1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp
;
109 u32 tail ____cacheline_aligned_in_smp
;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq
, cq
;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask
, cq_ring_mask
;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries
, cq_ring_entries
;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
190 struct io_mapped_ubuf
{
193 struct bio_vec
*bvec
;
194 unsigned int nr_bvecs
;
195 unsigned long acct_pages
;
198 struct fixed_file_table
{
202 struct fixed_file_ref_node
{
203 struct percpu_ref refs
;
204 struct list_head node
;
205 struct list_head file_list
;
206 struct fixed_file_data
*file_data
;
207 struct llist_node llist
;
211 struct fixed_file_data
{
212 struct fixed_file_table
*table
;
213 struct io_ring_ctx
*ctx
;
215 struct fixed_file_ref_node
*node
;
216 struct percpu_ref refs
;
217 struct completion done
;
218 struct list_head ref_list
;
223 struct list_head list
;
229 struct io_restriction
{
230 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
231 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
232 u8 sqe_flags_allowed
;
233 u8 sqe_flags_required
;
241 /* ctx's that are using this sqd */
242 struct list_head ctx_list
;
243 struct list_head ctx_new_list
;
244 struct mutex ctx_lock
;
246 struct task_struct
*thread
;
247 struct wait_queue_head wait
;
249 unsigned sq_thread_idle
;
254 struct percpu_ref refs
;
255 } ____cacheline_aligned_in_smp
;
259 unsigned int compat
: 1;
260 unsigned int limit_mem
: 1;
261 unsigned int cq_overflow_flushed
: 1;
262 unsigned int drain_next
: 1;
263 unsigned int eventfd_async
: 1;
264 unsigned int restricted
: 1;
265 unsigned int sqo_dead
: 1;
268 * Ring buffer of indices into array of io_uring_sqe, which is
269 * mmapped by the application using the IORING_OFF_SQES offset.
271 * This indirection could e.g. be used to assign fixed
272 * io_uring_sqe entries to operations and only submit them to
273 * the queue when needed.
275 * The kernel modifies neither the indices array nor the entries
279 unsigned cached_sq_head
;
282 unsigned sq_thread_idle
;
283 unsigned cached_sq_dropped
;
284 unsigned cached_cq_overflow
;
285 unsigned long sq_check_overflow
;
287 struct list_head defer_list
;
288 struct list_head timeout_list
;
289 struct list_head cq_overflow_list
;
291 struct io_uring_sqe
*sq_sqes
;
292 } ____cacheline_aligned_in_smp
;
294 struct io_rings
*rings
;
300 * For SQPOLL usage - we hold a reference to the parent task, so we
301 * have access to the ->files
303 struct task_struct
*sqo_task
;
305 /* Only used for accounting purposes */
306 struct mm_struct
*mm_account
;
308 #ifdef CONFIG_BLK_CGROUP
309 struct cgroup_subsys_state
*sqo_blkcg_css
;
312 struct io_sq_data
*sq_data
; /* if using sq thread polling */
314 struct wait_queue_head sqo_sq_wait
;
315 struct list_head sqd_list
;
318 * If used, fixed file set. Writers must ensure that ->refs is dead,
319 * readers must ensure that ->refs is alive as long as the file* is
320 * used. Only updated through io_uring_register(2).
322 struct fixed_file_data
*file_data
;
323 unsigned nr_user_files
;
325 /* if used, fixed mapped user buffers */
326 unsigned nr_user_bufs
;
327 struct io_mapped_ubuf
*user_bufs
;
329 struct user_struct
*user
;
331 const struct cred
*creds
;
335 unsigned int sessionid
;
338 struct completion ref_comp
;
339 struct completion sq_thread_comp
;
341 /* if all else fails... */
342 struct io_kiocb
*fallback_req
;
344 #if defined(CONFIG_UNIX)
345 struct socket
*ring_sock
;
348 struct idr io_buffer_idr
;
350 struct idr personality_idr
;
353 unsigned cached_cq_tail
;
356 atomic_t cq_timeouts
;
357 unsigned cq_last_tm_flush
;
358 unsigned long cq_check_overflow
;
359 struct wait_queue_head cq_wait
;
360 struct fasync_struct
*cq_fasync
;
361 struct eventfd_ctx
*cq_ev_fd
;
362 } ____cacheline_aligned_in_smp
;
365 struct mutex uring_lock
;
366 wait_queue_head_t wait
;
367 } ____cacheline_aligned_in_smp
;
370 spinlock_t completion_lock
;
373 * ->iopoll_list is protected by the ctx->uring_lock for
374 * io_uring instances that don't use IORING_SETUP_SQPOLL.
375 * For SQPOLL, only the single threaded io_sq_thread() will
376 * manipulate the list, hence no extra locking is needed there.
378 struct list_head iopoll_list
;
379 struct hlist_head
*cancel_hash
;
380 unsigned cancel_hash_bits
;
381 bool poll_multi_file
;
383 spinlock_t inflight_lock
;
384 struct list_head inflight_list
;
385 } ____cacheline_aligned_in_smp
;
387 struct delayed_work file_put_work
;
388 struct llist_head file_put_llist
;
390 struct work_struct exit_work
;
391 struct io_restriction restrictions
;
395 * First field must be the file pointer in all the
396 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
398 struct io_poll_iocb
{
400 struct wait_queue_head
*head
;
404 struct wait_queue_entry wait
;
407 struct io_poll_remove
{
414 struct file
*put_file
;
418 struct io_timeout_data
{
419 struct io_kiocb
*req
;
420 struct hrtimer timer
;
421 struct timespec64 ts
;
422 enum hrtimer_mode mode
;
427 struct sockaddr __user
*addr
;
428 int __user
*addr_len
;
430 unsigned long nofile
;
450 struct list_head list
;
451 /* head of the link, used by linked timeouts only */
452 struct io_kiocb
*head
;
455 struct io_timeout_rem
{
460 struct timespec64 ts
;
465 /* NOTE: kiocb has the file as the first member, so don't do it here */
473 struct sockaddr __user
*addr
;
480 struct user_msghdr __user
*umsg
;
486 struct io_buffer
*kbuf
;
492 bool ignore_nonblock
;
493 struct filename
*filename
;
495 unsigned long nofile
;
498 struct io_files_update
{
524 struct epoll_event event
;
528 struct file
*file_out
;
529 struct file
*file_in
;
536 struct io_provide_buf
{
550 const char __user
*filename
;
551 struct statx __user
*buffer
;
563 struct filename
*oldpath
;
564 struct filename
*newpath
;
572 struct filename
*filename
;
575 struct io_completion
{
577 struct list_head list
;
581 struct io_async_connect
{
582 struct sockaddr_storage address
;
585 struct io_async_msghdr
{
586 struct iovec fast_iov
[UIO_FASTIOV
];
588 struct sockaddr __user
*uaddr
;
590 struct sockaddr_storage addr
;
594 struct iovec fast_iov
[UIO_FASTIOV
];
595 const struct iovec
*free_iovec
;
596 struct iov_iter iter
;
598 struct wait_page_queue wpq
;
602 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
603 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
604 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
605 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
606 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
607 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
613 REQ_F_LINK_TIMEOUT_BIT
,
615 REQ_F_NEED_CLEANUP_BIT
,
617 REQ_F_BUFFER_SELECTED_BIT
,
618 REQ_F_NO_FILE_TABLE_BIT
,
619 REQ_F_WORK_INITIALIZED_BIT
,
620 REQ_F_LTIMEOUT_ACTIVE_BIT
,
622 /* not a real bit, just to check we're not overflowing the space */
628 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
629 /* drain existing IO first */
630 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
632 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
633 /* doesn't sever on completion < 0 */
634 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
636 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
637 /* IOSQE_BUFFER_SELECT */
638 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
640 /* fail rest of links */
641 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
642 /* on inflight list */
643 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
644 /* read/write uses file position */
645 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
646 /* must not punt to workers */
647 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
648 /* has or had linked timeout */
649 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
651 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
653 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
654 /* already went through poll handler */
655 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
656 /* buffer already selected */
657 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
658 /* doesn't need file table for this request */
659 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
660 /* io_wq_work is initialized */
661 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
662 /* linked timeout is active, i.e. prepared by link's head */
663 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
667 struct io_poll_iocb poll
;
668 struct io_poll_iocb
*double_poll
;
672 * NOTE! Each of the iocb union members has the file pointer
673 * as the first entry in their struct definition. So you can
674 * access the file pointer through any of the sub-structs,
675 * or directly as just 'ki_filp' in this struct.
681 struct io_poll_iocb poll
;
682 struct io_poll_remove poll_remove
;
683 struct io_accept accept
;
685 struct io_cancel cancel
;
686 struct io_timeout timeout
;
687 struct io_timeout_rem timeout_rem
;
688 struct io_connect connect
;
689 struct io_sr_msg sr_msg
;
691 struct io_close close
;
692 struct io_files_update files_update
;
693 struct io_fadvise fadvise
;
694 struct io_madvise madvise
;
695 struct io_epoll epoll
;
696 struct io_splice splice
;
697 struct io_provide_buf pbuf
;
698 struct io_statx statx
;
699 struct io_shutdown shutdown
;
700 struct io_rename rename
;
701 struct io_unlink unlink
;
702 /* use only after cleaning per-op data, see io_clean_op() */
703 struct io_completion
compl;
706 /* opcode allocated if it needs to store data for async defer */
709 /* polled IO has completed */
715 struct io_ring_ctx
*ctx
;
718 struct task_struct
*task
;
721 struct io_kiocb
*link
;
722 struct percpu_ref
*fixed_file_refs
;
725 * 1. used with ctx->iopoll_list with reads/writes
726 * 2. to track reqs with ->files (see io_op_def::file_table)
728 struct list_head inflight_entry
;
729 struct callback_head task_work
;
730 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
731 struct hlist_node hash_node
;
732 struct async_poll
*apoll
;
733 struct io_wq_work work
;
736 struct io_defer_entry
{
737 struct list_head list
;
738 struct io_kiocb
*req
;
742 #define IO_IOPOLL_BATCH 8
744 struct io_comp_state
{
746 struct list_head list
;
747 struct io_ring_ctx
*ctx
;
750 struct io_submit_state
{
751 struct blk_plug plug
;
754 * io_kiocb alloc cache
756 void *reqs
[IO_IOPOLL_BATCH
];
757 unsigned int free_reqs
;
762 * Batch completion logic
764 struct io_comp_state comp
;
767 * File reference cache
771 unsigned int file_refs
;
772 unsigned int ios_left
;
776 /* needs req->file assigned */
777 unsigned needs_file
: 1;
778 /* don't fail if file grab fails */
779 unsigned needs_file_no_error
: 1;
780 /* hash wq insertion if file is a regular file */
781 unsigned hash_reg_file
: 1;
782 /* unbound wq insertion if file is a non-regular file */
783 unsigned unbound_nonreg_file
: 1;
784 /* opcode is not supported by this kernel */
785 unsigned not_supported
: 1;
786 /* set if opcode supports polled "wait" */
788 unsigned pollout
: 1;
789 /* op supports buffer selection */
790 unsigned buffer_select
: 1;
791 /* must always have async data allocated */
792 unsigned needs_async_data
: 1;
793 /* should block plug */
795 /* size of async data needed, if any */
796 unsigned short async_size
;
800 static const struct io_op_def io_op_defs
[] = {
801 [IORING_OP_NOP
] = {},
802 [IORING_OP_READV
] = {
804 .unbound_nonreg_file
= 1,
807 .needs_async_data
= 1,
809 .async_size
= sizeof(struct io_async_rw
),
810 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
812 [IORING_OP_WRITEV
] = {
815 .unbound_nonreg_file
= 1,
817 .needs_async_data
= 1,
819 .async_size
= sizeof(struct io_async_rw
),
820 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
823 [IORING_OP_FSYNC
] = {
825 .work_flags
= IO_WQ_WORK_BLKCG
,
827 [IORING_OP_READ_FIXED
] = {
829 .unbound_nonreg_file
= 1,
832 .async_size
= sizeof(struct io_async_rw
),
833 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
835 [IORING_OP_WRITE_FIXED
] = {
838 .unbound_nonreg_file
= 1,
841 .async_size
= sizeof(struct io_async_rw
),
842 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
|
845 [IORING_OP_POLL_ADD
] = {
847 .unbound_nonreg_file
= 1,
849 [IORING_OP_POLL_REMOVE
] = {},
850 [IORING_OP_SYNC_FILE_RANGE
] = {
852 .work_flags
= IO_WQ_WORK_BLKCG
,
854 [IORING_OP_SENDMSG
] = {
856 .unbound_nonreg_file
= 1,
858 .needs_async_data
= 1,
859 .async_size
= sizeof(struct io_async_msghdr
),
860 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
862 [IORING_OP_RECVMSG
] = {
864 .unbound_nonreg_file
= 1,
867 .needs_async_data
= 1,
868 .async_size
= sizeof(struct io_async_msghdr
),
869 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
871 [IORING_OP_TIMEOUT
] = {
872 .needs_async_data
= 1,
873 .async_size
= sizeof(struct io_timeout_data
),
874 .work_flags
= IO_WQ_WORK_MM
,
876 [IORING_OP_TIMEOUT_REMOVE
] = {
877 /* used by timeout updates' prep() */
878 .work_flags
= IO_WQ_WORK_MM
,
880 [IORING_OP_ACCEPT
] = {
882 .unbound_nonreg_file
= 1,
884 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
,
886 [IORING_OP_ASYNC_CANCEL
] = {},
887 [IORING_OP_LINK_TIMEOUT
] = {
888 .needs_async_data
= 1,
889 .async_size
= sizeof(struct io_timeout_data
),
890 .work_flags
= IO_WQ_WORK_MM
,
892 [IORING_OP_CONNECT
] = {
894 .unbound_nonreg_file
= 1,
896 .needs_async_data
= 1,
897 .async_size
= sizeof(struct io_async_connect
),
898 .work_flags
= IO_WQ_WORK_MM
,
900 [IORING_OP_FALLOCATE
] = {
902 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
904 [IORING_OP_OPENAT
] = {
905 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
|
906 IO_WQ_WORK_FS
| IO_WQ_WORK_MM
,
908 [IORING_OP_CLOSE
] = {
910 .needs_file_no_error
= 1,
911 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
,
913 [IORING_OP_FILES_UPDATE
] = {
914 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
,
916 [IORING_OP_STATX
] = {
917 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
|
918 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
922 .unbound_nonreg_file
= 1,
926 .async_size
= sizeof(struct io_async_rw
),
927 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
929 [IORING_OP_WRITE
] = {
931 .unbound_nonreg_file
= 1,
934 .async_size
= sizeof(struct io_async_rw
),
935 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
938 [IORING_OP_FADVISE
] = {
940 .work_flags
= IO_WQ_WORK_BLKCG
,
942 [IORING_OP_MADVISE
] = {
943 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
947 .unbound_nonreg_file
= 1,
949 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
953 .unbound_nonreg_file
= 1,
956 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
958 [IORING_OP_OPENAT2
] = {
959 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_FS
|
960 IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
962 [IORING_OP_EPOLL_CTL
] = {
963 .unbound_nonreg_file
= 1,
964 .work_flags
= IO_WQ_WORK_FILES
,
966 [IORING_OP_SPLICE
] = {
969 .unbound_nonreg_file
= 1,
970 .work_flags
= IO_WQ_WORK_BLKCG
,
972 [IORING_OP_PROVIDE_BUFFERS
] = {},
973 [IORING_OP_REMOVE_BUFFERS
] = {},
977 .unbound_nonreg_file
= 1,
979 [IORING_OP_SHUTDOWN
] = {
982 [IORING_OP_RENAMEAT
] = {
983 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
984 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
986 [IORING_OP_UNLINKAT
] = {
987 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
988 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
992 enum io_mem_account
{
997 static void __io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
998 struct task_struct
*task
);
1000 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
);
1001 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
1002 struct io_ring_ctx
*ctx
);
1004 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
1005 struct io_comp_state
*cs
);
1006 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
1007 static void io_put_req(struct io_kiocb
*req
);
1008 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1009 static void io_double_put_req(struct io_kiocb
*req
);
1010 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1011 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
1012 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1013 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
1014 struct io_uring_files_update
*ip
,
1016 static void __io_clean_op(struct io_kiocb
*req
);
1017 static struct file
*io_file_get(struct io_submit_state
*state
,
1018 struct io_kiocb
*req
, int fd
, bool fixed
);
1019 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
);
1020 static void io_file_put_work(struct work_struct
*work
);
1022 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
1023 struct iovec
**iovec
, struct iov_iter
*iter
,
1025 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
1026 const struct iovec
*fast_iov
,
1027 struct iov_iter
*iter
, bool force
);
1029 static struct kmem_cache
*req_cachep
;
1031 static const struct file_operations io_uring_fops
;
1033 struct sock
*io_uring_get_socket(struct file
*file
)
1035 #if defined(CONFIG_UNIX)
1036 if (file
->f_op
== &io_uring_fops
) {
1037 struct io_ring_ctx
*ctx
= file
->private_data
;
1039 return ctx
->ring_sock
->sk
;
1044 EXPORT_SYMBOL(io_uring_get_socket
);
1046 #define io_for_each_link(pos, head) \
1047 for (pos = (head); pos; pos = pos->link)
1049 static inline void io_clean_op(struct io_kiocb
*req
)
1051 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
|
1056 static inline void io_set_resource_node(struct io_kiocb
*req
)
1058 struct io_ring_ctx
*ctx
= req
->ctx
;
1060 if (!req
->fixed_file_refs
) {
1061 req
->fixed_file_refs
= &ctx
->file_data
->node
->refs
;
1062 percpu_ref_get(req
->fixed_file_refs
);
1066 static bool io_match_task(struct io_kiocb
*head
,
1067 struct task_struct
*task
,
1068 struct files_struct
*files
)
1070 struct io_kiocb
*req
;
1072 if (task
&& head
->task
!= task
)
1077 io_for_each_link(req
, head
) {
1078 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
1079 (req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1080 req
->work
.identity
->files
== files
)
1086 static void io_sq_thread_drop_mm_files(void)
1088 struct files_struct
*files
= current
->files
;
1089 struct mm_struct
*mm
= current
->mm
;
1092 kthread_unuse_mm(mm
);
1097 struct nsproxy
*nsproxy
= current
->nsproxy
;
1100 current
->files
= NULL
;
1101 current
->nsproxy
= NULL
;
1102 task_unlock(current
);
1103 put_files_struct(files
);
1104 put_nsproxy(nsproxy
);
1108 static int __io_sq_thread_acquire_files(struct io_ring_ctx
*ctx
)
1110 if (current
->flags
& PF_EXITING
)
1113 if (!current
->files
) {
1114 struct files_struct
*files
;
1115 struct nsproxy
*nsproxy
;
1117 task_lock(ctx
->sqo_task
);
1118 files
= ctx
->sqo_task
->files
;
1120 task_unlock(ctx
->sqo_task
);
1123 atomic_inc(&files
->count
);
1124 get_nsproxy(ctx
->sqo_task
->nsproxy
);
1125 nsproxy
= ctx
->sqo_task
->nsproxy
;
1126 task_unlock(ctx
->sqo_task
);
1129 current
->files
= files
;
1130 current
->nsproxy
= nsproxy
;
1131 task_unlock(current
);
1136 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1138 struct mm_struct
*mm
;
1140 if (current
->flags
& PF_EXITING
)
1145 /* Should never happen */
1146 if (unlikely(!(ctx
->flags
& IORING_SETUP_SQPOLL
)))
1149 task_lock(ctx
->sqo_task
);
1150 mm
= ctx
->sqo_task
->mm
;
1151 if (unlikely(!mm
|| !mmget_not_zero(mm
)))
1153 task_unlock(ctx
->sqo_task
);
1163 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx
*ctx
,
1164 struct io_kiocb
*req
)
1166 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1169 if (def
->work_flags
& IO_WQ_WORK_MM
) {
1170 ret
= __io_sq_thread_acquire_mm(ctx
);
1175 if (def
->needs_file
|| (def
->work_flags
& IO_WQ_WORK_FILES
)) {
1176 ret
= __io_sq_thread_acquire_files(ctx
);
1184 static void io_sq_thread_associate_blkcg(struct io_ring_ctx
*ctx
,
1185 struct cgroup_subsys_state
**cur_css
)
1188 #ifdef CONFIG_BLK_CGROUP
1189 /* puts the old one when swapping */
1190 if (*cur_css
!= ctx
->sqo_blkcg_css
) {
1191 kthread_associate_blkcg(ctx
->sqo_blkcg_css
);
1192 *cur_css
= ctx
->sqo_blkcg_css
;
1197 static void io_sq_thread_unassociate_blkcg(void)
1199 #ifdef CONFIG_BLK_CGROUP
1200 kthread_associate_blkcg(NULL
);
1204 static inline void req_set_fail_links(struct io_kiocb
*req
)
1206 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1207 req
->flags
|= REQ_F_FAIL_LINK
;
1211 * None of these are dereferenced, they are simply used to check if any of
1212 * them have changed. If we're under current and check they are still the
1213 * same, we're fine to grab references to them for actual out-of-line use.
1215 static void io_init_identity(struct io_identity
*id
)
1217 id
->files
= current
->files
;
1218 id
->mm
= current
->mm
;
1219 #ifdef CONFIG_BLK_CGROUP
1221 id
->blkcg_css
= blkcg_css();
1224 id
->creds
= current_cred();
1225 id
->nsproxy
= current
->nsproxy
;
1226 id
->fs
= current
->fs
;
1227 id
->fsize
= rlimit(RLIMIT_FSIZE
);
1229 id
->loginuid
= current
->loginuid
;
1230 id
->sessionid
= current
->sessionid
;
1232 refcount_set(&id
->count
, 1);
1235 static inline void __io_req_init_async(struct io_kiocb
*req
)
1237 memset(&req
->work
, 0, sizeof(req
->work
));
1238 req
->flags
|= REQ_F_WORK_INITIALIZED
;
1242 * Note: must call io_req_init_async() for the first time you
1243 * touch any members of io_wq_work.
1245 static inline void io_req_init_async(struct io_kiocb
*req
)
1247 struct io_uring_task
*tctx
= current
->io_uring
;
1249 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
1252 __io_req_init_async(req
);
1254 /* Grab a ref if this isn't our static identity */
1255 req
->work
.identity
= tctx
->identity
;
1256 if (tctx
->identity
!= &tctx
->__identity
)
1257 refcount_inc(&req
->work
.identity
->count
);
1260 static inline bool io_async_submit(struct io_ring_ctx
*ctx
)
1262 return ctx
->flags
& IORING_SETUP_SQPOLL
;
1265 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1267 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1269 complete(&ctx
->ref_comp
);
1272 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1274 return !req
->timeout
.off
;
1277 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1279 struct io_ring_ctx
*ctx
;
1282 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1286 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
1287 if (!ctx
->fallback_req
)
1291 * Use 5 bits less than the max cq entries, that should give us around
1292 * 32 entries per hash list if totally full and uniformly spread.
1294 hash_bits
= ilog2(p
->cq_entries
);
1298 ctx
->cancel_hash_bits
= hash_bits
;
1299 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1301 if (!ctx
->cancel_hash
)
1303 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1305 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1306 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1309 ctx
->flags
= p
->flags
;
1310 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1311 INIT_LIST_HEAD(&ctx
->sqd_list
);
1312 init_waitqueue_head(&ctx
->cq_wait
);
1313 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1314 init_completion(&ctx
->ref_comp
);
1315 init_completion(&ctx
->sq_thread_comp
);
1316 idr_init(&ctx
->io_buffer_idr
);
1317 idr_init(&ctx
->personality_idr
);
1318 mutex_init(&ctx
->uring_lock
);
1319 init_waitqueue_head(&ctx
->wait
);
1320 spin_lock_init(&ctx
->completion_lock
);
1321 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1322 INIT_LIST_HEAD(&ctx
->defer_list
);
1323 INIT_LIST_HEAD(&ctx
->timeout_list
);
1324 spin_lock_init(&ctx
->inflight_lock
);
1325 INIT_LIST_HEAD(&ctx
->inflight_list
);
1326 INIT_DELAYED_WORK(&ctx
->file_put_work
, io_file_put_work
);
1327 init_llist_head(&ctx
->file_put_llist
);
1330 if (ctx
->fallback_req
)
1331 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1332 kfree(ctx
->cancel_hash
);
1337 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1339 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1340 struct io_ring_ctx
*ctx
= req
->ctx
;
1342 return seq
!= ctx
->cached_cq_tail
1343 + READ_ONCE(ctx
->cached_cq_overflow
);
1349 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1351 struct io_rings
*rings
= ctx
->rings
;
1353 /* order cqe stores with ring update */
1354 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1357 static void io_put_identity(struct io_uring_task
*tctx
, struct io_kiocb
*req
)
1359 if (req
->work
.identity
== &tctx
->__identity
)
1361 if (refcount_dec_and_test(&req
->work
.identity
->count
))
1362 kfree(req
->work
.identity
);
1365 static void io_req_clean_work(struct io_kiocb
*req
)
1367 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1370 req
->flags
&= ~REQ_F_WORK_INITIALIZED
;
1372 if (req
->work
.flags
& IO_WQ_WORK_MM
) {
1373 mmdrop(req
->work
.identity
->mm
);
1374 req
->work
.flags
&= ~IO_WQ_WORK_MM
;
1376 #ifdef CONFIG_BLK_CGROUP
1377 if (req
->work
.flags
& IO_WQ_WORK_BLKCG
) {
1378 css_put(req
->work
.identity
->blkcg_css
);
1379 req
->work
.flags
&= ~IO_WQ_WORK_BLKCG
;
1382 if (req
->work
.flags
& IO_WQ_WORK_CREDS
) {
1383 put_cred(req
->work
.identity
->creds
);
1384 req
->work
.flags
&= ~IO_WQ_WORK_CREDS
;
1386 if (req
->work
.flags
& IO_WQ_WORK_FS
) {
1387 struct fs_struct
*fs
= req
->work
.identity
->fs
;
1389 spin_lock(&req
->work
.identity
->fs
->lock
);
1392 spin_unlock(&req
->work
.identity
->fs
->lock
);
1395 req
->work
.flags
&= ~IO_WQ_WORK_FS
;
1398 io_put_identity(req
->task
->io_uring
, req
);
1402 * Create a private copy of io_identity, since some fields don't match
1403 * the current context.
