 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8881 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8882 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8884 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8889 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8890 struct task_struct
*task
,
8891 struct files_struct
*files
)
8893 struct io_defer_entry
*de
;
8896 spin_lock_irq(&ctx
->completion_lock
);
8897 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8898 if (io_match_task(de
->req
, task
, files
)) {
8899 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8903 spin_unlock_irq(&ctx
->completion_lock
);
8904 if (list_empty(&list
))
8907 while (!list_empty(&list
)) {
8908 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8909 list_del_init(&de
->list
);
8910 req_set_fail_links(de
->req
);
8911 io_put_req(de
->req
);
8912 io_req_complete(de
->req
, -ECANCELED
);
8918 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
8919 struct task_struct
*task
,
8920 struct files_struct
*files
)
8922 struct io_task_cancel cancel
= { .task
= task
, .files
= files
, };
8925 enum io_wq_cancel cret
;
8929 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
,
8931 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8934 /* SQPOLL thread does its own polling */
8935 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && !files
) ||
8936 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
8937 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8938 io_iopoll_try_reap_events(ctx
);
8943 ret
|= io_cancel_defer_files(ctx
, task
, files
);
8944 ret
|= io_poll_remove_all(ctx
, task
, files
);
8945 ret
|= io_kill_timeouts(ctx
, task
, files
);
8946 ret
|= io_run_task_work();
8947 io_cqring_overflow_flush(ctx
, true, task
, files
);
8954 static int io_uring_count_inflight(struct io_ring_ctx
*ctx
,
8955 struct task_struct
*task
,
8956 struct files_struct
*files
)
8958 struct io_kiocb
*req
;
8961 spin_lock_irq(&ctx
->inflight_lock
);
8962 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
)
8963 cnt
+= io_match_task(req
, task
, files
);
8964 spin_unlock_irq(&ctx
->inflight_lock
);
8968 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
8969 struct task_struct
*task
,
8970 struct files_struct
*files
)
8972 while (!list_empty_careful(&ctx
->inflight_list
)) {
8976 inflight
= io_uring_count_inflight(ctx
, task
, files
);
8980 io_uring_try_cancel_requests(ctx
, task
, files
);
8983 io_sq_thread_unpark(ctx
->sq_data
);
8984 prepare_to_wait(&task
->io_uring
->wait
, &wait
,
8985 TASK_UNINTERRUPTIBLE
);
8986 if (inflight
== io_uring_count_inflight(ctx
, task
, files
))
8988 finish_wait(&task
->io_uring
->wait
, &wait
);
8990 io_sq_thread_park(ctx
->sq_data
);
8994 static void io_disable_sqo_submit(struct io_ring_ctx
*ctx
)
8996 mutex_lock(&ctx
->uring_lock
);
8998 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
8999 io_sq_offload_start(ctx
);
9000 mutex_unlock(&ctx
->uring_lock
);
9002 /* make sure callers enter the ring to get error */
9004 io_ring_set_wakeup_flag(ctx
);
9008 * We need to iteratively cancel requests, in case a request has dependent
9009 * hard links. These persist even for failure of cancelations, hence keep
9010 * looping until none are found.
9012 static void io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
9013 struct files_struct
*files
)
9015 struct task_struct
*task
= current
;
9017 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
9018 io_disable_sqo_submit(ctx
);
9019 task
= ctx
->sq_data
->thread
;
9020 atomic_inc(&task
->io_uring
->in_idle
);
9021 io_sq_thread_park(ctx
->sq_data
);
9024 io_uring_cancel_files(ctx
, task
, files
);
9026 io_uring_try_cancel_requests(ctx
, task
, NULL
);
9028 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
9029 atomic_dec(&task
->io_uring
->in_idle
);
9030 io_sq_thread_unpark(ctx
->sq_data
);
9035 * Note that this task has used io_uring. We use it for cancelation purposes.
