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;
267 * Ring buffer of indices into array of io_uring_sqe, which is
268 * mmapped by the application using the IORING_OFF_SQES offset.
270 * This indirection could e.g. be used to assign fixed
271 * io_uring_sqe entries to operations and only submit them to
272 * the queue when needed.
274 * The kernel modifies neither the indices array nor the entries
278 unsigned cached_sq_head
;
281 unsigned sq_thread_idle
;
282 unsigned cached_sq_dropped
;
283 unsigned cached_cq_overflow
;
284 unsigned long sq_check_overflow
;
286 struct list_head defer_list
;
287 struct list_head timeout_list
;
288 struct list_head cq_overflow_list
;
290 wait_queue_head_t inflight_wait
;
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 long cq_check_overflow
;
358 struct wait_queue_head cq_wait
;
359 struct fasync_struct
*cq_fasync
;
360 struct eventfd_ctx
*cq_ev_fd
;
361 } ____cacheline_aligned_in_smp
;
364 struct mutex uring_lock
;
365 wait_queue_head_t wait
;
366 } ____cacheline_aligned_in_smp
;
369 spinlock_t completion_lock
;
372 * ->iopoll_list is protected by the ctx->uring_lock for
373 * io_uring instances that don't use IORING_SETUP_SQPOLL.
374 * For SQPOLL, only the single threaded io_sq_thread() will
375 * manipulate the list, hence no extra locking is needed there.
377 struct list_head iopoll_list
;
378 struct hlist_head
*cancel_hash
;
379 unsigned cancel_hash_bits
;
380 bool poll_multi_file
;
382 spinlock_t inflight_lock
;
383 struct list_head inflight_list
;
384 } ____cacheline_aligned_in_smp
;
386 struct delayed_work file_put_work
;
387 struct llist_head file_put_llist
;
389 struct work_struct exit_work
;
390 struct io_restriction restrictions
;
394 * First field must be the file pointer in all the
395 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
397 struct io_poll_iocb
{
399 struct wait_queue_head
*head
;
403 struct wait_queue_entry wait
;
406 struct io_poll_remove
{
413 struct file
*put_file
;
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
{
460 /* NOTE: kiocb has the file as the first member, so don't do it here */
468 struct sockaddr __user
*addr
;
475 struct user_msghdr __user
*umsg
;
481 struct io_buffer
*kbuf
;
487 bool ignore_nonblock
;
488 struct filename
*filename
;
490 unsigned long nofile
;
493 struct io_files_update
{
519 struct epoll_event event
;
523 struct file
*file_out
;
524 struct file
*file_in
;
531 struct io_provide_buf
{
545 const char __user
*filename
;
546 struct statx __user
*buffer
;
558 struct filename
*oldpath
;
559 struct filename
*newpath
;
567 struct filename
*filename
;
570 struct io_completion
{
572 struct list_head list
;
576 struct io_async_connect
{
577 struct sockaddr_storage address
;
580 struct io_async_msghdr
{
581 struct iovec fast_iov
[UIO_FASTIOV
];
583 struct sockaddr __user
*uaddr
;
585 struct sockaddr_storage addr
;
589 struct iovec fast_iov
[UIO_FASTIOV
];
590 const struct iovec
*free_iovec
;
591 struct iov_iter iter
;
593 struct wait_page_queue wpq
;
597 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
598 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
599 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
600 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
601 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
602 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
608 REQ_F_LINK_TIMEOUT_BIT
,
610 REQ_F_NEED_CLEANUP_BIT
,
612 REQ_F_BUFFER_SELECTED_BIT
,
613 REQ_F_NO_FILE_TABLE_BIT
,
614 REQ_F_WORK_INITIALIZED_BIT
,
615 REQ_F_LTIMEOUT_ACTIVE_BIT
,
617 /* not a real bit, just to check we're not overflowing the space */
623 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
624 /* drain existing IO first */
625 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
627 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
628 /* doesn't sever on completion < 0 */
629 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
631 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
632 /* IOSQE_BUFFER_SELECT */
633 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
635 /* fail rest of links */
636 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
637 /* on inflight list */
638 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
639 /* read/write uses file position */
640 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
641 /* must not punt to workers */
642 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
643 /* has or had linked timeout */
644 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
646 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
648 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
649 /* already went through poll handler */
650 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
651 /* buffer already selected */
652 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
653 /* doesn't need file table for this request */
654 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
655 /* io_wq_work is initialized */
656 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
657 /* linked timeout is active, i.e. prepared by link's head */
658 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
662 struct io_poll_iocb poll
;
663 struct io_poll_iocb
*double_poll
;
667 * NOTE! Each of the iocb union members has the file pointer
668 * as the first entry in their struct definition. So you can
669 * access the file pointer through any of the sub-structs,
670 * or directly as just 'ki_filp' in this struct.
676 struct io_poll_iocb poll
;
677 struct io_poll_remove poll_remove
;
678 struct io_accept accept
;
680 struct io_cancel cancel
;
681 struct io_timeout timeout
;
682 struct io_timeout_rem timeout_rem
;
683 struct io_connect connect
;
684 struct io_sr_msg sr_msg
;
686 struct io_close close
;
687 struct io_files_update files_update
;
688 struct io_fadvise fadvise
;
689 struct io_madvise madvise
;
690 struct io_epoll epoll
;
691 struct io_splice splice
;
692 struct io_provide_buf pbuf
;
693 struct io_statx statx
;
694 struct io_shutdown shutdown
;
695 struct io_rename rename
;
696 struct io_unlink unlink
;
697 /* use only after cleaning per-op data, see io_clean_op() */
698 struct io_completion
compl;
701 /* opcode allocated if it needs to store data for async defer */
704 /* polled IO has completed */
710 struct io_ring_ctx
*ctx
;
713 struct task_struct
*task
;
716 struct io_kiocb
*link
;
717 struct percpu_ref
*fixed_file_refs
;
720 * 1. used with ctx->iopoll_list with reads/writes
721 * 2. to track reqs with ->files (see io_op_def::file_table)
723 struct list_head inflight_entry
;
724 struct callback_head task_work
;
725 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
726 struct hlist_node hash_node
;
727 struct async_poll
*apoll
;
728 struct io_wq_work work
;
731 struct io_defer_entry
{
732 struct list_head list
;
733 struct io_kiocb
*req
;
737 #define IO_IOPOLL_BATCH 8
739 struct io_comp_state
{
741 struct list_head list
;
742 struct io_ring_ctx
*ctx
;
745 struct io_submit_state
{
746 struct blk_plug plug
;
749 * io_kiocb alloc cache
751 void *reqs
[IO_IOPOLL_BATCH
];
752 unsigned int free_reqs
;
757 * Batch completion logic
759 struct io_comp_state comp
;
762 * File reference cache
766 unsigned int has_refs
;
767 unsigned int ios_left
;
771 /* needs req->file assigned */
772 unsigned needs_file
: 1;
773 /* don't fail if file grab fails */
774 unsigned needs_file_no_error
: 1;
775 /* hash wq insertion if file is a regular file */
776 unsigned hash_reg_file
: 1;
777 /* unbound wq insertion if file is a non-regular file */
778 unsigned unbound_nonreg_file
: 1;
779 /* opcode is not supported by this kernel */
780 unsigned not_supported
: 1;
781 /* set if opcode supports polled "wait" */
783 unsigned pollout
: 1;
784 /* op supports buffer selection */
785 unsigned buffer_select
: 1;
786 /* must always have async data allocated */
787 unsigned needs_async_data
: 1;
788 /* should block plug */
790 /* size of async data needed, if any */
791 unsigned short async_size
;
795 static const struct io_op_def io_op_defs
[] = {
796 [IORING_OP_NOP
] = {},
797 [IORING_OP_READV
] = {
799 .unbound_nonreg_file
= 1,
802 .needs_async_data
= 1,
804 .async_size
= sizeof(struct io_async_rw
),
805 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
807 [IORING_OP_WRITEV
] = {
810 .unbound_nonreg_file
= 1,
812 .needs_async_data
= 1,
814 .async_size
= sizeof(struct io_async_rw
),
815 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
818 [IORING_OP_FSYNC
] = {
820 .work_flags
= IO_WQ_WORK_BLKCG
,
822 [IORING_OP_READ_FIXED
] = {
824 .unbound_nonreg_file
= 1,
827 .async_size
= sizeof(struct io_async_rw
),
828 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
830 [IORING_OP_WRITE_FIXED
] = {
833 .unbound_nonreg_file
= 1,
836 .async_size
= sizeof(struct io_async_rw
),
837 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
|
840 [IORING_OP_POLL_ADD
] = {
842 .unbound_nonreg_file
= 1,
844 [IORING_OP_POLL_REMOVE
] = {},
845 [IORING_OP_SYNC_FILE_RANGE
] = {
847 .work_flags
= IO_WQ_WORK_BLKCG
,
849 [IORING_OP_SENDMSG
] = {
851 .unbound_nonreg_file
= 1,
853 .needs_async_data
= 1,
854 .async_size
= sizeof(struct io_async_msghdr
),
855 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
858 [IORING_OP_RECVMSG
] = {
860 .unbound_nonreg_file
= 1,
863 .needs_async_data
= 1,
864 .async_size
= sizeof(struct io_async_msghdr
),
865 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
868 [IORING_OP_TIMEOUT
] = {
869 .needs_async_data
= 1,
870 .async_size
= sizeof(struct io_timeout_data
),
871 .work_flags
= IO_WQ_WORK_MM
,
873 [IORING_OP_TIMEOUT_REMOVE
] = {},
874 [IORING_OP_ACCEPT
] = {
876 .unbound_nonreg_file
= 1,
878 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
,
880 [IORING_OP_ASYNC_CANCEL
] = {},
881 [IORING_OP_LINK_TIMEOUT
] = {
882 .needs_async_data
= 1,
883 .async_size
= sizeof(struct io_timeout_data
),
884 .work_flags
= IO_WQ_WORK_MM
,
886 [IORING_OP_CONNECT
] = {
888 .unbound_nonreg_file
= 1,
890 .needs_async_data
= 1,
891 .async_size
= sizeof(struct io_async_connect
),
892 .work_flags
= IO_WQ_WORK_MM
,
894 [IORING_OP_FALLOCATE
] = {
896 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
898 [IORING_OP_OPENAT
] = {
899 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
|
900 IO_WQ_WORK_FS
| IO_WQ_WORK_MM
,
902 [IORING_OP_CLOSE
] = {
904 .needs_file_no_error
= 1,
905 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
,
907 [IORING_OP_FILES_UPDATE
] = {
908 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
,
910 [IORING_OP_STATX
] = {
911 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
|
912 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
916 .unbound_nonreg_file
= 1,
920 .async_size
= sizeof(struct io_async_rw
),
921 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
923 [IORING_OP_WRITE
] = {
925 .unbound_nonreg_file
= 1,
928 .async_size
= sizeof(struct io_async_rw
),
929 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
932 [IORING_OP_FADVISE
] = {
934 .work_flags
= IO_WQ_WORK_BLKCG
,
936 [IORING_OP_MADVISE
] = {
937 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
941 .unbound_nonreg_file
= 1,
943 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
947 .unbound_nonreg_file
= 1,
950 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
952 [IORING_OP_OPENAT2
] = {
953 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_FS
|
954 IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
956 [IORING_OP_EPOLL_CTL
] = {
957 .unbound_nonreg_file
= 1,
958 .work_flags
= IO_WQ_WORK_FILES
,
960 [IORING_OP_SPLICE
] = {
963 .unbound_nonreg_file
= 1,
964 .work_flags
= IO_WQ_WORK_BLKCG
,
966 [IORING_OP_PROVIDE_BUFFERS
] = {},
967 [IORING_OP_REMOVE_BUFFERS
] = {},
971 .unbound_nonreg_file
= 1,
973 [IORING_OP_SHUTDOWN
] = {
976 [IORING_OP_RENAMEAT
] = {
977 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
978 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
980 [IORING_OP_UNLINKAT
] = {
981 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
982 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
986 enum io_mem_account
{
991 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
992 struct io_comp_state
*cs
);
993 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
994 static void io_put_req(struct io_kiocb
*req
);
995 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
996 static void io_double_put_req(struct io_kiocb
*req
);
997 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
998 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
999 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1000 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
1001 struct io_uring_files_update
*ip
,
1003 static void __io_clean_op(struct io_kiocb
*req
);
1004 static struct file
*io_file_get(struct io_submit_state
*state
,
1005 struct io_kiocb
*req
, int fd
, bool fixed
);
1006 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
);
1007 static void io_file_put_work(struct work_struct
*work
);
1009 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
1010 struct iovec
**iovec
, struct iov_iter
*iter
,
1012 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
1013 const struct iovec
*fast_iov
,
1014 struct iov_iter
*iter
, bool force
);
1016 static struct kmem_cache
*req_cachep
;
1018 static const struct file_operations io_uring_fops
;
1020 struct sock
*io_uring_get_socket(struct file
*file
)
1022 #if defined(CONFIG_UNIX)
1023 if (file
->f_op
== &io_uring_fops
) {
1024 struct io_ring_ctx
*ctx
= file
->private_data
;
1026 return ctx
->ring_sock
->sk
;
1031 EXPORT_SYMBOL(io_uring_get_socket
);
1033 #define io_for_each_link(pos, head) \
1034 for (pos = (head); pos; pos = pos->link)
1036 static inline void io_clean_op(struct io_kiocb
*req
)
1038 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
|
1043 static bool io_match_task(struct io_kiocb
*head
,
1044 struct task_struct
*task
,
1045 struct files_struct
*files
)
1047 struct io_kiocb
*req
;
1049 if (task
&& head
->task
!= task
)
1054 io_for_each_link(req
, head
) {
1055 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
1056 (req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1057 req
->work
.identity
->files
== files
)
1063 static void io_sq_thread_drop_mm_files(void)
1065 struct files_struct
*files
= current
->files
;
1066 struct mm_struct
*mm
= current
->mm
;
1069 kthread_unuse_mm(mm
);
1074 struct nsproxy
*nsproxy
= current
->nsproxy
;
1077 current
->files
= NULL
;
1078 current
->nsproxy
= NULL
;
1079 task_unlock(current
);
1080 put_files_struct(files
);
1081 put_nsproxy(nsproxy
);
1085 static int __io_sq_thread_acquire_files(struct io_ring_ctx
*ctx
)
1087 if (!current
->files
) {
1088 struct files_struct
*files
;
1089 struct nsproxy
*nsproxy
;
1091 task_lock(ctx
->sqo_task
);
1092 files
= ctx
->sqo_task
->files
;
1094 task_unlock(ctx
->sqo_task
);
1097 atomic_inc(&files
->count
);
1098 get_nsproxy(ctx
->sqo_task
->nsproxy
);
1099 nsproxy
= ctx
->sqo_task
->nsproxy
;
1100 task_unlock(ctx
->sqo_task
);
1103 current
->files
= files
;
1104 current
->nsproxy
= nsproxy
;
1105 task_unlock(current
);
1110 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1112 struct mm_struct
*mm
;
1117 /* Should never happen */
1118 if (unlikely(!(ctx
->flags
& IORING_SETUP_SQPOLL
)))
1121 task_lock(ctx
->sqo_task
);
1122 mm
= ctx
->sqo_task
->mm
;
1123 if (unlikely(!mm
|| !mmget_not_zero(mm
)))
1125 task_unlock(ctx
->sqo_task
);
1135 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx
*ctx
,
1136 struct io_kiocb
*req
)
1138 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1141 if (def
->work_flags
& IO_WQ_WORK_MM
) {
1142 ret
= __io_sq_thread_acquire_mm(ctx
);
1147 if (def
->needs_file
|| (def
->work_flags
& IO_WQ_WORK_FILES
)) {
1148 ret
= __io_sq_thread_acquire_files(ctx
);
1156 static void io_sq_thread_associate_blkcg(struct io_ring_ctx
*ctx
,
1157 struct cgroup_subsys_state
**cur_css
)
1160 #ifdef CONFIG_BLK_CGROUP
1161 /* puts the old one when swapping */
1162 if (*cur_css
!= ctx
->sqo_blkcg_css
) {
1163 kthread_associate_blkcg(ctx
->sqo_blkcg_css
);
1164 *cur_css
= ctx
->sqo_blkcg_css
;
1169 static void io_sq_thread_unassociate_blkcg(void)
1171 #ifdef CONFIG_BLK_CGROUP
1172 kthread_associate_blkcg(NULL
);
1176 static inline void req_set_fail_links(struct io_kiocb
*req
)
1178 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1179 req
->flags
|= REQ_F_FAIL_LINK
;
1183 * None of these are dereferenced, they are simply used to check if any of
1184 * them have changed. If we're under current and check they are still the
1185 * same, we're fine to grab references to them for actual out-of-line use.
1187 static void io_init_identity(struct io_identity
*id
)
1189 id
->files
= current
->files
;
1190 id
->mm
= current
->mm
;
1191 #ifdef CONFIG_BLK_CGROUP
1193 id
->blkcg_css
= blkcg_css();
1196 id
->creds
= current_cred();
1197 id
->nsproxy
= current
->nsproxy
;
1198 id
->fs
= current
->fs
;
1199 id
->fsize
= rlimit(RLIMIT_FSIZE
);
1201 id
->loginuid
= current
->loginuid
;
1202 id
->sessionid
= current
->sessionid
;
1204 refcount_set(&id
->count
, 1);
1207 static inline void __io_req_init_async(struct io_kiocb
*req
)
1209 memset(&req
->work
, 0, sizeof(req
->work
));
1210 req
->flags
|= REQ_F_WORK_INITIALIZED
;
1214 * Note: must call io_req_init_async() for the first time you
1215 * touch any members of io_wq_work.
1217 static inline void io_req_init_async(struct io_kiocb
*req
)
1219 struct io_uring_task
*tctx
= current
->io_uring
;
1221 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
1224 __io_req_init_async(req
);
1226 /* Grab a ref if this isn't our static identity */
1227 req
->work
.identity
= tctx
->identity
;
1228 if (tctx
->identity
!= &tctx
->__identity
)
1229 refcount_inc(&req
->work
.identity
->count
);
1232 static inline bool io_async_submit(struct io_ring_ctx
*ctx
)
1234 return ctx
->flags
& IORING_SETUP_SQPOLL
;
1237 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1239 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1241 complete(&ctx
->ref_comp
);
1244 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1246 return !req
->timeout
.off
;
1249 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1251 struct io_ring_ctx
*ctx
;
1254 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1258 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
1259 if (!ctx
->fallback_req
)
1263 * Use 5 bits less than the max cq entries, that should give us around
1264 * 32 entries per hash list if totally full and uniformly spread.
1266 hash_bits
= ilog2(p
->cq_entries
);
1270 ctx
->cancel_hash_bits
= hash_bits
;
1271 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1273 if (!ctx
->cancel_hash
)
1275 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1277 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1278 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1281 ctx
->flags
= p
->flags
;
1282 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1283 INIT_LIST_HEAD(&ctx
->sqd_list
);
1284 init_waitqueue_head(&ctx
->cq_wait
);
1285 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1286 init_completion(&ctx
->ref_comp
);
1287 init_completion(&ctx
->sq_thread_comp
);
1288 idr_init(&ctx
->io_buffer_idr
);
1289 idr_init(&ctx
->personality_idr
);
1290 mutex_init(&ctx
->uring_lock
);
1291 init_waitqueue_head(&ctx
->wait
);
1292 spin_lock_init(&ctx
->completion_lock
);
1293 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1294 INIT_LIST_HEAD(&ctx
->defer_list
);
1295 INIT_LIST_HEAD(&ctx
->timeout_list
);
1296 init_waitqueue_head(&ctx
->inflight_wait
);
1297 spin_lock_init(&ctx
->inflight_lock
);
1298 INIT_LIST_HEAD(&ctx
->inflight_list
);
1299 INIT_DELAYED_WORK(&ctx
->file_put_work
, io_file_put_work
);
1300 init_llist_head(&ctx
->file_put_llist
);
1303 if (ctx
->fallback_req
)
1304 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1305 kfree(ctx
->cancel_hash
);
1310 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1312 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1313 struct io_ring_ctx
*ctx
= req
->ctx
;
1315 return seq
!= ctx
->cached_cq_tail
1316 + READ_ONCE(ctx
->cached_cq_overflow
);
1322 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1324 struct io_rings
*rings
= ctx
->rings
;
1326 /* order cqe stores with ring update */
1327 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1329 if (wq_has_sleeper(&ctx
->cq_wait
)) {
1330 wake_up_interruptible(&ctx
->cq_wait
);
1331 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1335 static void io_put_identity(struct io_uring_task
*tctx
, struct io_kiocb
*req
)
1337 if (req
->work
.identity
== &tctx
->__identity
)
1339 if (refcount_dec_and_test(&req
->work
.identity
->count
))
1340 kfree(req
->work
.identity
);
1343 static void io_req_clean_work(struct io_kiocb
*req
)
1345 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1348 req
->flags
&= ~REQ_F_WORK_INITIALIZED
;
1350 if (req
->work
.flags
& IO_WQ_WORK_MM
) {
1351 mmdrop(req
->work
.identity
->mm
);
1352 req
->work
.flags
&= ~IO_WQ_WORK_MM
;
1354 #ifdef CONFIG_BLK_CGROUP
1355 if (req
->work
.flags
& IO_WQ_WORK_BLKCG
) {
1356 css_put(req
->work
.identity
->blkcg_css
);
1357 req
->work
.flags
&= ~IO_WQ_WORK_BLKCG
;
1360 if (req
->work
.flags
& IO_WQ_WORK_CREDS
) {
1361 put_cred(req
->work
.identity
->creds
);
1362 req
->work
.flags
&= ~IO_WQ_WORK_CREDS
;
1364 if (req
->work
.flags
& IO_WQ_WORK_FS
) {
1365 struct fs_struct
*fs
= req
->work
.identity
->fs
;
1367 spin_lock(&req
->work
.identity
->fs
->lock
);
1370 spin_unlock(&req
->work
.identity
->fs
->lock
);
1373 req
->work
.flags
&= ~IO_WQ_WORK_FS
;
1376 io_put_identity(req
->task
->io_uring
, req
);
1380 * Create a private copy of io_identity, since some fields don't match
1381 * the current context.
1383 static bool io_identity_cow(struct io_kiocb
*req
)
1385 struct io_uring_task
*tctx
= current
->io_uring
;
1386 const struct cred
*creds
= NULL
;
1387 struct io_identity
*id
;
1389 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1390 creds
= req
->work
.identity
->creds
;
1392 id
= kmemdup(req
->work
.identity
, sizeof(*id
), GFP_KERNEL
);
1393 if (unlikely(!id
)) {
1394 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1399 * We can safely just re-init the creds we copied Either the field
1400 * matches the current one, or we haven't grabbed it yet. The only
1401 * exception is ->creds, through registered personalities, so handle
1402 * that one separately.
1404 io_init_identity(id
);
1406 req
->work
.identity
->creds
= creds
;
1408 /* add one for this request */
1409 refcount_inc(&id
->count
);
1411 /* drop tctx and req identity references, if needed */
1412 if (tctx
->identity
!= &tctx
->__identity
&&
1413 refcount_dec_and_test(&tctx
->identity
->count
))
1414 kfree(tctx
->identity
);
1415 if (req
->work
.identity
!= &tctx
->__identity
&&
1416 refcount_dec_and_test(&req
->work
.identity
->count
))
1417 kfree(req
->work
.identity
);
1419 req
->work
.identity
= id
;
1420 tctx
->identity
= id
;
1424 static bool io_grab_identity(struct io_kiocb
*req
)
1426 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1427 struct io_identity
*id
= req
->work
.identity
;
1428 struct io_ring_ctx
*ctx
= req
->ctx
;
1430 if (def
->work_flags
& IO_WQ_WORK_FSIZE
) {
1431 if (id
->fsize
!= rlimit(RLIMIT_FSIZE
))
1433 req
->work
.flags
|= IO_WQ_WORK_FSIZE
;
1436 if (!(req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1437 (def
->work_flags
& IO_WQ_WORK_FILES
) &&
1438 !(req
->flags
& REQ_F_NO_FILE_TABLE
)) {
1439 if (id
->files
!= current
->files
||
1440 id
->nsproxy
!= current
->nsproxy
)
1442 atomic_inc(&id
->files
->count
);
1443 get_nsproxy(id
->nsproxy
);
1444 req
->flags
|= REQ_F_INFLIGHT
;
1446 spin_lock_irq(&ctx
->inflight_lock
);
1447 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1448 spin_unlock_irq(&ctx
->inflight_lock
);
1449 req
->work
.flags
|= IO_WQ_WORK_FILES
;
1451 #ifdef CONFIG_BLK_CGROUP
1452 if (!(req
->work
.flags
& IO_WQ_WORK_BLKCG
) &&
1453 (def
->work_flags
& IO_WQ_WORK_BLKCG
)) {
1455 if (id
->blkcg_css
!= blkcg_css()) {
1460 * This should be rare, either the cgroup is dying or the task
1461 * is moving cgroups. Just punt to root for the handful of ios.