1405 static bool io_identity_cow(struct io_kiocb
*req
)
1407 struct io_uring_task
*tctx
= current
->io_uring
;
1408 const struct cred
*creds
= NULL
;
1409 struct io_identity
*id
;
1411 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1412 creds
= req
->work
.identity
->creds
;
1414 id
= kmemdup(req
->work
.identity
, sizeof(*id
), GFP_KERNEL
);
1415 if (unlikely(!id
)) {
1416 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1421 * We can safely just re-init the creds we copied Either the field
1422 * matches the current one, or we haven't grabbed it yet. The only
1423 * exception is ->creds, through registered personalities, so handle
1424 * that one separately.
1426 io_init_identity(id
);
1430 /* add one for this request */
1431 refcount_inc(&id
->count
);
1433 /* drop tctx and req identity references, if needed */
1434 if (tctx
->identity
!= &tctx
->__identity
&&
1435 refcount_dec_and_test(&tctx
->identity
->count
))
1436 kfree(tctx
->identity
);
1437 if (req
->work
.identity
!= &tctx
->__identity
&&
1438 refcount_dec_and_test(&req
->work
.identity
->count
))
1439 kfree(req
->work
.identity
);
1441 req
->work
.identity
= id
;
1442 tctx
->identity
= id
;
1446 static bool io_grab_identity(struct io_kiocb
*req
)
1448 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1449 struct io_identity
*id
= req
->work
.identity
;
1450 struct io_ring_ctx
*ctx
= req
->ctx
;
1452 if (def
->work_flags
& IO_WQ_WORK_FSIZE
) {
1453 if (id
->fsize
!= rlimit(RLIMIT_FSIZE
))
1455 req
->work
.flags
|= IO_WQ_WORK_FSIZE
;
1457 #ifdef CONFIG_BLK_CGROUP
1458 if (!(req
->work
.flags
& IO_WQ_WORK_BLKCG
) &&
1459 (def
->work_flags
& IO_WQ_WORK_BLKCG
)) {
1461 if (id
->blkcg_css
!= blkcg_css()) {
1466 * This should be rare, either the cgroup is dying or the task
1467 * is moving cgroups. Just punt to root for the handful of ios.
1469 if (css_tryget_online(id
->blkcg_css
))
1470 req
->work
.flags
|= IO_WQ_WORK_BLKCG
;
1474 if (!(req
->work
.flags
& IO_WQ_WORK_CREDS
)) {
1475 if (id
->creds
!= current_cred())
1477 get_cred(id
->creds
);
1478 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
1481 if (!uid_eq(current
->loginuid
, id
->loginuid
) ||
1482 current
->sessionid
!= id
->sessionid
)
1485 if (!(req
->work
.flags
& IO_WQ_WORK_FS
) &&
1486 (def
->work_flags
& IO_WQ_WORK_FS
)) {
1487 if (current
->fs
!= id
->fs
)
1489 spin_lock(&id
->fs
->lock
);
1490 if (!id
->fs
->in_exec
) {
1492 req
->work
.flags
|= IO_WQ_WORK_FS
;
1494 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1496 spin_unlock(¤t
->fs
->lock
);
1498 if (!(req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1499 (def
->work_flags
& IO_WQ_WORK_FILES
) &&
1500 !(req
->flags
& REQ_F_NO_FILE_TABLE
)) {
1501 if (id
->files
!= current
->files
||
1502 id
->nsproxy
!= current
->nsproxy
)
1504 atomic_inc(&id
->files
->count
);
1505 get_nsproxy(id
->nsproxy
);
1506 req
->flags
|= REQ_F_INFLIGHT
;
1508 spin_lock_irq(&ctx
->inflight_lock
);
1509 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1510 spin_unlock_irq(&ctx
->inflight_lock
);
1511 req
->work
.flags
|= IO_WQ_WORK_FILES
;
1513 if (!(req
->work
.flags
& IO_WQ_WORK_MM
) &&
1514 (def
->work_flags
& IO_WQ_WORK_MM
)) {
1515 if (id
->mm
!= current
->mm
)
1518 req
->work
.flags
|= IO_WQ_WORK_MM
;
1524 static void io_prep_async_work(struct io_kiocb
*req
)
1526 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1527 struct io_ring_ctx
*ctx
= req
->ctx
;
1529 io_req_init_async(req
);
1531 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1532 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1534 if (req
->flags
& REQ_F_ISREG
) {
1535 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1536 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1538 if (def
->unbound_nonreg_file
)
1539 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1542 /* if we fail grabbing identity, we must COW, regrab, and retry */
1543 if (io_grab_identity(req
))
1546 if (!io_identity_cow(req
))
1549 /* can't fail at this point */
1550 if (!io_grab_identity(req
))
1554 static void io_prep_async_link(struct io_kiocb
*req
)
1556 struct io_kiocb
*cur
;
1558 io_for_each_link(cur
, req
)
1559 io_prep_async_work(cur
);
1562 static struct io_kiocb
*__io_queue_async_work(struct io_kiocb
*req
)
1564 struct io_ring_ctx
*ctx
= req
->ctx
;
1565 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1567 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1568 &req
->work
, req
->flags
);
1569 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1573 static void io_queue_async_work(struct io_kiocb
*req
)
1575 struct io_kiocb
*link
;
1577 /* init ->work of the whole link before punting */
1578 io_prep_async_link(req
);
1579 link
= __io_queue_async_work(req
);
1582 io_queue_linked_timeout(link
);
1585 static void io_kill_timeout(struct io_kiocb
*req
)
1587 struct io_timeout_data
*io
= req
->async_data
;
1590 ret
= hrtimer_try_to_cancel(&io
->timer
);
1592 atomic_set(&req
->ctx
->cq_timeouts
,
1593 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1594 list_del_init(&req
->timeout
.list
);
1595 io_cqring_fill_event(req
, 0);
1596 io_put_req_deferred(req
, 1);
1601 * Returns true if we found and killed one or more timeouts
1603 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
1604 struct files_struct
*files
)
1606 struct io_kiocb
*req
, *tmp
;
1609 spin_lock_irq(&ctx
->completion_lock
);
1610 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1611 if (io_match_task(req
, tsk
, files
)) {
1612 io_kill_timeout(req
);
1616 spin_unlock_irq(&ctx
->completion_lock
);
1617 return canceled
!= 0;
1620 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1623 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1624 struct io_defer_entry
, list
);
1625 struct io_kiocb
*link
;
1627 if (req_need_defer(de
->req
, de
->seq
))
1629 list_del_init(&de
->list
);
1630 /* punt-init is done before queueing for defer */
1631 link
= __io_queue_async_work(de
->req
);
1633 __io_queue_linked_timeout(link
);
1634 /* drop submission reference */
1635 io_put_req_deferred(link
, 1);
1638 } while (!list_empty(&ctx
->defer_list
));
1641 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1645 if (list_empty(&ctx
->timeout_list
))
1648 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1651 u32 events_needed
, events_got
;
1652 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1653 struct io_kiocb
, timeout
.list
);
1655 if (io_is_timeout_noseq(req
))
1659 * Since seq can easily wrap around over time, subtract
1660 * the last seq at which timeouts were flushed before comparing.
1661 * Assuming not more than 2^31-1 events have happened since,
1662 * these subtractions won't have wrapped, so we can check if
1663 * target is in [last_seq, current_seq] by comparing the two.
1665 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1666 events_got
= seq
- ctx
->cq_last_tm_flush
;
1667 if (events_got
< events_needed
)
1670 list_del_init(&req
->timeout
.list
);
1671 io_kill_timeout(req
);
1672 } while (!list_empty(&ctx
->timeout_list
));
1674 ctx
->cq_last_tm_flush
= seq
;
1677 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1679 io_flush_timeouts(ctx
);
1680 __io_commit_cqring(ctx
);
1682 if (unlikely(!list_empty(&ctx
->defer_list
)))
1683 __io_queue_deferred(ctx
);
1686 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1688 struct io_rings
*r
= ctx
->rings
;
1690 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1693 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1695 struct io_rings
*rings
= ctx
->rings
;
1698 tail
= ctx
->cached_cq_tail
;
1700 * writes to the cq entry need to come after reading head; the
1701 * control dependency is enough as we're using WRITE_ONCE to
1704 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1707 ctx
->cached_cq_tail
++;
1708 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1711 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1715 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1717 if (!ctx
->eventfd_async
)
1719 return io_wq_current_is_worker();
1722 static inline unsigned __io_cqring_events(struct io_ring_ctx
*ctx
)
1724 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1727 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1729 /* see waitqueue_active() comment */
1732 if (waitqueue_active(&ctx
->wait
))
1733 wake_up(&ctx
->wait
);
1734 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1735 wake_up(&ctx
->sq_data
->wait
);
1736 if (io_should_trigger_evfd(ctx
))
1737 eventfd_signal(ctx
->cq_ev_fd
, 1);
1738 if (waitqueue_active(&ctx
->cq_wait
)) {
1739 wake_up_interruptible(&ctx
->cq_wait
);
1740 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1744 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1746 /* see waitqueue_active() comment */
1749 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1750 if (waitqueue_active(&ctx
->wait
))
1751 wake_up(&ctx
->wait
);
1753 if (io_should_trigger_evfd(ctx
))
1754 eventfd_signal(ctx
->cq_ev_fd
, 1);
1755 if (waitqueue_active(&ctx
->cq_wait
)) {
1756 wake_up_interruptible(&ctx
->cq_wait
);
1757 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1761 /* Returns true if there are no backlogged entries after the flush */
1762 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1763 struct task_struct
*tsk
,
1764 struct files_struct
*files
)
1766 struct io_rings
*rings
= ctx
->rings
;
1767 struct io_kiocb
*req
, *tmp
;
1768 struct io_uring_cqe
*cqe
;
1769 unsigned long flags
;
1773 if (!force
&& __io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1776 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1777 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1778 if (!io_match_task(req
, tsk
, files
))
1781 cqe
= io_get_cqring(ctx
);
1785 list_move(&req
->compl.list
, &list
);
1787 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1788 WRITE_ONCE(cqe
->res
, req
->result
);
1789 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1791 ctx
->cached_cq_overflow
++;
1792 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1793 ctx
->cached_cq_overflow
);
1797 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1799 clear_bit(0, &ctx
->sq_check_overflow
);
1800 clear_bit(0, &ctx
->cq_check_overflow
);
1801 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1804 io_commit_cqring(ctx
);
1805 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1806 io_cqring_ev_posted(ctx
);
1808 while (!list_empty(&list
)) {
1809 req
= list_first_entry(&list
, struct io_kiocb
, compl.list
);
1810 list_del(&req
->compl.list
);
1817 static void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1818 struct task_struct
*tsk
,
1819 struct files_struct
*files
)
1821 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1822 /* iopoll syncs against uring_lock, not completion_lock */
1823 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1824 mutex_lock(&ctx
->uring_lock
);
1825 __io_cqring_overflow_flush(ctx
, force
, tsk
, files
);
1826 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1827 mutex_unlock(&ctx
->uring_lock
);
1831 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1833 struct io_ring_ctx
*ctx
= req
->ctx
;
1834 struct io_uring_cqe
*cqe
;
1836 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1839 * If we can't get a cq entry, userspace overflowed the
1840 * submission (by quite a lot). Increment the overflow count in
1843 cqe
= io_get_cqring(ctx
);
1845 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1846 WRITE_ONCE(cqe
->res
, res
);
1847 WRITE_ONCE(cqe
->flags
, cflags
);
1848 } else if (ctx
->cq_overflow_flushed
||
1849 atomic_read(&req
->task
->io_uring
->in_idle
)) {
1851 * If we're in ring overflow flush mode, or in task cancel mode,
1852 * then we cannot store the request for later flushing, we need
1853 * to drop it on the floor.
1855 ctx
->cached_cq_overflow
++;
1856 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1858 if (list_empty(&ctx
->cq_overflow_list
)) {
1859 set_bit(0, &ctx
->sq_check_overflow
);
1860 set_bit(0, &ctx
->cq_check_overflow
);
1861 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1865 req
->compl.cflags
= cflags
;
1866 refcount_inc(&req
->refs
);
1867 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1871 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1873 __io_cqring_fill_event(req
, res
, 0);
1876 static void io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1878 struct io_ring_ctx
*ctx
= req
->ctx
;
1879 unsigned long flags
;
1881 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1882 __io_cqring_fill_event(req
, res
, cflags
);
1883 io_commit_cqring(ctx
);
1884 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1886 io_cqring_ev_posted(ctx
);
1889 static void io_submit_flush_completions(struct io_comp_state
*cs
)
1891 struct io_ring_ctx
*ctx
= cs
->ctx
;
1893 spin_lock_irq(&ctx
->completion_lock
);
1894 while (!list_empty(&cs
->list
)) {
1895 struct io_kiocb
*req
;
1897 req
= list_first_entry(&cs
->list
, struct io_kiocb
, compl.list
);
1898 list_del(&req
->compl.list
);
1899 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
1902 * io_free_req() doesn't care about completion_lock unless one
1903 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1904 * because of a potential deadlock with req->work.fs->lock
1906 if (req
->flags
& (REQ_F_FAIL_LINK
|REQ_F_LINK_TIMEOUT
1907 |REQ_F_WORK_INITIALIZED
)) {
1908 spin_unlock_irq(&ctx
->completion_lock
);
1910 spin_lock_irq(&ctx
->completion_lock
);
1915 io_commit_cqring(ctx
);
1916 spin_unlock_irq(&ctx
->completion_lock
);
1918 io_cqring_ev_posted(ctx
);
1922 static void __io_req_complete(struct io_kiocb
*req
, long res
, unsigned cflags
,
1923 struct io_comp_state
*cs
)
1926 io_cqring_add_event(req
, res
, cflags
);
1931 req
->compl.cflags
= cflags
;
1932 list_add_tail(&req
->compl.list
, &cs
->list
);
1934 io_submit_flush_completions(cs
);
1938 static void io_req_complete(struct io_kiocb
*req
, long res
)
1940 __io_req_complete(req
, res
, 0, NULL
);
1943 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1945 return req
== (struct io_kiocb
*)
1946 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1949 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1951 struct io_kiocb
*req
;
1953 req
= ctx
->fallback_req
;
1954 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1960 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1961 struct io_submit_state
*state
)
1963 if (!state
->free_reqs
) {
1964 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1968 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1969 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
1972 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1973 * retry single alloc to be on the safe side.
1975 if (unlikely(ret
<= 0)) {
1976 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1977 if (!state
->reqs
[0])
1981 state
->free_reqs
= ret
;
1985 return state
->reqs
[state
->free_reqs
];
1987 return io_get_fallback_req(ctx
);
1990 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
1997 static void io_dismantle_req(struct io_kiocb
*req
)
2001 if (req
->async_data
)
2002 kfree(req
->async_data
);
2004 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
2005 if (req
->fixed_file_refs
)
2006 percpu_ref_put(req
->fixed_file_refs
);
2007 io_req_clean_work(req
);
2010 static void __io_free_req(struct io_kiocb
*req
)
2012 struct io_uring_task
*tctx
= req
->task
->io_uring
;
2013 struct io_ring_ctx
*ctx
= req
->ctx
;
2015 io_dismantle_req(req
);
2017 percpu_counter_dec(&tctx
->inflight
);
2018 if (atomic_read(&tctx
->in_idle
))
2019 wake_up(&tctx
->wait
);
2020 put_task_struct(req
->task
);
2022 if (likely(!io_is_fallback_req(req
)))
2023 kmem_cache_free(req_cachep
, req
);
2025 clear_bit_unlock(0, (unsigned long *) &ctx
->fallback_req
);
2026 percpu_ref_put(&ctx
->refs
);
2029 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2031 struct io_kiocb
*nxt
= req
->link
;
2033 req
->link
= nxt
->link
;
2037 static void io_kill_linked_timeout(struct io_kiocb
*req
)
2039 struct io_ring_ctx
*ctx
= req
->ctx
;
2040 struct io_kiocb
*link
;
2041 bool cancelled
= false;
2042 unsigned long flags
;
2044 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2048 * Can happen if a linked timeout fired and link had been like
2049 * req -> link t-out -> link t-out [-> ...]
2051 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
2052 struct io_timeout_data
*io
= link
->async_data
;
2055 io_remove_next_linked(req
);
2056 link
->timeout
.head
= NULL
;
2057 ret
= hrtimer_try_to_cancel(&io
->timer
);
2059 io_cqring_fill_event(link
, -ECANCELED
);
2060 io_commit_cqring(ctx
);
2064 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
2065 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2068 io_cqring_ev_posted(ctx
);
2074 static void io_fail_links(struct io_kiocb
*req
)
2076 struct io_kiocb
*link
, *nxt
;
2077 struct io_ring_ctx
*ctx
= req
->ctx
;
2078 unsigned long flags
;
2080 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2088 trace_io_uring_fail_link(req
, link
);
2089 io_cqring_fill_event(link
, -ECANCELED
);
2092 * It's ok to free under spinlock as they're not linked anymore,
2093 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2096 if (link
->flags
& REQ_F_WORK_INITIALIZED
)
2097 io_put_req_deferred(link
, 2);
2099 io_double_put_req(link
);
2102 io_commit_cqring(ctx
);
2103 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2105 io_cqring_ev_posted(ctx
);
2108 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2110 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
2111 io_kill_linked_timeout(req
);
2114 * If LINK is set, we have dependent requests in this chain. If we
2115 * didn't fail this request, queue the first one up, moving any other
2116 * dependencies to the next request. In case of failure, fail the rest
2119 if (likely(!(req
->flags
& REQ_F_FAIL_LINK
))) {
2120 struct io_kiocb
*nxt
= req
->link
;
2129 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2131 if (likely(!(req
->link
) && !(req
->flags
& REQ_F_LINK_TIMEOUT
)))
2133 return __io_req_find_next(req
);
2136 static int io_req_task_work_add(struct io_kiocb
*req
)
2138 struct task_struct
*tsk
= req
->task
;
2139 struct io_ring_ctx
*ctx
= req
->ctx
;
2140 enum task_work_notify_mode notify
;
2143 if (tsk
->flags
& PF_EXITING
)
2147 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2148 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2149 * processing task_work. There's no reliable way to tell if TWA_RESUME
2153 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
))
2154 notify
= TWA_SIGNAL
;
2156 ret
= task_work_add(tsk
, &req
->task_work
, notify
);
2158 wake_up_process(tsk
);
2163 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
2165 struct io_ring_ctx
*ctx
= req
->ctx
;
2167 spin_lock_irq(&ctx
->completion_lock
);
2168 io_cqring_fill_event(req
, error
);
2169 io_commit_cqring(ctx
);
2170 spin_unlock_irq(&ctx
->completion_lock
);
2172 io_cqring_ev_posted(ctx
);
2173 req_set_fail_links(req
);
2174 io_double_put_req(req
);
2177 static void io_req_task_cancel(struct callback_head
*cb
)
2179 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2180 struct io_ring_ctx
*ctx
= req
->ctx
;
2182 __io_req_task_cancel(req
, -ECANCELED
);
2183 percpu_ref_put(&ctx
->refs
);
2186 static void __io_req_task_submit(struct io_kiocb
*req
)
2188 struct io_ring_ctx
*ctx
= req
->ctx
;
2190 mutex_lock(&ctx
->uring_lock
);
2191 if (!ctx
->sqo_dead
&&
2192 !__io_sq_thread_acquire_mm(ctx
) &&
2193 !__io_sq_thread_acquire_files(ctx
))
2194 __io_queue_sqe(req
, NULL
);
2196 __io_req_task_cancel(req
, -EFAULT
);
2197 mutex_unlock(&ctx
->uring_lock
);
2200 static void io_req_task_submit(struct callback_head
*cb
)
2202 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2203 struct io_ring_ctx
*ctx
= req
->ctx
;
2205 __io_req_task_submit(req
);
2206 percpu_ref_put(&ctx
->refs
);
2209 static void io_req_task_queue(struct io_kiocb
*req
)
2213 init_task_work(&req
->task_work
, io_req_task_submit
);
2214 percpu_ref_get(&req
->ctx
->refs
);
2216 ret
= io_req_task_work_add(req
);
2217 if (unlikely(ret
)) {
2218 struct task_struct
*tsk
;
2220 init_task_work(&req
->task_work
, io_req_task_cancel
);
2221 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2222 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2223 wake_up_process(tsk
);
2227 static inline void io_queue_next(struct io_kiocb
*req
)
2229 struct io_kiocb
*nxt
= io_req_find_next(req
);
2232 io_req_task_queue(nxt
);
2235 static void io_free_req(struct io_kiocb
*req
)
2242 void *reqs
[IO_IOPOLL_BATCH
];
2245 struct task_struct
*task
;
2249 static inline void io_init_req_batch(struct req_batch
*rb
)
2256 static void __io_req_free_batch_flush(struct io_ring_ctx
*ctx
,
2257 struct req_batch
*rb
)
2259 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
2260 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
2264 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2265 struct req_batch
*rb
)
2268 __io_req_free_batch_flush(ctx
, rb
);
2270 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2272 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2273 if (atomic_read(&tctx
->in_idle
))
2274 wake_up(&tctx
->wait
);
2275 put_task_struct_many(rb
->task
, rb
->task_refs
);
2280 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
)
2282 if (unlikely(io_is_fallback_req(req
))) {
2288 if (req
->task
!= rb
->task
) {
2290 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2292 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2293 if (atomic_read(&tctx
->in_idle
))
2294 wake_up(&tctx
->wait
);
2295 put_task_struct_many(rb
->task
, rb
->task_refs
);
2297 rb
->task
= req
->task
;
2302 io_dismantle_req(req
);
2303 rb
->reqs
[rb
->to_free
++] = req
;
2304 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
2305 __io_req_free_batch_flush(req
->ctx
, rb
);
2309 * Drop reference to request, return next in chain (if there is one) if this
2310 * was the last reference to this request.
2312 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2314 struct io_kiocb
*nxt
= NULL
;
2316 if (refcount_dec_and_test(&req
->refs
)) {
2317 nxt
= io_req_find_next(req
);
2323 static void io_put_req(struct io_kiocb
*req
)
2325 if (refcount_dec_and_test(&req
->refs
))
2329 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2331 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2336 static void io_free_req_deferred(struct io_kiocb
*req
)
2340 init_task_work(&req
->task_work
, io_put_req_deferred_cb
);
2341 ret
= io_req_task_work_add(req
);
2342 if (unlikely(ret
)) {
2343 struct task_struct
*tsk
;
2345 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2346 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2347 wake_up_process(tsk
);
2351 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2353 if (refcount_sub_and_test(refs
, &req
->refs
))
2354 io_free_req_deferred(req
);
2357 static struct io_wq_work
*io_steal_work(struct io_kiocb
*req
)
2359 struct io_kiocb
*nxt
;
2362 * A ref is owned by io-wq in which context we're. So, if that's the
2363 * last one, it's safe to steal next work. False negatives are Ok,
2364 * it just will be re-punted async in io_put_work()
2366 if (refcount_read(&req
->refs
) != 1)
2369 nxt
= io_req_find_next(req
);
2370 return nxt
? &nxt
->work
: NULL
;
2373 static void io_double_put_req(struct io_kiocb
*req
)
2375 /* drop both submit and complete references */
2376 if (refcount_sub_and_test(2, &req
->refs
))
2380 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2382 /* See comment at the top of this file */
2384 return __io_cqring_events(ctx
);
2387 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2389 struct io_rings
*rings
= ctx
->rings
;
2391 /* make sure SQ entry isn't read before tail */
2392 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2395 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2397 unsigned int cflags
;
2399 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2400 cflags
|= IORING_CQE_F_BUFFER
;
2401 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2406 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2408 struct io_buffer
*kbuf
;
2410 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2411 return io_put_kbuf(req
, kbuf
);
2414 static inline bool io_run_task_work(void)
2417 * Not safe to run on exiting task, and the task_work handling will
2418 * not add work to such a task.
2420 if (unlikely(current
->flags
& PF_EXITING
))
2422 if (current
->task_works
) {
2423 __set_current_state(TASK_RUNNING
);
2431 static void io_iopoll_queue(struct list_head
*again
)
2433 struct io_kiocb
*req
;
2436 req
= list_first_entry(again
, struct io_kiocb
, inflight_entry
);
2437 list_del(&req
->inflight_entry
);
2438 __io_complete_rw(req
, -EAGAIN
, 0, NULL
);
2439 } while (!list_empty(again
));
2443 * Find and free completed poll iocbs
2445 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2446 struct list_head
*done
)
2448 struct req_batch rb
;
2449 struct io_kiocb
*req
;
2452 /* order with ->result store in io_complete_rw_iopoll() */
2455 io_init_req_batch(&rb
);
2456 while (!list_empty(done
)) {
2459 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2460 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2462 req
->iopoll_completed
= 0;
2463 list_move_tail(&req
->inflight_entry
, &again
);
2466 list_del(&req
->inflight_entry
);
2468 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2469 cflags
= io_put_rw_kbuf(req
);
2471 __io_cqring_fill_event(req
, req
->result
, cflags
);
2474 if (refcount_dec_and_test(&req
->refs
))
2475 io_req_free_batch(&rb
, req
);
2478 io_commit_cqring(ctx
);
2479 io_cqring_ev_posted_iopoll(ctx
);
2480 io_req_free_batch_finish(ctx
, &rb
);
2482 if (!list_empty(&again
))
2483 io_iopoll_queue(&again
);
2486 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2489 struct io_kiocb
*req
, *tmp
;
2495 * Only spin for completions if we don't have multiple devices hanging
2496 * off our complete list, and we're under the requested amount.