9037 static int io_uring_add_task_file(struct io_ring_ctx
*ctx
, struct file
*file
)
9039 struct io_uring_task
*tctx
= current
->io_uring
;
9042 if (unlikely(!tctx
)) {
9043 ret
= io_uring_alloc_task_context(current
);
9046 tctx
= current
->io_uring
;
9048 if (tctx
->last
!= file
) {
9049 void *old
= xa_load(&tctx
->xa
, (unsigned long)file
);
9053 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)file
,
9064 * This is race safe in that the task itself is doing this, hence it
9065 * cannot be going through the exit/cancel paths at the same time.
9066 * This cannot be modified while exit/cancel is running.
9068 if (!tctx
->sqpoll
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9069 tctx
->sqpoll
= true;
9075 * Remove this io_uring_file -> task mapping.
9077 static void io_uring_del_task_file(struct file
*file
)
9079 struct io_uring_task
*tctx
= current
->io_uring
;
9081 if (tctx
->last
== file
)
9083 file
= xa_erase(&tctx
->xa
, (unsigned long)file
);
9088 static void io_uring_remove_task_files(struct io_uring_task
*tctx
)
9091 unsigned long index
;
9093 xa_for_each(&tctx
->xa
, index
, file
)
9094 io_uring_del_task_file(file
);
9097 void __io_uring_files_cancel(struct files_struct
*files
)
9099 struct io_uring_task
*tctx
= current
->io_uring
;
9101 unsigned long index
;
9103 /* make sure overflow events are dropped */
9104 atomic_inc(&tctx
->in_idle
);
9105 xa_for_each(&tctx
->xa
, index
, file
)
9106 io_uring_cancel_task_requests(file
->private_data
, files
);
9107 atomic_dec(&tctx
->in_idle
);
9110 io_uring_remove_task_files(tctx
);
9113 static s64
tctx_inflight(struct io_uring_task
*tctx
)
9115 unsigned long index
;
9119 inflight
= percpu_counter_sum(&tctx
->inflight
);
9124 * If we have SQPOLL rings, then we need to iterate and find them, and
9125 * add the pending count for those.
9127 xa_for_each(&tctx
->xa
, index
, file
) {
9128 struct io_ring_ctx
*ctx
= file
->private_data
;
9130 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9131 struct io_uring_task
*__tctx
= ctx
->sqo_task
->io_uring
;
9133 inflight
+= percpu_counter_sum(&__tctx
->inflight
);
9141 * Find any io_uring fd that this task has registered or done IO on, and cancel
9144 void __io_uring_task_cancel(void)
9146 struct io_uring_task
*tctx
= current
->io_uring
;
9150 /* make sure overflow events are dropped */
9151 atomic_inc(&tctx
->in_idle
);
9153 /* trigger io_disable_sqo_submit() */
9155 __io_uring_files_cancel(NULL
);
9158 /* read completions before cancelations */
9159 inflight
= tctx_inflight(tctx
);
9162 __io_uring_files_cancel(NULL
);
9164 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9167 * If we've seen completions, retry without waiting. This
9168 * avoids a race where a completion comes in before we did
9169 * prepare_to_wait().
9171 if (inflight
== tctx_inflight(tctx
))
9173 finish_wait(&tctx
->wait
, &wait
);
9176 atomic_dec(&tctx
->in_idle
);
9178 io_uring_remove_task_files(tctx
);
9181 static int io_uring_flush(struct file
*file
, void *data
)
9183 struct io_uring_task
*tctx
= current
->io_uring
;
9184 struct io_ring_ctx
*ctx
= file
->private_data
;
9186 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
))
9187 io_uring_cancel_task_requests(ctx
, NULL
);
9192 /* we should have cancelled and erased it before PF_EXITING */
9193 WARN_ON_ONCE((current
->flags
& PF_EXITING
) &&
9194 xa_load(&tctx
->xa
, (unsigned long)file
));
9197 * fput() is pending, will be 2 if the only other ref is our potential
9198 * task file note. If the task is exiting, drop regardless of count.