1463 if (css_tryget_online(id
->blkcg_css
))
1464 req
->work
.flags
|= IO_WQ_WORK_BLKCG
;
1468 if (!(req
->work
.flags
& IO_WQ_WORK_CREDS
)) {
1469 if (id
->creds
!= current_cred())
1471 get_cred(id
->creds
);
1472 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
1475 if (!uid_eq(current
->loginuid
, id
->loginuid
) ||
1476 current
->sessionid
!= id
->sessionid
)
1479 if (!(req
->work
.flags
& IO_WQ_WORK_FS
) &&
1480 (def
->work_flags
& IO_WQ_WORK_FS
)) {
1481 if (current
->fs
!= id
->fs
)
1483 spin_lock(&id
->fs
->lock
);
1484 if (!id
->fs
->in_exec
) {
1486 req
->work
.flags
|= IO_WQ_WORK_FS
;
1488 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1490 spin_unlock(¤t
->fs
->lock
);
1496 static void io_prep_async_work(struct io_kiocb
*req
)
1498 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1499 struct io_ring_ctx
*ctx
= req
->ctx
;
1500 struct io_identity
*id
;
1502 io_req_init_async(req
);
1503 id
= req
->work
.identity
;
1505 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1506 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1508 if (req
->flags
& REQ_F_ISREG
) {
1509 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1510 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1512 if (def
->unbound_nonreg_file
)
1513 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1516 /* ->mm can never change on us */
1517 if (!(req
->work
.flags
& IO_WQ_WORK_MM
) &&
1518 (def
->work_flags
& IO_WQ_WORK_MM
)) {
1520 req
->work
.flags
|= IO_WQ_WORK_MM
;
1523 /* if we fail grabbing identity, we must COW, regrab, and retry */
1524 if (io_grab_identity(req
))
1527 if (!io_identity_cow(req
))
1530 /* can't fail at this point */
1531 if (!io_grab_identity(req
))
1535 static void io_prep_async_link(struct io_kiocb
*req
)
1537 struct io_kiocb
*cur
;
1539 io_for_each_link(cur
, req
)
1540 io_prep_async_work(cur
);
1543 static struct io_kiocb
*__io_queue_async_work(struct io_kiocb
*req
)
1545 struct io_ring_ctx
*ctx
= req
->ctx
;
1546 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1548 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1549 &req
->work
, req
->flags
);
1550 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1554 static void io_queue_async_work(struct io_kiocb
*req
)
1556 struct io_kiocb
*link
;
1558 /* init ->work of the whole link before punting */
1559 io_prep_async_link(req
);
1560 link
= __io_queue_async_work(req
);
1563 io_queue_linked_timeout(link
);
1566 static void io_kill_timeout(struct io_kiocb
*req
)
1568 struct io_timeout_data
*io
= req
->async_data
;
1571 ret
= hrtimer_try_to_cancel(&io
->timer
);
1573 atomic_set(&req
->ctx
->cq_timeouts
,
1574 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1575 list_del_init(&req
->timeout
.list
);
1576 io_cqring_fill_event(req
, 0);
1577 io_put_req_deferred(req
, 1);
1582 * Returns true if we found and killed one or more timeouts
1584 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
1585 struct files_struct
*files
)
1587 struct io_kiocb
*req
, *tmp
;
1590 spin_lock_irq(&ctx
->completion_lock
);
1591 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1592 if (io_match_task(req
, tsk
, files
)) {
1593 io_kill_timeout(req
);
1597 spin_unlock_irq(&ctx
->completion_lock
);
1598 return canceled
!= 0;
1601 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1604 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1605 struct io_defer_entry
, list
);
1606 struct io_kiocb
*link
;
1608 if (req_need_defer(de
->req
, de
->seq
))
1610 list_del_init(&de
->list
);
1611 /* punt-init is done before queueing for defer */
1612 link
= __io_queue_async_work(de
->req
);
1614 __io_queue_linked_timeout(link
);
1615 /* drop submission reference */
1616 io_put_req_deferred(link
, 1);
1619 } while (!list_empty(&ctx
->defer_list
));
1622 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1624 while (!list_empty(&ctx
->timeout_list
)) {
1625 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1626 struct io_kiocb
, timeout
.list
);
1628 if (io_is_timeout_noseq(req
))
1630 if (req
->timeout
.target_seq
!= ctx
->cached_cq_tail
1631 - atomic_read(&ctx
->cq_timeouts
))
1634 list_del_init(&req
->timeout
.list
);
1635 io_kill_timeout(req
);
1639 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1641 io_flush_timeouts(ctx
);
1642 __io_commit_cqring(ctx
);
1644 if (unlikely(!list_empty(&ctx
->defer_list
)))
1645 __io_queue_deferred(ctx
);
1648 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1650 struct io_rings
*r
= ctx
->rings
;
1652 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1655 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1657 struct io_rings
*rings
= ctx
->rings
;
1660 tail
= ctx
->cached_cq_tail
;
1662 * writes to the cq entry need to come after reading head; the
1663 * control dependency is enough as we're using WRITE_ONCE to
1666 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1669 ctx
->cached_cq_tail
++;
1670 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1673 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1677 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1679 if (!ctx
->eventfd_async
)
1681 return io_wq_current_is_worker();
1684 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1686 if (waitqueue_active(&ctx
->wait
))
1687 wake_up(&ctx
->wait
);
1688 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1689 wake_up(&ctx
->sq_data
->wait
);
1690 if (io_should_trigger_evfd(ctx
))
1691 eventfd_signal(ctx
->cq_ev_fd
, 1);
1694 static void io_cqring_mark_overflow(struct io_ring_ctx
*ctx
)
1696 if (list_empty(&ctx
->cq_overflow_list
)) {
1697 clear_bit(0, &ctx
->sq_check_overflow
);
1698 clear_bit(0, &ctx
->cq_check_overflow
);
1699 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1703 /* Returns true if there are no backlogged entries after the flush */
1704 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1705 struct task_struct
*tsk
,
1706 struct files_struct
*files
)
1708 struct io_rings
*rings
= ctx
->rings
;
1709 struct io_kiocb
*req
, *tmp
;
1710 struct io_uring_cqe
*cqe
;
1711 unsigned long flags
;
1715 if (list_empty_careful(&ctx
->cq_overflow_list
))
1717 if ((ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
) ==
1718 rings
->cq_ring_entries
))
1722 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1724 /* if force is set, the ring is going away. always drop after that */
1726 ctx
->cq_overflow_flushed
= 1;
1729 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1730 if (!io_match_task(req
, tsk
, files
))
1733 cqe
= io_get_cqring(ctx
);
1737 list_move(&req
->compl.list
, &list
);
1739 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1740 WRITE_ONCE(cqe
->res
, req
->result
);
1741 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1743 ctx
->cached_cq_overflow
++;
1744 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1745 ctx
->cached_cq_overflow
);
1749 io_commit_cqring(ctx
);
1750 io_cqring_mark_overflow(ctx
);
1752 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1753 io_cqring_ev_posted(ctx
);
1755 while (!list_empty(&list
)) {
1756 req
= list_first_entry(&list
, struct io_kiocb
, compl.list
);
1757 list_del(&req
->compl.list
);
1764 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1766 struct io_ring_ctx
*ctx
= req
->ctx
;
1767 struct io_uring_cqe
*cqe
;
1769 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1772 * If we can't get a cq entry, userspace overflowed the
1773 * submission (by quite a lot). Increment the overflow count in
1776 cqe
= io_get_cqring(ctx
);
1778 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1779 WRITE_ONCE(cqe
->res
, res
);
1780 WRITE_ONCE(cqe
->flags
, cflags
);
1781 } else if (ctx
->cq_overflow_flushed
||
1782 atomic_read(&req
->task
->io_uring
->in_idle
)) {
1784 * If we're in ring overflow flush mode, or in task cancel mode,
1785 * then we cannot store the request for later flushing, we need
1786 * to drop it on the floor.
1788 ctx
->cached_cq_overflow
++;
1789 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1791 if (list_empty(&ctx
->cq_overflow_list
)) {
1792 set_bit(0, &ctx
->sq_check_overflow
);
1793 set_bit(0, &ctx
->cq_check_overflow
);
1794 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1798 req
->compl.cflags
= cflags
;
1799 refcount_inc(&req
->refs
);
1800 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1804 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1806 __io_cqring_fill_event(req
, res
, 0);
1809 static void io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1811 struct io_ring_ctx
*ctx
= req
->ctx
;
1812 unsigned long flags
;
1814 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1815 __io_cqring_fill_event(req
, res
, cflags
);
1816 io_commit_cqring(ctx
);
1817 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1819 io_cqring_ev_posted(ctx
);
1822 static void io_submit_flush_completions(struct io_comp_state
*cs
)
1824 struct io_ring_ctx
*ctx
= cs
->ctx
;
1826 spin_lock_irq(&ctx
->completion_lock
);
1827 while (!list_empty(&cs
->list
)) {
1828 struct io_kiocb
*req
;
1830 req
= list_first_entry(&cs
->list
, struct io_kiocb
, compl.list
);
1831 list_del(&req
->compl.list
);
1832 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
1835 * io_free_req() doesn't care about completion_lock unless one
1836 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1837 * because of a potential deadlock with req->work.fs->lock
1839 if (req
->flags
& (REQ_F_FAIL_LINK
|REQ_F_LINK_TIMEOUT
1840 |REQ_F_WORK_INITIALIZED
)) {
1841 spin_unlock_irq(&ctx
->completion_lock
);
1843 spin_lock_irq(&ctx
->completion_lock
);
1848 io_commit_cqring(ctx
);
1849 spin_unlock_irq(&ctx
->completion_lock
);
1851 io_cqring_ev_posted(ctx
);
1855 static void __io_req_complete(struct io_kiocb
*req
, long res
, unsigned cflags
,
1856 struct io_comp_state
*cs
)
1859 io_cqring_add_event(req
, res
, cflags
);
1864 req
->compl.cflags
= cflags
;
1865 list_add_tail(&req
->compl.list
, &cs
->list
);
1867 io_submit_flush_completions(cs
);
1871 static void io_req_complete(struct io_kiocb
*req
, long res
)
1873 __io_req_complete(req
, res
, 0, NULL
);
1876 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1878 return req
== (struct io_kiocb
*)
1879 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1882 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1884 struct io_kiocb
*req
;
1886 req
= ctx
->fallback_req
;
1887 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1893 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1894 struct io_submit_state
*state
)
1896 if (!state
->free_reqs
) {
1897 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1901 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1902 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
1905 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1906 * retry single alloc to be on the safe side.
1908 if (unlikely(ret
<= 0)) {
1909 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1910 if (!state
->reqs
[0])
1914 state
->free_reqs
= ret
;
1918 return state
->reqs
[state
->free_reqs
];
1920 return io_get_fallback_req(ctx
);
1923 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
1927 percpu_ref_put(req
->fixed_file_refs
);
1932 static void io_dismantle_req(struct io_kiocb
*req
)
1936 if (req
->async_data
)
1937 kfree(req
->async_data
);
1939 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
1941 io_req_clean_work(req
);
1944 static void __io_free_req(struct io_kiocb
*req
)
1946 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1947 struct io_ring_ctx
*ctx
= req
->ctx
;
1949 io_dismantle_req(req
);
1951 percpu_counter_dec(&tctx
->inflight
);
1952 if (atomic_read(&tctx
->in_idle
))
1953 wake_up(&tctx
->wait
);
1954 put_task_struct(req
->task
);
1956 if (likely(!io_is_fallback_req(req
)))
1957 kmem_cache_free(req_cachep
, req
);
1959 clear_bit_unlock(0, (unsigned long *) &ctx
->fallback_req
);
1960 percpu_ref_put(&ctx
->refs
);
1963 static inline void io_remove_next_linked(struct io_kiocb
*req
)
1965 struct io_kiocb
*nxt
= req
->link
;
1967 req
->link
= nxt
->link
;
1971 static void io_kill_linked_timeout(struct io_kiocb
*req
)
1973 struct io_ring_ctx
*ctx
= req
->ctx
;
1974 struct io_kiocb
*link
;
1975 bool cancelled
= false;
1976 unsigned long flags
;
1978 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1982 * Can happen if a linked timeout fired and link had been like
1983 * req -> link t-out -> link t-out [-> ...]
1985 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
1986 struct io_timeout_data
*io
= link
->async_data
;
1989 io_remove_next_linked(req
);
1990 link
->timeout
.head
= NULL
;
1991 ret
= hrtimer_try_to_cancel(&io
->timer
);
1993 io_cqring_fill_event(link
, -ECANCELED
);
1994 io_commit_cqring(ctx
);
1998 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1999 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2002 io_cqring_ev_posted(ctx
);
2008 static void io_fail_links(struct io_kiocb
*req
)
2010 struct io_kiocb
*link
, *nxt
;
2011 struct io_ring_ctx
*ctx
= req
->ctx
;
2012 unsigned long flags
;
2014 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2022 trace_io_uring_fail_link(req
, link
);
2023 io_cqring_fill_event(link
, -ECANCELED
);
2026 * It's ok to free under spinlock as they're not linked anymore,
2027 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2030 if (link
->flags
& REQ_F_WORK_INITIALIZED
)
2031 io_put_req_deferred(link
, 2);
2033 io_double_put_req(link
);
2036 io_commit_cqring(ctx
);
2037 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2039 io_cqring_ev_posted(ctx
);
2042 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2044 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
2045 io_kill_linked_timeout(req
);
2048 * If LINK is set, we have dependent requests in this chain. If we
2049 * didn't fail this request, queue the first one up, moving any other
2050 * dependencies to the next request. In case of failure, fail the rest
2053 if (likely(!(req
->flags
& REQ_F_FAIL_LINK
))) {
2054 struct io_kiocb
*nxt
= req
->link
;
2063 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2065 if (likely(!(req
->link
) && !(req
->flags
& REQ_F_LINK_TIMEOUT
)))
2067 return __io_req_find_next(req
);
2070 static int io_req_task_work_add(struct io_kiocb
*req
, bool twa_signal_ok
)
2072 struct task_struct
*tsk
= req
->task
;
2073 struct io_ring_ctx
*ctx
= req
->ctx
;
2074 enum task_work_notify_mode notify
;
2077 if (tsk
->flags
& PF_EXITING
)
2081 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2082 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2083 * processing task_work. There's no reliable way to tell if TWA_RESUME
2087 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) && twa_signal_ok
)
2088 notify
= TWA_SIGNAL
;
2090 ret
= task_work_add(tsk
, &req
->task_work
, notify
);
2092 wake_up_process(tsk
);
2097 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
2099 struct io_ring_ctx
*ctx
= req
->ctx
;
2101 spin_lock_irq(&ctx
->completion_lock
);
2102 io_cqring_fill_event(req
, error
);
2103 io_commit_cqring(ctx
);
2104 spin_unlock_irq(&ctx
->completion_lock
);
2106 io_cqring_ev_posted(ctx
);
2107 req_set_fail_links(req
);
2108 io_double_put_req(req
);
2111 static void io_req_task_cancel(struct callback_head
*cb
)
2113 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2114 struct io_ring_ctx
*ctx
= req
->ctx
;
2116 __io_req_task_cancel(req
, -ECANCELED
);
2117 percpu_ref_put(&ctx
->refs
);
2120 static void __io_req_task_submit(struct io_kiocb
*req
)
2122 struct io_ring_ctx
*ctx
= req
->ctx
;
2124 if (!__io_sq_thread_acquire_mm(ctx
) &&
2125 !__io_sq_thread_acquire_files(ctx
)) {
2126 mutex_lock(&ctx
->uring_lock
);
2127 __io_queue_sqe(req
, NULL
);
2128 mutex_unlock(&ctx
->uring_lock
);
2130 __io_req_task_cancel(req
, -EFAULT
);
2134 static void io_req_task_submit(struct callback_head
*cb
)
2136 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2137 struct io_ring_ctx
*ctx
= req
->ctx
;
2139 __io_req_task_submit(req
);
2140 percpu_ref_put(&ctx
->refs
);
2143 static void io_req_task_queue(struct io_kiocb
*req
)
2147 init_task_work(&req
->task_work
, io_req_task_submit
);
2148 percpu_ref_get(&req
->ctx
->refs
);
2150 ret
= io_req_task_work_add(req
, true);
2151 if (unlikely(ret
)) {
2152 struct task_struct
*tsk
;
2154 init_task_work(&req
->task_work
, io_req_task_cancel
);
2155 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2156 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2157 wake_up_process(tsk
);
2161 static inline void io_queue_next(struct io_kiocb
*req
)
2163 struct io_kiocb
*nxt
= io_req_find_next(req
);
2166 io_req_task_queue(nxt
);
2169 static void io_free_req(struct io_kiocb
*req
)
2176 void *reqs
[IO_IOPOLL_BATCH
];
2179 struct task_struct
*task
;
2183 static inline void io_init_req_batch(struct req_batch
*rb
)
2190 static void __io_req_free_batch_flush(struct io_ring_ctx
*ctx
,
2191 struct req_batch
*rb
)
2193 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
2194 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
2198 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2199 struct req_batch
*rb
)
2202 __io_req_free_batch_flush(ctx
, rb
);
2204 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2206 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2207 put_task_struct_many(rb
->task
, rb
->task_refs
);
2212 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
)
2214 if (unlikely(io_is_fallback_req(req
))) {
2220 if (req
->task
!= rb
->task
) {
2222 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2224 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2225 put_task_struct_many(rb
->task
, rb
->task_refs
);
2227 rb
->task
= req
->task
;
2232 io_dismantle_req(req
);
2233 rb
->reqs
[rb
->to_free
++] = req
;
2234 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
2235 __io_req_free_batch_flush(req
->ctx
, rb
);
2239 * Drop reference to request, return next in chain (if there is one) if this
2240 * was the last reference to this request.
2242 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2244 struct io_kiocb
*nxt
= NULL
;
2246 if (refcount_dec_and_test(&req
->refs
)) {
2247 nxt
= io_req_find_next(req
);
2253 static void io_put_req(struct io_kiocb
*req
)
2255 if (refcount_dec_and_test(&req
->refs
))
2259 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2261 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2266 static void io_free_req_deferred(struct io_kiocb
*req
)
2270 init_task_work(&req
->task_work
, io_put_req_deferred_cb
);
2271 ret
= io_req_task_work_add(req
, true);
2272 if (unlikely(ret
)) {
2273 struct task_struct
*tsk
;
2275 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2276 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2277 wake_up_process(tsk
);
2281 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2283 if (refcount_sub_and_test(refs
, &req
->refs
))
2284 io_free_req_deferred(req
);
2287 static struct io_wq_work
*io_steal_work(struct io_kiocb
*req
)
2289 struct io_kiocb
*nxt
;
2292 * A ref is owned by io-wq in which context we're. So, if that's the
2293 * last one, it's safe to steal next work. False negatives are Ok,
2294 * it just will be re-punted async in io_put_work()
2296 if (refcount_read(&req
->refs
) != 1)
2299 nxt
= io_req_find_next(req
);
2300 return nxt
? &nxt
->work
: NULL
;
2303 static void io_double_put_req(struct io_kiocb
*req
)
2305 /* drop both submit and complete references */
2306 if (refcount_sub_and_test(2, &req
->refs
))
2310 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
, bool noflush
)
2312 struct io_rings
*rings
= ctx
->rings
;
2314 if (test_bit(0, &ctx
->cq_check_overflow
)) {
2316 * noflush == true is from the waitqueue handler, just ensure
2317 * we wake up the task, and the next invocation will flush the
2318 * entries. We cannot safely to it from here.
2320 if (noflush
&& !list_empty(&ctx
->cq_overflow_list
))
2323 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2326 /* See comment at the top of this file */
2328 return ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
);
2331 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2333 struct io_rings
*rings
= ctx
->rings
;
2335 /* make sure SQ entry isn't read before tail */
2336 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2339 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2341 unsigned int cflags
;
2343 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2344 cflags
|= IORING_CQE_F_BUFFER
;
2345 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2350 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2352 struct io_buffer
*kbuf
;
2354 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2355 return io_put_kbuf(req
, kbuf
);
2358 static inline bool io_run_task_work(void)
2361 * Not safe to run on exiting task, and the task_work handling will
2362 * not add work to such a task.
2364 if (unlikely(current
->flags
& PF_EXITING
))
2366 if (current
->task_works
) {
2367 __set_current_state(TASK_RUNNING
);
2375 static void io_iopoll_queue(struct list_head
*again
)
2377 struct io_kiocb
*req
;
2380 req
= list_first_entry(again
, struct io_kiocb
, inflight_entry
);
2381 list_del(&req
->inflight_entry
);
2382 __io_complete_rw(req
, -EAGAIN
, 0, NULL
);
2383 } while (!list_empty(again
));
2387 * Find and free completed poll iocbs
2389 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2390 struct list_head
*done
)
2392 struct req_batch rb
;
2393 struct io_kiocb
*req
;
2396 /* order with ->result store in io_complete_rw_iopoll() */
2399 io_init_req_batch(&rb
);
2400 while (!list_empty(done
)) {
2403 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2404 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2406 req
->iopoll_completed
= 0;
2407 list_move_tail(&req
->inflight_entry
, &again
);
2410 list_del(&req
->inflight_entry
);
2412 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2413 cflags
= io_put_rw_kbuf(req
);
2415 __io_cqring_fill_event(req
, req
->result
, cflags
);
2418 if (refcount_dec_and_test(&req
->refs
))
2419 io_req_free_batch(&rb
, req
);
2422 io_commit_cqring(ctx
);
2423 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
2424 io_cqring_ev_posted(ctx
);
2425 io_req_free_batch_finish(ctx
, &rb
);
2427 if (!list_empty(&again
))
2428 io_iopoll_queue(&again
);
2431 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2434 struct io_kiocb
*req
, *tmp
;
2440 * Only spin for completions if we don't have multiple devices hanging
2441 * off our complete list, and we're under the requested amount.
2443 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2446 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2447 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2450 * Move completed and retryable entries to our local lists.
2451 * If we find a request that requires polling, break out
2452 * and complete those lists first, if we have entries there.
2454 if (READ_ONCE(req
->iopoll_completed
)) {
2455 list_move_tail(&req
->inflight_entry
, &done
);
2458 if (!list_empty(&done
))
2461 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2465 /* iopoll may have completed current req */
2466 if (READ_ONCE(req
->iopoll_completed
))
2467 list_move_tail(&req
->inflight_entry
, &done
);
2474 if (!list_empty(&done
))
2475 io_iopoll_complete(ctx
, nr_events
, &done
);
2481 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2482 * non-spinning poll check - we'll still enter the driver poll loop, but only
2483 * as a non-spinning completion check.
2485 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2488 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2491 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2494 if (*nr_events
>= min
)
2502 * We can't just wait for polled events to come to us, we have to actively
2503 * find and complete them.
2505 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2507 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2510 mutex_lock(&ctx
->uring_lock
);
2511 while (!list_empty(&ctx
->iopoll_list
)) {
2512 unsigned int nr_events
= 0;
2514 io_do_iopoll(ctx
, &nr_events
, 0);
2516 /* let it sleep and repeat later if can't complete a request */
2520 * Ensure we allow local-to-the-cpu processing to take place,
2521 * in this case we need to ensure that we reap all events.
2522 * Also let task_work, etc. to progress by releasing the mutex
2524 if (need_resched()) {
2525 mutex_unlock(&ctx
->uring_lock
);
2527 mutex_lock(&ctx
->uring_lock
);
2530 mutex_unlock(&ctx
->uring_lock
);
2533 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2535 unsigned int nr_events
= 0;
2536 int iters
= 0, ret
= 0;
2539 * We disallow the app entering submit/complete with polling, but we
2540 * still need to lock the ring to prevent racing with polled issue
2541 * that got punted to a workqueue.
2543 mutex_lock(&ctx
->uring_lock
);
2546 * Don't enter poll loop if we already have events pending.
2547 * If we do, we can potentially be spinning for commands that
2548 * already triggered a CQE (eg in error).
2550 if (io_cqring_events(ctx
, false))
2554 * If a submit got punted to a workqueue, we can have the
2555 * application entering polling for a command before it gets
2556 * issued. That app will hold the uring_lock for the duration
2557 * of the poll right here, so we need to take a breather every
2558 * now and then to ensure that the issue has a chance to add
2559 * the poll to the issued list. Otherwise we can spin here
2560 * forever, while the workqueue is stuck trying to acquire the
2563 if (!(++iters
& 7)) {
2564 mutex_unlock(&ctx
->uring_lock
);
2566 mutex_lock(&ctx
->uring_lock
);
2569 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2573 } while (min
&& !nr_events
&& !need_resched());
2575 mutex_unlock(&ctx
->uring_lock
);
2579 static void kiocb_end_write(struct io_kiocb
*req
)
2582 * Tell lockdep we inherited freeze protection from submission
2585 if (req
->flags
& REQ_F_ISREG
) {
2586 struct inode
*inode
= file_inode(req
->file
);
2588 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2590 file_end_write(req
->file
);
2593 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
,
2594 struct io_comp_state
*cs
)
2596 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2599 if (kiocb
->ki_flags
& IOCB_WRITE
)
2600 kiocb_end_write(req
);
2602 if (res
!= req
->result
)
2603 req_set_fail_links(req
);
2604 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2605 cflags
= io_put_rw_kbuf(req
);
2606 __io_req_complete(req
, res
, cflags
, cs
);
2610 static bool io_resubmit_prep(struct io_kiocb
*req
, int error
)
2612 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2613 ssize_t ret
= -ECANCELED
;
2614 struct iov_iter iter
;
2622 switch (req
->opcode
) {
2623 case IORING_OP_READV
:
2624 case IORING_OP_READ_FIXED
:
2625 case IORING_OP_READ
:
2628 case IORING_OP_WRITEV
:
2629 case IORING_OP_WRITE_FIXED
:
2630 case IORING_OP_WRITE
:
2634 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2639 if (!req
->async_data
) {
2640 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2643 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2651 req_set_fail_links(req
);
2656 static bool io_rw_reissue(struct io_kiocb
*req
, long res
)
2659 umode_t mode
= file_inode(req
->file
)->i_mode
;
2662 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2664 if ((res
!= -EAGAIN
&& res
!= -EOPNOTSUPP
) || io_wq_current_is_worker())
2667 ret
= io_sq_thread_acquire_mm_files(req
->ctx
, req
);
2669 if (io_resubmit_prep(req
, ret
)) {
2670 refcount_inc(&req
->refs
);
2671 io_queue_async_work(req
);
2679 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2680 struct io_comp_state
*cs
)
2682 if (!io_rw_reissue(req
, res
))
2683 io_complete_rw_common(&req
->rw
.kiocb
, res
, cs
);
2686 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2688 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2690 __io_complete_rw(req
, res
, res2
, NULL
);
2693 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2695 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2697 if (kiocb
->ki_flags
& IOCB_WRITE
)
2698 kiocb_end_write(req
);
2700 if (res
!= -EAGAIN
&& res
!= req
->result
)
2701 req_set_fail_links(req
);
2703 WRITE_ONCE(req
->result
, res
);
2704 /* order with io_poll_complete() checking ->result */
2706 WRITE_ONCE(req
->iopoll_completed
, 1);
2710 * After the iocb has been issued, it's safe to be found on the poll list.