2498 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2501 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2502 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2505 * Move completed and retryable entries to our local lists.
2506 * If we find a request that requires polling, break out
2507 * and complete those lists first, if we have entries there.
2509 if (READ_ONCE(req
->iopoll_completed
)) {
2510 list_move_tail(&req
->inflight_entry
, &done
);
2513 if (!list_empty(&done
))
2516 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2520 /* iopoll may have completed current req */
2521 if (READ_ONCE(req
->iopoll_completed
))
2522 list_move_tail(&req
->inflight_entry
, &done
);
2529 if (!list_empty(&done
))
2530 io_iopoll_complete(ctx
, nr_events
, &done
);
2536 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2537 * non-spinning poll check - we'll still enter the driver poll loop, but only
2538 * as a non-spinning completion check.
2540 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2543 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2546 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2549 if (*nr_events
>= min
)
2557 * We can't just wait for polled events to come to us, we have to actively
2558 * find and complete them.
2560 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2562 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2565 mutex_lock(&ctx
->uring_lock
);
2566 while (!list_empty(&ctx
->iopoll_list
)) {
2567 unsigned int nr_events
= 0;
2569 io_do_iopoll(ctx
, &nr_events
, 0);
2571 /* let it sleep and repeat later if can't complete a request */
2575 * Ensure we allow local-to-the-cpu processing to take place,
2576 * in this case we need to ensure that we reap all events.
2577 * Also let task_work, etc. to progress by releasing the mutex
2579 if (need_resched()) {
2580 mutex_unlock(&ctx
->uring_lock
);
2582 mutex_lock(&ctx
->uring_lock
);
2585 mutex_unlock(&ctx
->uring_lock
);
2588 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2590 unsigned int nr_events
= 0;
2591 int iters
= 0, ret
= 0;
2594 * We disallow the app entering submit/complete with polling, but we
2595 * still need to lock the ring to prevent racing with polled issue
2596 * that got punted to a workqueue.
2598 mutex_lock(&ctx
->uring_lock
);
2601 * Don't enter poll loop if we already have events pending.
2602 * If we do, we can potentially be spinning for commands that
2603 * already triggered a CQE (eg in error).
2605 if (test_bit(0, &ctx
->cq_check_overflow
))
2606 __io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2607 if (io_cqring_events(ctx
))
2611 * If a submit got punted to a workqueue, we can have the
2612 * application entering polling for a command before it gets
2613 * issued. That app will hold the uring_lock for the duration
2614 * of the poll right here, so we need to take a breather every
2615 * now and then to ensure that the issue has a chance to add
2616 * the poll to the issued list. Otherwise we can spin here
2617 * forever, while the workqueue is stuck trying to acquire the
2620 if (!(++iters
& 7)) {
2621 mutex_unlock(&ctx
->uring_lock
);
2623 mutex_lock(&ctx
->uring_lock
);
2626 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2630 } while (min
&& !nr_events
&& !need_resched());
2632 mutex_unlock(&ctx
->uring_lock
);
2636 static void kiocb_end_write(struct io_kiocb
*req
)
2639 * Tell lockdep we inherited freeze protection from submission
2642 if (req
->flags
& REQ_F_ISREG
) {
2643 struct inode
*inode
= file_inode(req
->file
);
2645 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2647 file_end_write(req
->file
);
2650 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
,
2651 struct io_comp_state
*cs
)
2653 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2656 if (kiocb
->ki_flags
& IOCB_WRITE
)
2657 kiocb_end_write(req
);
2659 if (res
!= req
->result
)
2660 req_set_fail_links(req
);
2661 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2662 cflags
= io_put_rw_kbuf(req
);
2663 __io_req_complete(req
, res
, cflags
, cs
);
2667 static bool io_resubmit_prep(struct io_kiocb
*req
, int error
)
2669 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2670 ssize_t ret
= -ECANCELED
;
2671 struct iov_iter iter
;
2679 switch (req
->opcode
) {
2680 case IORING_OP_READV
:
2681 case IORING_OP_READ_FIXED
:
2682 case IORING_OP_READ
:
2685 case IORING_OP_WRITEV
:
2686 case IORING_OP_WRITE_FIXED
:
2687 case IORING_OP_WRITE
:
2691 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2696 if (!req
->async_data
) {
2697 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2700 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2708 req_set_fail_links(req
);
2713 static bool io_rw_reissue(struct io_kiocb
*req
, long res
)
2716 umode_t mode
= file_inode(req
->file
)->i_mode
;
2719 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2721 if ((res
!= -EAGAIN
&& res
!= -EOPNOTSUPP
) || io_wq_current_is_worker())
2724 lockdep_assert_held(&req
->ctx
->uring_lock
);
2726 ret
= io_sq_thread_acquire_mm_files(req
->ctx
, req
);
2728 if (io_resubmit_prep(req
, ret
)) {
2729 refcount_inc(&req
->refs
);
2730 io_queue_async_work(req
);
2738 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2739 struct io_comp_state
*cs
)
2741 if (!io_rw_reissue(req
, res
))
2742 io_complete_rw_common(&req
->rw
.kiocb
, res
, cs
);
2745 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2747 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2749 __io_complete_rw(req
, res
, res2
, NULL
);
2752 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2754 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2756 if (kiocb
->ki_flags
& IOCB_WRITE
)
2757 kiocb_end_write(req
);
2759 if (res
!= -EAGAIN
&& res
!= req
->result
)
2760 req_set_fail_links(req
);
2762 WRITE_ONCE(req
->result
, res
);
2763 /* order with io_poll_complete() checking ->result */
2765 WRITE_ONCE(req
->iopoll_completed
, 1);
2769 * After the iocb has been issued, it's safe to be found on the poll list.
2770 * Adding the kiocb to the list AFTER submission ensures that we don't
2771 * find it from a io_iopoll_getevents() thread before the issuer is done
2772 * accessing the kiocb cookie.
2774 static void io_iopoll_req_issued(struct io_kiocb
*req
, bool in_async
)
2776 struct io_ring_ctx
*ctx
= req
->ctx
;
2779 * Track whether we have multiple files in our lists. This will impact
2780 * how we do polling eventually, not spinning if we're on potentially
2781 * different devices.
2783 if (list_empty(&ctx
->iopoll_list
)) {
2784 ctx
->poll_multi_file
= false;
2785 } else if (!ctx
->poll_multi_file
) {
2786 struct io_kiocb
*list_req
;
2788 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2790 if (list_req
->file
!= req
->file
)
2791 ctx
->poll_multi_file
= true;
2795 * For fast devices, IO may have already completed. If it has, add
2796 * it to the front so we find it first.
2798 if (READ_ONCE(req
->iopoll_completed
))
2799 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2801 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2804 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2805 * task context or in io worker task context. If current task context is
2806 * sq thread, we don't need to check whether should wake up sq thread.
2808 if (in_async
&& (ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2809 wq_has_sleeper(&ctx
->sq_data
->wait
))
2810 wake_up(&ctx
->sq_data
->wait
);
2813 static inline void __io_state_file_put(struct io_submit_state
*state
)
2815 fput_many(state
->file
, state
->file_refs
);
2816 state
->file_refs
= 0;
2819 static inline void io_state_file_put(struct io_submit_state
*state
)
2821 if (state
->file_refs
)
2822 __io_state_file_put(state
);
2826 * Get as many references to a file as we have IOs left in this submission,
2827 * assuming most submissions are for one file, or at least that each file
2828 * has more than one submission.
2830 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2835 if (state
->file_refs
) {
2836 if (state
->fd
== fd
) {
2840 __io_state_file_put(state
);
2842 state
->file
= fget_many(fd
, state
->ios_left
);
2843 if (unlikely(!state
->file
))
2847 state
->file_refs
= state
->ios_left
- 1;
2851 static bool io_bdev_nowait(struct block_device
*bdev
)
2853 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2857 * If we tracked the file through the SCM inflight mechanism, we could support
2858 * any file. For now, just ensure that anything potentially problematic is done
2861 static bool io_file_supports_async(struct file
*file
, int rw
)
2863 umode_t mode
= file_inode(file
)->i_mode
;
2865 if (S_ISBLK(mode
)) {
2866 if (IS_ENABLED(CONFIG_BLOCK
) &&
2867 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2871 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2873 if (S_ISREG(mode
)) {
2874 if (IS_ENABLED(CONFIG_BLOCK
) &&
2875 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2876 file
->f_op
!= &io_uring_fops
)
2881 /* any ->read/write should understand O_NONBLOCK */
2882 if (file
->f_flags
& O_NONBLOCK
)
2885 if (!(file
->f_mode
& FMODE_NOWAIT
))
2889 return file
->f_op
->read_iter
!= NULL
;
2891 return file
->f_op
->write_iter
!= NULL
;
2894 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2896 struct io_ring_ctx
*ctx
= req
->ctx
;
2897 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2901 if (S_ISREG(file_inode(req
->file
)->i_mode
))
2902 req
->flags
|= REQ_F_ISREG
;
2904 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2905 if (kiocb
->ki_pos
== -1 && !(req
->file
->f_mode
& FMODE_STREAM
)) {
2906 req
->flags
|= REQ_F_CUR_POS
;
2907 kiocb
->ki_pos
= req
->file
->f_pos
;
2909 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2910 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2911 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2915 ioprio
= READ_ONCE(sqe
->ioprio
);
2917 ret
= ioprio_check_cap(ioprio
);
2921 kiocb
->ki_ioprio
= ioprio
;
2923 kiocb
->ki_ioprio
= get_current_ioprio();
2925 /* don't allow async punt if RWF_NOWAIT was requested */
2926 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2927 req
->flags
|= REQ_F_NOWAIT
;
2929 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2930 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2931 !kiocb
->ki_filp
->f_op
->iopoll
)
2934 kiocb
->ki_flags
|= IOCB_HIPRI
;
2935 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2936 req
->iopoll_completed
= 0;
2938 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2940 kiocb
->ki_complete
= io_complete_rw
;
2943 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2944 req
->rw
.len
= READ_ONCE(sqe
->len
);
2945 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2949 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2955 case -ERESTARTNOINTR
:
2956 case -ERESTARTNOHAND
:
2957 case -ERESTART_RESTARTBLOCK
:
2959 * We can't just restart the syscall, since previously
2960 * submitted sqes may already be in progress. Just fail this
2966 kiocb
->ki_complete(kiocb
, ret
, 0);
2970 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2971 struct io_comp_state
*cs
)
2973 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2974 struct io_async_rw
*io
= req
->async_data
;
2976 /* add previously done IO, if any */
2977 if (io
&& io
->bytes_done
> 0) {
2979 ret
= io
->bytes_done
;
2981 ret
+= io
->bytes_done
;
2984 if (req
->flags
& REQ_F_CUR_POS
)
2985 req
->file
->f_pos
= kiocb
->ki_pos
;
2986 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2987 __io_complete_rw(req
, ret
, 0, cs
);
2989 io_rw_done(kiocb
, ret
);
2992 static ssize_t
io_import_fixed(struct io_kiocb
*req
, int rw
,
2993 struct iov_iter
*iter
)
2995 struct io_ring_ctx
*ctx
= req
->ctx
;
2996 size_t len
= req
->rw
.len
;
2997 struct io_mapped_ubuf
*imu
;
2998 u16 index
, buf_index
= req
->buf_index
;
3002 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3004 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3005 imu
= &ctx
->user_bufs
[index
];
3006 buf_addr
= req
->rw
.addr
;
3009 if (buf_addr
+ len
< buf_addr
)
3011 /* not inside the mapped region */
3012 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
3016 * May not be a start of buffer, set size appropriately
3017 * and advance us to the beginning.
3019 offset
= buf_addr
- imu
->ubuf
;
3020 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3024 * Don't use iov_iter_advance() here, as it's really slow for
3025 * using the latter parts of a big fixed buffer - it iterates
3026 * over each segment manually. We can cheat a bit here, because
3029 * 1) it's a BVEC iter, we set it up
3030 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3031 * first and last bvec
3033 * So just find our index, and adjust the iterator afterwards.
3034 * If the offset is within the first bvec (or the whole first
3035 * bvec, just use iov_iter_advance(). This makes it easier
3036 * since we can just skip the first segment, which may not
3037 * be PAGE_SIZE aligned.
3039 const struct bio_vec
*bvec
= imu
->bvec
;
3041 if (offset
<= bvec
->bv_len
) {
3042 iov_iter_advance(iter
, offset
);
3044 unsigned long seg_skip
;
3046 /* skip first vec */
3047 offset
-= bvec
->bv_len
;
3048 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3050 iter
->bvec
= bvec
+ seg_skip
;
3051 iter
->nr_segs
-= seg_skip
;
3052 iter
->count
-= bvec
->bv_len
+ offset
;
3053 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3060 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3063 mutex_unlock(&ctx
->uring_lock
);
3066 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3069 * "Normal" inline submissions always hold the uring_lock, since we
3070 * grab it from the system call. Same is true for the SQPOLL offload.
3071 * The only exception is when we've detached the request and issue it
3072 * from an async worker thread, grab the lock for that case.
3075 mutex_lock(&ctx
->uring_lock
);
3078 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3079 int bgid
, struct io_buffer
*kbuf
,
3082 struct io_buffer
*head
;
3084 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3087 io_ring_submit_lock(req
->ctx
, needs_lock
);
3089 lockdep_assert_held(&req
->ctx
->uring_lock
);
3091 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
3093 if (!list_empty(&head
->list
)) {
3094 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3096 list_del(&kbuf
->list
);
3099 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
3101 if (*len
> kbuf
->len
)
3104 kbuf
= ERR_PTR(-ENOBUFS
);
3107 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3112 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3115 struct io_buffer
*kbuf
;
3118 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3119 bgid
= req
->buf_index
;
3120 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3123 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3124 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3125 return u64_to_user_ptr(kbuf
->addr
);
3128 #ifdef CONFIG_COMPAT
3129 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3132 struct compat_iovec __user
*uiov
;
3133 compat_ssize_t clen
;
3137 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3138 if (!access_ok(uiov
, sizeof(*uiov
)))
3140 if (__get_user(clen
, &uiov
->iov_len
))
3146 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3148 return PTR_ERR(buf
);
3149 iov
[0].iov_base
= buf
;
3150 iov
[0].iov_len
= (compat_size_t
) len
;
3155 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3158 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3162 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3165 len
= iov
[0].iov_len
;
3168 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3170 return PTR_ERR(buf
);
3171 iov
[0].iov_base
= buf
;
3172 iov
[0].iov_len
= len
;
3176 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3179 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3180 struct io_buffer
*kbuf
;
3182 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3183 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3184 iov
[0].iov_len
= kbuf
->len
;
3187 if (req
->rw
.len
!= 1)
3190 #ifdef CONFIG_COMPAT
3191 if (req
->ctx
->compat
)
3192 return io_compat_import(req
, iov
, needs_lock
);
3195 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3198 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
3199 struct iovec
**iovec
, struct iov_iter
*iter
,
3202 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3203 size_t sqe_len
= req
->rw
.len
;
3207 opcode
= req
->opcode
;
3208 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3210 return io_import_fixed(req
, rw
, iter
);
3213 /* buffer index only valid with fixed read/write, or buffer select */
3214 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3217 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3218 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3219 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3221 return PTR_ERR(buf
);
3222 req
->rw
.len
= sqe_len
;
3225 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3230 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3231 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3233 ret
= (*iovec
)->iov_len
;
3234 iov_iter_init(iter
, rw
, *iovec
, 1, ret
);
3240 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3244 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3246 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3250 * For files that don't have ->read_iter() and ->write_iter(), handle them
3251 * by looping over ->read() or ->write() manually.
3253 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3255 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3256 struct file
*file
= req
->file
;
3260 * Don't support polled IO through this interface, and we can't
3261 * support non-blocking either. For the latter, this just causes
3262 * the kiocb to be handled from an async context.
3264 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3266 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3269 while (iov_iter_count(iter
)) {
3273 if (!iov_iter_is_bvec(iter
)) {
3274 iovec
= iov_iter_iovec(iter
);
3276 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3277 iovec
.iov_len
= req
->rw
.len
;
3281 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3282 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3284 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3285 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3294 if (nr
!= iovec
.iov_len
)
3298 iov_iter_advance(iter
, nr
);
3304 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3305 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3307 struct io_async_rw
*rw
= req
->async_data
;
3309 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3310 rw
->free_iovec
= iovec
;
3312 /* can only be fixed buffers, no need to do anything */
3313 if (iov_iter_is_bvec(iter
))
3316 unsigned iov_off
= 0;
3318 rw
->iter
.iov
= rw
->fast_iov
;
3319 if (iter
->iov
!= fast_iov
) {
3320 iov_off
= iter
->iov
- fast_iov
;
3321 rw
->iter
.iov
+= iov_off
;
3323 if (rw
->fast_iov
!= fast_iov
)
3324 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3325 sizeof(struct iovec
) * iter
->nr_segs
);
3327 req
->flags
|= REQ_F_NEED_CLEANUP
;
3331 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3333 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3334 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3335 return req
->async_data
== NULL
;
3338 static int io_alloc_async_data(struct io_kiocb
*req
)
3340 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3343 return __io_alloc_async_data(req
);
3346 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3347 const struct iovec
*fast_iov
,
3348 struct iov_iter
*iter
, bool force
)
3350 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3352 if (!req
->async_data
) {
3353 if (__io_alloc_async_data(req
))
3356 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3361 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3363 struct io_async_rw
*iorw
= req
->async_data
;
3364 struct iovec
*iov
= iorw
->fast_iov
;
3367 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3368 if (unlikely(ret
< 0))
3371 iorw
->bytes_done
= 0;
3372 iorw
->free_iovec
= iov
;
3374 req
->flags
|= REQ_F_NEED_CLEANUP
;
3378 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3382 ret
= io_prep_rw(req
, sqe
);
3386 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3389 /* either don't need iovec imported or already have it */
3390 if (!req
->async_data
)
3392 return io_rw_prep_async(req
, READ
);
3396 * This is our waitqueue callback handler, registered through lock_page_async()
3397 * when we initially tried to do the IO with the iocb armed our waitqueue.
3398 * This gets called when the page is unlocked, and we generally expect that to
3399 * happen when the page IO is completed and the page is now uptodate. This will
3400 * queue a task_work based retry of the operation, attempting to copy the data
3401 * again. If the latter fails because the page was NOT uptodate, then we will
3402 * do a thread based blocking retry of the operation. That's the unexpected
3405 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3406 int sync
, void *arg
)
3408 struct wait_page_queue
*wpq
;
3409 struct io_kiocb
*req
= wait
->private;
3410 struct wait_page_key
*key
= arg
;
3413 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3415 if (!wake_page_match(wpq
, key
))
3418 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3419 list_del_init(&wait
->entry
);
3421 init_task_work(&req
->task_work
, io_req_task_submit
);
3422 percpu_ref_get(&req
->ctx
->refs
);
3424 /* submit ref gets dropped, acquire a new one */
3425 refcount_inc(&req
->refs
);
3426 ret
= io_req_task_work_add(req
);
3427 if (unlikely(ret
)) {
3428 struct task_struct
*tsk
;
3430 /* queue just for cancelation */
3431 init_task_work(&req
->task_work
, io_req_task_cancel
);
3432 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
3433 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
3434 wake_up_process(tsk
);
3440 * This controls whether a given IO request should be armed for async page
3441 * based retry. If we return false here, the request is handed to the async
3442 * worker threads for retry. If we're doing buffered reads on a regular file,
3443 * we prepare a private wait_page_queue entry and retry the operation. This
3444 * will either succeed because the page is now uptodate and unlocked, or it
3445 * will register a callback when the page is unlocked at IO completion. Through
3446 * that callback, io_uring uses task_work to setup a retry of the operation.
3447 * That retry will attempt the buffered read again. The retry will generally
3448 * succeed, or in rare cases where it fails, we then fall back to using the
3449 * async worker threads for a blocking retry.
3451 static bool io_rw_should_retry(struct io_kiocb
*req
)
3453 struct io_async_rw
*rw
= req
->async_data
;
3454 struct wait_page_queue
*wait
= &rw
->wpq
;
3455 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3457 /* never retry for NOWAIT, we just complete with -EAGAIN */
3458 if (req
->flags
& REQ_F_NOWAIT
)
3461 /* Only for buffered IO */
3462 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3466 * just use poll if we can, and don't attempt if the fs doesn't
3467 * support callback based unlocks
3469 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3472 wait
->wait
.func
= io_async_buf_func
;
3473 wait
->wait
.private = req
;
3474 wait
->wait
.flags
= 0;
3475 INIT_LIST_HEAD(&wait
->wait
.entry
);
3476 kiocb
->ki_flags
|= IOCB_WAITQ
;
3477 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3478 kiocb
->ki_waitq
= wait
;
3482 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3484 if (req
->file
->f_op
->read_iter
)
3485 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3486 else if (req
->file
->f_op
->read
)
3487 return loop_rw_iter(READ
, req
, iter
);
3492 static int io_read(struct io_kiocb
*req
, bool force_nonblock
,
3493 struct io_comp_state
*cs
)
3495 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3496 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3497 struct iov_iter __iter
, *iter
= &__iter
;
3498 struct io_async_rw
*rw
= req
->async_data
;
3499 ssize_t io_size
, ret
, ret2
;
3506 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3510 io_size
= iov_iter_count(iter
);
3511 req
->result
= io_size
;
3514 /* Ensure we clear previously set non-block flag */
3515 if (!force_nonblock
)
3516 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3518 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3521 /* If the file doesn't support async, just async punt */
3522 no_async
= force_nonblock
&& !io_file_supports_async(req
->file
, READ
);
3526 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3530 ret
= io_iter_do_read(req
, iter
);
3534 } else if (ret
== -EIOCBQUEUED
) {
3537 } else if (ret
== -EAGAIN
) {
3538 /* IOPOLL retry should happen for io-wq threads */
3539 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3541 /* no retry on NONBLOCK marked file */
3542 if (req
->file
->f_flags
& O_NONBLOCK
)
3544 /* some cases will consume bytes even on error returns */
3545 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3548 } else if (ret
< 0) {
3549 /* make sure -ERESTARTSYS -> -EINTR is done */
3553 /* read it all, or we did blocking attempt. no retry. */
3554 if (!iov_iter_count(iter
) || !force_nonblock
||
3555 (req
->file
->f_flags
& O_NONBLOCK
) || !(req
->flags
& REQ_F_ISREG
))
3560 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3567 rw
= req
->async_data
;
3568 /* it's copied and will be cleaned with ->io */
3570 /* now use our persistent iterator, if we aren't already */
3573 rw
->bytes_done
+= ret
;
3574 /* if we can retry, do so with the callbacks armed */
3575 if (!io_rw_should_retry(req
)) {
3576 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3581 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3582 * get -EIOCBQUEUED, then we'll get a notification when the desired
3583 * page gets unlocked. We can also get a partial read here, and if we
3584 * do, then just retry at the new offset.
3586 ret
= io_iter_do_read(req
, iter
);
3587 if (ret
== -EIOCBQUEUED
) {
3590 } else if (ret
> 0 && ret
< io_size
) {
3591 /* we got some bytes, but not all. retry. */
3595 kiocb_done(kiocb
, ret
, cs
);
3598 /* it's reportedly faster than delegating the null check to kfree() */
3604 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3608 ret
= io_prep_rw(req
, sqe
);
3612 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3615 /* either don't need iovec imported or already have it */
3616 if (!req
->async_data
)
3618 return io_rw_prep_async(req
, WRITE
);
3621 static int io_write(struct io_kiocb
*req
, bool force_nonblock
,
3622 struct io_comp_state
*cs
)
3624 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3625 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3626 struct iov_iter __iter
, *iter
= &__iter
;
3627 struct io_async_rw
*rw
= req
->async_data
;
3628 ssize_t ret
, ret2
, io_size
;
3634 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3638 io_size
= iov_iter_count(iter
);
3639 req
->result
= io_size
;
3641 /* Ensure we clear previously set non-block flag */
3642 if (!force_nonblock
)
3643 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3645 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3647 /* If the file doesn't support async, just async punt */
3648 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3651 /* file path doesn't support NOWAIT for non-direct_IO */
3652 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3653 (req
->flags
& REQ_F_ISREG
))
3656 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3661 * Open-code file_start_write here to grab freeze protection,
3662 * which will be released by another thread in
3663 * io_complete_rw(). Fool lockdep by telling it the lock got
3664 * released so that it doesn't complain about the held lock when
3665 * we return to userspace.
3667 if (req
->flags
& REQ_F_ISREG
) {
3668 sb_start_write(file_inode(req
->file
)->i_sb
);
3669 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3672 kiocb
->ki_flags
|= IOCB_WRITE
;
3674 if (req
->file
->f_op
->write_iter
)
3675 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3676 else if (req
->file
->f_op
->write
)
3677 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3682 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3683 * retry them without IOCB_NOWAIT.