9200 if (atomic_long_read(&file
->f_count
) != 2)
9203 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9204 /* there is only one file note, which is owned by sqo_task */
9205 WARN_ON_ONCE(ctx
->sqo_task
!= current
&&
9206 xa_load(&tctx
->xa
, (unsigned long)file
));
9207 /* sqo_dead check is for when this happens after cancellation */
9208 WARN_ON_ONCE(ctx
->sqo_task
== current
&& !ctx
->sqo_dead
&&
9209 !xa_load(&tctx
->xa
, (unsigned long)file
));
9211 io_disable_sqo_submit(ctx
);
9214 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) || ctx
->sqo_task
== current
)
9215 io_uring_del_task_file(file
);
9219 static void *io_uring_validate_mmap_request(struct file
*file
,
9220 loff_t pgoff
, size_t sz
)
9222 struct io_ring_ctx
*ctx
= file
->private_data
;
9223 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9228 case IORING_OFF_SQ_RING
:
9229 case IORING_OFF_CQ_RING
:
9232 case IORING_OFF_SQES
:
9236 return ERR_PTR(-EINVAL
);
9239 page
= virt_to_head_page(ptr
);
9240 if (sz
> page_size(page
))
9241 return ERR_PTR(-EINVAL
);
9248 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9250 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9254 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9256 return PTR_ERR(ptr
);
9258 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9259 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9262 #else /* !CONFIG_MMU */
9264 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9266 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9269 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9271 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9274 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9275 unsigned long addr
, unsigned long len
,
9276 unsigned long pgoff
, unsigned long flags
)
9280 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9282 return PTR_ERR(ptr
);
9284 return (unsigned long) ptr
;
9287 #endif /* !CONFIG_MMU */
9289 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9295 if (!io_sqring_full(ctx
))
9298 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9300 if (unlikely(ctx
->sqo_dead
)) {
9305 if (!io_sqring_full(ctx
))
9309 } while (!signal_pending(current
));
9311 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9316 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9317 struct __kernel_timespec __user
**ts
,
9318 const sigset_t __user
**sig
)
9320 struct io_uring_getevents_arg arg
;
9323 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9324 * is just a pointer to the sigset_t.
9326 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9327 *sig
= (const sigset_t __user
*) argp
;
9333 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9334 * timespec and sigset_t pointers if good.
9336 if (*argsz
!= sizeof(arg
))
9338 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9340 *sig
= u64_to_user_ptr(arg
.sigmask
);
9341 *argsz
= arg
.sigmask_sz
;
9342 *ts
= u64_to_user_ptr(arg
.ts
);
9346 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9347 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9350 struct io_ring_ctx
*ctx
;
9357 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9358 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
))
9366 if (f
.file
->f_op
!= &io_uring_fops
)
9370 ctx
= f
.file
->private_data
;
9371 if (!percpu_ref_tryget(&ctx
->refs
))
9375 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9379 * For SQ polling, the thread will do all submissions and completions.
9380 * Just return the requested submit count, and wake the thread if
9384 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9385 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
9388 if (unlikely(ctx
->sqo_dead
))
9390 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9391 wake_up(&ctx
->sq_data
->wait
);
9392 if (flags
& IORING_ENTER_SQ_WAIT
) {
9393 ret
= io_sqpoll_wait_sq(ctx
);
9397 submitted
= to_submit
;
9398 } else if (to_submit
) {
9399 ret
= io_uring_add_task_file(ctx
, f
.file
);
9402 mutex_lock(&ctx
->uring_lock
);
9403 submitted
= io_submit_sqes(ctx
, to_submit
);
9404 mutex_unlock(&ctx
->uring_lock
);
9406 if (submitted
!= to_submit
)
9409 if (flags
& IORING_ENTER_GETEVENTS
) {
9410 const sigset_t __user
*sig
;
9411 struct __kernel_timespec __user
*ts
;
9413 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9417 min_complete
= min(min_complete
, ctx
->cq_entries
);
9420 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9421 * space applications don't need to do io completion events
9422 * polling again, they can rely on io_sq_thread to do polling
9423 * work, which can reduce cpu usage and uring_lock contention.