2711 * Adding the kiocb to the list AFTER submission ensures that we don't
2712 * find it from a io_iopoll_getevents() thread before the issuer is done
2713 * accessing the kiocb cookie.
2715 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2717 struct io_ring_ctx
*ctx
= req
->ctx
;
2720 * Track whether we have multiple files in our lists. This will impact
2721 * how we do polling eventually, not spinning if we're on potentially
2722 * different devices.
2724 if (list_empty(&ctx
->iopoll_list
)) {
2725 ctx
->poll_multi_file
= false;
2726 } else if (!ctx
->poll_multi_file
) {
2727 struct io_kiocb
*list_req
;
2729 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2731 if (list_req
->file
!= req
->file
)
2732 ctx
->poll_multi_file
= true;
2736 * For fast devices, IO may have already completed. If it has, add
2737 * it to the front so we find it first.
2739 if (READ_ONCE(req
->iopoll_completed
))
2740 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2742 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2744 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2745 wq_has_sleeper(&ctx
->sq_data
->wait
))
2746 wake_up(&ctx
->sq_data
->wait
);
2749 static void __io_state_file_put(struct io_submit_state
*state
)
2751 if (state
->has_refs
)
2752 fput_many(state
->file
, state
->has_refs
);
2756 static inline void io_state_file_put(struct io_submit_state
*state
)
2759 __io_state_file_put(state
);
2763 * Get as many references to a file as we have IOs left in this submission,
2764 * assuming most submissions are for one file, or at least that each file
2765 * has more than one submission.
2767 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2773 if (state
->fd
== fd
) {
2777 __io_state_file_put(state
);
2779 state
->file
= fget_many(fd
, state
->ios_left
);
2784 state
->has_refs
= state
->ios_left
- 1;
2788 static bool io_bdev_nowait(struct block_device
*bdev
)
2791 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2798 * If we tracked the file through the SCM inflight mechanism, we could support
2799 * any file. For now, just ensure that anything potentially problematic is done
2802 static bool io_file_supports_async(struct file
*file
, int rw
)
2804 umode_t mode
= file_inode(file
)->i_mode
;
2806 if (S_ISBLK(mode
)) {
2807 if (io_bdev_nowait(file
->f_inode
->i_bdev
))
2811 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2813 if (S_ISREG(mode
)) {
2814 if (io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2815 file
->f_op
!= &io_uring_fops
)
2820 /* any ->read/write should understand O_NONBLOCK */
2821 if (file
->f_flags
& O_NONBLOCK
)
2824 if (!(file
->f_mode
& FMODE_NOWAIT
))
2828 return file
->f_op
->read_iter
!= NULL
;
2830 return file
->f_op
->write_iter
!= NULL
;
2833 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2835 struct io_ring_ctx
*ctx
= req
->ctx
;
2836 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2840 if (S_ISREG(file_inode(req
->file
)->i_mode
))
2841 req
->flags
|= REQ_F_ISREG
;
2843 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2844 if (kiocb
->ki_pos
== -1 && !(req
->file
->f_mode
& FMODE_STREAM
)) {
2845 req
->flags
|= REQ_F_CUR_POS
;
2846 kiocb
->ki_pos
= req
->file
->f_pos
;
2848 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2849 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2850 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2854 ioprio
= READ_ONCE(sqe
->ioprio
);
2856 ret
= ioprio_check_cap(ioprio
);
2860 kiocb
->ki_ioprio
= ioprio
;
2862 kiocb
->ki_ioprio
= get_current_ioprio();
2864 /* don't allow async punt if RWF_NOWAIT was requested */
2865 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2866 req
->flags
|= REQ_F_NOWAIT
;
2868 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2869 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2870 !kiocb
->ki_filp
->f_op
->iopoll
)
2873 kiocb
->ki_flags
|= IOCB_HIPRI
;
2874 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2875 req
->iopoll_completed
= 0;
2877 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2879 kiocb
->ki_complete
= io_complete_rw
;
2882 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2883 req
->rw
.len
= READ_ONCE(sqe
->len
);
2884 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2888 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2894 case -ERESTARTNOINTR
:
2895 case -ERESTARTNOHAND
:
2896 case -ERESTART_RESTARTBLOCK
:
2898 * We can't just restart the syscall, since previously
2899 * submitted sqes may already be in progress. Just fail this
2905 kiocb
->ki_complete(kiocb
, ret
, 0);
2909 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2910 struct io_comp_state
*cs
)
2912 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2913 struct io_async_rw
*io
= req
->async_data
;
2915 /* add previously done IO, if any */
2916 if (io
&& io
->bytes_done
> 0) {
2918 ret
= io
->bytes_done
;
2920 ret
+= io
->bytes_done
;
2923 if (req
->flags
& REQ_F_CUR_POS
)
2924 req
->file
->f_pos
= kiocb
->ki_pos
;
2925 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2926 __io_complete_rw(req
, ret
, 0, cs
);
2928 io_rw_done(kiocb
, ret
);
2931 static ssize_t
io_import_fixed(struct io_kiocb
*req
, int rw
,
2932 struct iov_iter
*iter
)
2934 struct io_ring_ctx
*ctx
= req
->ctx
;
2935 size_t len
= req
->rw
.len
;
2936 struct io_mapped_ubuf
*imu
;
2937 u16 index
, buf_index
= req
->buf_index
;
2941 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
2943 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
2944 imu
= &ctx
->user_bufs
[index
];
2945 buf_addr
= req
->rw
.addr
;
2948 if (buf_addr
+ len
< buf_addr
)
2950 /* not inside the mapped region */
2951 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
2955 * May not be a start of buffer, set size appropriately
2956 * and advance us to the beginning.
2958 offset
= buf_addr
- imu
->ubuf
;
2959 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
2963 * Don't use iov_iter_advance() here, as it's really slow for
2964 * using the latter parts of a big fixed buffer - it iterates
2965 * over each segment manually. We can cheat a bit here, because
2968 * 1) it's a BVEC iter, we set it up
2969 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2970 * first and last bvec
2972 * So just find our index, and adjust the iterator afterwards.
2973 * If the offset is within the first bvec (or the whole first
2974 * bvec, just use iov_iter_advance(). This makes it easier
2975 * since we can just skip the first segment, which may not
2976 * be PAGE_SIZE aligned.
2978 const struct bio_vec
*bvec
= imu
->bvec
;
2980 if (offset
<= bvec
->bv_len
) {
2981 iov_iter_advance(iter
, offset
);
2983 unsigned long seg_skip
;
2985 /* skip first vec */
2986 offset
-= bvec
->bv_len
;
2987 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
2989 iter
->bvec
= bvec
+ seg_skip
;
2990 iter
->nr_segs
-= seg_skip
;
2991 iter
->count
-= bvec
->bv_len
+ offset
;
2992 iter
->iov_offset
= offset
& ~PAGE_MASK
;
2999 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3002 mutex_unlock(&ctx
->uring_lock
);
3005 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3008 * "Normal" inline submissions always hold the uring_lock, since we
3009 * grab it from the system call. Same is true for the SQPOLL offload.
3010 * The only exception is when we've detached the request and issue it
3011 * from an async worker thread, grab the lock for that case.
3014 mutex_lock(&ctx
->uring_lock
);
3017 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3018 int bgid
, struct io_buffer
*kbuf
,
3021 struct io_buffer
*head
;
3023 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3026 io_ring_submit_lock(req
->ctx
, needs_lock
);
3028 lockdep_assert_held(&req
->ctx
->uring_lock
);
3030 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
3032 if (!list_empty(&head
->list
)) {
3033 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3035 list_del(&kbuf
->list
);
3038 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
3040 if (*len
> kbuf
->len
)
3043 kbuf
= ERR_PTR(-ENOBUFS
);
3046 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3051 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3054 struct io_buffer
*kbuf
;
3057 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3058 bgid
= req
->buf_index
;
3059 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3062 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3063 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3064 return u64_to_user_ptr(kbuf
->addr
);
3067 #ifdef CONFIG_COMPAT
3068 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3071 struct compat_iovec __user
*uiov
;
3072 compat_ssize_t clen
;
3076 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3077 if (!access_ok(uiov
, sizeof(*uiov
)))
3079 if (__get_user(clen
, &uiov
->iov_len
))
3085 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3087 return PTR_ERR(buf
);
3088 iov
[0].iov_base
= buf
;
3089 iov
[0].iov_len
= (compat_size_t
) len
;
3094 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3097 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3101 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3104 len
= iov
[0].iov_len
;
3107 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3109 return PTR_ERR(buf
);
3110 iov
[0].iov_base
= buf
;
3111 iov
[0].iov_len
= len
;
3115 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3118 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3119 struct io_buffer
*kbuf
;
3121 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3122 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3123 iov
[0].iov_len
= kbuf
->len
;
3128 else if (req
->rw
.len
> 1)
3131 #ifdef CONFIG_COMPAT
3132 if (req
->ctx
->compat
)
3133 return io_compat_import(req
, iov
, needs_lock
);
3136 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3139 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
3140 struct iovec
**iovec
, struct iov_iter
*iter
,
3143 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3144 size_t sqe_len
= req
->rw
.len
;
3148 opcode
= req
->opcode
;
3149 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3151 return io_import_fixed(req
, rw
, iter
);
3154 /* buffer index only valid with fixed read/write, or buffer select */
3155 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3158 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3159 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3160 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3162 return PTR_ERR(buf
);
3163 req
->rw
.len
= sqe_len
;
3166 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3171 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3172 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3174 ret
= (*iovec
)->iov_len
;
3175 iov_iter_init(iter
, rw
, *iovec
, 1, ret
);
3181 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3185 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3187 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3191 * For files that don't have ->read_iter() and ->write_iter(), handle them
3192 * by looping over ->read() or ->write() manually.
3194 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3196 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3197 struct file
*file
= req
->file
;
3201 * Don't support polled IO through this interface, and we can't
3202 * support non-blocking either. For the latter, this just causes
3203 * the kiocb to be handled from an async context.
3205 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3207 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3210 while (iov_iter_count(iter
)) {
3214 if (!iov_iter_is_bvec(iter
)) {
3215 iovec
= iov_iter_iovec(iter
);
3217 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3218 iovec
.iov_len
= req
->rw
.len
;
3222 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3223 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3225 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3226 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3235 if (nr
!= iovec
.iov_len
)
3239 iov_iter_advance(iter
, nr
);
3245 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3246 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3248 struct io_async_rw
*rw
= req
->async_data
;
3250 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3251 rw
->free_iovec
= iovec
;
3253 /* can only be fixed buffers, no need to do anything */
3254 if (iter
->type
== ITER_BVEC
)
3257 unsigned iov_off
= 0;
3259 rw
->iter
.iov
= rw
->fast_iov
;
3260 if (iter
->iov
!= fast_iov
) {
3261 iov_off
= iter
->iov
- fast_iov
;
3262 rw
->iter
.iov
+= iov_off
;
3264 if (rw
->fast_iov
!= fast_iov
)
3265 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3266 sizeof(struct iovec
) * iter
->nr_segs
);
3268 req
->flags
|= REQ_F_NEED_CLEANUP
;
3272 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3274 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3275 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3276 return req
->async_data
== NULL
;
3279 static int io_alloc_async_data(struct io_kiocb
*req
)
3281 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3284 return __io_alloc_async_data(req
);
3287 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3288 const struct iovec
*fast_iov
,
3289 struct iov_iter
*iter
, bool force
)
3291 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3293 if (!req
->async_data
) {
3294 if (__io_alloc_async_data(req
))
3297 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3302 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3304 struct io_async_rw
*iorw
= req
->async_data
;
3305 struct iovec
*iov
= iorw
->fast_iov
;
3308 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3309 if (unlikely(ret
< 0))
3312 iorw
->bytes_done
= 0;
3313 iorw
->free_iovec
= iov
;
3315 req
->flags
|= REQ_F_NEED_CLEANUP
;
3319 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3323 ret
= io_prep_rw(req
, sqe
);
3327 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3330 /* either don't need iovec imported or already have it */
3331 if (!req
->async_data
)
3333 return io_rw_prep_async(req
, READ
);
3337 * This is our waitqueue callback handler, registered through lock_page_async()
3338 * when we initially tried to do the IO with the iocb armed our waitqueue.
3339 * This gets called when the page is unlocked, and we generally expect that to
3340 * happen when the page IO is completed and the page is now uptodate. This will
3341 * queue a task_work based retry of the operation, attempting to copy the data
3342 * again. If the latter fails because the page was NOT uptodate, then we will
3343 * do a thread based blocking retry of the operation. That's the unexpected
3346 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3347 int sync
, void *arg
)
3349 struct wait_page_queue
*wpq
;
3350 struct io_kiocb
*req
= wait
->private;
3351 struct wait_page_key
*key
= arg
;
3354 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3356 if (!wake_page_match(wpq
, key
))
3359 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3360 list_del_init(&wait
->entry
);
3362 init_task_work(&req
->task_work
, io_req_task_submit
);
3363 percpu_ref_get(&req
->ctx
->refs
);
3365 /* submit ref gets dropped, acquire a new one */
3366 refcount_inc(&req
->refs
);
3367 ret
= io_req_task_work_add(req
, true);
3368 if (unlikely(ret
)) {
3369 struct task_struct
*tsk
;
3371 /* queue just for cancelation */
3372 init_task_work(&req
->task_work
, io_req_task_cancel
);
3373 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
3374 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
3375 wake_up_process(tsk
);
3381 * This controls whether a given IO request should be armed for async page
3382 * based retry. If we return false here, the request is handed to the async
3383 * worker threads for retry. If we're doing buffered reads on a regular file,
3384 * we prepare a private wait_page_queue entry and retry the operation. This
3385 * will either succeed because the page is now uptodate and unlocked, or it
3386 * will register a callback when the page is unlocked at IO completion. Through
3387 * that callback, io_uring uses task_work to setup a retry of the operation.
3388 * That retry will attempt the buffered read again. The retry will generally
3389 * succeed, or in rare cases where it fails, we then fall back to using the
3390 * async worker threads for a blocking retry.
3392 static bool io_rw_should_retry(struct io_kiocb
*req
)
3394 struct io_async_rw
*rw
= req
->async_data
;
3395 struct wait_page_queue
*wait
= &rw
->wpq
;
3396 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3398 /* never retry for NOWAIT, we just complete with -EAGAIN */
3399 if (req
->flags
& REQ_F_NOWAIT
)
3402 /* Only for buffered IO */
3403 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3407 * just use poll if we can, and don't attempt if the fs doesn't
3408 * support callback based unlocks
3410 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3413 wait
->wait
.func
= io_async_buf_func
;
3414 wait
->wait
.private = req
;
3415 wait
->wait
.flags
= 0;
3416 INIT_LIST_HEAD(&wait
->wait
.entry
);
3417 kiocb
->ki_flags
|= IOCB_WAITQ
;
3418 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3419 kiocb
->ki_waitq
= wait
;
3423 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3425 if (req
->file
->f_op
->read_iter
)
3426 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3427 else if (req
->file
->f_op
->read
)
3428 return loop_rw_iter(READ
, req
, iter
);
3433 static int io_read(struct io_kiocb
*req
, bool force_nonblock
,
3434 struct io_comp_state
*cs
)
3436 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3437 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3438 struct iov_iter __iter
, *iter
= &__iter
;
3439 struct io_async_rw
*rw
= req
->async_data
;
3440 ssize_t io_size
, ret
, ret2
;
3447 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3451 io_size
= iov_iter_count(iter
);
3452 req
->result
= io_size
;
3455 /* Ensure we clear previously set non-block flag */
3456 if (!force_nonblock
)
3457 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3459 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3462 /* If the file doesn't support async, just async punt */
3463 no_async
= force_nonblock
&& !io_file_supports_async(req
->file
, READ
);
3467 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3471 ret
= io_iter_do_read(req
, iter
);
3475 } else if (ret
== -EIOCBQUEUED
) {
3478 } else if (ret
== -EAGAIN
) {
3479 /* IOPOLL retry should happen for io-wq threads */
3480 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3482 /* no retry on NONBLOCK marked file */
3483 if (req
->file
->f_flags
& O_NONBLOCK
)
3485 /* some cases will consume bytes even on error returns */
3486 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3489 } else if (ret
< 0) {
3490 /* make sure -ERESTARTSYS -> -EINTR is done */
3494 /* read it all, or we did blocking attempt. no retry. */
3495 if (!iov_iter_count(iter
) || !force_nonblock
||
3496 (req
->file
->f_flags
& O_NONBLOCK
))
3501 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3508 rw
= req
->async_data
;
3509 /* it's copied and will be cleaned with ->io */
3511 /* now use our persistent iterator, if we aren't already */
3514 rw
->bytes_done
+= ret
;
3515 /* if we can retry, do so with the callbacks armed */
3516 if (!io_rw_should_retry(req
)) {
3517 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3522 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3523 * get -EIOCBQUEUED, then we'll get a notification when the desired
3524 * page gets unlocked. We can also get a partial read here, and if we
3525 * do, then just retry at the new offset.
3527 ret
= io_iter_do_read(req
, iter
);
3528 if (ret
== -EIOCBQUEUED
) {
3531 } else if (ret
> 0 && ret
< io_size
) {
3532 /* we got some bytes, but not all. retry. */
3536 kiocb_done(kiocb
, ret
, cs
);
3539 /* it's reportedly faster than delegating the null check to kfree() */
3545 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3549 ret
= io_prep_rw(req
, sqe
);
3553 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3556 /* either don't need iovec imported or already have it */
3557 if (!req
->async_data
)
3559 return io_rw_prep_async(req
, WRITE
);
3562 static int io_write(struct io_kiocb
*req
, bool force_nonblock
,
3563 struct io_comp_state
*cs
)
3565 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3566 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3567 struct iov_iter __iter
, *iter
= &__iter
;
3568 struct io_async_rw
*rw
= req
->async_data
;
3569 ssize_t ret
, ret2
, io_size
;
3575 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3579 io_size
= iov_iter_count(iter
);
3580 req
->result
= io_size
;
3582 /* Ensure we clear previously set non-block flag */
3583 if (!force_nonblock
)
3584 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3586 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3588 /* If the file doesn't support async, just async punt */
3589 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3592 /* file path doesn't support NOWAIT for non-direct_IO */
3593 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3594 (req
->flags
& REQ_F_ISREG
))
3597 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3602 * Open-code file_start_write here to grab freeze protection,
3603 * which will be released by another thread in
3604 * io_complete_rw(). Fool lockdep by telling it the lock got
3605 * released so that it doesn't complain about the held lock when
3606 * we return to userspace.
3608 if (req
->flags
& REQ_F_ISREG
) {
3609 sb_start_write(file_inode(req
->file
)->i_sb
);
3610 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3613 kiocb
->ki_flags
|= IOCB_WRITE
;
3615 if (req
->file
->f_op
->write_iter
)
3616 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3617 else if (req
->file
->f_op
->write
)
3618 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3623 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3624 * retry them without IOCB_NOWAIT.
3626 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3628 /* no retry on NONBLOCK marked file */
3629 if (ret2
== -EAGAIN
&& (req
->file
->f_flags
& O_NONBLOCK
))
3631 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3632 /* IOPOLL retry should happen for io-wq threads */
3633 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3636 kiocb_done(kiocb
, ret2
, cs
);
3639 /* some cases will consume bytes even on error returns */
3640 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3641 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3646 /* it's reportedly faster than delegating the null check to kfree() */
3652 static int io_renameat_prep(struct io_kiocb
*req
,
3653 const struct io_uring_sqe
*sqe
)
3655 struct io_rename
*ren
= &req
->rename
;
3656 const char __user
*oldf
, *newf
;
3658 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3661 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3662 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3663 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3664 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3665 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3667 ren
->oldpath
= getname(oldf
);
3668 if (IS_ERR(ren
->oldpath
))
3669 return PTR_ERR(ren
->oldpath
);
3671 ren
->newpath
= getname(newf
);
3672 if (IS_ERR(ren
->newpath
)) {
3673 putname(ren
->oldpath
);
3674 return PTR_ERR(ren
->newpath
);
3677 req
->flags
|= REQ_F_NEED_CLEANUP
;
3681 static int io_renameat(struct io_kiocb
*req
, bool force_nonblock
)
3683 struct io_rename
*ren
= &req
->rename
;
3689 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3690 ren
->newpath
, ren
->flags
);
3692 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3694 req_set_fail_links(req
);
3695 io_req_complete(req
, ret
);
3699 static int io_unlinkat_prep(struct io_kiocb
*req
,
3700 const struct io_uring_sqe
*sqe
)
3702 struct io_unlink
*un
= &req
->unlink
;
3703 const char __user
*fname
;
3705 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3708 un
->dfd
= READ_ONCE(sqe
->fd
);
3710 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3711 if (un
->flags
& ~AT_REMOVEDIR
)
3714 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3715 un
->filename
= getname(fname
);
3716 if (IS_ERR(un
->filename
))
3717 return PTR_ERR(un
->filename
);
3719 req
->flags
|= REQ_F_NEED_CLEANUP
;
3723 static int io_unlinkat(struct io_kiocb
*req
, bool force_nonblock
)
3725 struct io_unlink
*un
= &req
->unlink
;
3731 if (un
->flags
& AT_REMOVEDIR
)
3732 ret
= do_rmdir(un
->dfd
, un
->filename
);
3734 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3736 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3738 req_set_fail_links(req
);
3739 io_req_complete(req
, ret
);
3743 static int io_shutdown_prep(struct io_kiocb
*req
,
3744 const struct io_uring_sqe
*sqe
)
3746 #if defined(CONFIG_NET)
3747 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3749 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3753 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3760 static int io_shutdown(struct io_kiocb
*req
, bool force_nonblock
)
3762 #if defined(CONFIG_NET)
3763 struct socket
*sock
;
3769 sock
= sock_from_file(req
->file
, &ret
);
3770 if (unlikely(!sock
))
3773 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3774 io_req_complete(req
, ret
);
3781 static int __io_splice_prep(struct io_kiocb
*req
,
3782 const struct io_uring_sqe
*sqe
)
3784 struct io_splice
* sp
= &req
->splice
;
3785 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3787 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3791 sp
->len
= READ_ONCE(sqe
->len
);
3792 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3794 if (unlikely(sp
->flags
& ~valid_flags
))
3797 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3798 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3801 req
->flags
|= REQ_F_NEED_CLEANUP
;
3803 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3805 * Splice operation will be punted aync, and here need to
3806 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3808 io_req_init_async(req
);
3809 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3815 static int io_tee_prep(struct io_kiocb
*req
,
3816 const struct io_uring_sqe
*sqe
)
3818 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3820 return __io_splice_prep(req
, sqe
);
3823 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3825 struct io_splice
*sp
= &req
->splice
;
3826 struct file
*in
= sp
->file_in
;
3827 struct file
*out
= sp
->file_out
;
3828 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3834 ret
= do_tee(in
, out
, sp
->len
, flags
);
3836 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3837 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3840 req_set_fail_links(req
);
3841 io_req_complete(req
, ret
);
3845 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3847 struct io_splice
* sp
= &req
->splice
;
3849 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3850 sp
->off_out
= READ_ONCE(sqe
->off
);
3851 return __io_splice_prep(req
, sqe
);
3854 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3856 struct io_splice
*sp
= &req
->splice
;
3857 struct file
*in
= sp
->file_in
;
3858 struct file
*out
= sp
->file_out
;
3859 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3860 loff_t
*poff_in
, *poff_out
;
3866 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3867 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3870 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3872 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3873 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3876 req_set_fail_links(req
);
3877 io_req_complete(req
, ret
);
3882 * IORING_OP_NOP just posts a completion event, nothing else.