3685 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3687 /* no retry on NONBLOCK marked file */
3688 if (ret2
== -EAGAIN
&& (req
->file
->f_flags
& O_NONBLOCK
))
3690 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3691 /* IOPOLL retry should happen for io-wq threads */
3692 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3695 kiocb_done(kiocb
, ret2
, cs
);
3698 /* some cases will consume bytes even on error returns */
3699 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3700 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3705 /* it's reportedly faster than delegating the null check to kfree() */
3711 static int io_renameat_prep(struct io_kiocb
*req
,
3712 const struct io_uring_sqe
*sqe
)
3714 struct io_rename
*ren
= &req
->rename
;
3715 const char __user
*oldf
, *newf
;
3717 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3720 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3721 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3722 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3723 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3724 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3726 ren
->oldpath
= getname(oldf
);
3727 if (IS_ERR(ren
->oldpath
))
3728 return PTR_ERR(ren
->oldpath
);
3730 ren
->newpath
= getname(newf
);
3731 if (IS_ERR(ren
->newpath
)) {
3732 putname(ren
->oldpath
);
3733 return PTR_ERR(ren
->newpath
);
3736 req
->flags
|= REQ_F_NEED_CLEANUP
;
3740 static int io_renameat(struct io_kiocb
*req
, bool force_nonblock
)
3742 struct io_rename
*ren
= &req
->rename
;
3748 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3749 ren
->newpath
, ren
->flags
);
3751 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3753 req_set_fail_links(req
);
3754 io_req_complete(req
, ret
);
3758 static int io_unlinkat_prep(struct io_kiocb
*req
,
3759 const struct io_uring_sqe
*sqe
)
3761 struct io_unlink
*un
= &req
->unlink
;
3762 const char __user
*fname
;
3764 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3767 un
->dfd
= READ_ONCE(sqe
->fd
);
3769 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3770 if (un
->flags
& ~AT_REMOVEDIR
)
3773 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3774 un
->filename
= getname(fname
);
3775 if (IS_ERR(un
->filename
))
3776 return PTR_ERR(un
->filename
);
3778 req
->flags
|= REQ_F_NEED_CLEANUP
;
3782 static int io_unlinkat(struct io_kiocb
*req
, bool force_nonblock
)
3784 struct io_unlink
*un
= &req
->unlink
;
3790 if (un
->flags
& AT_REMOVEDIR
)
3791 ret
= do_rmdir(un
->dfd
, un
->filename
);
3793 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3795 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3797 req_set_fail_links(req
);
3798 io_req_complete(req
, ret
);
3802 static int io_shutdown_prep(struct io_kiocb
*req
,
3803 const struct io_uring_sqe
*sqe
)
3805 #if defined(CONFIG_NET)
3806 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3808 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3812 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3819 static int io_shutdown(struct io_kiocb
*req
, bool force_nonblock
)
3821 #if defined(CONFIG_NET)
3822 struct socket
*sock
;
3828 sock
= sock_from_file(req
->file
);
3829 if (unlikely(!sock
))
3832 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3834 req_set_fail_links(req
);
3835 io_req_complete(req
, ret
);
3842 static int __io_splice_prep(struct io_kiocb
*req
,
3843 const struct io_uring_sqe
*sqe
)
3845 struct io_splice
* sp
= &req
->splice
;
3846 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3848 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3852 sp
->len
= READ_ONCE(sqe
->len
);
3853 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3855 if (unlikely(sp
->flags
& ~valid_flags
))
3858 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3859 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3862 req
->flags
|= REQ_F_NEED_CLEANUP
;
3864 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3866 * Splice operation will be punted aync, and here need to
3867 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3869 io_req_init_async(req
);
3870 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3876 static int io_tee_prep(struct io_kiocb
*req
,
3877 const struct io_uring_sqe
*sqe
)
3879 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3881 return __io_splice_prep(req
, sqe
);
3884 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3886 struct io_splice
*sp
= &req
->splice
;
3887 struct file
*in
= sp
->file_in
;
3888 struct file
*out
= sp
->file_out
;
3889 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3895 ret
= do_tee(in
, out
, sp
->len
, flags
);
3897 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3898 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3901 req_set_fail_links(req
);
3902 io_req_complete(req
, ret
);
3906 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3908 struct io_splice
* sp
= &req
->splice
;
3910 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3911 sp
->off_out
= READ_ONCE(sqe
->off
);
3912 return __io_splice_prep(req
, sqe
);
3915 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3917 struct io_splice
*sp
= &req
->splice
;
3918 struct file
*in
= sp
->file_in
;
3919 struct file
*out
= sp
->file_out
;
3920 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3921 loff_t
*poff_in
, *poff_out
;
3927 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3928 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3931 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3933 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3934 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3937 req_set_fail_links(req
);
3938 io_req_complete(req
, ret
);
3943 * IORING_OP_NOP just posts a completion event, nothing else.
3945 static int io_nop(struct io_kiocb
*req
, struct io_comp_state
*cs
)
3947 struct io_ring_ctx
*ctx
= req
->ctx
;
3949 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3952 __io_req_complete(req
, 0, 0, cs
);
3956 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3958 struct io_ring_ctx
*ctx
= req
->ctx
;
3963 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3965 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3968 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3969 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3972 req
->sync
.off
= READ_ONCE(sqe
->off
);
3973 req
->sync
.len
= READ_ONCE(sqe
->len
);
3977 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3979 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3982 /* fsync always requires a blocking context */
3986 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3987 end
> 0 ? end
: LLONG_MAX
,
3988 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3990 req_set_fail_links(req
);
3991 io_req_complete(req
, ret
);
3995 static int io_fallocate_prep(struct io_kiocb
*req
,
3996 const struct io_uring_sqe
*sqe
)
3998 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
4000 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4003 req
->sync
.off
= READ_ONCE(sqe
->off
);
4004 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4005 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4009 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
4013 /* fallocate always requiring blocking context */
4016 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4019 req_set_fail_links(req
);
4020 io_req_complete(req
, ret
);
4024 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4026 const char __user
*fname
;
4029 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4031 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4034 /* open.how should be already initialised */
4035 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4036 req
->open
.how
.flags
|= O_LARGEFILE
;
4038 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4039 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4040 req
->open
.filename
= getname(fname
);
4041 if (IS_ERR(req
->open
.filename
)) {
4042 ret
= PTR_ERR(req
->open
.filename
);
4043 req
->open
.filename
= NULL
;
4046 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4047 req
->open
.ignore_nonblock
= false;
4048 req
->flags
|= REQ_F_NEED_CLEANUP
;
4052 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4056 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4058 mode
= READ_ONCE(sqe
->len
);
4059 flags
= READ_ONCE(sqe
->open_flags
);
4060 req
->open
.how
= build_open_how(flags
, mode
);
4061 return __io_openat_prep(req
, sqe
);
4064 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4066 struct open_how __user
*how
;
4070 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4072 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4073 len
= READ_ONCE(sqe
->len
);
4074 if (len
< OPEN_HOW_SIZE_VER0
)
4077 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4082 return __io_openat_prep(req
, sqe
);
4085 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
4087 struct open_flags op
;
4091 if (force_nonblock
&& !req
->open
.ignore_nonblock
)
4094 ret
= build_open_flags(&req
->open
.how
, &op
);
4098 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4102 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4105 ret
= PTR_ERR(file
);
4107 * A work-around to ensure that /proc/self works that way
4108 * that it should - if we get -EOPNOTSUPP back, then assume
4109 * that proc_self_get_link() failed us because we're in async
4110 * context. We should be safe to retry this from the task
4111 * itself with force_nonblock == false set, as it should not
4112 * block on lookup. Would be nice to know this upfront and
4113 * avoid the async dance, but doesn't seem feasible.
4115 if (ret
== -EOPNOTSUPP
&& io_wq_current_is_worker()) {
4116 req
->open
.ignore_nonblock
= true;
4117 refcount_inc(&req
->refs
);
4118 io_req_task_queue(req
);
4122 fsnotify_open(file
);
4123 fd_install(ret
, file
);
4126 putname(req
->open
.filename
);
4127 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4129 req_set_fail_links(req
);
4130 io_req_complete(req
, ret
);
4134 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
4136 return io_openat2(req
, force_nonblock
);
4139 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4140 const struct io_uring_sqe
*sqe
)
4142 struct io_provide_buf
*p
= &req
->pbuf
;
4145 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
4148 tmp
= READ_ONCE(sqe
->fd
);
4149 if (!tmp
|| tmp
> USHRT_MAX
)
4152 memset(p
, 0, sizeof(*p
));
4154 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4158 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4159 int bgid
, unsigned nbufs
)
4163 /* shouldn't happen */
4167 /* the head kbuf is the list itself */
4168 while (!list_empty(&buf
->list
)) {
4169 struct io_buffer
*nxt
;
4171 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4172 list_del(&nxt
->list
);
4179 idr_remove(&ctx
->io_buffer_idr
, bgid
);
4184 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4185 struct io_comp_state
*cs
)
4187 struct io_provide_buf
*p
= &req
->pbuf
;
4188 struct io_ring_ctx
*ctx
= req
->ctx
;
4189 struct io_buffer
*head
;
4192 io_ring_submit_lock(ctx
, !force_nonblock
);
4194 lockdep_assert_held(&ctx
->uring_lock
);
4197 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4199 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4201 req_set_fail_links(req
);
4203 /* need to hold the lock to complete IOPOLL requests */
4204 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4205 __io_req_complete(req
, ret
, 0, cs
);
4206 io_ring_submit_unlock(ctx
, !force_nonblock
);
4208 io_ring_submit_unlock(ctx
, !force_nonblock
);
4209 __io_req_complete(req
, ret
, 0, cs
);
4214 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4215 const struct io_uring_sqe
*sqe
)
4217 struct io_provide_buf
*p
= &req
->pbuf
;
4220 if (sqe
->ioprio
|| sqe
->rw_flags
)
4223 tmp
= READ_ONCE(sqe
->fd
);
4224 if (!tmp
|| tmp
> USHRT_MAX
)
4227 p
->addr
= READ_ONCE(sqe
->addr
);
4228 p
->len
= READ_ONCE(sqe
->len
);
4230 if (!access_ok(u64_to_user_ptr(p
->addr
), (p
->len
* p
->nbufs
)))
4233 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4234 tmp
= READ_ONCE(sqe
->off
);
4235 if (tmp
> USHRT_MAX
)
4241 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4243 struct io_buffer
*buf
;
4244 u64 addr
= pbuf
->addr
;
4245 int i
, bid
= pbuf
->bid
;
4247 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4248 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
4253 buf
->len
= pbuf
->len
;
4258 INIT_LIST_HEAD(&buf
->list
);
4261 list_add_tail(&buf
->list
, &(*head
)->list
);
4265 return i
? i
: -ENOMEM
;
4268 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4269 struct io_comp_state
*cs
)
4271 struct io_provide_buf
*p
= &req
->pbuf
;
4272 struct io_ring_ctx
*ctx
= req
->ctx
;
4273 struct io_buffer
*head
, *list
;
4276 io_ring_submit_lock(ctx
, !force_nonblock
);
4278 lockdep_assert_held(&ctx
->uring_lock
);
4280 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4282 ret
= io_add_buffers(p
, &head
);
4287 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
4290 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4296 req_set_fail_links(req
);
4298 /* need to hold the lock to complete IOPOLL requests */
4299 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4300 __io_req_complete(req
, ret
, 0, cs
);
4301 io_ring_submit_unlock(ctx
, !force_nonblock
);
4303 io_ring_submit_unlock(ctx
, !force_nonblock
);
4304 __io_req_complete(req
, ret
, 0, cs
);
4309 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4310 const struct io_uring_sqe
*sqe
)
4312 #if defined(CONFIG_EPOLL)
4313 if (sqe
->ioprio
|| sqe
->buf_index
)
4315 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4318 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4319 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4320 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4322 if (ep_op_has_event(req
->epoll
.op
)) {
4323 struct epoll_event __user
*ev
;
4325 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4326 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4336 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
,
4337 struct io_comp_state
*cs
)
4339 #if defined(CONFIG_EPOLL)
4340 struct io_epoll
*ie
= &req
->epoll
;
4343 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4344 if (force_nonblock
&& ret
== -EAGAIN
)
4348 req_set_fail_links(req
);
4349 __io_req_complete(req
, ret
, 0, cs
);
4356 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4358 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4359 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4361 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4364 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4365 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4366 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4373 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
4375 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4376 struct io_madvise
*ma
= &req
->madvise
;
4382 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4384 req_set_fail_links(req
);
4385 io_req_complete(req
, ret
);
4392 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4394 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4396 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4399 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4400 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4401 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4405 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
4407 struct io_fadvise
*fa
= &req
->fadvise
;
4410 if (force_nonblock
) {
4411 switch (fa
->advice
) {
4412 case POSIX_FADV_NORMAL
:
4413 case POSIX_FADV_RANDOM
:
4414 case POSIX_FADV_SEQUENTIAL
:
4421 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4423 req_set_fail_links(req
);
4424 io_req_complete(req
, ret
);
4428 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4430 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4432 if (sqe
->ioprio
|| sqe
->buf_index
)
4434 if (req
->flags
& REQ_F_FIXED_FILE
)
4437 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4438 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4439 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4440 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4441 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4446 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
4448 struct io_statx
*ctx
= &req
->statx
;
4451 if (force_nonblock
) {
4452 /* only need file table for an actual valid fd */
4453 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4454 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4458 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4462 req_set_fail_links(req
);
4463 io_req_complete(req
, ret
);
4467 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4470 * If we queue this for async, it must not be cancellable. That would
4471 * leave the 'file' in an undeterminate state, and here need to modify
4472 * io_wq_work.flags, so initialize io_wq_work firstly.
4474 io_req_init_async(req
);
4476 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4478 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4479 sqe
->rw_flags
|| sqe
->buf_index
)
4481 if (req
->flags
& REQ_F_FIXED_FILE
)
4484 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4485 if ((req
->file
&& req
->file
->f_op
== &io_uring_fops
))
4488 req
->close
.put_file
= NULL
;
4492 static int io_close(struct io_kiocb
*req
, bool force_nonblock
,
4493 struct io_comp_state
*cs
)
4495 struct io_close
*close
= &req
->close
;
4498 /* might be already done during nonblock submission */
4499 if (!close
->put_file
) {
4500 ret
= close_fd_get_file(close
->fd
, &close
->put_file
);
4502 return (ret
== -ENOENT
) ? -EBADF
: ret
;
4505 /* if the file has a flush method, be safe and punt to async */
4506 if (close
->put_file
->f_op
->flush
&& force_nonblock
) {
4507 /* not safe to cancel at this point */
4508 req
->work
.flags
|= IO_WQ_WORK_NO_CANCEL
;
4509 /* was never set, but play safe */
4510 req
->flags
&= ~REQ_F_NOWAIT
;
4511 /* avoid grabbing files - we don't need the files */
4512 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4516 /* No ->flush() or already async, safely close from here */
4517 ret
= filp_close(close
->put_file
, req
->work
.identity
->files
);
4519 req_set_fail_links(req
);
4520 fput(close
->put_file
);
4521 close
->put_file
= NULL
;
4522 __io_req_complete(req
, ret
, 0, cs
);
4526 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4528 struct io_ring_ctx
*ctx
= req
->ctx
;
4533 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4535 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4538 req
->sync
.off
= READ_ONCE(sqe
->off
);
4539 req
->sync
.len
= READ_ONCE(sqe
->len
);
4540 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4544 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
4548 /* sync_file_range always requires a blocking context */
4552 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4555 req_set_fail_links(req
);
4556 io_req_complete(req
, ret
);
4560 #if defined(CONFIG_NET)
4561 static int io_setup_async_msg(struct io_kiocb
*req
,
4562 struct io_async_msghdr
*kmsg
)
4564 struct io_async_msghdr
*async_msg
= req
->async_data
;
4568 if (io_alloc_async_data(req
)) {
4569 if (kmsg
->iov
!= kmsg
->fast_iov
)
4573 async_msg
= req
->async_data
;
4574 req
->flags
|= REQ_F_NEED_CLEANUP
;
4575 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4579 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4580 struct io_async_msghdr
*iomsg
)
4582 iomsg
->iov
= iomsg
->fast_iov
;
4583 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4584 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4585 req
->sr_msg
.msg_flags
, &iomsg
->iov
);
4588 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4590 struct io_async_msghdr
*async_msg
= req
->async_data
;
4591 struct io_sr_msg
*sr
= &req
->sr_msg
;
4594 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4597 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4598 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4599 sr
->len
= READ_ONCE(sqe
->len
);
4601 #ifdef CONFIG_COMPAT
4602 if (req
->ctx
->compat
)
4603 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4606 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4608 ret
= io_sendmsg_copy_hdr(req
, async_msg
);
4610 req
->flags
|= REQ_F_NEED_CLEANUP
;
4614 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4615 struct io_comp_state
*cs
)
4617 struct io_async_msghdr iomsg
, *kmsg
;
4618 struct socket
*sock
;
4622 sock
= sock_from_file(req
->file
);
4623 if (unlikely(!sock
))
4626 if (req
->async_data
) {
4627 kmsg
= req
->async_data
;
4628 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4629 /* if iov is set, it's allocated already */
4631 kmsg
->iov
= kmsg
->fast_iov
;
4632 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4634 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4640 flags
= req
->sr_msg
.msg_flags
;
4641 if (flags
& MSG_DONTWAIT
)
4642 req
->flags
|= REQ_F_NOWAIT
;
4643 else if (force_nonblock
)
4644 flags
|= MSG_DONTWAIT
;
4646 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4647 if (force_nonblock
&& ret
== -EAGAIN
)
4648 return io_setup_async_msg(req
, kmsg
);
4649 if (ret
== -ERESTARTSYS
)
4652 if (kmsg
->iov
!= kmsg
->fast_iov
)
4654 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4656 req_set_fail_links(req
);
4657 __io_req_complete(req
, ret
, 0, cs
);
4661 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4662 struct io_comp_state
*cs
)
4664 struct io_sr_msg
*sr
= &req
->sr_msg
;
4667 struct socket
*sock
;
4671 sock
= sock_from_file(req
->file
);
4672 if (unlikely(!sock
))
4675 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4679 msg
.msg_name
= NULL
;
4680 msg
.msg_control
= NULL
;
4681 msg
.msg_controllen
= 0;
4682 msg
.msg_namelen
= 0;
4684 flags
= req
->sr_msg
.msg_flags
;
4685 if (flags
& MSG_DONTWAIT
)
4686 req
->flags
|= REQ_F_NOWAIT
;
4687 else if (force_nonblock
)
4688 flags
|= MSG_DONTWAIT
;
4690 msg
.msg_flags
= flags
;
4691 ret
= sock_sendmsg(sock
, &msg
);
4692 if (force_nonblock
&& ret
== -EAGAIN
)
4694 if (ret
== -ERESTARTSYS
)
4698 req_set_fail_links(req
);
4699 __io_req_complete(req
, ret
, 0, cs
);
4703 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4704 struct io_async_msghdr
*iomsg
)
4706 struct io_sr_msg
*sr
= &req
->sr_msg
;
4707 struct iovec __user
*uiov
;
4711 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4712 &iomsg
->uaddr
, &uiov
, &iov_len
);
4716 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4719 if (copy_from_user(iomsg
->iov
, uiov
, sizeof(*uiov
)))
4721 sr
->len
= iomsg
->iov
[0].iov_len
;
4722 iov_iter_init(&iomsg
->msg
.msg_iter
, READ
, iomsg
->iov
, 1,
4726 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4727 &iomsg
->iov
, &iomsg
->msg
.msg_iter
,
4736 #ifdef CONFIG_COMPAT
4737 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4738 struct io_async_msghdr
*iomsg
)
4740 struct compat_msghdr __user
*msg_compat
;
4741 struct io_sr_msg
*sr
= &req
->sr_msg
;
4742 struct compat_iovec __user
*uiov
;
4747 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4748 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4753 uiov
= compat_ptr(ptr
);
4754 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4755 compat_ssize_t clen
;
4759 if (!access_ok(uiov
, sizeof(*uiov
)))
4761 if (__get_user(clen
, &uiov
->iov_len
))
4766 iomsg
->iov
[0].iov_len
= clen
;
4769 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4770 UIO_FASTIOV
, &iomsg
->iov
,
4771 &iomsg
->msg
.msg_iter
, true);
4780 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4781 struct io_async_msghdr
*iomsg
)
4783 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4784 iomsg
->iov
= iomsg
->fast_iov
;
4786 #ifdef CONFIG_COMPAT
4787 if (req
->ctx
->compat
)
4788 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4791 return __io_recvmsg_copy_hdr(req
, iomsg
);
4794 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4797 struct io_sr_msg
*sr
= &req
->sr_msg
;
4798 struct io_buffer
*kbuf
;
4800 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4805 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4809 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4811 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4814 static int io_recvmsg_prep(struct io_kiocb
*req
,
4815 const struct io_uring_sqe
*sqe
)
4817 struct io_async_msghdr
*async_msg
= req
->async_data
;
4818 struct io_sr_msg
*sr
= &req
->sr_msg
;
4821 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4824 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4825 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4826 sr
->len
= READ_ONCE(sqe
->len
);
4827 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4829 #ifdef CONFIG_COMPAT
4830 if (req
->ctx
->compat
)
4831 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4834 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4836 ret
= io_recvmsg_copy_hdr(req
, async_msg
);
4838 req
->flags
|= REQ_F_NEED_CLEANUP
;
4842 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
4843 struct io_comp_state
*cs
)
4845 struct io_async_msghdr iomsg
, *kmsg
;
4846 struct socket
*sock
;
4847 struct io_buffer
*kbuf
;
4849 int ret
, cflags
= 0;
4851 sock
= sock_from_file(req
->file
);
4852 if (unlikely(!sock
))
4855 if (req
->async_data
) {
4856 kmsg
= req
->async_data
;
4857 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4858 /* if iov is set, it's allocated already */
4860 kmsg
->iov
= kmsg
->fast_iov
;
4861 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4863 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4869 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4870 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4872 return PTR_ERR(kbuf
);
4873 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4874 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4875 1, req
->sr_msg
.len
);
4878 flags
= req
->sr_msg
.msg_flags
;
4879 if (flags
& MSG_DONTWAIT
)
4880 req
->flags
|= REQ_F_NOWAIT
;
4881 else if (force_nonblock
)
4882 flags
|= MSG_DONTWAIT
;
4884 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4885 kmsg
->uaddr
, flags
);
4886 if (force_nonblock
&& ret
== -EAGAIN
)
4887 return io_setup_async_msg(req
, kmsg
);
4888 if (ret
== -ERESTARTSYS
)
4891 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4892 cflags
= io_put_recv_kbuf(req
);
4893 if (kmsg
->iov
!= kmsg
->fast_iov
)
4895 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4897 req_set_fail_links(req
);
4898 __io_req_complete(req
, ret
, cflags
, cs
);
4902 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
4903 struct io_comp_state
*cs
)
4905 struct io_buffer
*kbuf
;
4906 struct io_sr_msg
*sr
= &req
->sr_msg
;
4908 void __user
*buf
= sr
->buf
;
4909 struct socket
*sock
;
4912 int ret
, cflags
= 0;
4914 sock
= sock_from_file(req
->file
);
4915 if (unlikely(!sock
))
4918 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4919 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4921 return PTR_ERR(kbuf
);
4922 buf
= u64_to_user_ptr(kbuf
->addr
);
4925 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4929 msg
.msg_name
= NULL
;
4930 msg
.msg_control
= NULL
;
4931 msg
.msg_controllen
= 0;
4932 msg
.msg_namelen
= 0;
4933 msg
.msg_iocb
= NULL
;
4936 flags
= req
->sr_msg
.msg_flags
;
4937 if (flags
& MSG_DONTWAIT
)
4938 req
->flags
|= REQ_F_NOWAIT
;
4939 else if (force_nonblock
)
4940 flags
|= MSG_DONTWAIT
;
4942 ret
= sock_recvmsg(sock
, &msg
, flags
);
4943 if (force_nonblock
&& ret
== -EAGAIN
)
4945 if (ret
== -ERESTARTSYS
)
4948 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4949 cflags
= io_put_recv_kbuf(req
);
4951 req_set_fail_links(req
);
4952 __io_req_complete(req
, ret
, cflags
, cs
);
4956 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4958 struct io_accept
*accept
= &req
->accept
;
4960 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4962 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4965 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4966 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4967 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4968 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4972 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
4973 struct io_comp_state
*cs
)
4975 struct io_accept
*accept
= &req
->accept
;
4976 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4979 if (req
->file
->f_flags
& O_NONBLOCK
)
4980 req
->flags
|= REQ_F_NOWAIT
;
4982 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4983 accept
->addr_len
, accept
->flags
,
4985 if (ret
== -EAGAIN
&& force_nonblock
)
4988 if (ret
== -ERESTARTSYS
)
4990 req_set_fail_links(req
);
4992 __io_req_complete(req
, ret
, 0, cs
);
4996 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4998 struct io_connect
*conn
= &req
->connect
;
4999 struct io_async_connect
*io
= req
->async_data
;
5001 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5003 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
5006 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5007 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5012 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
5016 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5017 struct io_comp_state
*cs
)
5019 struct io_async_connect __io
, *io
;
5020 unsigned file_flags
;
5023 if (req
->async_data
) {
5024 io
= req
->async_data
;
5026 ret
= move_addr_to_kernel(req
->connect
.addr
,
5027 req
->connect
.addr_len
,
5034 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5036 ret
= __sys_connect_file(req
->file
, &io
->address
,
5037 req
->connect
.addr_len
, file_flags
);
5038 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5039 if (req
->async_data
)
5041 if (io_alloc_async_data(req
)) {
5045 io
= req
->async_data
;
5046 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5049 if (ret
== -ERESTARTSYS
)
5053 req_set_fail_links(req
);
5054 __io_req_complete(req
, ret
, 0, cs
);
5057 #else /* !CONFIG_NET */
5058 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5063 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
5064 struct io_comp_state
*cs
)
5069 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
5070 struct io_comp_state
*cs
)
5075 static int io_recvmsg_prep(struct io_kiocb
*req
,
5076 const struct io_uring_sqe
*sqe
)
5081 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
5082 struct io_comp_state
*cs
)
5087 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
5088 struct io_comp_state
*cs
)
5093 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5098 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
5099 struct io_comp_state
*cs
)
5104 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5109 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5110 struct io_comp_state
*cs
)
5114 #endif /* CONFIG_NET */
5116 struct io_poll_table
{
5117 struct poll_table_struct pt
;
5118 struct io_kiocb
*req
;
5122 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5123 __poll_t mask
, task_work_func_t func
)
5127 /* for instances that support it check for an event match first: */
5128 if (mask
&& !(mask
& poll
->events
))
5131 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5133 list_del_init(&poll
->wait
.entry
);
5136 init_task_work(&req
->task_work
, func
);
5137 percpu_ref_get(&req
->ctx
->refs
);
5140 * If this fails, then the task is exiting. When a task exits, the
5141 * work gets canceled, so just cancel this request as well instead
5142 * of executing it. We can't safely execute it anyway, as we may not
5143 * have the needed state needed for it anyway.