9425 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9426 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9427 ret
= io_iopoll_check(ctx
, min_complete
);
9429 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9434 percpu_ref_put(&ctx
->refs
);
9437 return submitted
? submitted
: ret
;
9440 #ifdef CONFIG_PROC_FS
9441 static int io_uring_show_cred(int id
, void *p
, void *data
)
9443 struct io_identity
*iod
= p
;
9444 const struct cred
*cred
= iod
->creds
;
9445 struct seq_file
*m
= data
;
9446 struct user_namespace
*uns
= seq_user_ns(m
);
9447 struct group_info
*gi
;
9452 seq_printf(m
, "%5d\n", id
);
9453 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9454 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9455 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9456 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9457 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9458 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9459 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9460 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9461 seq_puts(m
, "\n\tGroups:\t");
9462 gi
= cred
->group_info
;
9463 for (g
= 0; g
< gi
->ngroups
; g
++) {
9464 seq_put_decimal_ull(m
, g
? " " : "",
9465 from_kgid_munged(uns
, gi
->gid
[g
]));
9467 seq_puts(m
, "\n\tCapEff:\t");
9468 cap
= cred
->cap_effective
;
9469 CAP_FOR_EACH_U32(__capi
)
9470 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9475 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9477 struct io_sq_data
*sq
= NULL
;
9482 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9483 * since fdinfo case grabs it in the opposite direction of normal use
9484 * cases. If we fail to get the lock, we just don't iterate any
9485 * structures that could be going away outside the io_uring mutex.
9487 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9489 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9492 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9493 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9494 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9495 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9496 struct fixed_file_table
*table
;
9499 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
9500 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
9502 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9504 seq_printf(m
, "%5u: <none>\n", i
);
9506 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9507 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9508 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9510 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9511 (unsigned int) buf
->len
);
9513 if (has_lock
&& !idr_is_empty(&ctx
->personality_idr
)) {
9514 seq_printf(m
, "Personalities:\n");
9515 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
9517 seq_printf(m
, "PollList:\n");
9518 spin_lock_irq(&ctx
->completion_lock
);
9519 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9520 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9521 struct io_kiocb
*req
;
9523 hlist_for_each_entry(req
, list
, hash_node
)
9524 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9525 req
->task
->task_works
!= NULL
);
9527 spin_unlock_irq(&ctx
->completion_lock
);
9529 mutex_unlock(&ctx
->uring_lock
);
9532 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9534 struct io_ring_ctx
*ctx
= f
->private_data
;
9536 if (percpu_ref_tryget(&ctx
->refs
)) {
9537 __io_uring_show_fdinfo(ctx
, m
);
9538 percpu_ref_put(&ctx
->refs
);
9543 static const struct file_operations io_uring_fops
= {
9544 .release
= io_uring_release
,
9545 .flush
= io_uring_flush
,
9546 .mmap
= io_uring_mmap
,
9548 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9549 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9551 .poll
= io_uring_poll
,
9552 .fasync
= io_uring_fasync
,
9553 #ifdef CONFIG_PROC_FS
9554 .show_fdinfo
= io_uring_show_fdinfo
,
9558 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9559 struct io_uring_params
*p
)
9561 struct io_rings
*rings
;
9562 size_t size
, sq_array_offset
;
9564 /* make sure these are sane, as we already accounted them */
9565 ctx
->sq_entries
= p
->sq_entries
;
9566 ctx
->cq_entries
= p
->cq_entries
;
9568 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9569 if (size
== SIZE_MAX
)
9572 rings
= io_mem_alloc(size
);
9577 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9578 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9579 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9580 rings
->sq_ring_entries
= p
->sq_entries
;
9581 rings
->cq_ring_entries
= p
->cq_entries
;
9582 ctx
->sq_mask
= rings
->sq_ring_mask
;
9583 ctx
->cq_mask
= rings
->cq_ring_mask
;
9585 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9586 if (size
== SIZE_MAX
) {
9587 io_mem_free(ctx
->rings
);
9592 ctx
->sq_sqes
= io_mem_alloc(size
);
9593 if (!ctx
->sq_sqes
) {
9594 io_mem_free(ctx
->rings
);
9602 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9606 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9610 ret
= io_uring_add_task_file(ctx
, file
);
9615 fd_install(fd
, file
);
9620 * Allocate an anonymous fd, this is what constitutes the application
9621 * visible backing of an io_uring instance. The application mmaps this
9622 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9623 * we have to tie this fd to a socket for file garbage collection purposes.