3884 static int io_nop(struct io_kiocb
*req
, struct io_comp_state
*cs
)
3886 struct io_ring_ctx
*ctx
= req
->ctx
;
3888 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3891 __io_req_complete(req
, 0, 0, cs
);
3895 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3897 struct io_ring_ctx
*ctx
= req
->ctx
;
3902 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3904 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3907 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3908 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3911 req
->sync
.off
= READ_ONCE(sqe
->off
);
3912 req
->sync
.len
= READ_ONCE(sqe
->len
);
3916 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3918 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3921 /* fsync always requires a blocking context */
3925 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3926 end
> 0 ? end
: LLONG_MAX
,
3927 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3929 req_set_fail_links(req
);
3930 io_req_complete(req
, ret
);
3934 static int io_fallocate_prep(struct io_kiocb
*req
,
3935 const struct io_uring_sqe
*sqe
)
3937 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3939 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3942 req
->sync
.off
= READ_ONCE(sqe
->off
);
3943 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3944 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3948 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
3952 /* fallocate always requiring blocking context */
3955 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3958 req_set_fail_links(req
);
3959 io_req_complete(req
, ret
);
3963 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3965 const char __user
*fname
;
3968 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3970 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3973 /* open.how should be already initialised */
3974 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
3975 req
->open
.how
.flags
|= O_LARGEFILE
;
3977 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3978 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3979 req
->open
.filename
= getname(fname
);
3980 if (IS_ERR(req
->open
.filename
)) {
3981 ret
= PTR_ERR(req
->open
.filename
);
3982 req
->open
.filename
= NULL
;
3985 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
3986 req
->open
.ignore_nonblock
= false;
3987 req
->flags
|= REQ_F_NEED_CLEANUP
;
3991 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3995 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3997 mode
= READ_ONCE(sqe
->len
);
3998 flags
= READ_ONCE(sqe
->open_flags
);
3999 req
->open
.how
= build_open_how(flags
, mode
);
4000 return __io_openat_prep(req
, sqe
);
4003 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4005 struct open_how __user
*how
;
4009 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4011 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4012 len
= READ_ONCE(sqe
->len
);
4013 if (len
< OPEN_HOW_SIZE_VER0
)
4016 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4021 return __io_openat_prep(req
, sqe
);
4024 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
4026 struct open_flags op
;
4030 if (force_nonblock
&& !req
->open
.ignore_nonblock
)
4033 ret
= build_open_flags(&req
->open
.how
, &op
);
4037 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4041 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4044 ret
= PTR_ERR(file
);
4046 * A work-around to ensure that /proc/self works that way
4047 * that it should - if we get -EOPNOTSUPP back, then assume
4048 * that proc_self_get_link() failed us because we're in async
4049 * context. We should be safe to retry this from the task
4050 * itself with force_nonblock == false set, as it should not
4051 * block on lookup. Would be nice to know this upfront and
4052 * avoid the async dance, but doesn't seem feasible.
4054 if (ret
== -EOPNOTSUPP
&& io_wq_current_is_worker()) {
4055 req
->open
.ignore_nonblock
= true;
4056 refcount_inc(&req
->refs
);
4057 io_req_task_queue(req
);
4061 fsnotify_open(file
);
4062 fd_install(ret
, file
);
4065 putname(req
->open
.filename
);
4066 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4068 req_set_fail_links(req
);
4069 io_req_complete(req
, ret
);
4073 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
4075 return io_openat2(req
, force_nonblock
);
4078 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4079 const struct io_uring_sqe
*sqe
)
4081 struct io_provide_buf
*p
= &req
->pbuf
;
4084 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
4087 tmp
= READ_ONCE(sqe
->fd
);
4088 if (!tmp
|| tmp
> USHRT_MAX
)
4091 memset(p
, 0, sizeof(*p
));
4093 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4097 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4098 int bgid
, unsigned nbufs
)
4102 /* shouldn't happen */
4106 /* the head kbuf is the list itself */
4107 while (!list_empty(&buf
->list
)) {
4108 struct io_buffer
*nxt
;
4110 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4111 list_del(&nxt
->list
);
4118 idr_remove(&ctx
->io_buffer_idr
, bgid
);
4123 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4124 struct io_comp_state
*cs
)
4126 struct io_provide_buf
*p
= &req
->pbuf
;
4127 struct io_ring_ctx
*ctx
= req
->ctx
;
4128 struct io_buffer
*head
;
4131 io_ring_submit_lock(ctx
, !force_nonblock
);
4133 lockdep_assert_held(&ctx
->uring_lock
);
4136 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4138 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4140 io_ring_submit_lock(ctx
, !force_nonblock
);
4142 req_set_fail_links(req
);
4143 __io_req_complete(req
, ret
, 0, cs
);
4147 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4148 const struct io_uring_sqe
*sqe
)
4150 struct io_provide_buf
*p
= &req
->pbuf
;
4153 if (sqe
->ioprio
|| sqe
->rw_flags
)
4156 tmp
= READ_ONCE(sqe
->fd
);
4157 if (!tmp
|| tmp
> USHRT_MAX
)
4160 p
->addr
= READ_ONCE(sqe
->addr
);
4161 p
->len
= READ_ONCE(sqe
->len
);
4163 if (!access_ok(u64_to_user_ptr(p
->addr
), (p
->len
* p
->nbufs
)))
4166 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4167 tmp
= READ_ONCE(sqe
->off
);
4168 if (tmp
> USHRT_MAX
)
4174 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4176 struct io_buffer
*buf
;
4177 u64 addr
= pbuf
->addr
;
4178 int i
, bid
= pbuf
->bid
;
4180 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4181 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
4186 buf
->len
= pbuf
->len
;
4191 INIT_LIST_HEAD(&buf
->list
);
4194 list_add_tail(&buf
->list
, &(*head
)->list
);
4198 return i
? i
: -ENOMEM
;
4201 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4202 struct io_comp_state
*cs
)
4204 struct io_provide_buf
*p
= &req
->pbuf
;
4205 struct io_ring_ctx
*ctx
= req
->ctx
;
4206 struct io_buffer
*head
, *list
;
4209 io_ring_submit_lock(ctx
, !force_nonblock
);
4211 lockdep_assert_held(&ctx
->uring_lock
);
4213 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4215 ret
= io_add_buffers(p
, &head
);
4220 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
4223 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4228 io_ring_submit_unlock(ctx
, !force_nonblock
);
4230 req_set_fail_links(req
);
4231 __io_req_complete(req
, ret
, 0, cs
);
4235 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4236 const struct io_uring_sqe
*sqe
)
4238 #if defined(CONFIG_EPOLL)
4239 if (sqe
->ioprio
|| sqe
->buf_index
)
4241 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4244 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4245 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4246 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4248 if (ep_op_has_event(req
->epoll
.op
)) {
4249 struct epoll_event __user
*ev
;
4251 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4252 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4262 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
,
4263 struct io_comp_state
*cs
)
4265 #if defined(CONFIG_EPOLL)
4266 struct io_epoll
*ie
= &req
->epoll
;
4269 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4270 if (force_nonblock
&& ret
== -EAGAIN
)
4274 req_set_fail_links(req
);
4275 __io_req_complete(req
, ret
, 0, cs
);
4282 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4284 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4285 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4287 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4290 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4291 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4292 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4299 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
4301 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4302 struct io_madvise
*ma
= &req
->madvise
;
4308 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4310 req_set_fail_links(req
);
4311 io_req_complete(req
, ret
);
4318 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4320 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4322 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4325 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4326 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4327 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4331 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
4333 struct io_fadvise
*fa
= &req
->fadvise
;
4336 if (force_nonblock
) {
4337 switch (fa
->advice
) {
4338 case POSIX_FADV_NORMAL
:
4339 case POSIX_FADV_RANDOM
:
4340 case POSIX_FADV_SEQUENTIAL
:
4347 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4349 req_set_fail_links(req
);
4350 io_req_complete(req
, ret
);
4354 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4356 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4358 if (sqe
->ioprio
|| sqe
->buf_index
)
4360 if (req
->flags
& REQ_F_FIXED_FILE
)
4363 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4364 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4365 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4366 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4367 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4372 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
4374 struct io_statx
*ctx
= &req
->statx
;
4377 if (force_nonblock
) {
4378 /* only need file table for an actual valid fd */
4379 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4380 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4384 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4388 req_set_fail_links(req
);
4389 io_req_complete(req
, ret
);
4393 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4396 * If we queue this for async, it must not be cancellable. That would
4397 * leave the 'file' in an undeterminate state, and here need to modify
4398 * io_wq_work.flags, so initialize io_wq_work firstly.
4400 io_req_init_async(req
);
4401 req
->work
.flags
|= IO_WQ_WORK_NO_CANCEL
;
4403 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4405 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4406 sqe
->rw_flags
|| sqe
->buf_index
)
4408 if (req
->flags
& REQ_F_FIXED_FILE
)
4411 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4412 if ((req
->file
&& req
->file
->f_op
== &io_uring_fops
))
4415 req
->close
.put_file
= NULL
;
4419 static int io_close(struct io_kiocb
*req
, bool force_nonblock
,
4420 struct io_comp_state
*cs
)
4422 struct io_close
*close
= &req
->close
;
4425 /* might be already done during nonblock submission */
4426 if (!close
->put_file
) {
4427 ret
= __close_fd_get_file(close
->fd
, &close
->put_file
);
4429 return (ret
== -ENOENT
) ? -EBADF
: ret
;
4432 /* if the file has a flush method, be safe and punt to async */
4433 if (close
->put_file
->f_op
->flush
&& force_nonblock
) {
4434 /* was never set, but play safe */
4435 req
->flags
&= ~REQ_F_NOWAIT
;
4436 /* avoid grabbing files - we don't need the files */
4437 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4441 /* No ->flush() or already async, safely close from here */
4442 ret
= filp_close(close
->put_file
, req
->work
.identity
->files
);
4444 req_set_fail_links(req
);
4445 fput(close
->put_file
);
4446 close
->put_file
= NULL
;
4447 __io_req_complete(req
, ret
, 0, cs
);
4451 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4453 struct io_ring_ctx
*ctx
= req
->ctx
;
4458 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4460 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4463 req
->sync
.off
= READ_ONCE(sqe
->off
);
4464 req
->sync
.len
= READ_ONCE(sqe
->len
);
4465 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4469 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
4473 /* sync_file_range always requires a blocking context */
4477 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4480 req_set_fail_links(req
);
4481 io_req_complete(req
, ret
);
4485 #if defined(CONFIG_NET)
4486 static int io_setup_async_msg(struct io_kiocb
*req
,
4487 struct io_async_msghdr
*kmsg
)
4489 struct io_async_msghdr
*async_msg
= req
->async_data
;
4493 if (io_alloc_async_data(req
)) {
4494 if (kmsg
->iov
!= kmsg
->fast_iov
)
4498 async_msg
= req
->async_data
;
4499 req
->flags
|= REQ_F_NEED_CLEANUP
;
4500 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4504 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4505 struct io_async_msghdr
*iomsg
)
4507 iomsg
->iov
= iomsg
->fast_iov
;
4508 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4509 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4510 req
->sr_msg
.msg_flags
, &iomsg
->iov
);
4513 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4515 struct io_async_msghdr
*async_msg
= req
->async_data
;
4516 struct io_sr_msg
*sr
= &req
->sr_msg
;
4519 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4522 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4523 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4524 sr
->len
= READ_ONCE(sqe
->len
);
4526 #ifdef CONFIG_COMPAT
4527 if (req
->ctx
->compat
)
4528 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4531 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4533 ret
= io_sendmsg_copy_hdr(req
, async_msg
);
4535 req
->flags
|= REQ_F_NEED_CLEANUP
;
4539 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4540 struct io_comp_state
*cs
)
4542 struct io_async_msghdr iomsg
, *kmsg
;
4543 struct socket
*sock
;
4547 sock
= sock_from_file(req
->file
, &ret
);
4548 if (unlikely(!sock
))
4551 if (req
->async_data
) {
4552 kmsg
= req
->async_data
;
4553 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4554 /* if iov is set, it's allocated already */
4556 kmsg
->iov
= kmsg
->fast_iov
;
4557 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4559 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4565 flags
= req
->sr_msg
.msg_flags
;
4566 if (flags
& MSG_DONTWAIT
)
4567 req
->flags
|= REQ_F_NOWAIT
;
4568 else if (force_nonblock
)
4569 flags
|= MSG_DONTWAIT
;
4571 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4572 if (force_nonblock
&& ret
== -EAGAIN
)
4573 return io_setup_async_msg(req
, kmsg
);
4574 if (ret
== -ERESTARTSYS
)
4577 if (kmsg
->iov
!= kmsg
->fast_iov
)
4579 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4581 req_set_fail_links(req
);
4582 __io_req_complete(req
, ret
, 0, cs
);
4586 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4587 struct io_comp_state
*cs
)
4589 struct io_sr_msg
*sr
= &req
->sr_msg
;
4592 struct socket
*sock
;
4596 sock
= sock_from_file(req
->file
, &ret
);
4597 if (unlikely(!sock
))
4600 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4604 msg
.msg_name
= NULL
;
4605 msg
.msg_control
= NULL
;
4606 msg
.msg_controllen
= 0;
4607 msg
.msg_namelen
= 0;
4609 flags
= req
->sr_msg
.msg_flags
;
4610 if (flags
& MSG_DONTWAIT
)
4611 req
->flags
|= REQ_F_NOWAIT
;
4612 else if (force_nonblock
)
4613 flags
|= MSG_DONTWAIT
;
4615 msg
.msg_flags
= flags
;
4616 ret
= sock_sendmsg(sock
, &msg
);
4617 if (force_nonblock
&& ret
== -EAGAIN
)
4619 if (ret
== -ERESTARTSYS
)
4623 req_set_fail_links(req
);
4624 __io_req_complete(req
, ret
, 0, cs
);
4628 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4629 struct io_async_msghdr
*iomsg
)
4631 struct io_sr_msg
*sr
= &req
->sr_msg
;
4632 struct iovec __user
*uiov
;
4636 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4637 &iomsg
->uaddr
, &uiov
, &iov_len
);
4641 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4644 if (copy_from_user(iomsg
->iov
, uiov
, sizeof(*uiov
)))
4646 sr
->len
= iomsg
->iov
[0].iov_len
;
4647 iov_iter_init(&iomsg
->msg
.msg_iter
, READ
, iomsg
->iov
, 1,
4651 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4652 &iomsg
->iov
, &iomsg
->msg
.msg_iter
,
4661 #ifdef CONFIG_COMPAT
4662 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4663 struct io_async_msghdr
*iomsg
)
4665 struct compat_msghdr __user
*msg_compat
;
4666 struct io_sr_msg
*sr
= &req
->sr_msg
;
4667 struct compat_iovec __user
*uiov
;
4672 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4673 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4678 uiov
= compat_ptr(ptr
);
4679 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4680 compat_ssize_t clen
;
4684 if (!access_ok(uiov
, sizeof(*uiov
)))
4686 if (__get_user(clen
, &uiov
->iov_len
))
4690 sr
->len
= iomsg
->iov
[0].iov_len
;
4693 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4694 UIO_FASTIOV
, &iomsg
->iov
,
4695 &iomsg
->msg
.msg_iter
, true);
4704 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4705 struct io_async_msghdr
*iomsg
)
4707 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4708 iomsg
->iov
= iomsg
->fast_iov
;
4710 #ifdef CONFIG_COMPAT
4711 if (req
->ctx
->compat
)
4712 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4715 return __io_recvmsg_copy_hdr(req
, iomsg
);
4718 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4721 struct io_sr_msg
*sr
= &req
->sr_msg
;
4722 struct io_buffer
*kbuf
;
4724 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4729 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4733 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4735 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4738 static int io_recvmsg_prep(struct io_kiocb
*req
,
4739 const struct io_uring_sqe
*sqe
)
4741 struct io_async_msghdr
*async_msg
= req
->async_data
;
4742 struct io_sr_msg
*sr
= &req
->sr_msg
;
4745 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4748 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4749 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4750 sr
->len
= READ_ONCE(sqe
->len
);
4751 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4753 #ifdef CONFIG_COMPAT
4754 if (req
->ctx
->compat
)
4755 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4758 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4760 ret
= io_recvmsg_copy_hdr(req
, async_msg
);
4762 req
->flags
|= REQ_F_NEED_CLEANUP
;
4766 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
4767 struct io_comp_state
*cs
)
4769 struct io_async_msghdr iomsg
, *kmsg
;
4770 struct socket
*sock
;
4771 struct io_buffer
*kbuf
;
4773 int ret
, cflags
= 0;
4775 sock
= sock_from_file(req
->file
, &ret
);
4776 if (unlikely(!sock
))
4779 if (req
->async_data
) {
4780 kmsg
= req
->async_data
;
4781 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4782 /* if iov is set, it's allocated already */
4784 kmsg
->iov
= kmsg
->fast_iov
;
4785 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4787 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4793 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4794 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4796 return PTR_ERR(kbuf
);
4797 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4798 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4799 1, req
->sr_msg
.len
);
4802 flags
= req
->sr_msg
.msg_flags
;
4803 if (flags
& MSG_DONTWAIT
)
4804 req
->flags
|= REQ_F_NOWAIT
;
4805 else if (force_nonblock
)
4806 flags
|= MSG_DONTWAIT
;
4808 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4809 kmsg
->uaddr
, flags
);
4810 if (force_nonblock
&& ret
== -EAGAIN
)
4811 return io_setup_async_msg(req
, kmsg
);
4812 if (ret
== -ERESTARTSYS
)
4815 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4816 cflags
= io_put_recv_kbuf(req
);
4817 if (kmsg
->iov
!= kmsg
->fast_iov
)
4819 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4821 req_set_fail_links(req
);
4822 __io_req_complete(req
, ret
, cflags
, cs
);
4826 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
4827 struct io_comp_state
*cs
)
4829 struct io_buffer
*kbuf
;
4830 struct io_sr_msg
*sr
= &req
->sr_msg
;
4832 void __user
*buf
= sr
->buf
;
4833 struct socket
*sock
;
4836 int ret
, cflags
= 0;
4838 sock
= sock_from_file(req
->file
, &ret
);
4839 if (unlikely(!sock
))
4842 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4843 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4845 return PTR_ERR(kbuf
);
4846 buf
= u64_to_user_ptr(kbuf
->addr
);
4849 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4853 msg
.msg_name
= NULL
;
4854 msg
.msg_control
= NULL
;
4855 msg
.msg_controllen
= 0;
4856 msg
.msg_namelen
= 0;
4857 msg
.msg_iocb
= NULL
;
4860 flags
= req
->sr_msg
.msg_flags
;
4861 if (flags
& MSG_DONTWAIT
)
4862 req
->flags
|= REQ_F_NOWAIT
;
4863 else if (force_nonblock
)
4864 flags
|= MSG_DONTWAIT
;
4866 ret
= sock_recvmsg(sock
, &msg
, flags
);
4867 if (force_nonblock
&& ret
== -EAGAIN
)
4869 if (ret
== -ERESTARTSYS
)
4872 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4873 cflags
= io_put_recv_kbuf(req
);
4875 req_set_fail_links(req
);
4876 __io_req_complete(req
, ret
, cflags
, cs
);
4880 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4882 struct io_accept
*accept
= &req
->accept
;
4884 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4886 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4889 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4890 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4891 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4892 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4896 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
4897 struct io_comp_state
*cs
)
4899 struct io_accept
*accept
= &req
->accept
;
4900 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4903 if (req
->file
->f_flags
& O_NONBLOCK
)
4904 req
->flags
|= REQ_F_NOWAIT
;
4906 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4907 accept
->addr_len
, accept
->flags
,
4909 if (ret
== -EAGAIN
&& force_nonblock
)
4912 if (ret
== -ERESTARTSYS
)
4914 req_set_fail_links(req
);
4916 __io_req_complete(req
, ret
, 0, cs
);
4920 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4922 struct io_connect
*conn
= &req
->connect
;
4923 struct io_async_connect
*io
= req
->async_data
;
4925 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4927 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4930 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4931 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4936 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
4940 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
4941 struct io_comp_state
*cs
)
4943 struct io_async_connect __io
, *io
;
4944 unsigned file_flags
;
4947 if (req
->async_data
) {
4948 io
= req
->async_data
;
4950 ret
= move_addr_to_kernel(req
->connect
.addr
,
4951 req
->connect
.addr_len
,
4958 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4960 ret
= __sys_connect_file(req
->file
, &io
->address
,
4961 req
->connect
.addr_len
, file_flags
);
4962 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4963 if (req
->async_data
)
4965 if (io_alloc_async_data(req
)) {
4969 io
= req
->async_data
;
4970 memcpy(req
->async_data
, &__io
, sizeof(__io
));
4973 if (ret
== -ERESTARTSYS
)
4977 req_set_fail_links(req
);
4978 __io_req_complete(req
, ret
, 0, cs
);
4981 #else /* !CONFIG_NET */
4982 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4987 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4988 struct io_comp_state
*cs
)
4993 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4994 struct io_comp_state
*cs
)
4999 static int io_recvmsg_prep(struct io_kiocb
*req
,
5000 const struct io_uring_sqe
*sqe
)
5005 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
5006 struct io_comp_state
*cs
)
5011 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
5012 struct io_comp_state
*cs
)
5017 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5022 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
5023 struct io_comp_state
*cs
)
5028 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5033 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5034 struct io_comp_state
*cs
)
5038 #endif /* CONFIG_NET */
5040 struct io_poll_table
{
5041 struct poll_table_struct pt
;
5042 struct io_kiocb
*req
;
5046 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5047 __poll_t mask
, task_work_func_t func
)
5052 /* for instances that support it check for an event match first: */
5053 if (mask
&& !(mask
& poll
->events
))
5056 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5058 list_del_init(&poll
->wait
.entry
);
5061 init_task_work(&req
->task_work
, func
);
5062 percpu_ref_get(&req
->ctx
->refs
);
5065 * If we using the signalfd wait_queue_head for this wakeup, then
5066 * it's not safe to use TWA_SIGNAL as we could be recursing on the
5067 * tsk->sighand->siglock on doing the wakeup. Should not be needed
5068 * either, as the normal wakeup will suffice.
5070 twa_signal_ok
= (poll
->head
!= &req
->task
->sighand
->signalfd_wqh
);
5073 * If this fails, then the task is exiting. When a task exits, the
5074 * work gets canceled, so just cancel this request as well instead
5075 * of executing it. We can't safely execute it anyway, as we may not
5076 * have the needed state needed for it anyway.
5078 ret
= io_req_task_work_add(req
, twa_signal_ok
);
5079 if (unlikely(ret
)) {
5080 struct task_struct
*tsk
;
5082 WRITE_ONCE(poll
->canceled
, true);
5083 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
5084 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
5085 wake_up_process(tsk
);
5090 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5091 __acquires(&req
->ctx
->completion_lock
)
5093 struct io_ring_ctx
*ctx
= req
->ctx
;
5095 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5096 struct poll_table_struct pt
= { ._key
= poll
->events
};
5098 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5101 spin_lock_irq(&ctx
->completion_lock
);
5102 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5103 add_wait_queue(poll
->head
, &poll
->wait
);
5110 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5112 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5113 if (req
->opcode
== IORING_OP_POLL_ADD
)
5114 return req
->async_data
;
5115 return req
->apoll
->double_poll
;
5118 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5120 if (req
->opcode
== IORING_OP_POLL_ADD
)
5122 return &req
->apoll
->poll
;
5125 static void io_poll_remove_double(struct io_kiocb
*req
)
5127 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5129 lockdep_assert_held(&req
->ctx
->completion_lock
);
5131 if (poll
&& poll
->head
) {
5132 struct wait_queue_head
*head
= poll
->head
;
5134 spin_lock(&head
->lock
);
5135 list_del_init(&poll
->wait
.entry
);
5136 if (poll
->wait
.private)
5137 refcount_dec(&req
->refs
);
5139 spin_unlock(&head
->lock
);
5143 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
5145 struct io_ring_ctx
*ctx
= req
->ctx
;
5147 io_poll_remove_double(req
);
5148 req
->poll
.done
= true;
5149 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
5150 io_commit_cqring(ctx
);
5153 static void io_poll_task_func(struct callback_head
*cb
)
5155 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5156 struct io_ring_ctx
*ctx
= req
->ctx
;
5157 struct io_kiocb
*nxt
;
5159 if (io_poll_rewait(req
, &req
->poll
)) {
5160 spin_unlock_irq(&ctx
->completion_lock
);
5162 hash_del(&req
->hash_node
);
5163 io_poll_complete(req
, req
->result
, 0);
5164 spin_unlock_irq(&ctx
->completion_lock
);
5166 nxt
= io_put_req_find_next(req
);
5167 io_cqring_ev_posted(ctx
);
5169 __io_req_task_submit(nxt
);
5172 percpu_ref_put(&ctx
->refs
);
5175 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5176 int sync
, void *key
)
5178 struct io_kiocb
*req
= wait
->private;
5179 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5180 __poll_t mask
= key_to_poll(key
);
5182 /* for instances that support it check for an event match first: */
5183 if (mask
&& !(mask
& poll
->events
))
5186 list_del_init(&wait
->entry
);
5188 if (poll
&& poll
->head
) {
5191 spin_lock(&poll
->head
->lock
);
5192 done
= list_empty(&poll
->wait
.entry
);
5194 list_del_init(&poll
->wait
.entry
);
5195 /* make sure double remove sees this as being gone */
5196 wait
->private = NULL
;
5197 spin_unlock(&poll
->head
->lock
);
5199 /* use wait func handler, so it matches the rq type */
5200 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5203 refcount_dec(&req
->refs
);
5207 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5208 wait_queue_func_t wake_func
)
5212 poll
->canceled
= false;
5213 poll
->events
= events
;
5214 INIT_LIST_HEAD(&poll
->wait
.entry
);
5215 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5218 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5219 struct wait_queue_head
*head
,
5220 struct io_poll_iocb
**poll_ptr
)
5222 struct io_kiocb
*req
= pt
->req
;
5225 * If poll->head is already set, it's because the file being polled
5226 * uses multiple waitqueues for poll handling (eg one for read, one
5227 * for write). Setup a separate io_poll_iocb if this happens.