5145 ret
= io_req_task_work_add(req
);
5146 if (unlikely(ret
)) {
5147 struct task_struct
*tsk
;
5149 WRITE_ONCE(poll
->canceled
, true);
5150 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
5151 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
5152 wake_up_process(tsk
);
5157 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5158 __acquires(&req
->ctx
->completion_lock
)
5160 struct io_ring_ctx
*ctx
= req
->ctx
;
5162 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5163 struct poll_table_struct pt
= { ._key
= poll
->events
};
5165 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5168 spin_lock_irq(&ctx
->completion_lock
);
5169 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5170 add_wait_queue(poll
->head
, &poll
->wait
);
5177 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5179 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5180 if (req
->opcode
== IORING_OP_POLL_ADD
)
5181 return req
->async_data
;
5182 return req
->apoll
->double_poll
;
5185 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5187 if (req
->opcode
== IORING_OP_POLL_ADD
)
5189 return &req
->apoll
->poll
;
5192 static void io_poll_remove_double(struct io_kiocb
*req
)
5194 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5196 lockdep_assert_held(&req
->ctx
->completion_lock
);
5198 if (poll
&& poll
->head
) {
5199 struct wait_queue_head
*head
= poll
->head
;
5201 spin_lock(&head
->lock
);
5202 list_del_init(&poll
->wait
.entry
);
5203 if (poll
->wait
.private)
5204 refcount_dec(&req
->refs
);
5206 spin_unlock(&head
->lock
);
5210 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
5212 struct io_ring_ctx
*ctx
= req
->ctx
;
5214 io_poll_remove_double(req
);
5215 req
->poll
.done
= true;
5216 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
5217 io_commit_cqring(ctx
);
5220 static void io_poll_task_func(struct callback_head
*cb
)
5222 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5223 struct io_ring_ctx
*ctx
= req
->ctx
;
5224 struct io_kiocb
*nxt
;
5226 if (io_poll_rewait(req
, &req
->poll
)) {
5227 spin_unlock_irq(&ctx
->completion_lock
);
5229 hash_del(&req
->hash_node
);
5230 io_poll_complete(req
, req
->result
, 0);
5231 spin_unlock_irq(&ctx
->completion_lock
);
5233 nxt
= io_put_req_find_next(req
);
5234 io_cqring_ev_posted(ctx
);
5236 __io_req_task_submit(nxt
);
5239 percpu_ref_put(&ctx
->refs
);
5242 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5243 int sync
, void *key
)
5245 struct io_kiocb
*req
= wait
->private;
5246 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5247 __poll_t mask
= key_to_poll(key
);
5249 /* for instances that support it check for an event match first: */
5250 if (mask
&& !(mask
& poll
->events
))
5253 list_del_init(&wait
->entry
);
5255 if (poll
&& poll
->head
) {
5258 spin_lock(&poll
->head
->lock
);
5259 done
= list_empty(&poll
->wait
.entry
);
5261 list_del_init(&poll
->wait
.entry
);
5262 /* make sure double remove sees this as being gone */
5263 wait
->private = NULL
;
5264 spin_unlock(&poll
->head
->lock
);
5266 /* use wait func handler, so it matches the rq type */
5267 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5270 refcount_dec(&req
->refs
);
5274 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5275 wait_queue_func_t wake_func
)
5279 poll
->canceled
= false;
5280 poll
->events
= events
;
5281 INIT_LIST_HEAD(&poll
->wait
.entry
);
5282 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5285 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5286 struct wait_queue_head
*head
,
5287 struct io_poll_iocb
**poll_ptr
)
5289 struct io_kiocb
*req
= pt
->req
;
5292 * If poll->head is already set, it's because the file being polled
5293 * uses multiple waitqueues for poll handling (eg one for read, one
5294 * for write). Setup a separate io_poll_iocb if this happens.
5296 if (unlikely(poll
->head
)) {
5297 struct io_poll_iocb
*poll_one
= poll
;
5299 /* already have a 2nd entry, fail a third attempt */
5301 pt
->error
= -EINVAL
;
5304 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5306 pt
->error
= -ENOMEM
;
5309 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5310 refcount_inc(&req
->refs
);
5311 poll
->wait
.private = req
;
5318 if (poll
->events
& EPOLLEXCLUSIVE
)
5319 add_wait_queue_exclusive(head
, &poll
->wait
);
5321 add_wait_queue(head
, &poll
->wait
);
5324 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5325 struct poll_table_struct
*p
)
5327 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5328 struct async_poll
*apoll
= pt
->req
->apoll
;
5330 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5333 static void io_async_task_func(struct callback_head
*cb
)
5335 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5336 struct async_poll
*apoll
= req
->apoll
;
5337 struct io_ring_ctx
*ctx
= req
->ctx
;
5339 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5341 if (io_poll_rewait(req
, &apoll
->poll
)) {
5342 spin_unlock_irq(&ctx
->completion_lock
);
5343 percpu_ref_put(&ctx
->refs
);
5347 /* If req is still hashed, it cannot have been canceled. Don't check. */
5348 if (hash_hashed(&req
->hash_node
))
5349 hash_del(&req
->hash_node
);
5351 io_poll_remove_double(req
);
5352 spin_unlock_irq(&ctx
->completion_lock
);
5354 if (!READ_ONCE(apoll
->poll
.canceled
))
5355 __io_req_task_submit(req
);
5357 __io_req_task_cancel(req
, -ECANCELED
);
5359 percpu_ref_put(&ctx
->refs
);
5360 kfree(apoll
->double_poll
);
5364 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5367 struct io_kiocb
*req
= wait
->private;
5368 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5370 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5373 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5376 static void io_poll_req_insert(struct io_kiocb
*req
)
5378 struct io_ring_ctx
*ctx
= req
->ctx
;
5379 struct hlist_head
*list
;
5381 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5382 hlist_add_head(&req
->hash_node
, list
);
5385 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5386 struct io_poll_iocb
*poll
,
5387 struct io_poll_table
*ipt
, __poll_t mask
,
5388 wait_queue_func_t wake_func
)
5389 __acquires(&ctx
->completion_lock
)
5391 struct io_ring_ctx
*ctx
= req
->ctx
;
5392 bool cancel
= false;
5394 INIT_HLIST_NODE(&req
->hash_node
);
5395 io_init_poll_iocb(poll
, mask
, wake_func
);
5396 poll
->file
= req
->file
;
5397 poll
->wait
.private = req
;
5399 ipt
->pt
._key
= mask
;
5401 ipt
->error
= -EINVAL
;
5403 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5405 spin_lock_irq(&ctx
->completion_lock
);
5406 if (likely(poll
->head
)) {
5407 spin_lock(&poll
->head
->lock
);
5408 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5414 if (mask
|| ipt
->error
)
5415 list_del_init(&poll
->wait
.entry
);
5417 WRITE_ONCE(poll
->canceled
, true);
5418 else if (!poll
->done
) /* actually waiting for an event */
5419 io_poll_req_insert(req
);
5420 spin_unlock(&poll
->head
->lock
);
5426 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5428 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5429 struct io_ring_ctx
*ctx
= req
->ctx
;
5430 struct async_poll
*apoll
;
5431 struct io_poll_table ipt
;
5435 if (!req
->file
|| !file_can_poll(req
->file
))
5437 if (req
->flags
& REQ_F_POLLED
)
5441 else if (def
->pollout
)
5445 /* if we can't nonblock try, then no point in arming a poll handler */
5446 if (!io_file_supports_async(req
->file
, rw
))
5449 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5450 if (unlikely(!apoll
))
5452 apoll
->double_poll
= NULL
;
5454 req
->flags
|= REQ_F_POLLED
;
5459 mask
|= POLLIN
| POLLRDNORM
;
5461 mask
|= POLLOUT
| POLLWRNORM
;
5463 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5464 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5465 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5468 mask
|= POLLERR
| POLLPRI
;
5470 ipt
.pt
._qproc
= io_async_queue_proc
;
5472 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5474 if (ret
|| ipt
.error
) {
5475 io_poll_remove_double(req
);
5476 spin_unlock_irq(&ctx
->completion_lock
);
5477 kfree(apoll
->double_poll
);
5481 spin_unlock_irq(&ctx
->completion_lock
);
5482 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5483 apoll
->poll
.events
);
5487 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5488 struct io_poll_iocb
*poll
)
5490 bool do_complete
= false;
5492 spin_lock(&poll
->head
->lock
);
5493 WRITE_ONCE(poll
->canceled
, true);
5494 if (!list_empty(&poll
->wait
.entry
)) {
5495 list_del_init(&poll
->wait
.entry
);
5498 spin_unlock(&poll
->head
->lock
);
5499 hash_del(&req
->hash_node
);
5503 static bool io_poll_remove_one(struct io_kiocb
*req
)
5507 io_poll_remove_double(req
);
5509 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5510 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5512 struct async_poll
*apoll
= req
->apoll
;
5514 /* non-poll requests have submit ref still */
5515 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5518 kfree(apoll
->double_poll
);
5524 io_cqring_fill_event(req
, -ECANCELED
);
5525 io_commit_cqring(req
->ctx
);
5526 req_set_fail_links(req
);
5527 io_put_req_deferred(req
, 1);
5534 * Returns true if we found and killed one or more poll requests
5536 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5537 struct files_struct
*files
)
5539 struct hlist_node
*tmp
;
5540 struct io_kiocb
*req
;
5543 spin_lock_irq(&ctx
->completion_lock
);
5544 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5545 struct hlist_head
*list
;
5547 list
= &ctx
->cancel_hash
[i
];
5548 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5549 if (io_match_task(req
, tsk
, files
))
5550 posted
+= io_poll_remove_one(req
);
5553 spin_unlock_irq(&ctx
->completion_lock
);
5556 io_cqring_ev_posted(ctx
);
5561 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5563 struct hlist_head
*list
;
5564 struct io_kiocb
*req
;
5566 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5567 hlist_for_each_entry(req
, list
, hash_node
) {
5568 if (sqe_addr
!= req
->user_data
)
5570 if (io_poll_remove_one(req
))
5578 static int io_poll_remove_prep(struct io_kiocb
*req
,
5579 const struct io_uring_sqe
*sqe
)
5581 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5583 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5587 req
->poll_remove
.addr
= READ_ONCE(sqe
->addr
);
5592 * Find a running poll command that matches one specified in sqe->addr,
5593 * and remove it if found.
5595 static int io_poll_remove(struct io_kiocb
*req
)
5597 struct io_ring_ctx
*ctx
= req
->ctx
;
5600 spin_lock_irq(&ctx
->completion_lock
);
5601 ret
= io_poll_cancel(ctx
, req
->poll_remove
.addr
);
5602 spin_unlock_irq(&ctx
->completion_lock
);
5605 req_set_fail_links(req
);
5606 io_req_complete(req
, ret
);
5610 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5613 struct io_kiocb
*req
= wait
->private;
5614 struct io_poll_iocb
*poll
= &req
->poll
;
5616 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5619 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5620 struct poll_table_struct
*p
)
5622 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5624 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5627 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5629 struct io_poll_iocb
*poll
= &req
->poll
;
5632 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5634 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5637 events
= READ_ONCE(sqe
->poll32_events
);
5639 events
= swahw32(events
);
5641 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5642 (events
& EPOLLEXCLUSIVE
);
5646 static int io_poll_add(struct io_kiocb
*req
)
5648 struct io_poll_iocb
*poll
= &req
->poll
;
5649 struct io_ring_ctx
*ctx
= req
->ctx
;
5650 struct io_poll_table ipt
;
5653 ipt
.pt
._qproc
= io_poll_queue_proc
;
5655 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5658 if (mask
) { /* no async, we'd stolen it */
5660 io_poll_complete(req
, mask
, 0);
5662 spin_unlock_irq(&ctx
->completion_lock
);
5665 io_cqring_ev_posted(ctx
);
5671 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5673 struct io_timeout_data
*data
= container_of(timer
,
5674 struct io_timeout_data
, timer
);
5675 struct io_kiocb
*req
= data
->req
;
5676 struct io_ring_ctx
*ctx
= req
->ctx
;
5677 unsigned long flags
;
5679 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5680 list_del_init(&req
->timeout
.list
);
5681 atomic_set(&req
->ctx
->cq_timeouts
,
5682 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5684 io_cqring_fill_event(req
, -ETIME
);
5685 io_commit_cqring(ctx
);
5686 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5688 io_cqring_ev_posted(ctx
);
5689 req_set_fail_links(req
);
5691 return HRTIMER_NORESTART
;
5694 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5697 struct io_timeout_data
*io
;
5698 struct io_kiocb
*req
;
5701 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5702 if (user_data
== req
->user_data
) {
5709 return ERR_PTR(ret
);
5711 io
= req
->async_data
;
5712 ret
= hrtimer_try_to_cancel(&io
->timer
);
5714 return ERR_PTR(-EALREADY
);
5715 list_del_init(&req
->timeout
.list
);
5719 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5721 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5724 return PTR_ERR(req
);
5726 req_set_fail_links(req
);
5727 io_cqring_fill_event(req
, -ECANCELED
);
5728 io_put_req_deferred(req
, 1);
5732 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5733 struct timespec64
*ts
, enum hrtimer_mode mode
)
5735 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5736 struct io_timeout_data
*data
;
5739 return PTR_ERR(req
);
5741 req
->timeout
.off
= 0; /* noseq */
5742 data
= req
->async_data
;
5743 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5744 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5745 data
->timer
.function
= io_timeout_fn
;
5746 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5750 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5751 const struct io_uring_sqe
*sqe
)
5753 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5755 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5757 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5759 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
5762 tr
->addr
= READ_ONCE(sqe
->addr
);
5763 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5764 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5765 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5767 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5769 } else if (tr
->flags
) {
5770 /* timeout removal doesn't support flags */
5778 * Remove or update an existing timeout command
5780 static int io_timeout_remove(struct io_kiocb
*req
)
5782 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5783 struct io_ring_ctx
*ctx
= req
->ctx
;
5786 spin_lock_irq(&ctx
->completion_lock
);
5787 if (req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
) {
5788 enum hrtimer_mode mode
= (tr
->flags
& IORING_TIMEOUT_ABS
)
5789 ? HRTIMER_MODE_ABS
: HRTIMER_MODE_REL
;
5791 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
5793 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5796 io_cqring_fill_event(req
, ret
);
5797 io_commit_cqring(ctx
);
5798 spin_unlock_irq(&ctx
->completion_lock
);
5799 io_cqring_ev_posted(ctx
);
5801 req_set_fail_links(req
);
5806 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5807 bool is_timeout_link
)
5809 struct io_timeout_data
*data
;
5811 u32 off
= READ_ONCE(sqe
->off
);
5813 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5815 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5817 if (off
&& is_timeout_link
)
5819 flags
= READ_ONCE(sqe
->timeout_flags
);
5820 if (flags
& ~IORING_TIMEOUT_ABS
)
5823 req
->timeout
.off
= off
;
5825 if (!req
->async_data
&& io_alloc_async_data(req
))
5828 data
= req
->async_data
;
5831 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5834 if (flags
& IORING_TIMEOUT_ABS
)
5835 data
->mode
= HRTIMER_MODE_ABS
;
5837 data
->mode
= HRTIMER_MODE_REL
;
5839 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5843 static int io_timeout(struct io_kiocb
*req
)
5845 struct io_ring_ctx
*ctx
= req
->ctx
;
5846 struct io_timeout_data
*data
= req
->async_data
;
5847 struct list_head
*entry
;
5848 u32 tail
, off
= req
->timeout
.off
;
5850 spin_lock_irq(&ctx
->completion_lock
);
5853 * sqe->off holds how many events that need to occur for this
5854 * timeout event to be satisfied. If it isn't set, then this is
5855 * a pure timeout request, sequence isn't used.
5857 if (io_is_timeout_noseq(req
)) {
5858 entry
= ctx
->timeout_list
.prev
;
5862 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5863 req
->timeout
.target_seq
= tail
+ off
;
5865 /* Update the last seq here in case io_flush_timeouts() hasn't.
5866 * This is safe because ->completion_lock is held, and submissions
5867 * and completions are never mixed in the same ->completion_lock section.
5869 ctx
->cq_last_tm_flush
= tail
;
5872 * Insertion sort, ensuring the first entry in the list is always
5873 * the one we need first.
5875 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5876 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5879 if (io_is_timeout_noseq(nxt
))
5881 /* nxt.seq is behind @tail, otherwise would've been completed */
5882 if (off
>= nxt
->timeout
.target_seq
- tail
)
5886 list_add(&req
->timeout
.list
, entry
);
5887 data
->timer
.function
= io_timeout_fn
;
5888 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5889 spin_unlock_irq(&ctx
->completion_lock
);
5893 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5895 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5897 return req
->user_data
== (unsigned long) data
;
5900 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5902 enum io_wq_cancel cancel_ret
;
5905 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5906 switch (cancel_ret
) {
5907 case IO_WQ_CANCEL_OK
:
5910 case IO_WQ_CANCEL_RUNNING
:
5913 case IO_WQ_CANCEL_NOTFOUND
:
5921 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5922 struct io_kiocb
*req
, __u64 sqe_addr
,
5925 unsigned long flags
;
5928 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5929 if (ret
!= -ENOENT
) {
5930 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5934 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5935 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5938 ret
= io_poll_cancel(ctx
, sqe_addr
);
5942 io_cqring_fill_event(req
, ret
);
5943 io_commit_cqring(ctx
);
5944 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5945 io_cqring_ev_posted(ctx
);
5948 req_set_fail_links(req
);
5952 static int io_async_cancel_prep(struct io_kiocb
*req
,
5953 const struct io_uring_sqe
*sqe
)
5955 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5957 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5959 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5962 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5966 static int io_async_cancel(struct io_kiocb
*req
)
5968 struct io_ring_ctx
*ctx
= req
->ctx
;
5970 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
5974 static int io_files_update_prep(struct io_kiocb
*req
,
5975 const struct io_uring_sqe
*sqe
)
5977 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
5979 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5981 if (sqe
->ioprio
|| sqe
->rw_flags
)
5984 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
5985 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
5986 if (!req
->files_update
.nr_args
)
5988 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
5992 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
,
5993 struct io_comp_state
*cs
)
5995 struct io_ring_ctx
*ctx
= req
->ctx
;
5996 struct io_uring_files_update up
;
6002 up
.offset
= req
->files_update
.offset
;
6003 up
.fds
= req
->files_update
.arg
;
6005 mutex_lock(&ctx
->uring_lock
);
6006 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
6007 mutex_unlock(&ctx
->uring_lock
);
6010 req_set_fail_links(req
);
6011 __io_req_complete(req
, ret
, 0, cs
);
6015 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6017 switch (req
->opcode
) {
6020 case IORING_OP_READV
:
6021 case IORING_OP_READ_FIXED
:
6022 case IORING_OP_READ
:
6023 return io_read_prep(req
, sqe
);
6024 case IORING_OP_WRITEV
:
6025 case IORING_OP_WRITE_FIXED
:
6026 case IORING_OP_WRITE
:
6027 return io_write_prep(req
, sqe
);
6028 case IORING_OP_POLL_ADD
:
6029 return io_poll_add_prep(req
, sqe
);
6030 case IORING_OP_POLL_REMOVE
:
6031 return io_poll_remove_prep(req
, sqe
);
6032 case IORING_OP_FSYNC
:
6033 return io_prep_fsync(req
, sqe
);
6034 case IORING_OP_SYNC_FILE_RANGE
:
6035 return io_prep_sfr(req
, sqe
);
6036 case IORING_OP_SENDMSG
:
6037 case IORING_OP_SEND
:
6038 return io_sendmsg_prep(req
, sqe
);
6039 case IORING_OP_RECVMSG
:
6040 case IORING_OP_RECV
:
6041 return io_recvmsg_prep(req
, sqe
);
6042 case IORING_OP_CONNECT
:
6043 return io_connect_prep(req
, sqe
);
6044 case IORING_OP_TIMEOUT
:
6045 return io_timeout_prep(req
, sqe
, false);
6046 case IORING_OP_TIMEOUT_REMOVE
:
6047 return io_timeout_remove_prep(req
, sqe
);
6048 case IORING_OP_ASYNC_CANCEL
:
6049 return io_async_cancel_prep(req
, sqe
);
6050 case IORING_OP_LINK_TIMEOUT
:
6051 return io_timeout_prep(req
, sqe
, true);
6052 case IORING_OP_ACCEPT
:
6053 return io_accept_prep(req
, sqe
);
6054 case IORING_OP_FALLOCATE
:
6055 return io_fallocate_prep(req
, sqe
);
6056 case IORING_OP_OPENAT
:
6057 return io_openat_prep(req
, sqe
);
6058 case IORING_OP_CLOSE
:
6059 return io_close_prep(req
, sqe
);
6060 case IORING_OP_FILES_UPDATE
:
6061 return io_files_update_prep(req
, sqe
);
6062 case IORING_OP_STATX
:
6063 return io_statx_prep(req
, sqe
);
6064 case IORING_OP_FADVISE
:
6065 return io_fadvise_prep(req
, sqe
);
6066 case IORING_OP_MADVISE
:
6067 return io_madvise_prep(req
, sqe
);
6068 case IORING_OP_OPENAT2
:
6069 return io_openat2_prep(req
, sqe
);
6070 case IORING_OP_EPOLL_CTL
:
6071 return io_epoll_ctl_prep(req
, sqe
);
6072 case IORING_OP_SPLICE
:
6073 return io_splice_prep(req
, sqe
);
6074 case IORING_OP_PROVIDE_BUFFERS
:
6075 return io_provide_buffers_prep(req
, sqe
);
6076 case IORING_OP_REMOVE_BUFFERS
:
6077 return io_remove_buffers_prep(req
, sqe
);
6079 return io_tee_prep(req
, sqe
);
6080 case IORING_OP_SHUTDOWN
:
6081 return io_shutdown_prep(req
, sqe
);
6082 case IORING_OP_RENAMEAT
:
6083 return io_renameat_prep(req
, sqe
);
6084 case IORING_OP_UNLINKAT
:
6085 return io_unlinkat_prep(req
, sqe
);
6088 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6093 static int io_req_defer_prep(struct io_kiocb
*req
,
6094 const struct io_uring_sqe
*sqe
)
6098 if (io_alloc_async_data(req
))
6100 return io_req_prep(req
, sqe
);
6103 static u32
io_get_sequence(struct io_kiocb
*req
)
6105 struct io_kiocb
*pos
;
6106 struct io_ring_ctx
*ctx
= req
->ctx
;
6107 u32 total_submitted
, nr_reqs
= 0;
6109 io_for_each_link(pos
, req
)
6112 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
6113 return total_submitted
- nr_reqs
;
6116 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6118 struct io_ring_ctx
*ctx
= req
->ctx
;
6119 struct io_defer_entry
*de
;
6123 /* Still need defer if there is pending req in defer list. */
6124 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6125 !(req
->flags
& REQ_F_IO_DRAIN
)))
6128 seq
= io_get_sequence(req
);
6129 /* Still a chance to pass the sequence check */
6130 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6133 if (!req
->async_data
) {
6134 ret
= io_req_defer_prep(req
, sqe
);
6138 io_prep_async_link(req
);
6139 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6143 spin_lock_irq(&ctx
->completion_lock
);
6144 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6145 spin_unlock_irq(&ctx
->completion_lock
);
6147 io_queue_async_work(req
);
6148 return -EIOCBQUEUED
;
6151 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6154 list_add_tail(&de
->list
, &ctx
->defer_list
);
6155 spin_unlock_irq(&ctx
->completion_lock
);
6156 return -EIOCBQUEUED
;
6159 static void io_req_drop_files(struct io_kiocb
*req
)
6161 struct io_ring_ctx
*ctx
= req
->ctx
;
6162 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6163 unsigned long flags
;
6165 put_files_struct(req
->work
.identity
->files
);
6166 put_nsproxy(req
->work
.identity
->nsproxy
);
6167 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
6168 list_del(&req
->inflight_entry
);
6169 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
6170 req
->flags
&= ~REQ_F_INFLIGHT
;
6171 req
->work
.flags
&= ~IO_WQ_WORK_FILES
;
6172 if (atomic_read(&tctx
->in_idle
))
6173 wake_up(&tctx
->wait
);
6176 static void __io_clean_op(struct io_kiocb
*req
)
6178 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6179 switch (req
->opcode
) {
6180 case IORING_OP_READV
:
6181 case IORING_OP_READ_FIXED
:
6182 case IORING_OP_READ
:
6183 kfree((void *)(unsigned long)req
->rw
.addr
);
6185 case IORING_OP_RECVMSG
:
6186 case IORING_OP_RECV
:
6187 kfree(req
->sr_msg
.kbuf
);
6190 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
6193 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6194 switch (req
->opcode
) {
6195 case IORING_OP_READV
:
6196 case IORING_OP_READ_FIXED
:
6197 case IORING_OP_READ
:
6198 case IORING_OP_WRITEV
:
6199 case IORING_OP_WRITE_FIXED
:
6200 case IORING_OP_WRITE
: {
6201 struct io_async_rw
*io
= req
->async_data
;
6203 kfree(io
->free_iovec
);
6206 case IORING_OP_RECVMSG
:
6207 case IORING_OP_SENDMSG
: {
6208 struct io_async_msghdr
*io
= req
->async_data
;
6209 if (io
->iov
!= io
->fast_iov
)
6213 case IORING_OP_SPLICE
:
6215 io_put_file(req
, req
->splice
.file_in
,
6216 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
6218 case IORING_OP_OPENAT
:
6219 case IORING_OP_OPENAT2
:
6220 if (req
->open
.filename
)
6221 putname(req
->open
.filename
);
6223 case IORING_OP_RENAMEAT
:
6224 putname(req
->rename
.oldpath
);
6225 putname(req
->rename
.newpath
);
6227 case IORING_OP_UNLINKAT
:
6228 putname(req
->unlink
.filename
);
6231 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6234 if (req
->flags
& REQ_F_INFLIGHT
)
6235 io_req_drop_files(req
);
6238 static int io_issue_sqe(struct io_kiocb
*req
, bool force_nonblock
,
6239 struct io_comp_state
*cs
)
6241 struct io_ring_ctx
*ctx
= req
->ctx
;
6244 switch (req
->opcode
) {
6246 ret
= io_nop(req
, cs
);
6248 case IORING_OP_READV
:
6249 case IORING_OP_READ_FIXED
:
6250 case IORING_OP_READ
:
6251 ret
= io_read(req
, force_nonblock
, cs
);
6253 case IORING_OP_WRITEV
:
6254 case IORING_OP_WRITE_FIXED
:
6255 case IORING_OP_WRITE
:
6256 ret
= io_write(req
, force_nonblock
, cs
);
6258 case IORING_OP_FSYNC
:
6259 ret
= io_fsync(req
, force_nonblock
);
6261 case IORING_OP_POLL_ADD
:
6262 ret
= io_poll_add(req
);
6264 case IORING_OP_POLL_REMOVE
:
6265 ret
= io_poll_remove(req
);
6267 case IORING_OP_SYNC_FILE_RANGE
:
6268 ret
= io_sync_file_range(req
, force_nonblock
);
6270 case IORING_OP_SENDMSG
:
6271 ret
= io_sendmsg(req
, force_nonblock
, cs
);
6273 case IORING_OP_SEND
:
6274 ret
= io_send(req
, force_nonblock
, cs
);
6276 case IORING_OP_RECVMSG
:
6277 ret
= io_recvmsg(req
, force_nonblock
, cs
);
6279 case IORING_OP_RECV
:
6280 ret
= io_recv(req
, force_nonblock
, cs
);
6282 case IORING_OP_TIMEOUT
:
6283 ret
= io_timeout(req
);
6285 case IORING_OP_TIMEOUT_REMOVE
:
6286 ret
= io_timeout_remove(req
);
6288 case IORING_OP_ACCEPT
:
6289 ret
= io_accept(req
, force_nonblock
, cs
);
6291 case IORING_OP_CONNECT
:
6292 ret
= io_connect(req
, force_nonblock
, cs
);
6294 case IORING_OP_ASYNC_CANCEL
:
6295 ret
= io_async_cancel(req
);
6297 case IORING_OP_FALLOCATE
:
6298 ret
= io_fallocate(req
, force_nonblock
);
6300 case IORING_OP_OPENAT
:
6301 ret
= io_openat(req
, force_nonblock
);
6303 case IORING_OP_CLOSE
:
6304 ret
= io_close(req
, force_nonblock
, cs
);
6306 case IORING_OP_FILES_UPDATE
:
6307 ret
= io_files_update(req
, force_nonblock
, cs
);
6309 case IORING_OP_STATX
:
6310 ret
= io_statx(req
, force_nonblock
);
6312 case IORING_OP_FADVISE
:
6313 ret
= io_fadvise(req
, force_nonblock
);
6315 case IORING_OP_MADVISE
:
6316 ret
= io_madvise(req
, force_nonblock
);
6318 case IORING_OP_OPENAT2
:
6319 ret
= io_openat2(req
, force_nonblock
);
6321 case IORING_OP_EPOLL_CTL
:
6322 ret
= io_epoll_ctl(req
, force_nonblock
, cs
);
6324 case IORING_OP_SPLICE
:
6325 ret
= io_splice(req
, force_nonblock
);
6327 case IORING_OP_PROVIDE_BUFFERS
:
6328 ret
= io_provide_buffers(req
, force_nonblock
, cs
);
6330 case IORING_OP_REMOVE_BUFFERS
:
6331 ret
= io_remove_buffers(req
, force_nonblock
, cs
);
6334 ret
= io_tee(req
, force_nonblock
);
6336 case IORING_OP_SHUTDOWN
:
6337 ret
= io_shutdown(req
, force_nonblock
);
6339 case IORING_OP_RENAMEAT
:
6340 ret
= io_renameat(req
, force_nonblock
);
6342 case IORING_OP_UNLINKAT
:
6343 ret
= io_unlinkat(req
, force_nonblock
);
6353 /* If the op doesn't have a file, we're not polling for it */
6354 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6355 const bool in_async
= io_wq_current_is_worker();
6357 /* workqueue context doesn't hold uring_lock, grab it now */
6359 mutex_lock(&ctx
->uring_lock
);
6361 io_iopoll_req_issued(req
, in_async
);
6364 mutex_unlock(&ctx
->uring_lock
);
6370 static struct io_wq_work
*io_wq_submit_work(struct io_wq_work
*work
)
6372 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6373 struct io_kiocb
*timeout
;
6376 timeout
= io_prep_linked_timeout(req
);
6378 io_queue_linked_timeout(timeout
);
6380 /* if NO_CANCEL is set, we must still run the work */
6381 if ((work
->flags
& (IO_WQ_WORK_CANCEL
|IO_WQ_WORK_NO_CANCEL
)) ==
6382 IO_WQ_WORK_CANCEL
) {
6388 ret
= io_issue_sqe(req
, false, NULL
);
6390 * We can get EAGAIN for polled IO even though we're
6391 * forcing a sync submission from here, since we can't
6392 * wait for request slots on the block side.