9625 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9628 #if defined(CONFIG_UNIX)
9631 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9634 return ERR_PTR(ret
);
9637 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9638 O_RDWR
| O_CLOEXEC
);
9639 #if defined(CONFIG_UNIX)
9641 sock_release(ctx
->ring_sock
);
9642 ctx
->ring_sock
= NULL
;
9644 ctx
->ring_sock
->file
= file
;
9650 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9651 struct io_uring_params __user
*params
)
9653 struct user_struct
*user
= NULL
;
9654 struct io_ring_ctx
*ctx
;
9661 if (entries
> IORING_MAX_ENTRIES
) {
9662 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9664 entries
= IORING_MAX_ENTRIES
;
9668 * Use twice as many entries for the CQ ring. It's possible for the
9669 * application to drive a higher depth than the size of the SQ ring,
9670 * since the sqes are only used at submission time. This allows for
9671 * some flexibility in overcommitting a bit. If the application has
9672 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9673 * of CQ ring entries manually.
9675 p
->sq_entries
= roundup_pow_of_two(entries
);
9676 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9678 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9679 * to a power-of-two, if it isn't already. We do NOT impose
9680 * any cq vs sq ring sizing.
9684 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9685 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9687 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9689 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9690 if (p
->cq_entries
< p
->sq_entries
)
9693 p
->cq_entries
= 2 * p
->sq_entries
;
9696 user
= get_uid(current_user());
9697 limit_mem
= !capable(CAP_IPC_LOCK
);
9700 ret
= __io_account_mem(user
,
9701 ring_pages(p
->sq_entries
, p
->cq_entries
));
9708 ctx
= io_ring_ctx_alloc(p
);
9711 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
9716 ctx
->compat
= in_compat_syscall();
9718 ctx
->creds
= get_current_cred();
9720 ctx
->loginuid
= current
->loginuid
;
9721 ctx
->sessionid
= current
->sessionid
;
9723 ctx
->sqo_task
= get_task_struct(current
);
9726 * This is just grabbed for accounting purposes. When a process exits,
9727 * the mm is exited and dropped before the files, hence we need to hang
9728 * on to this mm purely for the purposes of being able to unaccount
9729 * memory (locked/pinned vm). It's not used for anything else.
9731 mmgrab(current
->mm
);
9732 ctx
->mm_account
= current
->mm
;
9734 #ifdef CONFIG_BLK_CGROUP
9736 * The sq thread will belong to the original cgroup it was inited in.
9737 * If the cgroup goes offline (e.g. disabling the io controller), then
9738 * issued bios will be associated with the closest cgroup later in the
9742 ctx
->sqo_blkcg_css
= blkcg_css();
9743 ret
= css_tryget_online(ctx
->sqo_blkcg_css
);
9746 /* don't init against a dying cgroup, have the user try again */
9747 ctx
->sqo_blkcg_css
= NULL
;
9754 * Account memory _before_ installing the file descriptor. Once
9755 * the descriptor is installed, it can get closed at any time. Also
9756 * do this before hitting the general error path, as ring freeing
9757 * will un-account as well.