5229 if (unlikely(poll
->head
)) {
5230 struct io_poll_iocb
*poll_one
= poll
;
5232 /* already have a 2nd entry, fail a third attempt */
5234 pt
->error
= -EINVAL
;
5237 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5239 pt
->error
= -ENOMEM
;
5242 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5243 refcount_inc(&req
->refs
);
5244 poll
->wait
.private = req
;
5251 if (poll
->events
& EPOLLEXCLUSIVE
)
5252 add_wait_queue_exclusive(head
, &poll
->wait
);
5254 add_wait_queue(head
, &poll
->wait
);
5257 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5258 struct poll_table_struct
*p
)
5260 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5261 struct async_poll
*apoll
= pt
->req
->apoll
;
5263 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5266 static void io_async_task_func(struct callback_head
*cb
)
5268 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5269 struct async_poll
*apoll
= req
->apoll
;
5270 struct io_ring_ctx
*ctx
= req
->ctx
;
5272 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5274 if (io_poll_rewait(req
, &apoll
->poll
)) {
5275 spin_unlock_irq(&ctx
->completion_lock
);
5276 percpu_ref_put(&ctx
->refs
);
5280 /* If req is still hashed, it cannot have been canceled. Don't check. */
5281 if (hash_hashed(&req
->hash_node
))
5282 hash_del(&req
->hash_node
);
5284 io_poll_remove_double(req
);
5285 spin_unlock_irq(&ctx
->completion_lock
);
5287 if (!READ_ONCE(apoll
->poll
.canceled
))
5288 __io_req_task_submit(req
);
5290 __io_req_task_cancel(req
, -ECANCELED
);
5292 percpu_ref_put(&ctx
->refs
);
5293 kfree(apoll
->double_poll
);
5297 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5300 struct io_kiocb
*req
= wait
->private;
5301 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5303 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5306 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5309 static void io_poll_req_insert(struct io_kiocb
*req
)
5311 struct io_ring_ctx
*ctx
= req
->ctx
;
5312 struct hlist_head
*list
;
5314 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5315 hlist_add_head(&req
->hash_node
, list
);
5318 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5319 struct io_poll_iocb
*poll
,
5320 struct io_poll_table
*ipt
, __poll_t mask
,
5321 wait_queue_func_t wake_func
)
5322 __acquires(&ctx
->completion_lock
)
5324 struct io_ring_ctx
*ctx
= req
->ctx
;
5325 bool cancel
= false;
5327 INIT_HLIST_NODE(&req
->hash_node
);
5328 io_init_poll_iocb(poll
, mask
, wake_func
);
5329 poll
->file
= req
->file
;
5330 poll
->wait
.private = req
;
5332 ipt
->pt
._key
= mask
;
5334 ipt
->error
= -EINVAL
;
5336 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5338 spin_lock_irq(&ctx
->completion_lock
);
5339 if (likely(poll
->head
)) {
5340 spin_lock(&poll
->head
->lock
);
5341 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5347 if (mask
|| ipt
->error
)
5348 list_del_init(&poll
->wait
.entry
);
5350 WRITE_ONCE(poll
->canceled
, true);
5351 else if (!poll
->done
) /* actually waiting for an event */
5352 io_poll_req_insert(req
);
5353 spin_unlock(&poll
->head
->lock
);
5359 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5361 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5362 struct io_ring_ctx
*ctx
= req
->ctx
;
5363 struct async_poll
*apoll
;
5364 struct io_poll_table ipt
;
5368 if (!req
->file
|| !file_can_poll(req
->file
))
5370 if (req
->flags
& REQ_F_POLLED
)
5374 else if (def
->pollout
)
5378 /* if we can't nonblock try, then no point in arming a poll handler */
5379 if (!io_file_supports_async(req
->file
, rw
))
5382 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5383 if (unlikely(!apoll
))
5385 apoll
->double_poll
= NULL
;
5387 req
->flags
|= REQ_F_POLLED
;
5392 mask
|= POLLIN
| POLLRDNORM
;
5394 mask
|= POLLOUT
| POLLWRNORM
;
5396 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5397 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5398 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5401 mask
|= POLLERR
| POLLPRI
;
5403 ipt
.pt
._qproc
= io_async_queue_proc
;
5405 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5407 if (ret
|| ipt
.error
) {
5408 io_poll_remove_double(req
);
5409 spin_unlock_irq(&ctx
->completion_lock
);
5410 kfree(apoll
->double_poll
);
5414 spin_unlock_irq(&ctx
->completion_lock
);
5415 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5416 apoll
->poll
.events
);
5420 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5421 struct io_poll_iocb
*poll
)
5423 bool do_complete
= false;
5425 spin_lock(&poll
->head
->lock
);
5426 WRITE_ONCE(poll
->canceled
, true);
5427 if (!list_empty(&poll
->wait
.entry
)) {
5428 list_del_init(&poll
->wait
.entry
);
5431 spin_unlock(&poll
->head
->lock
);
5432 hash_del(&req
->hash_node
);
5436 static bool io_poll_remove_one(struct io_kiocb
*req
)
5440 io_poll_remove_double(req
);
5442 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5443 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5445 struct async_poll
*apoll
= req
->apoll
;
5447 /* non-poll requests have submit ref still */
5448 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5451 kfree(apoll
->double_poll
);
5457 io_cqring_fill_event(req
, -ECANCELED
);
5458 io_commit_cqring(req
->ctx
);
5459 req_set_fail_links(req
);
5460 io_put_req_deferred(req
, 1);
5467 * Returns true if we found and killed one or more poll requests
5469 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5470 struct files_struct
*files
)
5472 struct hlist_node
*tmp
;
5473 struct io_kiocb
*req
;
5476 spin_lock_irq(&ctx
->completion_lock
);
5477 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5478 struct hlist_head
*list
;
5480 list
= &ctx
->cancel_hash
[i
];
5481 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5482 if (io_match_task(req
, tsk
, files
))
5483 posted
+= io_poll_remove_one(req
);
5486 spin_unlock_irq(&ctx
->completion_lock
);
5489 io_cqring_ev_posted(ctx
);
5494 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5496 struct hlist_head
*list
;
5497 struct io_kiocb
*req
;
5499 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5500 hlist_for_each_entry(req
, list
, hash_node
) {
5501 if (sqe_addr
!= req
->user_data
)
5503 if (io_poll_remove_one(req
))
5511 static int io_poll_remove_prep(struct io_kiocb
*req
,
5512 const struct io_uring_sqe
*sqe
)
5514 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5516 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5520 req
->poll_remove
.addr
= READ_ONCE(sqe
->addr
);
5525 * Find a running poll command that matches one specified in sqe->addr,
5526 * and remove it if found.
5528 static int io_poll_remove(struct io_kiocb
*req
)
5530 struct io_ring_ctx
*ctx
= req
->ctx
;
5533 spin_lock_irq(&ctx
->completion_lock
);
5534 ret
= io_poll_cancel(ctx
, req
->poll_remove
.addr
);
5535 spin_unlock_irq(&ctx
->completion_lock
);
5538 req_set_fail_links(req
);
5539 io_req_complete(req
, ret
);
5543 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5546 struct io_kiocb
*req
= wait
->private;
5547 struct io_poll_iocb
*poll
= &req
->poll
;
5549 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5552 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5553 struct poll_table_struct
*p
)
5555 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5557 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5560 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5562 struct io_poll_iocb
*poll
= &req
->poll
;
5565 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5567 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5570 events
= READ_ONCE(sqe
->poll32_events
);
5572 events
= swahw32(events
);
5574 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5575 (events
& EPOLLEXCLUSIVE
);
5579 static int io_poll_add(struct io_kiocb
*req
)
5581 struct io_poll_iocb
*poll
= &req
->poll
;
5582 struct io_ring_ctx
*ctx
= req
->ctx
;
5583 struct io_poll_table ipt
;
5586 ipt
.pt
._qproc
= io_poll_queue_proc
;
5588 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5591 if (mask
) { /* no async, we'd stolen it */
5593 io_poll_complete(req
, mask
, 0);
5595 spin_unlock_irq(&ctx
->completion_lock
);
5598 io_cqring_ev_posted(ctx
);
5604 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5606 struct io_timeout_data
*data
= container_of(timer
,
5607 struct io_timeout_data
, timer
);
5608 struct io_kiocb
*req
= data
->req
;
5609 struct io_ring_ctx
*ctx
= req
->ctx
;
5610 unsigned long flags
;
5612 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5613 list_del_init(&req
->timeout
.list
);
5614 atomic_set(&req
->ctx
->cq_timeouts
,
5615 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5617 io_cqring_fill_event(req
, -ETIME
);
5618 io_commit_cqring(ctx
);
5619 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5621 io_cqring_ev_posted(ctx
);
5622 req_set_fail_links(req
);
5624 return HRTIMER_NORESTART
;
5627 static int __io_timeout_cancel(struct io_kiocb
*req
)
5629 struct io_timeout_data
*io
= req
->async_data
;
5632 ret
= hrtimer_try_to_cancel(&io
->timer
);
5635 list_del_init(&req
->timeout
.list
);
5637 req_set_fail_links(req
);
5638 io_cqring_fill_event(req
, -ECANCELED
);
5639 io_put_req_deferred(req
, 1);
5643 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5645 struct io_kiocb
*req
;
5648 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5649 if (user_data
== req
->user_data
) {
5658 return __io_timeout_cancel(req
);
5661 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5662 const struct io_uring_sqe
*sqe
)
5664 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5666 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5668 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->timeout_flags
)
5671 req
->timeout_rem
.addr
= READ_ONCE(sqe
->addr
);
5676 * Remove or update an existing timeout command
5678 static int io_timeout_remove(struct io_kiocb
*req
)
5680 struct io_ring_ctx
*ctx
= req
->ctx
;
5683 spin_lock_irq(&ctx
->completion_lock
);
5684 ret
= io_timeout_cancel(ctx
, req
->timeout_rem
.addr
);
5686 io_cqring_fill_event(req
, ret
);
5687 io_commit_cqring(ctx
);
5688 spin_unlock_irq(&ctx
->completion_lock
);
5689 io_cqring_ev_posted(ctx
);
5691 req_set_fail_links(req
);
5696 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5697 bool is_timeout_link
)
5699 struct io_timeout_data
*data
;
5701 u32 off
= READ_ONCE(sqe
->off
);
5703 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5705 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5707 if (off
&& is_timeout_link
)
5709 flags
= READ_ONCE(sqe
->timeout_flags
);
5710 if (flags
& ~IORING_TIMEOUT_ABS
)
5713 req
->timeout
.off
= off
;
5715 if (!req
->async_data
&& io_alloc_async_data(req
))
5718 data
= req
->async_data
;
5721 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5724 if (flags
& IORING_TIMEOUT_ABS
)
5725 data
->mode
= HRTIMER_MODE_ABS
;
5727 data
->mode
= HRTIMER_MODE_REL
;
5729 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5733 static int io_timeout(struct io_kiocb
*req
)
5735 struct io_ring_ctx
*ctx
= req
->ctx
;
5736 struct io_timeout_data
*data
= req
->async_data
;
5737 struct list_head
*entry
;
5738 u32 tail
, off
= req
->timeout
.off
;
5740 spin_lock_irq(&ctx
->completion_lock
);
5743 * sqe->off holds how many events that need to occur for this
5744 * timeout event to be satisfied. If it isn't set, then this is
5745 * a pure timeout request, sequence isn't used.
5747 if (io_is_timeout_noseq(req
)) {
5748 entry
= ctx
->timeout_list
.prev
;
5752 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5753 req
->timeout
.target_seq
= tail
+ off
;
5756 * Insertion sort, ensuring the first entry in the list is always
5757 * the one we need first.
5759 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5760 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5763 if (io_is_timeout_noseq(nxt
))
5765 /* nxt.seq is behind @tail, otherwise would've been completed */
5766 if (off
>= nxt
->timeout
.target_seq
- tail
)
5770 list_add(&req
->timeout
.list
, entry
);
5771 data
->timer
.function
= io_timeout_fn
;
5772 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5773 spin_unlock_irq(&ctx
->completion_lock
);
5777 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5779 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5781 return req
->user_data
== (unsigned long) data
;
5784 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5786 enum io_wq_cancel cancel_ret
;
5789 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5790 switch (cancel_ret
) {
5791 case IO_WQ_CANCEL_OK
:
5794 case IO_WQ_CANCEL_RUNNING
:
5797 case IO_WQ_CANCEL_NOTFOUND
:
5805 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5806 struct io_kiocb
*req
, __u64 sqe_addr
,
5809 unsigned long flags
;
5812 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5813 if (ret
!= -ENOENT
) {
5814 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5818 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5819 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5822 ret
= io_poll_cancel(ctx
, sqe_addr
);
5826 io_cqring_fill_event(req
, ret
);
5827 io_commit_cqring(ctx
);
5828 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5829 io_cqring_ev_posted(ctx
);
5832 req_set_fail_links(req
);
5836 static int io_async_cancel_prep(struct io_kiocb
*req
,
5837 const struct io_uring_sqe
*sqe
)
5839 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5841 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5843 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5846 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5850 static int io_async_cancel(struct io_kiocb
*req
)
5852 struct io_ring_ctx
*ctx
= req
->ctx
;
5854 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
5858 static int io_files_update_prep(struct io_kiocb
*req
,
5859 const struct io_uring_sqe
*sqe
)
5861 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
5863 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5865 if (sqe
->ioprio
|| sqe
->rw_flags
)
5868 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
5869 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
5870 if (!req
->files_update
.nr_args
)
5872 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
5876 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
,
5877 struct io_comp_state
*cs
)
5879 struct io_ring_ctx
*ctx
= req
->ctx
;
5880 struct io_uring_files_update up
;
5886 up
.offset
= req
->files_update
.offset
;
5887 up
.fds
= req
->files_update
.arg
;
5889 mutex_lock(&ctx
->uring_lock
);
5890 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
5891 mutex_unlock(&ctx
->uring_lock
);
5894 req_set_fail_links(req
);
5895 __io_req_complete(req
, ret
, 0, cs
);
5899 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5901 switch (req
->opcode
) {
5904 case IORING_OP_READV
:
5905 case IORING_OP_READ_FIXED
:
5906 case IORING_OP_READ
:
5907 return io_read_prep(req
, sqe
);
5908 case IORING_OP_WRITEV
:
5909 case IORING_OP_WRITE_FIXED
:
5910 case IORING_OP_WRITE
:
5911 return io_write_prep(req
, sqe
);
5912 case IORING_OP_POLL_ADD
:
5913 return io_poll_add_prep(req
, sqe
);
5914 case IORING_OP_POLL_REMOVE
:
5915 return io_poll_remove_prep(req
, sqe
);
5916 case IORING_OP_FSYNC
:
5917 return io_prep_fsync(req
, sqe
);
5918 case IORING_OP_SYNC_FILE_RANGE
:
5919 return io_prep_sfr(req
, sqe
);
5920 case IORING_OP_SENDMSG
:
5921 case IORING_OP_SEND
:
5922 return io_sendmsg_prep(req
, sqe
);
5923 case IORING_OP_RECVMSG
:
5924 case IORING_OP_RECV
:
5925 return io_recvmsg_prep(req
, sqe
);
5926 case IORING_OP_CONNECT
:
5927 return io_connect_prep(req
, sqe
);
5928 case IORING_OP_TIMEOUT
:
5929 return io_timeout_prep(req
, sqe
, false);
5930 case IORING_OP_TIMEOUT_REMOVE
:
5931 return io_timeout_remove_prep(req
, sqe
);
5932 case IORING_OP_ASYNC_CANCEL
:
5933 return io_async_cancel_prep(req
, sqe
);
5934 case IORING_OP_LINK_TIMEOUT
:
5935 return io_timeout_prep(req
, sqe
, true);
5936 case IORING_OP_ACCEPT
:
5937 return io_accept_prep(req
, sqe
);
5938 case IORING_OP_FALLOCATE
:
5939 return io_fallocate_prep(req
, sqe
);
5940 case IORING_OP_OPENAT
:
5941 return io_openat_prep(req
, sqe
);
5942 case IORING_OP_CLOSE
:
5943 return io_close_prep(req
, sqe
);
5944 case IORING_OP_FILES_UPDATE
:
5945 return io_files_update_prep(req
, sqe
);
5946 case IORING_OP_STATX
:
5947 return io_statx_prep(req
, sqe
);
5948 case IORING_OP_FADVISE
:
5949 return io_fadvise_prep(req
, sqe
);
5950 case IORING_OP_MADVISE
:
5951 return io_madvise_prep(req
, sqe
);
5952 case IORING_OP_OPENAT2
:
5953 return io_openat2_prep(req
, sqe
);
5954 case IORING_OP_EPOLL_CTL
:
5955 return io_epoll_ctl_prep(req
, sqe
);
5956 case IORING_OP_SPLICE
:
5957 return io_splice_prep(req
, sqe
);
5958 case IORING_OP_PROVIDE_BUFFERS
:
5959 return io_provide_buffers_prep(req
, sqe
);
5960 case IORING_OP_REMOVE_BUFFERS
:
5961 return io_remove_buffers_prep(req
, sqe
);
5963 return io_tee_prep(req
, sqe
);
5964 case IORING_OP_SHUTDOWN
:
5965 return io_shutdown_prep(req
, sqe
);
5966 case IORING_OP_RENAMEAT
:
5967 return io_renameat_prep(req
, sqe
);
5968 case IORING_OP_UNLINKAT
:
5969 return io_unlinkat_prep(req
, sqe
);
5972 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5977 static int io_req_defer_prep(struct io_kiocb
*req
,
5978 const struct io_uring_sqe
*sqe
)
5982 if (io_alloc_async_data(req
))
5984 return io_req_prep(req
, sqe
);
5987 static u32
io_get_sequence(struct io_kiocb
*req
)
5989 struct io_kiocb
*pos
;
5990 struct io_ring_ctx
*ctx
= req
->ctx
;
5991 u32 total_submitted
, nr_reqs
= 0;
5993 io_for_each_link(pos
, req
)
5996 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
5997 return total_submitted
- nr_reqs
;
6000 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6002 struct io_ring_ctx
*ctx
= req
->ctx
;
6003 struct io_defer_entry
*de
;
6007 /* Still need defer if there is pending req in defer list. */
6008 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6009 !(req
->flags
& REQ_F_IO_DRAIN
)))
6012 seq
= io_get_sequence(req
);
6013 /* Still a chance to pass the sequence check */
6014 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6017 if (!req
->async_data
) {
6018 ret
= io_req_defer_prep(req
, sqe
);
6022 io_prep_async_link(req
);
6023 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6027 spin_lock_irq(&ctx
->completion_lock
);
6028 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6029 spin_unlock_irq(&ctx
->completion_lock
);
6031 io_queue_async_work(req
);
6032 return -EIOCBQUEUED
;
6035 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6038 list_add_tail(&de
->list
, &ctx
->defer_list
);
6039 spin_unlock_irq(&ctx
->completion_lock
);
6040 return -EIOCBQUEUED
;
6043 static void io_req_drop_files(struct io_kiocb
*req
)
6045 struct io_ring_ctx
*ctx
= req
->ctx
;
6046 unsigned long flags
;
6048 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
6049 list_del(&req
->inflight_entry
);
6050 if (waitqueue_active(&ctx
->inflight_wait
))
6051 wake_up(&ctx
->inflight_wait
);
6052 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
6053 req
->flags
&= ~REQ_F_INFLIGHT
;
6054 put_files_struct(req
->work
.identity
->files
);
6055 put_nsproxy(req
->work
.identity
->nsproxy
);
6056 req
->work
.flags
&= ~IO_WQ_WORK_FILES
;
6059 static void __io_clean_op(struct io_kiocb
*req
)
6061 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6062 switch (req
->opcode
) {
6063 case IORING_OP_READV
:
6064 case IORING_OP_READ_FIXED
:
6065 case IORING_OP_READ
:
6066 kfree((void *)(unsigned long)req
->rw
.addr
);
6068 case IORING_OP_RECVMSG
:
6069 case IORING_OP_RECV
:
6070 kfree(req
->sr_msg
.kbuf
);
6073 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
6076 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6077 switch (req
->opcode
) {
6078 case IORING_OP_READV
:
6079 case IORING_OP_READ_FIXED
:
6080 case IORING_OP_READ
:
6081 case IORING_OP_WRITEV
:
6082 case IORING_OP_WRITE_FIXED
:
6083 case IORING_OP_WRITE
: {
6084 struct io_async_rw
*io
= req
->async_data
;
6086 kfree(io
->free_iovec
);
6089 case IORING_OP_RECVMSG
:
6090 case IORING_OP_SENDMSG
: {
6091 struct io_async_msghdr
*io
= req
->async_data
;
6092 if (io
->iov
!= io
->fast_iov
)
6096 case IORING_OP_SPLICE
:
6098 io_put_file(req
, req
->splice
.file_in
,
6099 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
6101 case IORING_OP_OPENAT
:
6102 case IORING_OP_OPENAT2
:
6103 if (req
->open
.filename
)
6104 putname(req
->open
.filename
);
6106 case IORING_OP_RENAMEAT
:
6107 putname(req
->rename
.oldpath
);
6108 putname(req
->rename
.newpath
);
6110 case IORING_OP_UNLINKAT
:
6111 putname(req
->unlink
.filename
);
6114 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6117 if (req
->flags
& REQ_F_INFLIGHT
)
6118 io_req_drop_files(req
);
6121 static int io_issue_sqe(struct io_kiocb
*req
, bool force_nonblock
,
6122 struct io_comp_state
*cs
)
6124 struct io_ring_ctx
*ctx
= req
->ctx
;
6127 switch (req
->opcode
) {
6129 ret
= io_nop(req
, cs
);
6131 case IORING_OP_READV
:
6132 case IORING_OP_READ_FIXED
:
6133 case IORING_OP_READ
:
6134 ret
= io_read(req
, force_nonblock
, cs
);
6136 case IORING_OP_WRITEV
:
6137 case IORING_OP_WRITE_FIXED
:
6138 case IORING_OP_WRITE
:
6139 ret
= io_write(req
, force_nonblock
, cs
);
6141 case IORING_OP_FSYNC
:
6142 ret
= io_fsync(req
, force_nonblock
);
6144 case IORING_OP_POLL_ADD
:
6145 ret
= io_poll_add(req
);
6147 case IORING_OP_POLL_REMOVE
:
6148 ret
= io_poll_remove(req
);
6150 case IORING_OP_SYNC_FILE_RANGE
:
6151 ret
= io_sync_file_range(req
, force_nonblock
);
6153 case IORING_OP_SENDMSG
:
6154 ret
= io_sendmsg(req
, force_nonblock
, cs
);
6156 case IORING_OP_SEND
:
6157 ret
= io_send(req
, force_nonblock
, cs
);
6159 case IORING_OP_RECVMSG
:
6160 ret
= io_recvmsg(req
, force_nonblock
, cs
);
6162 case IORING_OP_RECV
:
6163 ret
= io_recv(req
, force_nonblock
, cs
);
6165 case IORING_OP_TIMEOUT
:
6166 ret
= io_timeout(req
);
6168 case IORING_OP_TIMEOUT_REMOVE
:
6169 ret
= io_timeout_remove(req
);
6171 case IORING_OP_ACCEPT
:
6172 ret
= io_accept(req
, force_nonblock
, cs
);
6174 case IORING_OP_CONNECT
:
6175 ret
= io_connect(req
, force_nonblock
, cs
);
6177 case IORING_OP_ASYNC_CANCEL
:
6178 ret
= io_async_cancel(req
);
6180 case IORING_OP_FALLOCATE
:
6181 ret
= io_fallocate(req
, force_nonblock
);
6183 case IORING_OP_OPENAT
:
6184 ret
= io_openat(req
, force_nonblock
);
6186 case IORING_OP_CLOSE
:
6187 ret
= io_close(req
, force_nonblock
, cs
);
6189 case IORING_OP_FILES_UPDATE
:
6190 ret
= io_files_update(req
, force_nonblock
, cs
);
6192 case IORING_OP_STATX
:
6193 ret
= io_statx(req
, force_nonblock
);
6195 case IORING_OP_FADVISE
:
6196 ret
= io_fadvise(req
, force_nonblock
);
6198 case IORING_OP_MADVISE
:
6199 ret
= io_madvise(req
, force_nonblock
);
6201 case IORING_OP_OPENAT2
:
6202 ret
= io_openat2(req
, force_nonblock
);
6204 case IORING_OP_EPOLL_CTL
:
6205 ret
= io_epoll_ctl(req
, force_nonblock
, cs
);
6207 case IORING_OP_SPLICE
:
6208 ret
= io_splice(req
, force_nonblock
);
6210 case IORING_OP_PROVIDE_BUFFERS
:
6211 ret
= io_provide_buffers(req
, force_nonblock
, cs
);
6213 case IORING_OP_REMOVE_BUFFERS
:
6214 ret
= io_remove_buffers(req
, force_nonblock
, cs
);
6217 ret
= io_tee(req
, force_nonblock
);
6219 case IORING_OP_SHUTDOWN
:
6220 ret
= io_shutdown(req
, force_nonblock
);
6222 case IORING_OP_RENAMEAT
:
6223 ret
= io_renameat(req
, force_nonblock
);
6225 case IORING_OP_UNLINKAT
:
6226 ret
= io_unlinkat(req
, force_nonblock
);
6236 /* If the op doesn't have a file, we're not polling for it */
6237 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6238 const bool in_async
= io_wq_current_is_worker();
6240 /* workqueue context doesn't hold uring_lock, grab it now */
6242 mutex_lock(&ctx
->uring_lock
);
6244 io_iopoll_req_issued(req
);
6247 mutex_unlock(&ctx
->uring_lock
);
6253 static struct io_wq_work
*io_wq_submit_work(struct io_wq_work
*work
)
6255 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6256 struct io_kiocb
*timeout
;
6259 timeout
= io_prep_linked_timeout(req
);
6261 io_queue_linked_timeout(timeout
);
6263 /* if NO_CANCEL is set, we must still run the work */
6264 if ((work
->flags
& (IO_WQ_WORK_CANCEL
|IO_WQ_WORK_NO_CANCEL
)) ==
6265 IO_WQ_WORK_CANCEL
) {
6271 ret
= io_issue_sqe(req
, false, NULL
);
6273 * We can get EAGAIN for polled IO even though we're
6274 * forcing a sync submission from here, since we can't
6275 * wait for request slots on the block side.