6401 struct io_ring_ctx
*lock_ctx
= NULL
;
6403 if (req
->ctx
->flags
& IORING_SETUP_IOPOLL
)
6404 lock_ctx
= req
->ctx
;
6407 * io_iopoll_complete() does not hold completion_lock to
6408 * complete polled io, so here for polled io, we can not call
6409 * io_req_complete() directly, otherwise there maybe concurrent
6410 * access to cqring, defer_list, etc, which is not safe. Given
6411 * that io_iopoll_complete() is always called under uring_lock,
6412 * so here for polled io, we also get uring_lock to complete
6416 mutex_lock(&lock_ctx
->uring_lock
);
6418 req_set_fail_links(req
);
6419 io_req_complete(req
, ret
);
6422 mutex_unlock(&lock_ctx
->uring_lock
);
6425 return io_steal_work(req
);
6428 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6431 struct fixed_file_table
*table
;
6433 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6434 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6437 static struct file
*io_file_get(struct io_submit_state
*state
,
6438 struct io_kiocb
*req
, int fd
, bool fixed
)
6440 struct io_ring_ctx
*ctx
= req
->ctx
;
6444 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6446 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6447 file
= io_file_from_index(ctx
, fd
);
6448 io_set_resource_node(req
);
6450 trace_io_uring_file_get(ctx
, fd
);
6451 file
= __io_file_get(state
, fd
);
6457 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6459 struct io_timeout_data
*data
= container_of(timer
,
6460 struct io_timeout_data
, timer
);
6461 struct io_kiocb
*prev
, *req
= data
->req
;
6462 struct io_ring_ctx
*ctx
= req
->ctx
;
6463 unsigned long flags
;
6465 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6466 prev
= req
->timeout
.head
;
6467 req
->timeout
.head
= NULL
;
6470 * We don't expect the list to be empty, that will only happen if we
6471 * race with the completion of the linked work.
6473 if (prev
&& refcount_inc_not_zero(&prev
->refs
))
6474 io_remove_next_linked(prev
);
6477 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6480 req_set_fail_links(prev
);
6481 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6484 io_req_complete(req
, -ETIME
);
6486 return HRTIMER_NORESTART
;
6489 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6492 * If the back reference is NULL, then our linked request finished
6493 * before we got a chance to setup the timer
6495 if (req
->timeout
.head
) {
6496 struct io_timeout_data
*data
= req
->async_data
;
6498 data
->timer
.function
= io_link_timeout_fn
;
6499 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6504 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6506 struct io_ring_ctx
*ctx
= req
->ctx
;
6508 spin_lock_irq(&ctx
->completion_lock
);
6509 __io_queue_linked_timeout(req
);
6510 spin_unlock_irq(&ctx
->completion_lock
);
6512 /* drop submission reference */
6516 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6518 struct io_kiocb
*nxt
= req
->link
;
6520 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6521 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6524 nxt
->timeout
.head
= req
;
6525 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6526 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6530 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
)
6532 struct io_kiocb
*linked_timeout
;
6533 const struct cred
*old_creds
= NULL
;
6537 linked_timeout
= io_prep_linked_timeout(req
);
6539 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
6540 (req
->work
.flags
& IO_WQ_WORK_CREDS
) &&
6541 req
->work
.identity
->creds
!= current_cred()) {
6543 revert_creds(old_creds
);
6544 if (old_creds
== req
->work
.identity
->creds
)
6545 old_creds
= NULL
; /* restored original creds */
6547 old_creds
= override_creds(req
->work
.identity
->creds
);
6550 ret
= io_issue_sqe(req
, true, cs
);
6553 * We async punt it if the file wasn't marked NOWAIT, or if the file
6554 * doesn't support non-blocking read/write attempts
6556 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6557 if (!io_arm_poll_handler(req
)) {
6559 * Queued up for async execution, worker will release
6560 * submit reference when the iocb is actually submitted.
6562 io_queue_async_work(req
);
6566 io_queue_linked_timeout(linked_timeout
);
6567 } else if (likely(!ret
)) {
6568 /* drop submission reference */
6569 req
= io_put_req_find_next(req
);
6571 io_queue_linked_timeout(linked_timeout
);
6574 if (!(req
->flags
& REQ_F_FORCE_ASYNC
))
6576 io_queue_async_work(req
);
6579 /* un-prep timeout, so it'll be killed as any other linked */
6580 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
6581 req_set_fail_links(req
);
6583 io_req_complete(req
, ret
);
6587 revert_creds(old_creds
);
6590 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6591 struct io_comp_state
*cs
)
6595 ret
= io_req_defer(req
, sqe
);
6597 if (ret
!= -EIOCBQUEUED
) {
6599 req_set_fail_links(req
);
6601 io_req_complete(req
, ret
);
6603 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6604 if (!req
->async_data
) {
6605 ret
= io_req_defer_prep(req
, sqe
);
6609 io_queue_async_work(req
);
6612 ret
= io_req_prep(req
, sqe
);
6616 __io_queue_sqe(req
, cs
);
6620 static inline void io_queue_link_head(struct io_kiocb
*req
,
6621 struct io_comp_state
*cs
)
6623 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
6625 io_req_complete(req
, -ECANCELED
);
6627 io_queue_sqe(req
, NULL
, cs
);
6630 struct io_submit_link
{
6631 struct io_kiocb
*head
;
6632 struct io_kiocb
*last
;
6635 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6636 struct io_submit_link
*link
, struct io_comp_state
*cs
)
6638 struct io_ring_ctx
*ctx
= req
->ctx
;
6642 * If we already have a head request, queue this one for async
6643 * submittal once the head completes. If we don't have a head but
6644 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6645 * submitted sync once the chain is complete. If none of those
6646 * conditions are true (normal request), then just queue it.
6649 struct io_kiocb
*head
= link
->head
;
6652 * Taking sequential execution of a link, draining both sides
6653 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6654 * requests in the link. So, it drains the head and the
6655 * next after the link request. The last one is done via
6656 * drain_next flag to persist the effect across calls.
6658 if (req
->flags
& REQ_F_IO_DRAIN
) {
6659 head
->flags
|= REQ_F_IO_DRAIN
;
6660 ctx
->drain_next
= 1;
6662 ret
= io_req_defer_prep(req
, sqe
);
6663 if (unlikely(ret
)) {
6664 /* fail even hard links since we don't submit */
6665 head
->flags
|= REQ_F_FAIL_LINK
;
6668 trace_io_uring_link(ctx
, req
, head
);
6669 link
->last
->link
= req
;
6672 /* last request of a link, enqueue the link */
6673 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6674 io_queue_link_head(head
, cs
);
6678 if (unlikely(ctx
->drain_next
)) {
6679 req
->flags
|= REQ_F_IO_DRAIN
;
6680 ctx
->drain_next
= 0;
6682 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6683 ret
= io_req_defer_prep(req
, sqe
);
6685 req
->flags
|= REQ_F_FAIL_LINK
;
6689 io_queue_sqe(req
, sqe
, cs
);
6697 * Batched submission is done, ensure local IO is flushed out.
6699 static void io_submit_state_end(struct io_submit_state
*state
)
6701 if (!list_empty(&state
->comp
.list
))
6702 io_submit_flush_completions(&state
->comp
);
6703 if (state
->plug_started
)
6704 blk_finish_plug(&state
->plug
);
6705 io_state_file_put(state
);
6706 if (state
->free_reqs
)
6707 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6711 * Start submission side cache.
6713 static void io_submit_state_start(struct io_submit_state
*state
,
6714 struct io_ring_ctx
*ctx
, unsigned int max_ios
)
6716 state
->plug_started
= false;
6718 INIT_LIST_HEAD(&state
->comp
.list
);
6719 state
->comp
.ctx
= ctx
;
6720 state
->free_reqs
= 0;
6721 state
->file_refs
= 0;
6722 state
->ios_left
= max_ios
;
6725 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6727 struct io_rings
*rings
= ctx
->rings
;
6730 * Ensure any loads from the SQEs are done at this point,
6731 * since once we write the new head, the application could
6732 * write new data to them.
6734 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6738 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6739 * that is mapped by userspace. This means that care needs to be taken to
6740 * ensure that reads are stable, as we cannot rely on userspace always
6741 * being a good citizen. If members of the sqe are validated and then later
6742 * used, it's important that those reads are done through READ_ONCE() to
6743 * prevent a re-load down the line.
6745 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6747 u32
*sq_array
= ctx
->sq_array
;
6751 * The cached sq head (or cq tail) serves two purposes:
6753 * 1) allows us to batch the cost of updating the user visible
6755 * 2) allows the kernel side to track the head on its own, even
6756 * though the application is the one updating it.
6758 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6759 if (likely(head
< ctx
->sq_entries
))
6760 return &ctx
->sq_sqes
[head
];
6762 /* drop invalid entries */
6763 ctx
->cached_sq_dropped
++;
6764 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6768 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6770 ctx
->cached_sq_head
++;
6774 * Check SQE restrictions (opcode and flags).
6776 * Returns 'true' if SQE is allowed, 'false' otherwise.
6778 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6779 struct io_kiocb
*req
,
6780 unsigned int sqe_flags
)
6782 if (!ctx
->restricted
)
6785 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6788 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6789 ctx
->restrictions
.sqe_flags_required
)
6792 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6793 ctx
->restrictions
.sqe_flags_required
))
6799 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6800 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6801 IOSQE_BUFFER_SELECT)
6803 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6804 const struct io_uring_sqe
*sqe
,
6805 struct io_submit_state
*state
)
6807 unsigned int sqe_flags
;
6810 req
->opcode
= READ_ONCE(sqe
->opcode
);
6811 req
->user_data
= READ_ONCE(sqe
->user_data
);
6812 req
->async_data
= NULL
;
6817 req
->fixed_file_refs
= NULL
;
6818 /* one is dropped after submission, the other at completion */
6819 refcount_set(&req
->refs
, 2);
6820 req
->task
= current
;
6823 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6826 if (unlikely(io_sq_thread_acquire_mm_files(ctx
, req
)))
6829 sqe_flags
= READ_ONCE(sqe
->flags
);
6830 /* enforce forwards compatibility on users */
6831 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6834 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6837 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6838 !io_op_defs
[req
->opcode
].buffer_select
)
6841 id
= READ_ONCE(sqe
->personality
);
6843 struct io_identity
*iod
;
6845 iod
= idr_find(&ctx
->personality_idr
, id
);
6848 refcount_inc(&iod
->count
);
6850 __io_req_init_async(req
);
6851 get_cred(iod
->creds
);
6852 req
->work
.identity
= iod
;
6853 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
6856 /* same numerical values with corresponding REQ_F_*, safe to copy */
6857 req
->flags
|= sqe_flags
;
6860 * Plug now if we have more than 1 IO left after this, and the target
6861 * is potentially a read/write to block based storage.
6863 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6864 io_op_defs
[req
->opcode
].plug
) {
6865 blk_start_plug(&state
->plug
);
6866 state
->plug_started
= true;
6870 if (io_op_defs
[req
->opcode
].needs_file
) {
6871 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6873 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6874 if (unlikely(!req
->file
&&
6875 !io_op_defs
[req
->opcode
].needs_file_no_error
))
6883 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6885 struct io_submit_state state
;
6886 struct io_submit_link link
;
6887 int i
, submitted
= 0;
6889 /* if we have a backlog and couldn't flush it all, return BUSY */
6890 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6891 if (!__io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6895 /* make sure SQ entry isn't read before tail */
6896 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6898 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6901 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6902 refcount_add(nr
, ¤t
->usage
);
6904 io_submit_state_start(&state
, ctx
, nr
);
6907 for (i
= 0; i
< nr
; i
++) {
6908 const struct io_uring_sqe
*sqe
;
6909 struct io_kiocb
*req
;
6912 sqe
= io_get_sqe(ctx
);
6913 if (unlikely(!sqe
)) {
6914 io_consume_sqe(ctx
);
6917 req
= io_alloc_req(ctx
, &state
);
6918 if (unlikely(!req
)) {
6920 submitted
= -EAGAIN
;
6923 io_consume_sqe(ctx
);
6924 /* will complete beyond this point, count as submitted */
6927 err
= io_init_req(ctx
, req
, sqe
, &state
);
6928 if (unlikely(err
)) {
6931 io_req_complete(req
, err
);
6935 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6936 true, io_async_submit(ctx
));
6937 err
= io_submit_sqe(req
, sqe
, &link
, &state
.comp
);
6942 if (unlikely(submitted
!= nr
)) {
6943 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6944 struct io_uring_task
*tctx
= current
->io_uring
;
6945 int unused
= nr
- ref_used
;
6947 percpu_ref_put_many(&ctx
->refs
, unused
);
6948 percpu_counter_sub(&tctx
->inflight
, unused
);
6949 put_task_struct_many(current
, unused
);
6952 io_queue_link_head(link
.head
, &state
.comp
);
6953 io_submit_state_end(&state
);
6955 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6956 io_commit_sqring(ctx
);
6961 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6963 /* Tell userspace we may need a wakeup call */
6964 spin_lock_irq(&ctx
->completion_lock
);
6965 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6966 spin_unlock_irq(&ctx
->completion_lock
);
6969 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6971 spin_lock_irq(&ctx
->completion_lock
);
6972 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6973 spin_unlock_irq(&ctx
->completion_lock
);
6976 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
6978 unsigned int to_submit
;
6981 to_submit
= io_sqring_entries(ctx
);
6982 /* if we're handling multiple rings, cap submit size for fairness */
6983 if (cap_entries
&& to_submit
> 8)
6986 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
6987 unsigned nr_events
= 0;
6989 mutex_lock(&ctx
->uring_lock
);
6990 if (!list_empty(&ctx
->iopoll_list
))
6991 io_do_iopoll(ctx
, &nr_events
, 0);
6993 if (to_submit
&& !ctx
->sqo_dead
&&
6994 likely(!percpu_ref_is_dying(&ctx
->refs
)))
6995 ret
= io_submit_sqes(ctx
, to_submit
);
6996 mutex_unlock(&ctx
->uring_lock
);
6999 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
7000 wake_up(&ctx
->sqo_sq_wait
);
7005 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7007 struct io_ring_ctx
*ctx
;
7008 unsigned sq_thread_idle
= 0;
7010 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7011 if (sq_thread_idle
< ctx
->sq_thread_idle
)
7012 sq_thread_idle
= ctx
->sq_thread_idle
;
7015 sqd
->sq_thread_idle
= sq_thread_idle
;
7018 static void io_sqd_init_new(struct io_sq_data
*sqd
)
7020 struct io_ring_ctx
*ctx
;
7022 while (!list_empty(&sqd
->ctx_new_list
)) {
7023 ctx
= list_first_entry(&sqd
->ctx_new_list
, struct io_ring_ctx
, sqd_list
);
7024 list_move_tail(&ctx
->sqd_list
, &sqd
->ctx_list
);
7025 complete(&ctx
->sq_thread_comp
);
7028 io_sqd_update_thread_idle(sqd
);
7031 static int io_sq_thread(void *data
)
7033 struct cgroup_subsys_state
*cur_css
= NULL
;
7034 struct files_struct
*old_files
= current
->files
;
7035 struct nsproxy
*old_nsproxy
= current
->nsproxy
;
7036 const struct cred
*old_cred
= NULL
;
7037 struct io_sq_data
*sqd
= data
;
7038 struct io_ring_ctx
*ctx
;
7039 unsigned long timeout
= 0;
7043 current
->files
= NULL
;
7044 current
->nsproxy
= NULL
;
7045 task_unlock(current
);
7047 while (!kthread_should_stop()) {
7049 bool cap_entries
, sqt_spin
, needs_sched
;
7052 * Any changes to the sqd lists are synchronized through the
7053 * kthread parking. This synchronizes the thread vs users,
7054 * the users are synchronized on the sqd->ctx_lock.
7056 if (kthread_should_park()) {
7059 * When sq thread is unparked, in case the previous park operation
7060 * comes from io_put_sq_data(), which means that sq thread is going
7061 * to be stopped, so here needs to have a check.
7063 if (kthread_should_stop())
7067 if (unlikely(!list_empty(&sqd
->ctx_new_list
))) {
7068 io_sqd_init_new(sqd
);
7069 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7073 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7074 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7075 if (current
->cred
!= ctx
->creds
) {
7077 revert_creds(old_cred
);
7078 old_cred
= override_creds(ctx
->creds
);
7080 io_sq_thread_associate_blkcg(ctx
, &cur_css
);
7082 current
->loginuid
= ctx
->loginuid
;
7083 current
->sessionid
= ctx
->sessionid
;
7086 ret
= __io_sq_thread(ctx
, cap_entries
);
7087 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7090 io_sq_thread_drop_mm_files();
7093 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7095 io_sq_thread_drop_mm_files();
7098 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7102 if (kthread_should_park())
7106 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7107 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7108 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7109 !list_empty_careful(&ctx
->iopoll_list
)) {
7110 needs_sched
= false;
7113 if (io_sqring_entries(ctx
)) {
7114 needs_sched
= false;
7120 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7121 io_ring_set_wakeup_flag(ctx
);
7124 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7125 io_ring_clear_wakeup_flag(ctx
);
7128 finish_wait(&sqd
->wait
, &wait
);
7129 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7133 io_sq_thread_drop_mm_files();
7136 io_sq_thread_unassociate_blkcg();
7138 revert_creds(old_cred
);
7141 current
->files
= old_files
;
7142 current
->nsproxy
= old_nsproxy
;
7143 task_unlock(current
);
7150 struct io_wait_queue
{
7151 struct wait_queue_entry wq
;
7152 struct io_ring_ctx
*ctx
;
7154 unsigned nr_timeouts
;
7157 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7159 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7162 * Wake up if we have enough events, or if a timeout occurred since we
7163 * started waiting. For timeouts, we always want to return to userspace,
7164 * regardless of event count.
7166 return io_cqring_events(ctx
) >= iowq
->to_wait
||
7167 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7170 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7171 int wake_flags
, void *key
)
7173 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7177 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7178 * the task, and the next invocation will do it.
7180 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->cq_check_overflow
))
7181 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7185 static int io_run_task_work_sig(void)
7187 if (io_run_task_work())
7189 if (!signal_pending(current
))
7191 if (test_tsk_thread_flag(current
, TIF_NOTIFY_SIGNAL
))
7192 return -ERESTARTSYS
;
7197 * Wait until events become available, if we don't already have some. The
7198 * application must reap them itself, as they reside on the shared cq ring.