9759 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
9761 ctx
->limit_mem
= limit_mem
;
9763 ret
= io_allocate_scq_urings(ctx
, p
);
9767 ret
= io_sq_offload_create(ctx
, p
);
9771 if (!(p
->flags
& IORING_SETUP_R_DISABLED
))
9772 io_sq_offload_start(ctx
);
9774 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9775 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9776 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9777 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9778 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9779 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9780 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9781 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9783 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9784 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9785 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9786 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9787 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9788 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9789 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9790 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9792 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9793 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9794 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9795 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9796 IORING_FEAT_EXT_ARG
;
9798 if (copy_to_user(params
, p
, sizeof(*p
))) {
9803 file
= io_uring_get_file(ctx
);
9805 ret
= PTR_ERR(file
);
9810 * Install ring fd as the very last thing, so we don't risk someone
9811 * having closed it before we finish setup
9813 ret
= io_uring_install_fd(ctx
, file
);
9815 io_disable_sqo_submit(ctx
);
9816 /* fput will clean it up */
9821 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9824 io_disable_sqo_submit(ctx
);
9825 io_ring_ctx_wait_and_kill(ctx
);
9830 * Sets up an aio uring context, and returns the fd. Applications asks for a
9831 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9832 * params structure passed in.
9834 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9836 struct io_uring_params p
;
9839 if (copy_from_user(&p
, params
, sizeof(p
)))
9841 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9846 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9847 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9848 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9849 IORING_SETUP_R_DISABLED
))
9852 return io_uring_create(entries
, &p
, params
);
9855 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9856 struct io_uring_params __user
*, params
)
9858 return io_uring_setup(entries
, params
);
9861 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9863 struct io_uring_probe
*p
;
9867 size
= struct_size(p
, ops
, nr_args
);
9868 if (size
== SIZE_MAX
)
9870 p
= kzalloc(size
, GFP_KERNEL
);
9875 if (copy_from_user(p
, arg
, size
))
9878 if (memchr_inv(p
, 0, size
))
9881 p
->last_op
= IORING_OP_LAST
- 1;
9882 if (nr_args
> IORING_OP_LAST
)
9883 nr_args
= IORING_OP_LAST
;
9885 for (i
= 0; i
< nr_args
; i
++) {
9887 if (!io_op_defs
[i
].not_supported
)
9888 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9893 if (copy_to_user(arg
, p
, size
))
9900 static int io_register_personality(struct io_ring_ctx
*ctx
)
9902 struct io_identity
*id
;
9905 id
= kmalloc(sizeof(*id
), GFP_KERNEL
);
9909 io_init_identity(id
);
9910 id
->creds
= get_current_cred();
9912 ret
= idr_alloc_cyclic(&ctx
->personality_idr
, id
, 1, USHRT_MAX
, GFP_KERNEL
);
9914 put_cred(id
->creds
);
9920 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9922 struct io_identity
*iod
;
9924 iod
= idr_remove(&ctx
->personality_idr
, id
);
9926 put_cred(iod
->creds
);
9927 if (refcount_dec_and_test(&iod
->count
))
9935 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9936 unsigned int nr_args
)
9938 struct io_uring_restriction
*res
;
9942 /* Restrictions allowed only if rings started disabled */
9943 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9946 /* We allow only a single restrictions registration */
9947 if (ctx
->restrictions
.registered
)
9950 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9953 size
= array_size(nr_args
, sizeof(*res
));
9954 if (size
== SIZE_MAX
)
9957 res
= memdup_user(arg
, size
);
9959 return PTR_ERR(res
);
9963 for (i
= 0; i
< nr_args
; i
++) {
9964 switch (res
[i
].opcode
) {
9965 case IORING_RESTRICTION_REGISTER_OP
:
9966 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9971 __set_bit(res
[i
].register_op
,
9972 ctx
->restrictions
.register_op
);
9974 case IORING_RESTRICTION_SQE_OP
:
9975 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9980 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9982 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9983 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9985 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9986 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9995 /* Reset all restrictions if an error happened */
9997 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9999 ctx
->restrictions
.registered
= true;
10005 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10007 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10010 if (ctx
->restrictions
.registered
)
10011 ctx
->restricted
= 1;
10013 io_sq_offload_start(ctx
);
10017 static bool io_register_op_must_quiesce(int op
)
10020 case IORING_UNREGISTER_FILES
:
10021 case IORING_REGISTER_FILES_UPDATE
:
10022 case IORING_REGISTER_PROBE
:
10023 case IORING_REGISTER_PERSONALITY
:
10024 case IORING_UNREGISTER_PERSONALITY
:
10031 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10032 void __user
*arg
, unsigned nr_args
)
10033 __releases(ctx
->uring_lock
)
10034 __acquires(ctx
->uring_lock
)
10039 * We're inside the ring mutex, if the ref is already dying, then
10040 * someone else killed the ctx or is already going through
10041 * io_uring_register().