6284 req_set_fail_links(req
);
6285 io_req_complete(req
, ret
);
6288 return io_steal_work(req
);
6291 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6294 struct fixed_file_table
*table
;
6296 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6297 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6300 static struct file
*io_file_get(struct io_submit_state
*state
,
6301 struct io_kiocb
*req
, int fd
, bool fixed
)
6303 struct io_ring_ctx
*ctx
= req
->ctx
;
6307 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6309 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6310 file
= io_file_from_index(ctx
, fd
);
6312 req
->fixed_file_refs
= &ctx
->file_data
->node
->refs
;
6313 percpu_ref_get(req
->fixed_file_refs
);
6316 trace_io_uring_file_get(ctx
, fd
);
6317 file
= __io_file_get(state
, fd
);
6323 static int io_req_set_file(struct io_submit_state
*state
, struct io_kiocb
*req
,
6326 req
->file
= io_file_get(state
, req
, fd
, req
->flags
& REQ_F_FIXED_FILE
);
6327 if (req
->file
|| io_op_defs
[req
->opcode
].needs_file_no_error
)
6332 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6334 struct io_timeout_data
*data
= container_of(timer
,
6335 struct io_timeout_data
, timer
);
6336 struct io_kiocb
*prev
, *req
= data
->req
;
6337 struct io_ring_ctx
*ctx
= req
->ctx
;
6338 unsigned long flags
;
6340 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6341 prev
= req
->timeout
.head
;
6342 req
->timeout
.head
= NULL
;
6345 * We don't expect the list to be empty, that will only happen if we
6346 * race with the completion of the linked work.
6348 if (prev
&& refcount_inc_not_zero(&prev
->refs
))
6349 io_remove_next_linked(prev
);
6352 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6355 req_set_fail_links(prev
);
6356 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6359 io_req_complete(req
, -ETIME
);
6361 return HRTIMER_NORESTART
;
6364 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6367 * If the back reference is NULL, then our linked request finished
6368 * before we got a chance to setup the timer
6370 if (req
->timeout
.head
) {
6371 struct io_timeout_data
*data
= req
->async_data
;
6373 data
->timer
.function
= io_link_timeout_fn
;
6374 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6379 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6381 struct io_ring_ctx
*ctx
= req
->ctx
;
6383 spin_lock_irq(&ctx
->completion_lock
);
6384 __io_queue_linked_timeout(req
);
6385 spin_unlock_irq(&ctx
->completion_lock
);
6387 /* drop submission reference */
6391 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6393 struct io_kiocb
*nxt
= req
->link
;
6395 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6396 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6399 nxt
->timeout
.head
= req
;
6400 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6401 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6405 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
)
6407 struct io_kiocb
*linked_timeout
;
6408 const struct cred
*old_creds
= NULL
;
6412 linked_timeout
= io_prep_linked_timeout(req
);
6414 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
6415 (req
->work
.flags
& IO_WQ_WORK_CREDS
) &&
6416 req
->work
.identity
->creds
!= current_cred()) {
6418 revert_creds(old_creds
);
6419 if (old_creds
== req
->work
.identity
->creds
)
6420 old_creds
= NULL
; /* restored original creds */
6422 old_creds
= override_creds(req
->work
.identity
->creds
);
6425 ret
= io_issue_sqe(req
, true, cs
);
6428 * We async punt it if the file wasn't marked NOWAIT, or if the file
6429 * doesn't support non-blocking read/write attempts
6431 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6432 if (!io_arm_poll_handler(req
)) {
6434 * Queued up for async execution, worker will release
6435 * submit reference when the iocb is actually submitted.
6437 io_queue_async_work(req
);
6441 io_queue_linked_timeout(linked_timeout
);
6442 } else if (likely(!ret
)) {
6443 /* drop submission reference */
6444 req
= io_put_req_find_next(req
);
6446 io_queue_linked_timeout(linked_timeout
);
6449 if (!(req
->flags
& REQ_F_FORCE_ASYNC
))
6451 io_queue_async_work(req
);
6454 /* un-prep timeout, so it'll be killed as any other linked */
6455 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
6456 req_set_fail_links(req
);
6458 io_req_complete(req
, ret
);
6462 revert_creds(old_creds
);
6465 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6466 struct io_comp_state
*cs
)
6470 ret
= io_req_defer(req
, sqe
);
6472 if (ret
!= -EIOCBQUEUED
) {
6474 req_set_fail_links(req
);
6476 io_req_complete(req
, ret
);
6478 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6479 if (!req
->async_data
) {
6480 ret
= io_req_defer_prep(req
, sqe
);
6484 io_queue_async_work(req
);
6487 ret
= io_req_prep(req
, sqe
);
6491 __io_queue_sqe(req
, cs
);
6495 static inline void io_queue_link_head(struct io_kiocb
*req
,
6496 struct io_comp_state
*cs
)
6498 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
6500 io_req_complete(req
, -ECANCELED
);
6502 io_queue_sqe(req
, NULL
, cs
);
6505 struct io_submit_link
{
6506 struct io_kiocb
*head
;
6507 struct io_kiocb
*last
;
6510 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6511 struct io_submit_link
*link
, struct io_comp_state
*cs
)
6513 struct io_ring_ctx
*ctx
= req
->ctx
;
6517 * If we already have a head request, queue this one for async
6518 * submittal once the head completes. If we don't have a head but
6519 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6520 * submitted sync once the chain is complete. If none of those
6521 * conditions are true (normal request), then just queue it.
6524 struct io_kiocb
*head
= link
->head
;
6527 * Taking sequential execution of a link, draining both sides
6528 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6529 * requests in the link. So, it drains the head and the
6530 * next after the link request. The last one is done via
6531 * drain_next flag to persist the effect across calls.
6533 if (req
->flags
& REQ_F_IO_DRAIN
) {
6534 head
->flags
|= REQ_F_IO_DRAIN
;
6535 ctx
->drain_next
= 1;
6537 ret
= io_req_defer_prep(req
, sqe
);
6538 if (unlikely(ret
)) {
6539 /* fail even hard links since we don't submit */
6540 head
->flags
|= REQ_F_FAIL_LINK
;
6543 trace_io_uring_link(ctx
, req
, head
);
6544 link
->last
->link
= req
;
6547 /* last request of a link, enqueue the link */
6548 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6549 io_queue_link_head(head
, cs
);
6553 if (unlikely(ctx
->drain_next
)) {
6554 req
->flags
|= REQ_F_IO_DRAIN
;
6555 ctx
->drain_next
= 0;
6557 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6558 ret
= io_req_defer_prep(req
, sqe
);
6560 req
->flags
|= REQ_F_FAIL_LINK
;
6564 io_queue_sqe(req
, sqe
, cs
);
6572 * Batched submission is done, ensure local IO is flushed out.
6574 static void io_submit_state_end(struct io_submit_state
*state
)
6576 if (!list_empty(&state
->comp
.list
))
6577 io_submit_flush_completions(&state
->comp
);
6578 if (state
->plug_started
)
6579 blk_finish_plug(&state
->plug
);
6580 io_state_file_put(state
);
6581 if (state
->free_reqs
)
6582 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6586 * Start submission side cache.
6588 static void io_submit_state_start(struct io_submit_state
*state
,
6589 struct io_ring_ctx
*ctx
, unsigned int max_ios
)
6591 state
->plug_started
= false;
6593 INIT_LIST_HEAD(&state
->comp
.list
);
6594 state
->comp
.ctx
= ctx
;
6595 state
->free_reqs
= 0;
6597 state
->ios_left
= max_ios
;
6600 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6602 struct io_rings
*rings
= ctx
->rings
;
6605 * Ensure any loads from the SQEs are done at this point,
6606 * since once we write the new head, the application could
6607 * write new data to them.
6609 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6613 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6614 * that is mapped by userspace. This means that care needs to be taken to
6615 * ensure that reads are stable, as we cannot rely on userspace always
6616 * being a good citizen. If members of the sqe are validated and then later
6617 * used, it's important that those reads are done through READ_ONCE() to
6618 * prevent a re-load down the line.
6620 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6622 u32
*sq_array
= ctx
->sq_array
;
6626 * The cached sq head (or cq tail) serves two purposes:
6628 * 1) allows us to batch the cost of updating the user visible
6630 * 2) allows the kernel side to track the head on its own, even
6631 * though the application is the one updating it.
6633 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6634 if (likely(head
< ctx
->sq_entries
))
6635 return &ctx
->sq_sqes
[head
];
6637 /* drop invalid entries */
6638 ctx
->cached_sq_dropped
++;
6639 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6643 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6645 ctx
->cached_sq_head
++;
6649 * Check SQE restrictions (opcode and flags).
6651 * Returns 'true' if SQE is allowed, 'false' otherwise.
6653 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6654 struct io_kiocb
*req
,
6655 unsigned int sqe_flags
)
6657 if (!ctx
->restricted
)
6660 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6663 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6664 ctx
->restrictions
.sqe_flags_required
)
6667 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6668 ctx
->restrictions
.sqe_flags_required
))
6674 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6675 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6676 IOSQE_BUFFER_SELECT)
6678 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6679 const struct io_uring_sqe
*sqe
,
6680 struct io_submit_state
*state
)
6682 unsigned int sqe_flags
;
6685 req
->opcode
= READ_ONCE(sqe
->opcode
);
6686 req
->user_data
= READ_ONCE(sqe
->user_data
);
6687 req
->async_data
= NULL
;
6692 /* one is dropped after submission, the other at completion */
6693 refcount_set(&req
->refs
, 2);
6694 req
->task
= current
;
6697 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6700 if (unlikely(io_sq_thread_acquire_mm_files(ctx
, req
)))
6703 sqe_flags
= READ_ONCE(sqe
->flags
);
6704 /* enforce forwards compatibility on users */
6705 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6708 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6711 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6712 !io_op_defs
[req
->opcode
].buffer_select
)
6715 id
= READ_ONCE(sqe
->personality
);
6717 struct io_identity
*iod
;
6719 iod
= idr_find(&ctx
->personality_idr
, id
);
6722 refcount_inc(&iod
->count
);
6724 __io_req_init_async(req
);
6725 get_cred(iod
->creds
);
6726 req
->work
.identity
= iod
;
6727 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
6730 /* same numerical values with corresponding REQ_F_*, safe to copy */
6731 req
->flags
|= sqe_flags
;
6734 * Plug now if we have more than 1 IO left after this, and the target
6735 * is potentially a read/write to block based storage.
6737 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6738 io_op_defs
[req
->opcode
].plug
) {
6739 blk_start_plug(&state
->plug
);
6740 state
->plug_started
= true;
6743 if (!io_op_defs
[req
->opcode
].needs_file
)
6746 ret
= io_req_set_file(state
, req
, READ_ONCE(sqe
->fd
));
6751 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6753 struct io_submit_state state
;
6754 struct io_submit_link link
;
6755 int i
, submitted
= 0;
6757 /* if we have a backlog and couldn't flush it all, return BUSY */
6758 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6759 if (!list_empty(&ctx
->cq_overflow_list
) &&
6760 !io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6764 /* make sure SQ entry isn't read before tail */
6765 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6767 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6770 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6771 refcount_add(nr
, ¤t
->usage
);
6773 io_submit_state_start(&state
, ctx
, nr
);
6776 for (i
= 0; i
< nr
; i
++) {
6777 const struct io_uring_sqe
*sqe
;
6778 struct io_kiocb
*req
;
6781 sqe
= io_get_sqe(ctx
);
6782 if (unlikely(!sqe
)) {
6783 io_consume_sqe(ctx
);
6786 req
= io_alloc_req(ctx
, &state
);
6787 if (unlikely(!req
)) {
6789 submitted
= -EAGAIN
;
6792 io_consume_sqe(ctx
);
6793 /* will complete beyond this point, count as submitted */
6796 err
= io_init_req(ctx
, req
, sqe
, &state
);
6797 if (unlikely(err
)) {
6800 io_req_complete(req
, err
);
6804 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6805 true, io_async_submit(ctx
));
6806 err
= io_submit_sqe(req
, sqe
, &link
, &state
.comp
);
6811 if (unlikely(submitted
!= nr
)) {
6812 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6813 struct io_uring_task
*tctx
= current
->io_uring
;
6814 int unused
= nr
- ref_used
;
6816 percpu_ref_put_many(&ctx
->refs
, unused
);
6817 percpu_counter_sub(&tctx
->inflight
, unused
);
6818 put_task_struct_many(current
, unused
);
6821 io_queue_link_head(link
.head
, &state
.comp
);
6822 io_submit_state_end(&state
);
6824 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6825 io_commit_sqring(ctx
);
6830 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6832 /* Tell userspace we may need a wakeup call */
6833 spin_lock_irq(&ctx
->completion_lock
);
6834 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6835 spin_unlock_irq(&ctx
->completion_lock
);
6838 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6840 spin_lock_irq(&ctx
->completion_lock
);
6841 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6842 spin_unlock_irq(&ctx
->completion_lock
);
6845 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
6847 unsigned int to_submit
;
6850 if (!list_empty(&ctx
->iopoll_list
)) {
6851 unsigned nr_events
= 0;
6853 mutex_lock(&ctx
->uring_lock
);
6854 if (!list_empty(&ctx
->iopoll_list
))
6855 io_do_iopoll(ctx
, &nr_events
, 0);
6856 mutex_unlock(&ctx
->uring_lock
);
6859 to_submit
= io_sqring_entries(ctx
);
6860 /* if we're handling multiple rings, cap submit size for fairness */
6861 if (cap_entries
&& to_submit
> 8)
6865 mutex_lock(&ctx
->uring_lock
);
6866 if (likely(!percpu_ref_is_dying(&ctx
->refs
)))
6867 ret
= io_submit_sqes(ctx
, to_submit
);
6868 mutex_unlock(&ctx
->uring_lock
);
6871 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
6872 wake_up(&ctx
->sqo_sq_wait
);
6877 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
6879 struct io_ring_ctx
*ctx
;
6880 unsigned sq_thread_idle
= 0;
6882 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6883 if (sq_thread_idle
< ctx
->sq_thread_idle
)
6884 sq_thread_idle
= ctx
->sq_thread_idle
;
6887 sqd
->sq_thread_idle
= sq_thread_idle
;
6890 static void io_sqd_init_new(struct io_sq_data
*sqd
)
6892 struct io_ring_ctx
*ctx
;
6894 while (!list_empty(&sqd
->ctx_new_list
)) {
6895 ctx
= list_first_entry(&sqd
->ctx_new_list
, struct io_ring_ctx
, sqd_list
);
6896 list_move_tail(&ctx
->sqd_list
, &sqd
->ctx_list
);
6897 complete(&ctx
->sq_thread_comp
);
6900 io_sqd_update_thread_idle(sqd
);
6903 static int io_sq_thread(void *data
)
6905 struct cgroup_subsys_state
*cur_css
= NULL
;
6906 struct files_struct
*old_files
= current
->files
;
6907 struct nsproxy
*old_nsproxy
= current
->nsproxy
;
6908 const struct cred
*old_cred
= NULL
;
6909 struct io_sq_data
*sqd
= data
;
6910 struct io_ring_ctx
*ctx
;
6911 unsigned long timeout
= 0;
6915 current
->files
= NULL
;
6916 current
->nsproxy
= NULL
;
6917 task_unlock(current
);
6919 while (!kthread_should_stop()) {
6921 bool cap_entries
, sqt_spin
, needs_sched
;
6924 * Any changes to the sqd lists are synchronized through the
6925 * kthread parking. This synchronizes the thread vs users,
6926 * the users are synchronized on the sqd->ctx_lock.
6928 if (kthread_should_park())
6931 if (unlikely(!list_empty(&sqd
->ctx_new_list
))) {
6932 io_sqd_init_new(sqd
);
6933 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6937 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
6938 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6939 if (current
->cred
!= ctx
->creds
) {
6941 revert_creds(old_cred
);
6942 old_cred
= override_creds(ctx
->creds
);
6944 io_sq_thread_associate_blkcg(ctx
, &cur_css
);
6946 current
->loginuid
= ctx
->loginuid
;
6947 current
->sessionid
= ctx
->sessionid
;
6950 ret
= __io_sq_thread(ctx
, cap_entries
);
6951 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
6954 io_sq_thread_drop_mm_files();
6957 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
6961 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6965 if (kthread_should_park())
6969 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
6970 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6971 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6972 !list_empty_careful(&ctx
->iopoll_list
)) {
6973 needs_sched
= false;
6976 if (io_sqring_entries(ctx
)) {
6977 needs_sched
= false;
6983 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6984 io_ring_set_wakeup_flag(ctx
);
6987 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6988 io_ring_clear_wakeup_flag(ctx
);
6991 finish_wait(&sqd
->wait
, &wait
);
6992 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6998 io_sq_thread_unassociate_blkcg();
7000 revert_creds(old_cred
);
7003 current
->files
= old_files
;
7004 current
->nsproxy
= old_nsproxy
;
7005 task_unlock(current
);
7012 struct io_wait_queue
{
7013 struct wait_queue_entry wq
;
7014 struct io_ring_ctx
*ctx
;
7016 unsigned nr_timeouts
;
7019 static inline bool io_should_wake(struct io_wait_queue
*iowq
, bool noflush
)
7021 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7024 * Wake up if we have enough events, or if a timeout occurred since we
7025 * started waiting. For timeouts, we always want to return to userspace,
7026 * regardless of event count.
7028 return io_cqring_events(ctx
, noflush
) >= iowq
->to_wait
||
7029 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7032 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7033 int wake_flags
, void *key
)
7035 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7038 /* use noflush == true, as we can't safely rely on locking context */
7039 if (!io_should_wake(iowq
, true))
7042 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7045 static int io_run_task_work_sig(void)
7047 if (io_run_task_work())
7049 if (!signal_pending(current
))
7051 if (current
->jobctl
& JOBCTL_TASK_WORK
) {
7052 spin_lock_irq(¤t
->sighand
->siglock
);
7053 current
->jobctl
&= ~JOBCTL_TASK_WORK
;
7054 recalc_sigpending();
7055 spin_unlock_irq(¤t
->sighand
->siglock
);
7062 * Wait until events become available, if we don't already have some. The
7063 * application must reap them itself, as they reside on the shared cq ring.
7065 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7066 const sigset_t __user
*sig
, size_t sigsz
,
7067 struct __kernel_timespec __user
*uts
)
7069 struct io_wait_queue iowq
= {
7072 .func
= io_wake_function
,
7073 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
7076 .to_wait
= min_events
,
7078 struct io_rings
*rings
= ctx
->rings
;
7079 struct timespec64 ts
;
7080 signed long timeout
= 0;
7084 if (io_cqring_events(ctx
, false) >= min_events
)
7086 if (!io_run_task_work())
7091 #ifdef CONFIG_COMPAT
7092 if (in_compat_syscall())
7093 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7097 ret
= set_user_sigmask(sig
, sigsz
);
7104 if (get_timespec64(&ts
, uts
))
7106 timeout
= timespec64_to_jiffies(&ts
);
7109 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7110 trace_io_uring_cqring_wait(ctx
, min_events
);
7112 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
7113 TASK_INTERRUPTIBLE
);
7114 /* make sure we run task_work before checking for signals */
7115 ret
= io_run_task_work_sig();
7120 if (io_should_wake(&iowq
, false))
7123 timeout
= schedule_timeout(timeout
);
7132 finish_wait(&ctx
->wait
, &iowq
.wq
);
7134 restore_saved_sigmask_unless(ret
== -EINTR
);
7136 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7139 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7141 #if defined(CONFIG_UNIX)
7142 if (ctx
->ring_sock
) {
7143 struct sock
*sock
= ctx
->ring_sock
->sk
;
7144 struct sk_buff
*skb
;
7146 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7152 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7155 file
= io_file_from_index(ctx
, i
);
7162 static void io_file_ref_kill(struct percpu_ref
*ref
)
7164 struct fixed_file_data
*data
;
7166 data
= container_of(ref
, struct fixed_file_data
, refs
);
7167 complete(&data
->done
);
7170 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7172 struct fixed_file_data
*data
= ctx
->file_data
;
7173 struct fixed_file_ref_node
*ref_node
= NULL
;
7174 unsigned nr_tables
, i
;
7179 spin_lock(&data
->lock
);
7180 ref_node
= data
->node
;
7181 spin_unlock(&data
->lock
);
7183 percpu_ref_kill(&ref_node
->refs
);
7185 percpu_ref_kill(&data
->refs
);
7187 /* wait for all refs nodes to complete */
7188 flush_delayed_work(&ctx
->file_put_work
);
7189 wait_for_completion(&data
->done
);
7191 __io_sqe_files_unregister(ctx
);
7192 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
7193 for (i
= 0; i
< nr_tables
; i
++)
7194 kfree(data
->table
[i
].files
);
7196 percpu_ref_exit(&data
->refs
);
7198 ctx
->file_data
= NULL
;
7199 ctx
->nr_user_files
= 0;
7203 static void io_put_sq_data(struct io_sq_data
*sqd
)
7205 if (refcount_dec_and_test(&sqd
->refs
)) {
7207 * The park is a bit of a work-around, without it we get
7208 * warning spews on shutdown with SQPOLL set and affinity
7209 * set to a single CPU.
7212 kthread_park(sqd
->thread
);
7213 kthread_stop(sqd
->thread
);
7220 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7222 struct io_ring_ctx
*ctx_attach
;
7223 struct io_sq_data
*sqd
;
7226 f
= fdget(p
->wq_fd
);
7228 return ERR_PTR(-ENXIO
);
7229 if (f
.file
->f_op
!= &io_uring_fops
) {
7231 return ERR_PTR(-EINVAL
);
7234 ctx_attach
= f
.file
->private_data
;
7235 sqd
= ctx_attach
->sq_data
;
7238 return ERR_PTR(-EINVAL
);
7241 refcount_inc(&sqd
->refs
);
7246 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
)
7248 struct io_sq_data
*sqd
;
7250 if (p
->flags
& IORING_SETUP_ATTACH_WQ
)
7251 return io_attach_sq_data(p
);
7253 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7255 return ERR_PTR(-ENOMEM
);
7257 refcount_set(&sqd
->refs
, 1);
7258 INIT_LIST_HEAD(&sqd
->ctx_list
);
7259 INIT_LIST_HEAD(&sqd
->ctx_new_list
);
7260 mutex_init(&sqd
->ctx_lock
);
7261 mutex_init(&sqd
->lock
);
7262 init_waitqueue_head(&sqd
->wait
);
7266 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7267 __releases(&sqd
->lock
)
7271 kthread_unpark(sqd
->thread
);
7272 mutex_unlock(&sqd
->lock
);
7275 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7276 __acquires(&sqd
->lock
)
7280 mutex_lock(&sqd
->lock
);
7281 kthread_park(sqd
->thread
);
7284 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
7286 struct io_sq_data
*sqd
= ctx
->sq_data
;
7291 * We may arrive here from the error branch in
7292 * io_sq_offload_create() where the kthread is created
7293 * without being waked up, thus wake it up now to make
7294 * sure the wait will complete.
7296 wake_up_process(sqd
->thread
);
7297 wait_for_completion(&ctx
->sq_thread_comp
);
7299 io_sq_thread_park(sqd
);
7302 mutex_lock(&sqd
->ctx_lock
);
7303 list_del(&ctx
->sqd_list
);
7304 io_sqd_update_thread_idle(sqd
);
7305 mutex_unlock(&sqd
->ctx_lock
);
7308 io_sq_thread_unpark(sqd
);
7310 io_put_sq_data(sqd
);
7311 ctx
->sq_data
= NULL
;
7315 static void io_finish_async(struct io_ring_ctx
*ctx
)
7317 io_sq_thread_stop(ctx
);
7320 io_wq_destroy(ctx
->io_wq
);
7325 #if defined(CONFIG_UNIX)
7327 * Ensure the UNIX gc is aware of our file set, so we are certain that
7328 * the io_uring can be safely unregistered on process exit, even if we have
7329 * loops in the file referencing.
7331 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7333 struct sock
*sk
= ctx
->ring_sock
->sk
;
7334 struct scm_fp_list
*fpl
;
7335 struct sk_buff
*skb
;
7338 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7342 skb
= alloc_skb(0, GFP_KERNEL
);
7351 fpl
->user
= get_uid(ctx
->user
);
7352 for (i
= 0; i
< nr
; i
++) {
7353 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7357 fpl
->fp
[nr_files
] = get_file(file
);
7358 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7363 fpl
->max
= SCM_MAX_FD
;
7364 fpl
->count
= nr_files
;
7365 UNIXCB(skb
).fp
= fpl
;
7366 skb
->destructor
= unix_destruct_scm
;
7367 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7368 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7370 for (i
= 0; i
< nr_files
; i
++)
7381 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7382 * causes regular reference counting to break down. We rely on the UNIX
7383 * garbage collection to take care of this problem for us.