7200 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7201 const sigset_t __user
*sig
, size_t sigsz
,
7202 struct __kernel_timespec __user
*uts
)
7204 struct io_wait_queue iowq
= {
7207 .func
= io_wake_function
,
7208 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
7211 .to_wait
= min_events
,
7213 struct io_rings
*rings
= ctx
->rings
;
7214 struct timespec64 ts
;
7215 signed long timeout
= 0;
7219 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7220 if (io_cqring_events(ctx
) >= min_events
)
7222 if (!io_run_task_work())
7227 #ifdef CONFIG_COMPAT
7228 if (in_compat_syscall())
7229 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7233 ret
= set_user_sigmask(sig
, sigsz
);
7240 if (get_timespec64(&ts
, uts
))
7242 timeout
= timespec64_to_jiffies(&ts
);
7245 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7246 trace_io_uring_cqring_wait(ctx
, min_events
);
7248 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7249 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
7250 TASK_INTERRUPTIBLE
);
7251 /* make sure we run task_work before checking for signals */
7252 ret
= io_run_task_work_sig();
7257 if (io_should_wake(&iowq
))
7259 if (test_bit(0, &ctx
->cq_check_overflow
))
7262 timeout
= schedule_timeout(timeout
);
7271 finish_wait(&ctx
->wait
, &iowq
.wq
);
7273 restore_saved_sigmask_unless(ret
== -EINTR
);
7275 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7278 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7280 #if defined(CONFIG_UNIX)
7281 if (ctx
->ring_sock
) {
7282 struct sock
*sock
= ctx
->ring_sock
->sk
;
7283 struct sk_buff
*skb
;
7285 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7291 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7294 file
= io_file_from_index(ctx
, i
);
7301 static void io_file_ref_kill(struct percpu_ref
*ref
)
7303 struct fixed_file_data
*data
;
7305 data
= container_of(ref
, struct fixed_file_data
, refs
);
7306 complete(&data
->done
);
7309 static void io_sqe_files_set_node(struct fixed_file_data
*file_data
,
7310 struct fixed_file_ref_node
*ref_node
)
7312 spin_lock_bh(&file_data
->lock
);
7313 file_data
->node
= ref_node
;
7314 list_add_tail(&ref_node
->node
, &file_data
->ref_list
);
7315 spin_unlock_bh(&file_data
->lock
);
7316 percpu_ref_get(&file_data
->refs
);
7319 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7321 struct fixed_file_data
*data
= ctx
->file_data
;
7322 struct fixed_file_ref_node
*backup_node
, *ref_node
= NULL
;
7323 unsigned nr_tables
, i
;
7328 backup_node
= alloc_fixed_file_ref_node(ctx
);
7332 spin_lock_bh(&data
->lock
);
7333 ref_node
= data
->node
;
7334 spin_unlock_bh(&data
->lock
);
7336 percpu_ref_kill(&ref_node
->refs
);
7338 percpu_ref_kill(&data
->refs
);
7340 /* wait for all refs nodes to complete */
7341 flush_delayed_work(&ctx
->file_put_work
);
7343 ret
= wait_for_completion_interruptible(&data
->done
);
7346 ret
= io_run_task_work_sig();
7348 percpu_ref_resurrect(&data
->refs
);
7349 reinit_completion(&data
->done
);
7350 io_sqe_files_set_node(data
, backup_node
);
7355 __io_sqe_files_unregister(ctx
);
7356 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
7357 for (i
= 0; i
< nr_tables
; i
++)
7358 kfree(data
->table
[i
].files
);
7360 percpu_ref_exit(&data
->refs
);
7362 ctx
->file_data
= NULL
;
7363 ctx
->nr_user_files
= 0;
7364 destroy_fixed_file_ref_node(backup_node
);
7368 static void io_put_sq_data(struct io_sq_data
*sqd
)
7370 if (refcount_dec_and_test(&sqd
->refs
)) {
7372 * The park is a bit of a work-around, without it we get
7373 * warning spews on shutdown with SQPOLL set and affinity
7374 * set to a single CPU.
7377 kthread_park(sqd
->thread
);
7378 kthread_stop(sqd
->thread
);
7385 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7387 struct io_ring_ctx
*ctx_attach
;
7388 struct io_sq_data
*sqd
;
7391 f
= fdget(p
->wq_fd
);
7393 return ERR_PTR(-ENXIO
);
7394 if (f
.file
->f_op
!= &io_uring_fops
) {
7396 return ERR_PTR(-EINVAL
);
7399 ctx_attach
= f
.file
->private_data
;
7400 sqd
= ctx_attach
->sq_data
;
7403 return ERR_PTR(-EINVAL
);
7406 refcount_inc(&sqd
->refs
);
7411 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
)
7413 struct io_sq_data
*sqd
;
7415 if (p
->flags
& IORING_SETUP_ATTACH_WQ
)
7416 return io_attach_sq_data(p
);
7418 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7420 return ERR_PTR(-ENOMEM
);
7422 refcount_set(&sqd
->refs
, 1);
7423 INIT_LIST_HEAD(&sqd
->ctx_list
);
7424 INIT_LIST_HEAD(&sqd
->ctx_new_list
);
7425 mutex_init(&sqd
->ctx_lock
);
7426 mutex_init(&sqd
->lock
);
7427 init_waitqueue_head(&sqd
->wait
);
7431 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7432 __releases(&sqd
->lock
)
7436 kthread_unpark(sqd
->thread
);
7437 mutex_unlock(&sqd
->lock
);
7440 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7441 __acquires(&sqd
->lock
)
7445 mutex_lock(&sqd
->lock
);
7446 kthread_park(sqd
->thread
);
7449 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
7451 struct io_sq_data
*sqd
= ctx
->sq_data
;
7456 * We may arrive here from the error branch in
7457 * io_sq_offload_create() where the kthread is created
7458 * without being waked up, thus wake it up now to make
7459 * sure the wait will complete.
7461 wake_up_process(sqd
->thread
);
7462 wait_for_completion(&ctx
->sq_thread_comp
);
7464 io_sq_thread_park(sqd
);
7467 mutex_lock(&sqd
->ctx_lock
);
7468 list_del(&ctx
->sqd_list
);
7469 io_sqd_update_thread_idle(sqd
);
7470 mutex_unlock(&sqd
->ctx_lock
);
7473 io_sq_thread_unpark(sqd
);
7475 io_put_sq_data(sqd
);
7476 ctx
->sq_data
= NULL
;
7480 static void io_finish_async(struct io_ring_ctx
*ctx
)
7482 io_sq_thread_stop(ctx
);
7485 io_wq_destroy(ctx
->io_wq
);
7490 #if defined(CONFIG_UNIX)
7492 * Ensure the UNIX gc is aware of our file set, so we are certain that
7493 * the io_uring can be safely unregistered on process exit, even if we have
7494 * loops in the file referencing.
7496 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7498 struct sock
*sk
= ctx
->ring_sock
->sk
;
7499 struct scm_fp_list
*fpl
;
7500 struct sk_buff
*skb
;
7503 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7507 skb
= alloc_skb(0, GFP_KERNEL
);
7516 fpl
->user
= get_uid(ctx
->user
);
7517 for (i
= 0; i
< nr
; i
++) {
7518 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7522 fpl
->fp
[nr_files
] = get_file(file
);
7523 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7528 fpl
->max
= SCM_MAX_FD
;
7529 fpl
->count
= nr_files
;
7530 UNIXCB(skb
).fp
= fpl
;
7531 skb
->destructor
= unix_destruct_scm
;
7532 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7533 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7535 for (i
= 0; i
< nr_files
; i
++)
7546 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7547 * causes regular reference counting to break down. We rely on the UNIX
7548 * garbage collection to take care of this problem for us.
7550 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7552 unsigned left
, total
;
7556 left
= ctx
->nr_user_files
;
7558 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7560 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7564 total
+= this_files
;
7570 while (total
< ctx
->nr_user_files
) {
7571 struct file
*file
= io_file_from_index(ctx
, total
);
7581 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7587 static int io_sqe_alloc_file_tables(struct fixed_file_data
*file_data
,
7588 unsigned nr_tables
, unsigned nr_files
)
7592 for (i
= 0; i
< nr_tables
; i
++) {
7593 struct fixed_file_table
*table
= &file_data
->table
[i
];
7594 unsigned this_files
;
7596 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7597 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7601 nr_files
-= this_files
;
7607 for (i
= 0; i
< nr_tables
; i
++) {
7608 struct fixed_file_table
*table
= &file_data
->table
[i
];
7609 kfree(table
->files
);
7614 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
7616 #if defined(CONFIG_UNIX)
7617 struct sock
*sock
= ctx
->ring_sock
->sk
;
7618 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7619 struct sk_buff
*skb
;
7622 __skb_queue_head_init(&list
);
7625 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7626 * remove this entry and rearrange the file array.
7628 skb
= skb_dequeue(head
);
7630 struct scm_fp_list
*fp
;
7632 fp
= UNIXCB(skb
).fp
;
7633 for (i
= 0; i
< fp
->count
; i
++) {
7636 if (fp
->fp
[i
] != file
)
7639 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7640 left
= fp
->count
- 1 - i
;
7642 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7643 left
* sizeof(struct file
*));
7650 __skb_queue_tail(&list
, skb
);
7660 __skb_queue_tail(&list
, skb
);
7662 skb
= skb_dequeue(head
);
7665 if (skb_peek(&list
)) {
7666 spin_lock_irq(&head
->lock
);
7667 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7668 __skb_queue_tail(head
, skb
);
7669 spin_unlock_irq(&head
->lock
);
7676 struct io_file_put
{
7677 struct list_head list
;
7681 static void __io_file_put_work(struct fixed_file_ref_node
*ref_node
)
7683 struct fixed_file_data
*file_data
= ref_node
->file_data
;
7684 struct io_ring_ctx
*ctx
= file_data
->ctx
;
7685 struct io_file_put
*pfile
, *tmp
;
7687 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
7688 list_del(&pfile
->list
);
7689 io_ring_file_put(ctx
, pfile
->file
);
7693 percpu_ref_exit(&ref_node
->refs
);
7695 percpu_ref_put(&file_data
->refs
);
7698 static void io_file_put_work(struct work_struct
*work
)
7700 struct io_ring_ctx
*ctx
;
7701 struct llist_node
*node
;
7703 ctx
= container_of(work
, struct io_ring_ctx
, file_put_work
.work
);
7704 node
= llist_del_all(&ctx
->file_put_llist
);
7707 struct fixed_file_ref_node
*ref_node
;
7708 struct llist_node
*next
= node
->next
;
7710 ref_node
= llist_entry(node
, struct fixed_file_ref_node
, llist
);
7711 __io_file_put_work(ref_node
);
7716 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
7718 struct fixed_file_ref_node
*ref_node
;
7719 struct fixed_file_data
*data
;
7720 struct io_ring_ctx
*ctx
;
7721 bool first_add
= false;
7724 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
7725 data
= ref_node
->file_data
;
7728 spin_lock_bh(&data
->lock
);
7729 ref_node
->done
= true;
7731 while (!list_empty(&data
->ref_list
)) {
7732 ref_node
= list_first_entry(&data
->ref_list
,
7733 struct fixed_file_ref_node
, node
);
7734 /* recycle ref nodes in order */
7735 if (!ref_node
->done
)
7737 list_del(&ref_node
->node
);
7738 first_add
|= llist_add(&ref_node
->llist
, &ctx
->file_put_llist
);
7740 spin_unlock_bh(&data
->lock
);
7742 if (percpu_ref_is_dying(&data
->refs
))
7746 mod_delayed_work(system_wq
, &ctx
->file_put_work
, 0);
7748 queue_delayed_work(system_wq
, &ctx
->file_put_work
, delay
);
7751 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
7752 struct io_ring_ctx
*ctx
)
7754 struct fixed_file_ref_node
*ref_node
;
7756 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7760 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
7765 INIT_LIST_HEAD(&ref_node
->node
);
7766 INIT_LIST_HEAD(&ref_node
->file_list
);
7767 ref_node
->file_data
= ctx
->file_data
;
7768 ref_node
->done
= false;
7772 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
7774 percpu_ref_exit(&ref_node
->refs
);
7778 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7781 __s32 __user
*fds
= (__s32 __user
*) arg
;
7782 unsigned nr_tables
, i
;
7784 int fd
, ret
= -ENOMEM
;
7785 struct fixed_file_ref_node
*ref_node
;
7786 struct fixed_file_data
*file_data
;
7792 if (nr_args
> IORING_MAX_FIXED_FILES
)
7795 file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL
);
7798 file_data
->ctx
= ctx
;
7799 init_completion(&file_data
->done
);
7800 INIT_LIST_HEAD(&file_data
->ref_list
);
7801 spin_lock_init(&file_data
->lock
);
7803 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7804 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7806 if (!file_data
->table
)
7809 if (percpu_ref_init(&file_data
->refs
, io_file_ref_kill
,
7810 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
7813 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7815 ctx
->file_data
= file_data
;
7817 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7818 struct fixed_file_table
*table
;
7821 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7825 /* allow sparse sets */
7835 * Don't allow io_uring instances to be registered. If UNIX
7836 * isn't enabled, then this causes a reference cycle and this
7837 * instance can never get freed. If UNIX is enabled we'll
7838 * handle it just fine, but there's still no point in allowing
7839 * a ring fd as it doesn't support regular read/write anyway.
7841 if (file
->f_op
== &io_uring_fops
) {
7845 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7846 index
= i
& IORING_FILE_TABLE_MASK
;
7847 table
->files
[index
] = file
;
7850 ret
= io_sqe_files_scm(ctx
);
7852 io_sqe_files_unregister(ctx
);
7856 ref_node
= alloc_fixed_file_ref_node(ctx
);
7858 io_sqe_files_unregister(ctx
);
7862 io_sqe_files_set_node(file_data
, ref_node
);
7865 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7866 file
= io_file_from_index(ctx
, i
);
7870 for (i
= 0; i
< nr_tables
; i
++)
7871 kfree(file_data
->table
[i
].files
);
7872 ctx
->nr_user_files
= 0;
7874 percpu_ref_exit(&file_data
->refs
);
7876 kfree(file_data
->table
);
7878 ctx
->file_data
= NULL
;
7882 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7885 #if defined(CONFIG_UNIX)
7886 struct sock
*sock
= ctx
->ring_sock
->sk
;
7887 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7888 struct sk_buff
*skb
;
7891 * See if we can merge this file into an existing skb SCM_RIGHTS
7892 * file set. If there's no room, fall back to allocating a new skb
7893 * and filling it in.
7895 spin_lock_irq(&head
->lock
);
7896 skb
= skb_peek(head
);
7898 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7900 if (fpl
->count
< SCM_MAX_FD
) {
7901 __skb_unlink(skb
, head
);
7902 spin_unlock_irq(&head
->lock
);
7903 fpl
->fp
[fpl
->count
] = get_file(file
);
7904 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7906 spin_lock_irq(&head
->lock
);
7907 __skb_queue_head(head
, skb
);
7912 spin_unlock_irq(&head
->lock
);
7919 return __io_sqe_files_scm(ctx
, 1, index
);
7925 static int io_queue_file_removal(struct fixed_file_data
*data
,
7928 struct io_file_put
*pfile
;
7929 struct fixed_file_ref_node
*ref_node
= data
->node
;
7931 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
7936 list_add(&pfile
->list
, &ref_node
->file_list
);
7941 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7942 struct io_uring_files_update
*up
,
7945 struct fixed_file_data
*data
= ctx
->file_data
;
7946 struct fixed_file_ref_node
*ref_node
;
7951 bool needs_switch
= false;
7953 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7955 if (done
> ctx
->nr_user_files
)
7958 ref_node
= alloc_fixed_file_ref_node(ctx
);
7963 fds
= u64_to_user_ptr(up
->fds
);
7965 struct fixed_file_table
*table
;
7969 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7973 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
7974 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7975 index
= i
& IORING_FILE_TABLE_MASK
;
7976 if (table
->files
[index
]) {
7977 file
= table
->files
[index
];
7978 err
= io_queue_file_removal(data
, file
);
7981 table
->files
[index
] = NULL
;
7982 needs_switch
= true;
7991 * Don't allow io_uring instances to be registered. If
7992 * UNIX isn't enabled, then this causes a reference
7993 * cycle and this instance can never get freed. If UNIX
7994 * is enabled we'll handle it just fine, but there's
7995 * still no point in allowing a ring fd as it doesn't
7996 * support regular read/write anyway.
7998 if (file
->f_op
== &io_uring_fops
) {
8003 table
->files
[index
] = file
;
8004 err
= io_sqe_file_register(ctx
, file
, i
);
8006 table
->files
[index
] = NULL
;
8017 percpu_ref_kill(&data
->node
->refs
);
8018 io_sqe_files_set_node(data
, ref_node
);
8020 destroy_fixed_file_ref_node(ref_node
);
8022 return done
? done
: err
;
8025 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
8028 struct io_uring_files_update up
;
8030 if (!ctx
->file_data
)
8034 if (copy_from_user(&up
, arg
, sizeof(up
)))
8039 return __io_sqe_files_update(ctx
, &up
, nr_args
);
8042 static void io_free_work(struct io_wq_work
*work
)
8044 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8046 /* Consider that io_steal_work() relies on this ref */
8050 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
8051 struct io_uring_params
*p
)
8053 struct io_wq_data data
;
8055 struct io_ring_ctx
*ctx_attach
;
8056 unsigned int concurrency
;
8059 data
.user
= ctx
->user
;
8060 data
.free_work
= io_free_work
;
8061 data
.do_work
= io_wq_submit_work
;
8063 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
8064 /* Do QD, or 4 * CPUS, whatever is smallest */
8065 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8067 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
8068 if (IS_ERR(ctx
->io_wq
)) {
8069 ret
= PTR_ERR(ctx
->io_wq
);
8075 f
= fdget(p
->wq_fd
);
8079 if (f
.file
->f_op
!= &io_uring_fops
) {
8084 ctx_attach
= f
.file
->private_data
;
8085 /* @io_wq is protected by holding the fd */
8086 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
8091 ctx
->io_wq
= ctx_attach
->io_wq
;
8097 static int io_uring_alloc_task_context(struct task_struct
*task
)
8099 struct io_uring_task
*tctx
;
8102 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
8103 if (unlikely(!tctx
))
8106 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8107 if (unlikely(ret
)) {
8113 init_waitqueue_head(&tctx
->wait
);
8115 atomic_set(&tctx
->in_idle
, 0);
8116 tctx
->sqpoll
= false;
8117 io_init_identity(&tctx
->__identity
);
8118 tctx
->identity
= &tctx
->__identity
;
8119 task
->io_uring
= tctx
;
8123 void __io_uring_free(struct task_struct
*tsk
)
8125 struct io_uring_task
*tctx
= tsk
->io_uring
;
8127 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8128 WARN_ON_ONCE(refcount_read(&tctx
->identity
->count
) != 1);
8129 if (tctx
->identity
!= &tctx
->__identity
)
8130 kfree(tctx
->identity
);
8131 percpu_counter_destroy(&tctx
->inflight
);
8133 tsk
->io_uring
= NULL
;
8136 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8137 struct io_uring_params
*p
)
8141 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8142 struct io_sq_data
*sqd
;
8145 if (!capable(CAP_SYS_ADMIN
) && !capable(CAP_SYS_NICE
))
8148 sqd
= io_get_sq_data(p
);
8155 io_sq_thread_park(sqd
);
8156 mutex_lock(&sqd
->ctx_lock
);
8157 list_add(&ctx
->sqd_list
, &sqd
->ctx_new_list
);
8158 mutex_unlock(&sqd
->ctx_lock
);
8159 io_sq_thread_unpark(sqd
);
8161 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8162 if (!ctx
->sq_thread_idle
)
8163 ctx
->sq_thread_idle
= HZ
;
8168 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8169 int cpu
= p
->sq_thread_cpu
;
8172 if (cpu
>= nr_cpu_ids
)
8174 if (!cpu_online(cpu
))
8177 sqd
->thread
= kthread_create_on_cpu(io_sq_thread
, sqd
,
8178 cpu
, "io_uring-sq");
8180 sqd
->thread
= kthread_create(io_sq_thread
, sqd
,
8183 if (IS_ERR(sqd
->thread
)) {
8184 ret
= PTR_ERR(sqd
->thread
);
8188 ret
= io_uring_alloc_task_context(sqd
->thread
);
8191 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8192 /* Can't have SQ_AFF without SQPOLL */
8198 ret
= io_init_wq_offload(ctx
, p
);
8204 io_finish_async(ctx
);
8208 static void io_sq_offload_start(struct io_ring_ctx
*ctx
)
8210 struct io_sq_data
*sqd
= ctx
->sq_data
;
8212 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && sqd
->thread
)
8213 wake_up_process(sqd
->thread
);
8216 static inline void __io_unaccount_mem(struct user_struct
*user
,
8217 unsigned long nr_pages
)
8219 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8222 static inline int __io_account_mem(struct user_struct
*user
,
8223 unsigned long nr_pages
)
8225 unsigned long page_limit
, cur_pages
, new_pages
;
8227 /* Don't allow more pages than we can safely lock */
8228 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8231 cur_pages
= atomic_long_read(&user
->locked_vm
);
8232 new_pages
= cur_pages
+ nr_pages
;
8233 if (new_pages
> page_limit
)
8235 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8236 new_pages
) != cur_pages
);
8241 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8242 enum io_mem_account acct
)
8245 __io_unaccount_mem(ctx
->user
, nr_pages
);
8247 if (ctx
->mm_account
) {
8248 if (acct
== ACCT_LOCKED
) {
8249 mmap_write_lock(ctx
->mm_account
);
8250 ctx
->mm_account
->locked_vm
-= nr_pages
;
8251 mmap_write_unlock(ctx
->mm_account
);
8252 }else if (acct
== ACCT_PINNED
) {
8253 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8258 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8259 enum io_mem_account acct
)
8263 if (ctx
->limit_mem
) {
8264 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8269 if (ctx
->mm_account
) {
8270 if (acct
== ACCT_LOCKED
) {
8271 mmap_write_lock(ctx
->mm_account
);
8272 ctx
->mm_account
->locked_vm
+= nr_pages
;
8273 mmap_write_unlock(ctx
->mm_account
);
8274 } else if (acct
== ACCT_PINNED
) {
8275 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8282 static void io_mem_free(void *ptr
)
8289 page
= virt_to_head_page(ptr
);
8290 if (put_page_testzero(page
))
8291 free_compound_page(page
);
8294 static void *io_mem_alloc(size_t size
)
8296 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8299 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8302 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8305 struct io_rings
*rings
;
8306 size_t off
, sq_array_size
;
8308 off
= struct_size(rings
, cqes
, cq_entries
);
8309 if (off
== SIZE_MAX
)
8313 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8321 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8322 if (sq_array_size
== SIZE_MAX
)
8325 if (check_add_overflow(off
, sq_array_size
, &off
))
8331 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
8335 pages
= (size_t)1 << get_order(
8336 rings_size(sq_entries
, cq_entries
, NULL
));
8337 pages
+= (size_t)1 << get_order(
8338 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
8343 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
8347 if (!ctx
->user_bufs
)
8350 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8351 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8353 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
8354 unpin_user_page(imu
->bvec
[j
].bv_page
);
8356 if (imu
->acct_pages
)
8357 io_unaccount_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8362 kfree(ctx
->user_bufs
);
8363 ctx
->user_bufs
= NULL
;
8364 ctx
->nr_user_bufs
= 0;
8368 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8369 void __user
*arg
, unsigned index
)
8371 struct iovec __user
*src
;
8373 #ifdef CONFIG_COMPAT
8375 struct compat_iovec __user
*ciovs
;
8376 struct compat_iovec ciov
;
8378 ciovs
= (struct compat_iovec __user
*) arg
;
8379 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8382 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8383 dst
->iov_len
= ciov
.iov_len
;
8387 src
= (struct iovec __user
*) arg
;
8388 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8394 * Not super efficient, but this is just a registration time. And we do cache
8395 * the last compound head, so generally we'll only do a full search if we don't
8398 * We check if the given compound head page has already been accounted, to
8399 * avoid double accounting it. This allows us to account the full size of the
8400 * page, not just the constituent pages of a huge page.
8402 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8403 int nr_pages
, struct page
*hpage
)
8407 /* check current page array */
8408 for (i
= 0; i
< nr_pages
; i
++) {
8409 if (!PageCompound(pages
[i
]))
8411 if (compound_head(pages
[i
]) == hpage
)
8415 /* check previously registered pages */
8416 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8417 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8419 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8420 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8422 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8430 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8431 int nr_pages
, struct io_mapped_ubuf
*imu
,
8432 struct page
**last_hpage
)
8436 for (i
= 0; i
< nr_pages
; i
++) {
8437 if (!PageCompound(pages
[i
])) {
8442 hpage
= compound_head(pages
[i
]);
8443 if (hpage
== *last_hpage
)
8445 *last_hpage
= hpage
;
8446 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8448 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8452 if (!imu
->acct_pages
)
8455 ret
= io_account_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8457 imu
->acct_pages
= 0;
8461 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8464 struct vm_area_struct
**vmas
= NULL
;
8465 struct page
**pages
= NULL
;
8466 struct page
*last_hpage
= NULL
;
8467 int i
, j
, got_pages
= 0;
8472 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8475 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8477 if (!ctx
->user_bufs
)
8480 for (i
= 0; i
< nr_args
; i
++) {
8481 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8482 unsigned long off
, start
, end
, ubuf
;
8487 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8492 * Don't impose further limits on the size and buffer
8493 * constraints here, we'll -EINVAL later when IO is
8494 * submitted if they are wrong.