10043 if (percpu_ref_is_dying(&ctx
->refs
))
10046 if (io_register_op_must_quiesce(opcode
)) {
10047 percpu_ref_kill(&ctx
->refs
);
10050 * Drop uring mutex before waiting for references to exit. If
10051 * another thread is currently inside io_uring_enter() it might
10052 * need to grab the uring_lock to make progress. If we hold it
10053 * here across the drain wait, then we can deadlock. It's safe
10054 * to drop the mutex here, since no new references will come in
10055 * after we've killed the percpu ref.
10057 mutex_unlock(&ctx
->uring_lock
);
10059 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10062 ret
= io_run_task_work_sig();
10067 mutex_lock(&ctx
->uring_lock
);
10070 percpu_ref_resurrect(&ctx
->refs
);
10075 if (ctx
->restricted
) {
10076 if (opcode
>= IORING_REGISTER_LAST
) {
10081 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
10088 case IORING_REGISTER_BUFFERS
:
10089 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
10091 case IORING_UNREGISTER_BUFFERS
:
10093 if (arg
|| nr_args
)
10095 ret
= io_sqe_buffer_unregister(ctx
);
10097 case IORING_REGISTER_FILES
:
10098 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
10100 case IORING_UNREGISTER_FILES
:
10102 if (arg
|| nr_args
)
10104 ret
= io_sqe_files_unregister(ctx
);
10106 case IORING_REGISTER_FILES_UPDATE
:
10107 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
10109 case IORING_REGISTER_EVENTFD
:
10110 case IORING_REGISTER_EVENTFD_ASYNC
:
10114 ret
= io_eventfd_register(ctx
, arg
);
10117 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10118 ctx
->eventfd_async
= 1;
10120 ctx
->eventfd_async
= 0;
10122 case IORING_UNREGISTER_EVENTFD
:
10124 if (arg
|| nr_args
)
10126 ret
= io_eventfd_unregister(ctx
);
10128 case IORING_REGISTER_PROBE
:
10130 if (!arg
|| nr_args
> 256)
10132 ret
= io_probe(ctx
, arg
, nr_args
);
10134 case IORING_REGISTER_PERSONALITY
:
10136 if (arg
|| nr_args
)
10138 ret
= io_register_personality(ctx
);
10140 case IORING_UNREGISTER_PERSONALITY
:
10144 ret
= io_unregister_personality(ctx
, nr_args
);
10146 case IORING_REGISTER_ENABLE_RINGS
:
10148 if (arg
|| nr_args
)
10150 ret
= io_register_enable_rings(ctx
);
10152 case IORING_REGISTER_RESTRICTIONS
:
10153 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10161 if (io_register_op_must_quiesce(opcode
)) {
10162 /* bring the ctx back to life */
10163 percpu_ref_reinit(&ctx
->refs
);
10165 reinit_completion(&ctx
->ref_comp
);
10170 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10171 void __user
*, arg
, unsigned int, nr_args
)
10173 struct io_ring_ctx
*ctx
;
10182 if (f
.file
->f_op
!= &io_uring_fops
)
10185 ctx
= f
.file
->private_data
;
10187 mutex_lock(&ctx
->uring_lock
);
10188 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10189 mutex_unlock(&ctx
->uring_lock
);
10190 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10191 ctx
->cq_ev_fd
!= NULL
, ret
);
10197 static int __init
io_uring_init(void)
10199 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10200 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10201 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10204 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10205 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10206 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10207 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10208 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10209 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10210 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10211 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10212 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10213 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10214 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10215 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10216 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10217 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10218 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10219 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10220 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10221 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10222 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10223 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10224 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10225 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10226 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10227 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10228 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10229 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10230 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10231 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10232 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10233 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10234 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10236 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10237 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
10238 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
10241 __initcall(io_uring_init
);