7385 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7387 unsigned left
, total
;
7391 left
= ctx
->nr_user_files
;
7393 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7395 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7399 total
+= this_files
;
7405 while (total
< ctx
->nr_user_files
) {
7406 struct file
*file
= io_file_from_index(ctx
, total
);
7416 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7422 static int io_sqe_alloc_file_tables(struct fixed_file_data
*file_data
,
7423 unsigned nr_tables
, unsigned nr_files
)
7427 for (i
= 0; i
< nr_tables
; i
++) {
7428 struct fixed_file_table
*table
= &file_data
->table
[i
];
7429 unsigned this_files
;
7431 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7432 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7436 nr_files
-= this_files
;
7442 for (i
= 0; i
< nr_tables
; i
++) {
7443 struct fixed_file_table
*table
= &file_data
->table
[i
];
7444 kfree(table
->files
);
7449 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
7451 #if defined(CONFIG_UNIX)
7452 struct sock
*sock
= ctx
->ring_sock
->sk
;
7453 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7454 struct sk_buff
*skb
;
7457 __skb_queue_head_init(&list
);
7460 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7461 * remove this entry and rearrange the file array.
7463 skb
= skb_dequeue(head
);
7465 struct scm_fp_list
*fp
;
7467 fp
= UNIXCB(skb
).fp
;
7468 for (i
= 0; i
< fp
->count
; i
++) {
7471 if (fp
->fp
[i
] != file
)
7474 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7475 left
= fp
->count
- 1 - i
;
7477 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7478 left
* sizeof(struct file
*));
7485 __skb_queue_tail(&list
, skb
);
7495 __skb_queue_tail(&list
, skb
);
7497 skb
= skb_dequeue(head
);
7500 if (skb_peek(&list
)) {
7501 spin_lock_irq(&head
->lock
);
7502 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7503 __skb_queue_tail(head
, skb
);
7504 spin_unlock_irq(&head
->lock
);
7511 struct io_file_put
{
7512 struct list_head list
;
7516 static void __io_file_put_work(struct fixed_file_ref_node
*ref_node
)
7518 struct fixed_file_data
*file_data
= ref_node
->file_data
;
7519 struct io_ring_ctx
*ctx
= file_data
->ctx
;
7520 struct io_file_put
*pfile
, *tmp
;
7522 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
7523 list_del(&pfile
->list
);
7524 io_ring_file_put(ctx
, pfile
->file
);
7528 percpu_ref_exit(&ref_node
->refs
);
7530 percpu_ref_put(&file_data
->refs
);
7533 static void io_file_put_work(struct work_struct
*work
)
7535 struct io_ring_ctx
*ctx
;
7536 struct llist_node
*node
;
7538 ctx
= container_of(work
, struct io_ring_ctx
, file_put_work
.work
);
7539 node
= llist_del_all(&ctx
->file_put_llist
);
7542 struct fixed_file_ref_node
*ref_node
;
7543 struct llist_node
*next
= node
->next
;
7545 ref_node
= llist_entry(node
, struct fixed_file_ref_node
, llist
);
7546 __io_file_put_work(ref_node
);
7551 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
7553 struct fixed_file_ref_node
*ref_node
;
7554 struct fixed_file_data
*data
;
7555 struct io_ring_ctx
*ctx
;
7556 bool first_add
= false;
7559 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
7560 data
= ref_node
->file_data
;
7563 spin_lock(&data
->lock
);
7564 ref_node
->done
= true;
7566 while (!list_empty(&data
->ref_list
)) {
7567 ref_node
= list_first_entry(&data
->ref_list
,
7568 struct fixed_file_ref_node
, node
);
7569 /* recycle ref nodes in order */
7570 if (!ref_node
->done
)
7572 list_del(&ref_node
->node
);
7573 first_add
|= llist_add(&ref_node
->llist
, &ctx
->file_put_llist
);
7575 spin_unlock(&data
->lock
);
7577 if (percpu_ref_is_dying(&data
->refs
))
7581 mod_delayed_work(system_wq
, &ctx
->file_put_work
, 0);
7583 queue_delayed_work(system_wq
, &ctx
->file_put_work
, delay
);
7586 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
7587 struct io_ring_ctx
*ctx
)
7589 struct fixed_file_ref_node
*ref_node
;
7591 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7593 return ERR_PTR(-ENOMEM
);
7595 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
7598 return ERR_PTR(-ENOMEM
);
7600 INIT_LIST_HEAD(&ref_node
->node
);
7601 INIT_LIST_HEAD(&ref_node
->file_list
);
7602 ref_node
->file_data
= ctx
->file_data
;
7603 ref_node
->done
= false;
7607 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
7609 percpu_ref_exit(&ref_node
->refs
);
7613 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7616 __s32 __user
*fds
= (__s32 __user
*) arg
;
7617 unsigned nr_tables
, i
;
7619 int fd
, ret
= -ENOMEM
;
7620 struct fixed_file_ref_node
*ref_node
;
7621 struct fixed_file_data
*file_data
;
7627 if (nr_args
> IORING_MAX_FIXED_FILES
)
7630 file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL
);
7633 file_data
->ctx
= ctx
;
7634 init_completion(&file_data
->done
);
7635 INIT_LIST_HEAD(&file_data
->ref_list
);
7636 spin_lock_init(&file_data
->lock
);
7638 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7639 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7641 if (!file_data
->table
)
7644 if (percpu_ref_init(&file_data
->refs
, io_file_ref_kill
,
7645 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
7648 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7650 ctx
->file_data
= file_data
;
7652 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7653 struct fixed_file_table
*table
;
7656 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7660 /* allow sparse sets */
7670 * Don't allow io_uring instances to be registered. If UNIX
7671 * isn't enabled, then this causes a reference cycle and this
7672 * instance can never get freed. If UNIX is enabled we'll
7673 * handle it just fine, but there's still no point in allowing
7674 * a ring fd as it doesn't support regular read/write anyway.
7676 if (file
->f_op
== &io_uring_fops
) {
7680 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7681 index
= i
& IORING_FILE_TABLE_MASK
;
7682 table
->files
[index
] = file
;
7685 ret
= io_sqe_files_scm(ctx
);
7687 io_sqe_files_unregister(ctx
);
7691 ref_node
= alloc_fixed_file_ref_node(ctx
);
7692 if (IS_ERR(ref_node
)) {
7693 io_sqe_files_unregister(ctx
);
7694 return PTR_ERR(ref_node
);
7697 file_data
->node
= ref_node
;
7698 spin_lock(&file_data
->lock
);
7699 list_add_tail(&ref_node
->node
, &file_data
->ref_list
);
7700 spin_unlock(&file_data
->lock
);
7701 percpu_ref_get(&file_data
->refs
);
7704 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7705 file
= io_file_from_index(ctx
, i
);
7709 for (i
= 0; i
< nr_tables
; i
++)
7710 kfree(file_data
->table
[i
].files
);
7711 ctx
->nr_user_files
= 0;
7713 percpu_ref_exit(&file_data
->refs
);
7715 kfree(file_data
->table
);
7717 ctx
->file_data
= NULL
;
7721 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7724 #if defined(CONFIG_UNIX)
7725 struct sock
*sock
= ctx
->ring_sock
->sk
;
7726 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7727 struct sk_buff
*skb
;
7730 * See if we can merge this file into an existing skb SCM_RIGHTS
7731 * file set. If there's no room, fall back to allocating a new skb
7732 * and filling it in.
7734 spin_lock_irq(&head
->lock
);
7735 skb
= skb_peek(head
);
7737 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7739 if (fpl
->count
< SCM_MAX_FD
) {
7740 __skb_unlink(skb
, head
);
7741 spin_unlock_irq(&head
->lock
);
7742 fpl
->fp
[fpl
->count
] = get_file(file
);
7743 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7745 spin_lock_irq(&head
->lock
);
7746 __skb_queue_head(head
, skb
);
7751 spin_unlock_irq(&head
->lock
);
7758 return __io_sqe_files_scm(ctx
, 1, index
);
7764 static int io_queue_file_removal(struct fixed_file_data
*data
,
7767 struct io_file_put
*pfile
;
7768 struct fixed_file_ref_node
*ref_node
= data
->node
;
7770 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
7775 list_add(&pfile
->list
, &ref_node
->file_list
);
7780 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7781 struct io_uring_files_update
*up
,
7784 struct fixed_file_data
*data
= ctx
->file_data
;
7785 struct fixed_file_ref_node
*ref_node
;
7790 bool needs_switch
= false;
7792 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7794 if (done
> ctx
->nr_user_files
)
7797 ref_node
= alloc_fixed_file_ref_node(ctx
);
7798 if (IS_ERR(ref_node
))
7799 return PTR_ERR(ref_node
);
7802 fds
= u64_to_user_ptr(up
->fds
);
7804 struct fixed_file_table
*table
;
7808 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7812 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
7813 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7814 index
= i
& IORING_FILE_TABLE_MASK
;
7815 if (table
->files
[index
]) {
7816 file
= table
->files
[index
];
7817 err
= io_queue_file_removal(data
, file
);
7820 table
->files
[index
] = NULL
;
7821 needs_switch
= true;
7830 * Don't allow io_uring instances to be registered. If
7831 * UNIX isn't enabled, then this causes a reference
7832 * cycle and this instance can never get freed. If UNIX
7833 * is enabled we'll handle it just fine, but there's
7834 * still no point in allowing a ring fd as it doesn't
7835 * support regular read/write anyway.
7837 if (file
->f_op
== &io_uring_fops
) {
7842 table
->files
[index
] = file
;
7843 err
= io_sqe_file_register(ctx
, file
, i
);
7845 table
->files
[index
] = NULL
;
7856 percpu_ref_kill(&data
->node
->refs
);
7857 spin_lock(&data
->lock
);
7858 list_add_tail(&ref_node
->node
, &data
->ref_list
);
7859 data
->node
= ref_node
;
7860 spin_unlock(&data
->lock
);
7861 percpu_ref_get(&ctx
->file_data
->refs
);
7863 destroy_fixed_file_ref_node(ref_node
);
7865 return done
? done
: err
;
7868 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
7871 struct io_uring_files_update up
;
7873 if (!ctx
->file_data
)
7877 if (copy_from_user(&up
, arg
, sizeof(up
)))
7882 return __io_sqe_files_update(ctx
, &up
, nr_args
);
7885 static void io_free_work(struct io_wq_work
*work
)
7887 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7889 /* Consider that io_steal_work() relies on this ref */
7893 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
7894 struct io_uring_params
*p
)
7896 struct io_wq_data data
;
7898 struct io_ring_ctx
*ctx_attach
;
7899 unsigned int concurrency
;
7902 data
.user
= ctx
->user
;
7903 data
.free_work
= io_free_work
;
7904 data
.do_work
= io_wq_submit_work
;
7906 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
7907 /* Do QD, or 4 * CPUS, whatever is smallest */
7908 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
7910 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
7911 if (IS_ERR(ctx
->io_wq
)) {
7912 ret
= PTR_ERR(ctx
->io_wq
);
7918 f
= fdget(p
->wq_fd
);
7922 if (f
.file
->f_op
!= &io_uring_fops
) {
7927 ctx_attach
= f
.file
->private_data
;
7928 /* @io_wq is protected by holding the fd */
7929 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
7934 ctx
->io_wq
= ctx_attach
->io_wq
;
7940 static int io_uring_alloc_task_context(struct task_struct
*task
)
7942 struct io_uring_task
*tctx
;
7945 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
7946 if (unlikely(!tctx
))
7949 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
7950 if (unlikely(ret
)) {
7956 init_waitqueue_head(&tctx
->wait
);
7958 atomic_set(&tctx
->in_idle
, 0);
7959 tctx
->sqpoll
= false;
7960 io_init_identity(&tctx
->__identity
);
7961 tctx
->identity
= &tctx
->__identity
;
7962 task
->io_uring
= tctx
;
7966 void __io_uring_free(struct task_struct
*tsk
)
7968 struct io_uring_task
*tctx
= tsk
->io_uring
;
7970 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
7971 WARN_ON_ONCE(refcount_read(&tctx
->identity
->count
) != 1);
7972 if (tctx
->identity
!= &tctx
->__identity
)
7973 kfree(tctx
->identity
);
7974 percpu_counter_destroy(&tctx
->inflight
);
7976 tsk
->io_uring
= NULL
;
7979 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
7980 struct io_uring_params
*p
)
7984 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7985 struct io_sq_data
*sqd
;
7988 if (!capable(CAP_SYS_ADMIN
) && !capable(CAP_SYS_NICE
))
7991 sqd
= io_get_sq_data(p
);
7998 io_sq_thread_park(sqd
);
7999 mutex_lock(&sqd
->ctx_lock
);
8000 list_add(&ctx
->sqd_list
, &sqd
->ctx_new_list
);
8001 mutex_unlock(&sqd
->ctx_lock
);
8002 io_sq_thread_unpark(sqd
);
8004 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8005 if (!ctx
->sq_thread_idle
)
8006 ctx
->sq_thread_idle
= HZ
;
8011 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8012 int cpu
= p
->sq_thread_cpu
;
8015 if (cpu
>= nr_cpu_ids
)
8017 if (!cpu_online(cpu
))
8020 sqd
->thread
= kthread_create_on_cpu(io_sq_thread
, sqd
,
8021 cpu
, "io_uring-sq");
8023 sqd
->thread
= kthread_create(io_sq_thread
, sqd
,
8026 if (IS_ERR(sqd
->thread
)) {
8027 ret
= PTR_ERR(sqd
->thread
);
8031 ret
= io_uring_alloc_task_context(sqd
->thread
);
8034 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8035 /* Can't have SQ_AFF without SQPOLL */
8041 ret
= io_init_wq_offload(ctx
, p
);
8047 io_finish_async(ctx
);
8051 static void io_sq_offload_start(struct io_ring_ctx
*ctx
)
8053 struct io_sq_data
*sqd
= ctx
->sq_data
;
8055 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && sqd
->thread
)
8056 wake_up_process(sqd
->thread
);
8059 static inline void __io_unaccount_mem(struct user_struct
*user
,
8060 unsigned long nr_pages
)
8062 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8065 static inline int __io_account_mem(struct user_struct
*user
,
8066 unsigned long nr_pages
)
8068 unsigned long page_limit
, cur_pages
, new_pages
;
8070 /* Don't allow more pages than we can safely lock */
8071 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8074 cur_pages
= atomic_long_read(&user
->locked_vm
);
8075 new_pages
= cur_pages
+ nr_pages
;
8076 if (new_pages
> page_limit
)
8078 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8079 new_pages
) != cur_pages
);
8084 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8085 enum io_mem_account acct
)
8088 __io_unaccount_mem(ctx
->user
, nr_pages
);
8090 if (ctx
->mm_account
) {
8091 if (acct
== ACCT_LOCKED
)
8092 ctx
->mm_account
->locked_vm
-= nr_pages
;
8093 else if (acct
== ACCT_PINNED
)
8094 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8098 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8099 enum io_mem_account acct
)
8103 if (ctx
->limit_mem
) {
8104 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8109 if (ctx
->mm_account
) {
8110 if (acct
== ACCT_LOCKED
)
8111 ctx
->mm_account
->locked_vm
+= nr_pages
;
8112 else if (acct
== ACCT_PINNED
)
8113 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8119 static void io_mem_free(void *ptr
)
8126 page
= virt_to_head_page(ptr
);
8127 if (put_page_testzero(page
))
8128 free_compound_page(page
);
8131 static void *io_mem_alloc(size_t size
)
8133 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8136 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8139 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8142 struct io_rings
*rings
;
8143 size_t off
, sq_array_size
;
8145 off
= struct_size(rings
, cqes
, cq_entries
);
8146 if (off
== SIZE_MAX
)
8150 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8158 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8159 if (sq_array_size
== SIZE_MAX
)
8162 if (check_add_overflow(off
, sq_array_size
, &off
))
8168 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
8172 pages
= (size_t)1 << get_order(
8173 rings_size(sq_entries
, cq_entries
, NULL
));
8174 pages
+= (size_t)1 << get_order(
8175 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
8180 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
8184 if (!ctx
->user_bufs
)
8187 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8188 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8190 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
8191 unpin_user_page(imu
->bvec
[j
].bv_page
);
8193 if (imu
->acct_pages
)
8194 io_unaccount_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8199 kfree(ctx
->user_bufs
);
8200 ctx
->user_bufs
= NULL
;
8201 ctx
->nr_user_bufs
= 0;
8205 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8206 void __user
*arg
, unsigned index
)
8208 struct iovec __user
*src
;
8210 #ifdef CONFIG_COMPAT
8212 struct compat_iovec __user
*ciovs
;
8213 struct compat_iovec ciov
;
8215 ciovs
= (struct compat_iovec __user
*) arg
;
8216 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8219 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8220 dst
->iov_len
= ciov
.iov_len
;
8224 src
= (struct iovec __user
*) arg
;
8225 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8231 * Not super efficient, but this is just a registration time. And we do cache
8232 * the last compound head, so generally we'll only do a full search if we don't
8235 * We check if the given compound head page has already been accounted, to
8236 * avoid double accounting it. This allows us to account the full size of the
8237 * page, not just the constituent pages of a huge page.
8239 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8240 int nr_pages
, struct page
*hpage
)
8244 /* check current page array */
8245 for (i
= 0; i
< nr_pages
; i
++) {
8246 if (!PageCompound(pages
[i
]))
8248 if (compound_head(pages
[i
]) == hpage
)
8252 /* check previously registered pages */
8253 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8254 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8256 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8257 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8259 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8267 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8268 int nr_pages
, struct io_mapped_ubuf
*imu
,
8269 struct page
**last_hpage
)
8273 for (i
= 0; i
< nr_pages
; i
++) {
8274 if (!PageCompound(pages
[i
])) {
8279 hpage
= compound_head(pages
[i
]);
8280 if (hpage
== *last_hpage
)
8282 *last_hpage
= hpage
;
8283 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8285 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8289 if (!imu
->acct_pages
)
8292 ret
= io_account_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8294 imu
->acct_pages
= 0;
8298 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8301 struct vm_area_struct
**vmas
= NULL
;
8302 struct page
**pages
= NULL
;
8303 struct page
*last_hpage
= NULL
;
8304 int i
, j
, got_pages
= 0;
8309 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8312 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8314 if (!ctx
->user_bufs
)
8317 for (i
= 0; i
< nr_args
; i
++) {
8318 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8319 unsigned long off
, start
, end
, ubuf
;
8324 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8329 * Don't impose further limits on the size and buffer
8330 * constraints here, we'll -EINVAL later when IO is
8331 * submitted if they are wrong.
8334 if (!iov
.iov_base
|| !iov
.iov_len
)
8337 /* arbitrary limit, but we need something */
8338 if (iov
.iov_len
> SZ_1G
)
8341 ubuf
= (unsigned long) iov
.iov_base
;
8342 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8343 start
= ubuf
>> PAGE_SHIFT
;
8344 nr_pages
= end
- start
;
8347 if (!pages
|| nr_pages
> got_pages
) {
8350 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
8352 vmas
= kvmalloc_array(nr_pages
,
8353 sizeof(struct vm_area_struct
*),
8355 if (!pages
|| !vmas
) {
8359 got_pages
= nr_pages
;
8362 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8369 mmap_read_lock(current
->mm
);
8370 pret
= pin_user_pages(ubuf
, nr_pages
,
8371 FOLL_WRITE
| FOLL_LONGTERM
,
8373 if (pret
== nr_pages
) {
8374 /* don't support file backed memory */
8375 for (j
= 0; j
< nr_pages
; j
++) {
8376 struct vm_area_struct
*vma
= vmas
[j
];
8379 !is_file_hugepages(vma
->vm_file
)) {
8385 ret
= pret
< 0 ? pret
: -EFAULT
;
8387 mmap_read_unlock(current
->mm
);
8390 * if we did partial map, or found file backed vmas,
8391 * release any pages we did get
8394 unpin_user_pages(pages
, pret
);
8399 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, &last_hpage
);
8401 unpin_user_pages(pages
, pret
);
8406 off
= ubuf
& ~PAGE_MASK
;
8408 for (j
= 0; j
< nr_pages
; j
++) {
8411 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8412 imu
->bvec
[j
].bv_page
= pages
[j
];
8413 imu
->bvec
[j
].bv_len
= vec_len
;
8414 imu
->bvec
[j
].bv_offset
= off
;
8418 /* store original address for later verification */
8420 imu
->len
= iov
.iov_len
;
8421 imu
->nr_bvecs
= nr_pages
;
8423 ctx
->nr_user_bufs
++;
8431 io_sqe_buffer_unregister(ctx
);
8435 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8437 __s32 __user
*fds
= arg
;
8443 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8446 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8447 if (IS_ERR(ctx
->cq_ev_fd
)) {
8448 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8449 ctx
->cq_ev_fd
= NULL
;
8456 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8458 if (ctx
->cq_ev_fd
) {
8459 eventfd_ctx_put(ctx
->cq_ev_fd
);
8460 ctx
->cq_ev_fd
= NULL
;
8467 static int __io_destroy_buffers(int id
, void *p
, void *data
)
8469 struct io_ring_ctx
*ctx
= data
;
8470 struct io_buffer
*buf
= p
;
8472 __io_remove_buffers(ctx
, buf
, id
, -1U);
8476 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8478 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
8479 idr_destroy(&ctx
->io_buffer_idr
);
8482 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8484 io_finish_async(ctx
);
8485 io_sqe_buffer_unregister(ctx
);
8487 if (ctx
->sqo_task
) {
8488 put_task_struct(ctx
->sqo_task
);
8489 ctx
->sqo_task
= NULL
;
8490 mmdrop(ctx
->mm_account
);
8491 ctx
->mm_account
= NULL
;
8494 #ifdef CONFIG_BLK_CGROUP
8495 if (ctx
->sqo_blkcg_css
)
8496 css_put(ctx
->sqo_blkcg_css
);
8499 io_sqe_files_unregister(ctx
);
8500 io_eventfd_unregister(ctx
);
8501 io_destroy_buffers(ctx
);
8502 idr_destroy(&ctx
->personality_idr
);
8504 #if defined(CONFIG_UNIX)
8505 if (ctx
->ring_sock
) {
8506 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8507 sock_release(ctx
->ring_sock
);
8511 io_mem_free(ctx
->rings
);
8512 io_mem_free(ctx
->sq_sqes
);
8514 percpu_ref_exit(&ctx
->refs
);
8515 free_uid(ctx
->user
);
8516 put_cred(ctx
->creds
);
8517 kfree(ctx
->cancel_hash
);
8518 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
8522 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8524 struct io_ring_ctx
*ctx
= file
->private_data
;
8527 poll_wait(file
, &ctx
->cq_wait
, wait
);
8529 * synchronizes with barrier from wq_has_sleeper call in
8533 if (!io_sqring_full(ctx
))
8534 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8535 if (io_cqring_events(ctx
, false))
8536 mask
|= EPOLLIN
| EPOLLRDNORM
;
8541 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8543 struct io_ring_ctx
*ctx
= file
->private_data
;
8545 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8548 static int io_remove_personalities(int id
, void *p
, void *data
)
8550 struct io_ring_ctx
*ctx
= data
;
8551 struct io_identity
*iod
;
8553 iod
= idr_remove(&ctx
->personality_idr
, id
);
8555 put_cred(iod
->creds
);
8556 if (refcount_dec_and_test(&iod
->count
))
8562 static void io_ring_exit_work(struct work_struct
*work
)
8564 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
8568 * If we're doing polled IO and end up having requests being
8569 * submitted async (out-of-line), then completions can come in while
8570 * we're waiting for refs to drop. We need to reap these manually,
8571 * as nobody else will be looking for them.
8575 io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8576 io_iopoll_try_reap_events(ctx
);
8577 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8578 io_ring_ctx_free(ctx
);
8581 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8583 mutex_lock(&ctx
->uring_lock
);
8584 percpu_ref_kill(&ctx
->refs
);
8585 mutex_unlock(&ctx
->uring_lock
);
8587 io_kill_timeouts(ctx
, NULL
, NULL
);
8588 io_poll_remove_all(ctx
, NULL
, NULL
);
8591 io_wq_cancel_all(ctx
->io_wq
);
8593 /* if we failed setting up the ctx, we might not have any rings */
8595 io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8596 io_iopoll_try_reap_events(ctx
);
8597 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
8600 * Do this upfront, so we won't have a grace period where the ring
8601 * is closed but resources aren't reaped yet. This can cause
8602 * spurious failure in setting up a new ring.
8604 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
8607 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8609 * Use system_unbound_wq to avoid spawning tons of event kworkers
8610 * if we're exiting a ton of rings at the same time. It just adds
8611 * noise and overhead, there's no discernable change in runtime
8612 * over using system_wq.
8614 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8617 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8619 struct io_ring_ctx
*ctx
= file
->private_data
;
8621 file
->private_data
= NULL
;
8622 io_ring_ctx_wait_and_kill(ctx
);
8626 struct io_task_cancel
{
8627 struct task_struct
*task
;
8628 struct files_struct
*files
;
8631 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8633 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8634 struct io_task_cancel
*cancel
= data
;
8637 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8638 unsigned long flags
;
8639 struct io_ring_ctx
*ctx
= req
->ctx
;
8641 /* protect against races with linked timeouts */
8642 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8643 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8644 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8646 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8651 static void io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8652 struct task_struct
*task
,
8653 struct files_struct
*files
)
8655 struct io_defer_entry
*de
= NULL
;
8658 spin_lock_irq(&ctx
->completion_lock
);
8659 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8660 if (io_match_task(de
->req
, task
, files
)) {
8661 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8665 spin_unlock_irq(&ctx
->completion_lock
);
8667 while (!list_empty(&list
)) {
8668 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8669 list_del_init(&de
->list
);
8670 req_set_fail_links(de
->req
);
8671 io_put_req(de
->req
);
8672 io_req_complete(de
->req
, -ECANCELED
);
8677 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
8678 struct task_struct
*task
,
8679 struct files_struct
*files
)
8681 while (!list_empty_careful(&ctx
->inflight_list
)) {
8682 struct io_task_cancel cancel
= { .task
= task
, .files
= NULL
, };
8683 struct io_kiocb
*req
;
8687 spin_lock_irq(&ctx
->inflight_lock
);
8688 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
) {
8689 if (req
->task
== task
&&
8690 (req
->work
.flags
& IO_WQ_WORK_FILES
) &&
8691 req
->work
.identity
->files
!= files
)
8697 prepare_to_wait(&ctx
->inflight_wait
, &wait
,
8698 TASK_UNINTERRUPTIBLE
);
8699 spin_unlock_irq(&ctx
->inflight_lock
);
8701 /* We need to keep going until we don't find a matching req */
8705 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
, &cancel
, true);
8706 io_poll_remove_all(ctx
, task
, files
);
8707 io_kill_timeouts(ctx
, task
, files
);
8708 /* cancellations _may_ trigger task work */
8711 finish_wait(&ctx
->inflight_wait
, &wait
);
8715 static void __io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
8716 struct task_struct
*task
)
8719 struct io_task_cancel cancel
= { .task
= task
, .files
= NULL
, };
8720 enum io_wq_cancel cret
;
8723 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
, &cancel
, true);
8724 if (cret
!= IO_WQ_CANCEL_NOTFOUND
)
8727 /* SQPOLL thread does its own polling */
8728 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
8729 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8730 io_iopoll_try_reap_events(ctx
);
8735 ret
|= io_poll_remove_all(ctx
, task
, NULL
);
8736 ret
|= io_kill_timeouts(ctx
, task
, NULL
);
8745 * We need to iteratively cancel requests, in case a request has dependent
8746 * hard links. These persist even for failure of cancelations, hence keep
8747 * looping until none are found.
8749 static void io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
8750 struct files_struct
*files
)
8752 struct task_struct
*task
= current
;
8754 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
8755 task
= ctx
->sq_data
->thread
;
8756 atomic_inc(&task
->io_uring
->in_idle
);
8757 io_sq_thread_park(ctx
->sq_data
);
8760 io_cancel_defer_files(ctx
, task
, files
);
8761 io_cqring_overflow_flush(ctx
, true, task
, files
);
8762 io_uring_cancel_files(ctx
, task
, files
);
8765 __io_uring_cancel_task_requests(ctx
, task
);
8767 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
8768 atomic_dec(&task
->io_uring
->in_idle
);
8770 * If the files that are going away are the ones in the thread
8771 * identity, clear them out.
8773 if (task
->io_uring
->identity
->files
== files
)
8774 task
->io_uring
->identity
->files
= NULL
;
8775 io_sq_thread_unpark(ctx
->sq_data
);
8780 * Note that this task has used io_uring. We use it for cancelation purposes.
8782 static int io_uring_add_task_file(struct io_ring_ctx
*ctx
, struct file
*file
)
8784 struct io_uring_task
*tctx
= current
->io_uring
;
8786 if (unlikely(!tctx
)) {
8789 ret
= io_uring_alloc_task_context(current
);
8792 tctx
= current
->io_uring
;
8794 if (tctx
->last
!= file
) {
8795 void *old
= xa_load(&tctx
->xa
, (unsigned long)file
);
8799 xa_store(&tctx
->xa
, (unsigned long)file
, file
, GFP_KERNEL
);
8805 * This is race safe in that the task itself is doing this, hence it
8806 * cannot be going through the exit/cancel paths at the same time.
8807 * This cannot be modified while exit/cancel is running.
8809 if (!tctx
->sqpoll
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
8810 tctx
->sqpoll
= true;
8816 * Remove this io_uring_file -> task mapping.
8818 static void io_uring_del_task_file(struct file
*file
)
8820 struct io_uring_task
*tctx
= current
->io_uring
;
8822 if (tctx
->last
== file
)
8824 file
= xa_erase(&tctx
->xa
, (unsigned long)file
);
8830 * Drop task note for this file if we're the only ones that hold it after
8833 static void io_uring_attempt_task_drop(struct file
*file
)
8835 if (!current
->io_uring
)
8838 * fput() is pending, will be 2 if the only other ref is our potential
8839 * task file note. If the task is exiting, drop regardless of count.
8841 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
) ||
8842 atomic_long_read(&file
->f_count
) == 2)
8843 io_uring_del_task_file(file
);
8846 void __io_uring_files_cancel(struct files_struct
*files
)
8848 struct io_uring_task
*tctx
= current
->io_uring
;
8850 unsigned long index
;
8852 /* make sure overflow events are dropped */
8853 atomic_inc(&tctx
->in_idle
);
8855 xa_for_each(&tctx
->xa
, index
, file
) {
8856 struct io_ring_ctx
*ctx
= file
->private_data
;
8858 io_uring_cancel_task_requests(ctx
, files
);
8860 io_uring_del_task_file(file
);
8863 atomic_dec(&tctx
->in_idle
);
8866 static s64
tctx_inflight(struct io_uring_task
*tctx
)
8868 unsigned long index
;
8872 inflight
= percpu_counter_sum(&tctx
->inflight
);
8877 * If we have SQPOLL rings, then we need to iterate and find them, and
8878 * add the pending count for those.
8880 xa_for_each(&tctx
->xa
, index
, file
) {
8881 struct io_ring_ctx
*ctx
= file
->private_data
;
8883 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8884 struct io_uring_task
*__tctx
= ctx
->sqo_task
->io_uring
;
8886 inflight
+= percpu_counter_sum(&__tctx
->inflight
);
8894 * Find any io_uring fd that this task has registered or done IO on, and cancel
8897 void __io_uring_task_cancel(void)
8899 struct io_uring_task
*tctx
= current
->io_uring
;
8903 /* make sure overflow events are dropped */
8904 atomic_inc(&tctx
->in_idle
);
8907 /* read completions before cancelations */
8908 inflight
= tctx_inflight(tctx
);
8911 __io_uring_files_cancel(NULL
);
8913 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
8916 * If we've seen completions, retry. This avoids a race where
8917 * a completion comes in before we did prepare_to_wait().
8919 if (inflight
!= tctx_inflight(tctx
))
8924 finish_wait(&tctx
->wait
, &wait
);
8925 atomic_dec(&tctx
->in_idle
);
8928 static int io_uring_flush(struct file
*file
, void *data
)
8930 io_uring_attempt_task_drop(file
);
8934 static void *io_uring_validate_mmap_request(struct file
*file
,
8935 loff_t pgoff
, size_t sz
)
8937 struct io_ring_ctx
*ctx
= file
->private_data
;
8938 loff_t offset
= pgoff
<< PAGE_SHIFT
;
8943 case IORING_OFF_SQ_RING
:
8944 case IORING_OFF_CQ_RING
:
8947 case IORING_OFF_SQES
:
8951 return ERR_PTR(-EINVAL
);
8954 page
= virt_to_head_page(ptr
);
8955 if (sz
> page_size(page
))
8956 return ERR_PTR(-EINVAL
);
8963 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
8965 size_t sz
= vma
->vm_end
- vma
->vm_start
;
8969 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
8971 return PTR_ERR(ptr
);
8973 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
8974 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
8977 #else /* !CONFIG_MMU */
8979 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
8981 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
8984 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
8986 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
8989 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
8990 unsigned long addr
, unsigned long len
,
8991 unsigned long pgoff
, unsigned long flags
)
8995 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
8997 return PTR_ERR(ptr
);
8999 return (unsigned long) ptr
;
9002 #endif /* !CONFIG_MMU */
9004 static void io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9009 if (!io_sqring_full(ctx
))
9012 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9014 if (!io_sqring_full(ctx
))
9018 } while (!signal_pending(current
));
9020 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9023 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9024 struct __kernel_timespec __user
**ts
,
9025 const sigset_t __user
**sig
)
9027 struct io_uring_getevents_arg arg
;
9030 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9031 * is just a pointer to the sigset_t.
9033 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9034 *sig
= (const sigset_t __user
*) argp
;
9040 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9041 * timespec and sigset_t pointers if good.
9043 if (*argsz
!= sizeof(arg
))
9045 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9047 *sig
= u64_to_user_ptr(arg
.sigmask
);
9048 *argsz
= arg
.sigmask_sz
;
9049 *ts
= u64_to_user_ptr(arg
.ts
);
9053 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9054 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9057 struct io_ring_ctx
*ctx
;
9064 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9065 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
))
9073 if (f
.file
->f_op
!= &io_uring_fops
)
9077 ctx
= f
.file
->private_data
;
9078 if (!percpu_ref_tryget(&ctx
->refs
))
9082 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9086 * For SQ polling, the thread will do all submissions and completions.
9087 * Just return the requested submit count, and wake the thread if
9091 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9092 if (!list_empty_careful(&ctx
->cq_overflow_list
))
9093 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
9094 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9095 wake_up(&ctx
->sq_data
->wait
);
9096 if (flags
& IORING_ENTER_SQ_WAIT
)
9097 io_sqpoll_wait_sq(ctx
);
9098 submitted
= to_submit
;
9099 } else if (to_submit
) {
9100 ret
= io_uring_add_task_file(ctx
, f
.file
);
9103 mutex_lock(&ctx
->uring_lock
);
9104 submitted
= io_submit_sqes(ctx
, to_submit
);
9105 mutex_unlock(&ctx
->uring_lock
);
9107 if (submitted
!= to_submit
)
9110 if (flags
& IORING_ENTER_GETEVENTS
) {
9111 const sigset_t __user
*sig
;
9112 struct __kernel_timespec __user
*ts
;
9114 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9118 min_complete
= min(min_complete
, ctx
->cq_entries
);
9121 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9122 * space applications don't need to do io completion events
9123 * polling again, they can rely on io_sq_thread to do polling
9124 * work, which can reduce cpu usage and uring_lock contention.
9126 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9127 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9128 ret
= io_iopoll_check(ctx
, min_complete
);
9130 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9135 percpu_ref_put(&ctx
->refs
);
9138 return submitted
? submitted
: ret
;
9141 #ifdef CONFIG_PROC_FS
9142 static int io_uring_show_cred(int id
, void *p
, void *data
)
9144 struct io_identity
*iod
= p
;
9145 const struct cred
*cred
= iod
->creds
;
9146 struct seq_file
*m
= data
;
9147 struct user_namespace
*uns
= seq_user_ns(m
);
9148 struct group_info
*gi
;
9153 seq_printf(m
, "%5d\n", id
);
9154 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9155 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9156 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9157 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9158 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9159 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9160 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9161 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9162 seq_puts(m
, "\n\tGroups:\t");
9163 gi
= cred
->group_info
;
9164 for (g
= 0; g
< gi
->ngroups
; g
++) {
9165 seq_put_decimal_ull(m
, g
? " " : "",
9166 from_kgid_munged(uns
, gi
->gid
[g
]));
9168 seq_puts(m
, "\n\tCapEff:\t");
9169 cap
= cred
->cap_effective
;
9170 CAP_FOR_EACH_U32(__capi
)
9171 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9176 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9178 struct io_sq_data
*sq
= NULL
;
9183 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9184 * since fdinfo case grabs it in the opposite direction of normal use
9185 * cases. If we fail to get the lock, we just don't iterate any
9186 * structures that could be going away outside the io_uring mutex.
9188 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9190 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9193 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9194 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9195 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9196 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9197 struct fixed_file_table
*table
;
9200 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
9201 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
9203 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9205 seq_printf(m
, "%5u: <none>\n", i
);
9207 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9208 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9209 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9211 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9212 (unsigned int) buf
->len
);
9214 if (has_lock
&& !idr_is_empty(&ctx
->personality_idr
)) {
9215 seq_printf(m
, "Personalities:\n");
9216 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
9218 seq_printf(m
, "PollList:\n");
9219 spin_lock_irq(&ctx
->completion_lock
);
9220 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9221 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9222 struct io_kiocb
*req
;
9224 hlist_for_each_entry(req
, list
, hash_node
)
9225 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9226 req
->task
->task_works
!= NULL
);
9228 spin_unlock_irq(&ctx
->completion_lock
);
9230 mutex_unlock(&ctx
->uring_lock
);
9233 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9235 struct io_ring_ctx
*ctx
= f
->private_data
;
9237 if (percpu_ref_tryget(&ctx
->refs
)) {
9238 __io_uring_show_fdinfo(ctx
, m
);
9239 percpu_ref_put(&ctx
->refs
);
9244 static const struct file_operations io_uring_fops
= {
9245 .release
= io_uring_release
,
9246 .flush
= io_uring_flush
,
9247 .mmap
= io_uring_mmap
,
9249 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9250 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9252 .poll
= io_uring_poll
,
9253 .fasync
= io_uring_fasync
,
9254 #ifdef CONFIG_PROC_FS
9255 .show_fdinfo
= io_uring_show_fdinfo
,
9259 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9260 struct io_uring_params
*p
)
9262 struct io_rings
*rings
;
9263 size_t size
, sq_array_offset
;
9265 /* make sure these are sane, as we already accounted them */
9266 ctx
->sq_entries
= p
->sq_entries
;
9267 ctx
->cq_entries
= p
->cq_entries
;
9269 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9270 if (size
== SIZE_MAX
)
9273 rings
= io_mem_alloc(size
);
9278 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9279 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9280 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9281 rings
->sq_ring_entries
= p
->sq_entries
;
9282 rings
->cq_ring_entries
= p
->cq_entries
;
9283 ctx
->sq_mask
= rings
->sq_ring_mask
;
9284 ctx
->cq_mask
= rings
->cq_ring_mask
;
9286 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9287 if (size
== SIZE_MAX
) {
9288 io_mem_free(ctx
->rings
);
9293 ctx
->sq_sqes
= io_mem_alloc(size
);
9294 if (!ctx
->sq_sqes
) {
9295 io_mem_free(ctx
->rings
);
9304 * Allocate an anonymous fd, this is what constitutes the application
9305 * visible backing of an io_uring instance. The application mmaps this
9306 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9307 * we have to tie this fd to a socket for file garbage collection purposes.
9309 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
9314 #if defined(CONFIG_UNIX)
9315 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9321 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9325 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9326 O_RDWR
| O_CLOEXEC
);
9330 ret
= PTR_ERR(file
);
9334 #if defined(CONFIG_UNIX)
9335 ctx
->ring_sock
->file
= file
;
9337 if (unlikely(io_uring_add_task_file(ctx
, file
))) {
9338 file
= ERR_PTR(-ENOMEM
);
9341 fd_install(ret
, file
);
9344 #if defined(CONFIG_UNIX)
9345 sock_release(ctx
->ring_sock
);
9346 ctx
->ring_sock
= NULL
;
9351 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9352 struct io_uring_params __user
*params
)
9354 struct user_struct
*user
= NULL
;
9355 struct io_ring_ctx
*ctx
;
9361 if (entries
> IORING_MAX_ENTRIES
) {
9362 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9364 entries
= IORING_MAX_ENTRIES
;
9368 * Use twice as many entries for the CQ ring. It's possible for the
9369 * application to drive a higher depth than the size of the SQ ring,
9370 * since the sqes are only used at submission time. This allows for
9371 * some flexibility in overcommitting a bit. If the application has
9372 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9373 * of CQ ring entries manually.
9375 p
->sq_entries
= roundup_pow_of_two(entries
);
9376 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9378 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9379 * to a power-of-two, if it isn't already. We do NOT impose
9380 * any cq vs sq ring sizing.
9382 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9383 if (p
->cq_entries
< p
->sq_entries
)
9385 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9386 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9388 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9391 p
->cq_entries
= 2 * p
->sq_entries
;
9394 user
= get_uid(current_user());
9395 limit_mem
= !capable(CAP_IPC_LOCK
);
9398 ret
= __io_account_mem(user
,
9399 ring_pages(p
->sq_entries
, p
->cq_entries
));
9406 ctx
= io_ring_ctx_alloc(p
);
9409 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
9414 ctx
->compat
= in_compat_syscall();
9416 ctx
->creds
= get_current_cred();
9418 ctx
->loginuid
= current
->loginuid
;
9419 ctx
->sessionid
= current
->sessionid
;
9421 ctx
->sqo_task
= get_task_struct(current
);
9424 * This is just grabbed for accounting purposes. When a process exits,
9425 * the mm is exited and dropped before the files, hence we need to hang
9426 * on to this mm purely for the purposes of being able to unaccount
9427 * memory (locked/pinned vm). It's not used for anything else.
9429 mmgrab(current
->mm
);
9430 ctx
->mm_account
= current
->mm
;
9432 #ifdef CONFIG_BLK_CGROUP
9434 * The sq thread will belong to the original cgroup it was inited in.
9435 * If the cgroup goes offline (e.g. disabling the io controller), then
9436 * issued bios will be associated with the closest cgroup later in the
9440 ctx
->sqo_blkcg_css
= blkcg_css();
9441 ret
= css_tryget_online(ctx
->sqo_blkcg_css
);
9444 /* don't init against a dying cgroup, have the user try again */
9445 ctx
->sqo_blkcg_css
= NULL
;
9452 * Account memory _before_ installing the file descriptor. Once
9453 * the descriptor is installed, it can get closed at any time. Also
9454 * do this before hitting the general error path, as ring freeing
9455 * will un-account as well.
9457 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
9459 ctx
->limit_mem
= limit_mem
;
9461 ret
= io_allocate_scq_urings(ctx
, p
);
9465 ret
= io_sq_offload_create(ctx
, p
);
9469 if (!(p
->flags
& IORING_SETUP_R_DISABLED
))
9470 io_sq_offload_start(ctx
);
9472 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9473 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9474 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9475 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9476 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9477 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9478 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9479 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9481 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9482 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9483 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9484 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9485 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9486 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9487 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9488 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9490 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9491 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9492 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9493 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9494 IORING_FEAT_EXT_ARG
;
9496 if (copy_to_user(params
, p
, sizeof(*p
))) {
9502 * Install ring fd as the very last thing, so we don't risk someone
9503 * having closed it before we finish setup
9505 ret
= io_uring_get_fd(ctx
);
9509 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9512 io_ring_ctx_wait_and_kill(ctx
);
9517 * Sets up an aio uring context, and returns the fd. Applications asks for a
9518 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9519 * params structure passed in.
9521 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9523 struct io_uring_params p
;
9526 if (copy_from_user(&p
, params
, sizeof(p
)))
9528 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9533 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9534 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9535 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9536 IORING_SETUP_R_DISABLED
))
9539 return io_uring_create(entries
, &p
, params
);
9542 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9543 struct io_uring_params __user
*, params
)
9545 return io_uring_setup(entries
, params
);
9548 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9550 struct io_uring_probe
*p
;
9554 size
= struct_size(p
, ops
, nr_args
);
9555 if (size
== SIZE_MAX
)
9557 p
= kzalloc(size
, GFP_KERNEL
);
9562 if (copy_from_user(p
, arg
, size
))
9565 if (memchr_inv(p
, 0, size
))
9568 p
->last_op
= IORING_OP_LAST
- 1;
9569 if (nr_args
> IORING_OP_LAST
)
9570 nr_args
= IORING_OP_LAST
;
9572 for (i
= 0; i
< nr_args
; i
++) {
9574 if (!io_op_defs
[i
].not_supported
)
9575 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9580 if (copy_to_user(arg
, p
, size
))
9587 static int io_register_personality(struct io_ring_ctx
*ctx
)
9589 struct io_identity
*id
;
9592 id
= kmalloc(sizeof(*id
), GFP_KERNEL
);
9596 io_init_identity(id
);
9597 id
->creds
= get_current_cred();
9599 ret
= idr_alloc_cyclic(&ctx
->personality_idr
, id
, 1, USHRT_MAX
, GFP_KERNEL
);
9601 put_cred(id
->creds
);
9607 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9609 struct io_identity
*iod
;
9611 iod
= idr_remove(&ctx
->personality_idr
, id
);
9613 put_cred(iod
->creds
);
9614 if (refcount_dec_and_test(&iod
->count
))
9622 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9623 unsigned int nr_args
)
9625 struct io_uring_restriction
*res
;
9629 /* Restrictions allowed only if rings started disabled */
9630 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9633 /* We allow only a single restrictions registration */
9634 if (ctx
->restrictions
.registered
)
9637 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9640 size
= array_size(nr_args
, sizeof(*res
));
9641 if (size
== SIZE_MAX
)
9644 res
= memdup_user(arg
, size
);
9646 return PTR_ERR(res
);
9650 for (i
= 0; i
< nr_args
; i
++) {
9651 switch (res
[i
].opcode
) {
9652 case IORING_RESTRICTION_REGISTER_OP
:
9653 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9658 __set_bit(res
[i
].register_op
,
9659 ctx
->restrictions
.register_op
);
9661 case IORING_RESTRICTION_SQE_OP
:
9662 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9667 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9669 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9670 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9672 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9673 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9682 /* Reset all restrictions if an error happened */
9684 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9686 ctx
->restrictions
.registered
= true;
9692 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
9694 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9697 if (ctx
->restrictions
.registered
)
9698 ctx
->restricted
= 1;
9700 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
9702 io_sq_offload_start(ctx
);
9707 static bool io_register_op_must_quiesce(int op
)
9710 case IORING_UNREGISTER_FILES
:
9711 case IORING_REGISTER_FILES_UPDATE
:
9712 case IORING_REGISTER_PROBE
:
9713 case IORING_REGISTER_PERSONALITY
:
9714 case IORING_UNREGISTER_PERSONALITY
:
9721 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
9722 void __user
*arg
, unsigned nr_args
)
9723 __releases(ctx
->uring_lock
)
9724 __acquires(ctx
->uring_lock
)
9729 * We're inside the ring mutex, if the ref is already dying, then
9730 * someone else killed the ctx or is already going through
9731 * io_uring_register().
9733 if (percpu_ref_is_dying(&ctx
->refs
))
9736 if (io_register_op_must_quiesce(opcode
)) {
9737 percpu_ref_kill(&ctx
->refs
);
9740 * Drop uring mutex before waiting for references to exit. If
9741 * another thread is currently inside io_uring_enter() it might
9742 * need to grab the uring_lock to make progress. If we hold it
9743 * here across the drain wait, then we can deadlock. It's safe
9744 * to drop the mutex here, since no new references will come in
9745 * after we've killed the percpu ref.
9747 mutex_unlock(&ctx
->uring_lock
);
9749 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
9752 ret
= io_run_task_work_sig();
9757 mutex_lock(&ctx
->uring_lock
);
9760 percpu_ref_resurrect(&ctx
->refs
);
9765 if (ctx
->restricted
) {
9766 if (opcode
>= IORING_REGISTER_LAST
) {
9771 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
9778 case IORING_REGISTER_BUFFERS
:
9779 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
9781 case IORING_UNREGISTER_BUFFERS
:
9785 ret
= io_sqe_buffer_unregister(ctx
);
9787 case IORING_REGISTER_FILES
:
9788 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
9790 case IORING_UNREGISTER_FILES
:
9794 ret
= io_sqe_files_unregister(ctx
);
9796 case IORING_REGISTER_FILES_UPDATE
:
9797 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
9799 case IORING_REGISTER_EVENTFD
:
9800 case IORING_REGISTER_EVENTFD_ASYNC
:
9804 ret
= io_eventfd_register(ctx
, arg
);
9807 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
9808 ctx
->eventfd_async
= 1;
9810 ctx
->eventfd_async
= 0;
9812 case IORING_UNREGISTER_EVENTFD
:
9816 ret
= io_eventfd_unregister(ctx
);
9818 case IORING_REGISTER_PROBE
:
9820 if (!arg
|| nr_args
> 256)
9822 ret
= io_probe(ctx
, arg
, nr_args
);
9824 case IORING_REGISTER_PERSONALITY
:
9828 ret
= io_register_personality(ctx
);
9830 case IORING_UNREGISTER_PERSONALITY
:
9834 ret
= io_unregister_personality(ctx
, nr_args
);
9836 case IORING_REGISTER_ENABLE_RINGS
:
9840 ret
= io_register_enable_rings(ctx
);
9842 case IORING_REGISTER_RESTRICTIONS
:
9843 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
9851 if (io_register_op_must_quiesce(opcode
)) {
9852 /* bring the ctx back to life */
9853 percpu_ref_reinit(&ctx
->refs
);
9855 reinit_completion(&ctx
->ref_comp
);
9860 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
9861 void __user
*, arg
, unsigned int, nr_args
)
9863 struct io_ring_ctx
*ctx
;
9872 if (f
.file
->f_op
!= &io_uring_fops
)
9875 ctx
= f
.file
->private_data
;
9877 mutex_lock(&ctx
->uring_lock
);
9878 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
9879 mutex_unlock(&ctx
->uring_lock
);
9880 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
9881 ctx
->cq_ev_fd
!= NULL
, ret
);
9887 static int __init
io_uring_init(void)
9889 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9890 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9891 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9894 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9895 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9896 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
9897 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
9898 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
9899 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
9900 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
9901 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
9902 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
9903 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
9904 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
9905 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
9906 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
9907 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
9908 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
9909 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
9910 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
9911 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
9912 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
9913 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
9914 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
9915 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
9916 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
9917 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
9918 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
9919 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
9920 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
9921 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
9922 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
9923 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
9924 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
9926 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
9927 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
9928 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
9931 __initcall(io_uring_init
);