8497 if (!iov
.iov_base
|| !iov
.iov_len
)
8500 /* arbitrary limit, but we need something */
8501 if (iov
.iov_len
> SZ_1G
)
8504 ubuf
= (unsigned long) iov
.iov_base
;
8505 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8506 start
= ubuf
>> PAGE_SHIFT
;
8507 nr_pages
= end
- start
;
8510 if (!pages
|| nr_pages
> got_pages
) {
8513 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
8515 vmas
= kvmalloc_array(nr_pages
,
8516 sizeof(struct vm_area_struct
*),
8518 if (!pages
|| !vmas
) {
8522 got_pages
= nr_pages
;
8525 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8532 mmap_read_lock(current
->mm
);
8533 pret
= pin_user_pages(ubuf
, nr_pages
,
8534 FOLL_WRITE
| FOLL_LONGTERM
,
8536 if (pret
== nr_pages
) {
8537 /* don't support file backed memory */
8538 for (j
= 0; j
< nr_pages
; j
++) {
8539 struct vm_area_struct
*vma
= vmas
[j
];
8542 !is_file_hugepages(vma
->vm_file
)) {
8548 ret
= pret
< 0 ? pret
: -EFAULT
;
8550 mmap_read_unlock(current
->mm
);
8553 * if we did partial map, or found file backed vmas,
8554 * release any pages we did get
8557 unpin_user_pages(pages
, pret
);
8562 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, &last_hpage
);
8564 unpin_user_pages(pages
, pret
);
8569 off
= ubuf
& ~PAGE_MASK
;
8571 for (j
= 0; j
< nr_pages
; j
++) {
8574 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8575 imu
->bvec
[j
].bv_page
= pages
[j
];
8576 imu
->bvec
[j
].bv_len
= vec_len
;
8577 imu
->bvec
[j
].bv_offset
= off
;
8581 /* store original address for later verification */
8583 imu
->len
= iov
.iov_len
;
8584 imu
->nr_bvecs
= nr_pages
;
8586 ctx
->nr_user_bufs
++;
8594 io_sqe_buffer_unregister(ctx
);
8598 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8600 __s32 __user
*fds
= arg
;
8606 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8609 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8610 if (IS_ERR(ctx
->cq_ev_fd
)) {
8611 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8612 ctx
->cq_ev_fd
= NULL
;
8619 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8621 if (ctx
->cq_ev_fd
) {
8622 eventfd_ctx_put(ctx
->cq_ev_fd
);
8623 ctx
->cq_ev_fd
= NULL
;
8630 static int __io_destroy_buffers(int id
, void *p
, void *data
)
8632 struct io_ring_ctx
*ctx
= data
;
8633 struct io_buffer
*buf
= p
;
8635 __io_remove_buffers(ctx
, buf
, id
, -1U);
8639 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8641 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
8642 idr_destroy(&ctx
->io_buffer_idr
);
8645 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8647 io_finish_async(ctx
);
8648 io_sqe_buffer_unregister(ctx
);
8650 if (ctx
->sqo_task
) {
8651 put_task_struct(ctx
->sqo_task
);
8652 ctx
->sqo_task
= NULL
;
8653 mmdrop(ctx
->mm_account
);
8654 ctx
->mm_account
= NULL
;
8657 #ifdef CONFIG_BLK_CGROUP
8658 if (ctx
->sqo_blkcg_css
)
8659 css_put(ctx
->sqo_blkcg_css
);
8662 io_sqe_files_unregister(ctx
);
8663 io_eventfd_unregister(ctx
);
8664 io_destroy_buffers(ctx
);
8665 idr_destroy(&ctx
->personality_idr
);
8667 #if defined(CONFIG_UNIX)
8668 if (ctx
->ring_sock
) {
8669 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8670 sock_release(ctx
->ring_sock
);
8674 io_mem_free(ctx
->rings
);
8675 io_mem_free(ctx
->sq_sqes
);
8677 percpu_ref_exit(&ctx
->refs
);
8678 free_uid(ctx
->user
);
8679 put_cred(ctx
->creds
);
8680 kfree(ctx
->cancel_hash
);
8681 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
8685 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8687 struct io_ring_ctx
*ctx
= file
->private_data
;
8690 poll_wait(file
, &ctx
->cq_wait
, wait
);
8692 * synchronizes with barrier from wq_has_sleeper call in
8696 if (!io_sqring_full(ctx
))
8697 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8698 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
8699 if (io_cqring_events(ctx
))
8700 mask
|= EPOLLIN
| EPOLLRDNORM
;
8705 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8707 struct io_ring_ctx
*ctx
= file
->private_data
;
8709 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8712 static int io_remove_personalities(int id
, void *p
, void *data
)
8714 struct io_ring_ctx
*ctx
= data
;
8715 struct io_identity
*iod
;
8717 iod
= idr_remove(&ctx
->personality_idr
, id
);
8719 put_cred(iod
->creds
);
8720 if (refcount_dec_and_test(&iod
->count
))
8726 static void io_ring_exit_work(struct work_struct
*work
)
8728 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
8732 * If we're doing polled IO and end up having requests being
8733 * submitted async (out-of-line), then completions can come in while
8734 * we're waiting for refs to drop. We need to reap these manually,
8735 * as nobody else will be looking for them.
8738 __io_uring_cancel_task_requests(ctx
, NULL
);
8739 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8740 io_ring_ctx_free(ctx
);
8743 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8745 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8747 return req
->ctx
== data
;
8750 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8752 mutex_lock(&ctx
->uring_lock
);
8753 percpu_ref_kill(&ctx
->refs
);
8755 if (WARN_ON_ONCE((ctx
->flags
& IORING_SETUP_SQPOLL
) && !ctx
->sqo_dead
))
8758 /* if force is set, the ring is going away. always drop after that */
8759 ctx
->cq_overflow_flushed
= 1;
8761 __io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8762 mutex_unlock(&ctx
->uring_lock
);
8764 io_kill_timeouts(ctx
, NULL
, NULL
);
8765 io_poll_remove_all(ctx
, NULL
, NULL
);
8768 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8770 /* if we failed setting up the ctx, we might not have any rings */
8771 io_iopoll_try_reap_events(ctx
);
8772 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
8775 * Do this upfront, so we won't have a grace period where the ring
8776 * is closed but resources aren't reaped yet. This can cause
8777 * spurious failure in setting up a new ring.
8779 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
8782 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8784 * Use system_unbound_wq to avoid spawning tons of event kworkers
8785 * if we're exiting a ton of rings at the same time. It just adds
8786 * noise and overhead, there's no discernable change in runtime
8787 * over using system_wq.
8789 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8792 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8794 struct io_ring_ctx
*ctx
= file
->private_data
;
8796 file
->private_data
= NULL
;
8797 io_ring_ctx_wait_and_kill(ctx
);
8801 struct io_task_cancel
{
8802 struct task_struct
*task
;
8803 struct files_struct
*files
;
8806 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8808 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8809 struct io_task_cancel
*cancel
= data
;
8812 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8813 unsigned long flags
;
8814 struct io_ring_ctx
*ctx
= req
->ctx
;
8816 /* protect against races with linked timeouts */
8817 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8818 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8819 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8821 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8826 static void io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8827 struct task_struct
*task
,
8828 struct files_struct
*files
)
8830 struct io_defer_entry
*de
= NULL
;
8833 spin_lock_irq(&ctx
->completion_lock
);
8834 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8835 if (io_match_task(de
->req
, task
, files
)) {
8836 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8840 spin_unlock_irq(&ctx
->completion_lock
);
8842 while (!list_empty(&list
)) {
8843 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8844 list_del_init(&de
->list
);
8845 req_set_fail_links(de
->req
);
8846 io_put_req(de
->req
);
8847 io_req_complete(de
->req
, -ECANCELED
);
8852 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
8853 struct task_struct
*task
,
8854 struct files_struct
*files
)
8856 while (!list_empty_careful(&ctx
->inflight_list
)) {
8857 struct io_task_cancel cancel
= { .task
= task
, .files
= files
};
8858 struct io_kiocb
*req
;
8862 spin_lock_irq(&ctx
->inflight_lock
);
8863 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
) {
8864 if (req
->task
!= task
||
8865 req
->work
.identity
->files
!= files
)
8871 prepare_to_wait(&task
->io_uring
->wait
, &wait
,
8872 TASK_UNINTERRUPTIBLE
);
8873 spin_unlock_irq(&ctx
->inflight_lock
);
8875 /* We need to keep going until we don't find a matching req */
8879 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
, &cancel
, true);
8880 io_poll_remove_all(ctx
, task
, files
);
8881 io_kill_timeouts(ctx
, task
, files
);
8882 /* cancellations _may_ trigger task work */
8885 finish_wait(&task
->io_uring
->wait
, &wait
);
8889 static void __io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
8890 struct task_struct
*task
)
8893 struct io_task_cancel cancel
= { .task
= task
, .files
= NULL
, };
8894 enum io_wq_cancel cret
;
8898 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
,
8900 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8903 /* SQPOLL thread does its own polling */
8904 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
8905 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8906 io_iopoll_try_reap_events(ctx
);
8911 ret
|= io_poll_remove_all(ctx
, task
, NULL
);
8912 ret
|= io_kill_timeouts(ctx
, task
, NULL
);
8913 ret
|= io_run_task_work();
8920 static void io_disable_sqo_submit(struct io_ring_ctx
*ctx
)
8922 mutex_lock(&ctx
->uring_lock
);
8924 mutex_unlock(&ctx
->uring_lock
);
8926 /* make sure callers enter the ring to get error */
8928 io_ring_set_wakeup_flag(ctx
);
8932 * We need to iteratively cancel requests, in case a request has dependent
8933 * hard links. These persist even for failure of cancelations, hence keep
8934 * looping until none are found.
8936 static void io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
8937 struct files_struct
*files
)
8939 struct task_struct
*task
= current
;
8941 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
8942 /* for SQPOLL only sqo_task has task notes */
8943 WARN_ON_ONCE(ctx
->sqo_task
!= current
);
8944 io_disable_sqo_submit(ctx
);
8945 task
= ctx
->sq_data
->thread
;
8946 atomic_inc(&task
->io_uring
->in_idle
);
8947 io_sq_thread_park(ctx
->sq_data
);
8950 io_cancel_defer_files(ctx
, task
, files
);
8951 io_cqring_overflow_flush(ctx
, true, task
, files
);
8954 __io_uring_cancel_task_requests(ctx
, task
);
8956 io_uring_cancel_files(ctx
, task
, files
);
8958 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
8959 atomic_dec(&task
->io_uring
->in_idle
);
8961 * If the files that are going away are the ones in the thread
8962 * identity, clear them out.
8964 if (task
->io_uring
->identity
->files
== files
)
8965 task
->io_uring
->identity
->files
= NULL
;
8966 io_sq_thread_unpark(ctx
->sq_data
);
8971 * Note that this task has used io_uring. We use it for cancelation purposes.
8973 static int io_uring_add_task_file(struct io_ring_ctx
*ctx
, struct file
*file
)
8975 struct io_uring_task
*tctx
= current
->io_uring
;
8978 if (unlikely(!tctx
)) {
8979 ret
= io_uring_alloc_task_context(current
);
8982 tctx
= current
->io_uring
;
8984 if (tctx
->last
!= file
) {
8985 void *old
= xa_load(&tctx
->xa
, (unsigned long)file
);
8989 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)file
,
9000 * This is race safe in that the task itself is doing this, hence it
9001 * cannot be going through the exit/cancel paths at the same time.
9002 * This cannot be modified while exit/cancel is running.
9004 if (!tctx
->sqpoll
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9005 tctx
->sqpoll
= true;
9011 * Remove this io_uring_file -> task mapping.
9013 static void io_uring_del_task_file(struct file
*file
)
9015 struct io_uring_task
*tctx
= current
->io_uring
;
9017 if (tctx
->last
== file
)
9019 file
= xa_erase(&tctx
->xa
, (unsigned long)file
);
9024 static void io_uring_remove_task_files(struct io_uring_task
*tctx
)
9027 unsigned long index
;
9029 xa_for_each(&tctx
->xa
, index
, file
)
9030 io_uring_del_task_file(file
);
9033 void __io_uring_files_cancel(struct files_struct
*files
)
9035 struct io_uring_task
*tctx
= current
->io_uring
;
9037 unsigned long index
;
9039 /* make sure overflow events are dropped */
9040 atomic_inc(&tctx
->in_idle
);
9041 xa_for_each(&tctx
->xa
, index
, file
)
9042 io_uring_cancel_task_requests(file
->private_data
, files
);
9043 atomic_dec(&tctx
->in_idle
);
9046 io_uring_remove_task_files(tctx
);
9049 static s64
tctx_inflight(struct io_uring_task
*tctx
)
9051 unsigned long index
;
9055 inflight
= percpu_counter_sum(&tctx
->inflight
);
9060 * If we have SQPOLL rings, then we need to iterate and find them, and
9061 * add the pending count for those.
9063 xa_for_each(&tctx
->xa
, index
, file
) {
9064 struct io_ring_ctx
*ctx
= file
->private_data
;
9066 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9067 struct io_uring_task
*__tctx
= ctx
->sqo_task
->io_uring
;
9069 inflight
+= percpu_counter_sum(&__tctx
->inflight
);
9077 * Find any io_uring fd that this task has registered or done IO on, and cancel
9080 void __io_uring_task_cancel(void)
9082 struct io_uring_task
*tctx
= current
->io_uring
;
9086 /* make sure overflow events are dropped */
9087 atomic_inc(&tctx
->in_idle
);
9089 /* trigger io_disable_sqo_submit() */
9091 __io_uring_files_cancel(NULL
);
9094 /* read completions before cancelations */
9095 inflight
= tctx_inflight(tctx
);
9098 __io_uring_files_cancel(NULL
);
9100 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9103 * If we've seen completions, retry. This avoids a race where
9104 * a completion comes in before we did prepare_to_wait().
9106 if (inflight
!= tctx_inflight(tctx
))
9109 finish_wait(&tctx
->wait
, &wait
);
9112 finish_wait(&tctx
->wait
, &wait
);
9113 atomic_dec(&tctx
->in_idle
);
9115 io_uring_remove_task_files(tctx
);
9118 static int io_uring_flush(struct file
*file
, void *data
)
9120 struct io_uring_task
*tctx
= current
->io_uring
;
9121 struct io_ring_ctx
*ctx
= file
->private_data
;
9126 /* we should have cancelled and erased it before PF_EXITING */
9127 WARN_ON_ONCE((current
->flags
& PF_EXITING
) &&
9128 xa_load(&tctx
->xa
, (unsigned long)file
));
9131 * fput() is pending, will be 2 if the only other ref is our potential
9132 * task file note. If the task is exiting, drop regardless of count.
9134 if (atomic_long_read(&file
->f_count
) != 2)
9137 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9138 /* there is only one file note, which is owned by sqo_task */
9139 WARN_ON_ONCE(ctx
->sqo_task
!= current
&&
9140 xa_load(&tctx
->xa
, (unsigned long)file
));
9141 /* sqo_dead check is for when this happens after cancellation */
9142 WARN_ON_ONCE(ctx
->sqo_task
== current
&& !ctx
->sqo_dead
&&
9143 !xa_load(&tctx
->xa
, (unsigned long)file
));
9145 io_disable_sqo_submit(ctx
);
9148 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) || ctx
->sqo_task
== current
)
9149 io_uring_del_task_file(file
);
9153 static void *io_uring_validate_mmap_request(struct file
*file
,
9154 loff_t pgoff
, size_t sz
)
9156 struct io_ring_ctx
*ctx
= file
->private_data
;
9157 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9162 case IORING_OFF_SQ_RING
:
9163 case IORING_OFF_CQ_RING
:
9166 case IORING_OFF_SQES
:
9170 return ERR_PTR(-EINVAL
);
9173 page
= virt_to_head_page(ptr
);
9174 if (sz
> page_size(page
))
9175 return ERR_PTR(-EINVAL
);
9182 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9184 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9188 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9190 return PTR_ERR(ptr
);
9192 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9193 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9196 #else /* !CONFIG_MMU */
9198 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9200 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9203 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9205 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9208 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9209 unsigned long addr
, unsigned long len
,
9210 unsigned long pgoff
, unsigned long flags
)
9214 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9216 return PTR_ERR(ptr
);
9218 return (unsigned long) ptr
;
9221 #endif /* !CONFIG_MMU */
9223 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9229 if (!io_sqring_full(ctx
))
9232 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9234 if (unlikely(ctx
->sqo_dead
)) {
9239 if (!io_sqring_full(ctx
))
9243 } while (!signal_pending(current
));
9245 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9250 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9251 struct __kernel_timespec __user
**ts
,
9252 const sigset_t __user
**sig
)
9254 struct io_uring_getevents_arg arg
;
9257 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9258 * is just a pointer to the sigset_t.
9260 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9261 *sig
= (const sigset_t __user
*) argp
;
9267 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9268 * timespec and sigset_t pointers if good.
9270 if (*argsz
!= sizeof(arg
))
9272 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9274 *sig
= u64_to_user_ptr(arg
.sigmask
);
9275 *argsz
= arg
.sigmask_sz
;
9276 *ts
= u64_to_user_ptr(arg
.ts
);
9280 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9281 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9284 struct io_ring_ctx
*ctx
;
9291 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9292 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
))
9300 if (f
.file
->f_op
!= &io_uring_fops
)
9304 ctx
= f
.file
->private_data
;
9305 if (!percpu_ref_tryget(&ctx
->refs
))
9309 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9313 * For SQ polling, the thread will do all submissions and completions.
9314 * Just return the requested submit count, and wake the thread if
9318 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9319 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
9322 if (unlikely(ctx
->sqo_dead
))
9324 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9325 wake_up(&ctx
->sq_data
->wait
);
9326 if (flags
& IORING_ENTER_SQ_WAIT
) {
9327 ret
= io_sqpoll_wait_sq(ctx
);
9331 submitted
= to_submit
;
9332 } else if (to_submit
) {
9333 ret
= io_uring_add_task_file(ctx
, f
.file
);
9336 mutex_lock(&ctx
->uring_lock
);
9337 submitted
= io_submit_sqes(ctx
, to_submit
);
9338 mutex_unlock(&ctx
->uring_lock
);
9340 if (submitted
!= to_submit
)
9343 if (flags
& IORING_ENTER_GETEVENTS
) {
9344 const sigset_t __user
*sig
;
9345 struct __kernel_timespec __user
*ts
;
9347 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9351 min_complete
= min(min_complete
, ctx
->cq_entries
);
9354 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9355 * space applications don't need to do io completion events
9356 * polling again, they can rely on io_sq_thread to do polling
9357 * work, which can reduce cpu usage and uring_lock contention.
9359 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9360 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9361 ret
= io_iopoll_check(ctx
, min_complete
);
9363 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9368 percpu_ref_put(&ctx
->refs
);
9371 return submitted
? submitted
: ret
;
9374 #ifdef CONFIG_PROC_FS
9375 static int io_uring_show_cred(int id
, void *p
, void *data
)
9377 struct io_identity
*iod
= p
;
9378 const struct cred
*cred
= iod
->creds
;
9379 struct seq_file
*m
= data
;
9380 struct user_namespace
*uns
= seq_user_ns(m
);
9381 struct group_info
*gi
;
9386 seq_printf(m
, "%5d\n", id
);
9387 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9388 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9389 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9390 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9391 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9392 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9393 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9394 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9395 seq_puts(m
, "\n\tGroups:\t");
9396 gi
= cred
->group_info
;
9397 for (g
= 0; g
< gi
->ngroups
; g
++) {
9398 seq_put_decimal_ull(m
, g
? " " : "",
9399 from_kgid_munged(uns
, gi
->gid
[g
]));
9401 seq_puts(m
, "\n\tCapEff:\t");
9402 cap
= cred
->cap_effective
;
9403 CAP_FOR_EACH_U32(__capi
)
9404 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9409 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9411 struct io_sq_data
*sq
= NULL
;
9416 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9417 * since fdinfo case grabs it in the opposite direction of normal use
9418 * cases. If we fail to get the lock, we just don't iterate any
9419 * structures that could be going away outside the io_uring mutex.
9421 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9423 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9426 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9427 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9428 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9429 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9430 struct fixed_file_table
*table
;
9433 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
9434 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
9436 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9438 seq_printf(m
, "%5u: <none>\n", i
);
9440 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9441 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9442 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9444 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9445 (unsigned int) buf
->len
);
9447 if (has_lock
&& !idr_is_empty(&ctx
->personality_idr
)) {
9448 seq_printf(m
, "Personalities:\n");
9449 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
9451 seq_printf(m
, "PollList:\n");
9452 spin_lock_irq(&ctx
->completion_lock
);
9453 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9454 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9455 struct io_kiocb
*req
;
9457 hlist_for_each_entry(req
, list
, hash_node
)
9458 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9459 req
->task
->task_works
!= NULL
);
9461 spin_unlock_irq(&ctx
->completion_lock
);
9463 mutex_unlock(&ctx
->uring_lock
);
9466 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9468 struct io_ring_ctx
*ctx
= f
->private_data
;
9470 if (percpu_ref_tryget(&ctx
->refs
)) {
9471 __io_uring_show_fdinfo(ctx
, m
);
9472 percpu_ref_put(&ctx
->refs
);
9477 static const struct file_operations io_uring_fops
= {
9478 .release
= io_uring_release
,
9479 .flush
= io_uring_flush
,
9480 .mmap
= io_uring_mmap
,
9482 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9483 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9485 .poll
= io_uring_poll
,
9486 .fasync
= io_uring_fasync
,
9487 #ifdef CONFIG_PROC_FS
9488 .show_fdinfo
= io_uring_show_fdinfo
,
9492 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9493 struct io_uring_params
*p
)
9495 struct io_rings
*rings
;
9496 size_t size
, sq_array_offset
;
9498 /* make sure these are sane, as we already accounted them */
9499 ctx
->sq_entries
= p
->sq_entries
;
9500 ctx
->cq_entries
= p
->cq_entries
;
9502 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9503 if (size
== SIZE_MAX
)
9506 rings
= io_mem_alloc(size
);
9511 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9512 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9513 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9514 rings
->sq_ring_entries
= p
->sq_entries
;
9515 rings
->cq_ring_entries
= p
->cq_entries
;
9516 ctx
->sq_mask
= rings
->sq_ring_mask
;
9517 ctx
->cq_mask
= rings
->cq_ring_mask
;
9519 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9520 if (size
== SIZE_MAX
) {
9521 io_mem_free(ctx
->rings
);
9526 ctx
->sq_sqes
= io_mem_alloc(size
);
9527 if (!ctx
->sq_sqes
) {
9528 io_mem_free(ctx
->rings
);
9536 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9540 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9544 ret
= io_uring_add_task_file(ctx
, file
);
9549 fd_install(fd
, file
);
9554 * Allocate an anonymous fd, this is what constitutes the application
9555 * visible backing of an io_uring instance. The application mmaps this
9556 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9557 * we have to tie this fd to a socket for file garbage collection purposes.
9559 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9562 #if defined(CONFIG_UNIX)
9565 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9568 return ERR_PTR(ret
);
9571 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9572 O_RDWR
| O_CLOEXEC
);
9573 #if defined(CONFIG_UNIX)
9575 sock_release(ctx
->ring_sock
);
9576 ctx
->ring_sock
= NULL
;
9578 ctx
->ring_sock
->file
= file
;
9584 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9585 struct io_uring_params __user
*params
)
9587 struct user_struct
*user
= NULL
;
9588 struct io_ring_ctx
*ctx
;
9595 if (entries
> IORING_MAX_ENTRIES
) {
9596 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9598 entries
= IORING_MAX_ENTRIES
;
9602 * Use twice as many entries for the CQ ring. It's possible for the
9603 * application to drive a higher depth than the size of the SQ ring,
9604 * since the sqes are only used at submission time. This allows for
9605 * some flexibility in overcommitting a bit. If the application has
9606 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9607 * of CQ ring entries manually.
9609 p
->sq_entries
= roundup_pow_of_two(entries
);
9610 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9612 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9613 * to a power-of-two, if it isn't already. We do NOT impose
9614 * any cq vs sq ring sizing.
9618 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9619 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9621 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9623 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9624 if (p
->cq_entries
< p
->sq_entries
)
9627 p
->cq_entries
= 2 * p
->sq_entries
;
9630 user
= get_uid(current_user());
9631 limit_mem
= !capable(CAP_IPC_LOCK
);
9634 ret
= __io_account_mem(user
,
9635 ring_pages(p
->sq_entries
, p
->cq_entries
));
9642 ctx
= io_ring_ctx_alloc(p
);
9645 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
9650 ctx
->compat
= in_compat_syscall();
9652 ctx
->creds
= get_current_cred();
9654 ctx
->loginuid
= current
->loginuid
;
9655 ctx
->sessionid
= current
->sessionid
;
9657 ctx
->sqo_task
= get_task_struct(current
);
9660 * This is just grabbed for accounting purposes. When a process exits,
9661 * the mm is exited and dropped before the files, hence we need to hang
9662 * on to this mm purely for the purposes of being able to unaccount
9663 * memory (locked/pinned vm). It's not used for anything else.
9665 mmgrab(current
->mm
);
9666 ctx
->mm_account
= current
->mm
;
9668 #ifdef CONFIG_BLK_CGROUP
9670 * The sq thread will belong to the original cgroup it was inited in.
9671 * If the cgroup goes offline (e.g. disabling the io controller), then
9672 * issued bios will be associated with the closest cgroup later in the
9676 ctx
->sqo_blkcg_css
= blkcg_css();
9677 ret
= css_tryget_online(ctx
->sqo_blkcg_css
);
9680 /* don't init against a dying cgroup, have the user try again */
9681 ctx
->sqo_blkcg_css
= NULL
;
9688 * Account memory _before_ installing the file descriptor. Once
9689 * the descriptor is installed, it can get closed at any time. Also
9690 * do this before hitting the general error path, as ring freeing
9691 * will un-account as well.
9693 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
9695 ctx
->limit_mem
= limit_mem
;
9697 ret
= io_allocate_scq_urings(ctx
, p
);
9701 ret
= io_sq_offload_create(ctx
, p
);
9705 if (!(p
->flags
& IORING_SETUP_R_DISABLED
))
9706 io_sq_offload_start(ctx
);
9708 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9709 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9710 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9711 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9712 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9713 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9714 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9715 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9717 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9718 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9719 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9720 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9721 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9722 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9723 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9724 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9726 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9727 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9728 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9729 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9730 IORING_FEAT_EXT_ARG
;
9732 if (copy_to_user(params
, p
, sizeof(*p
))) {
9737 file
= io_uring_get_file(ctx
);
9739 ret
= PTR_ERR(file
);
9744 * Install ring fd as the very last thing, so we don't risk someone
9745 * having closed it before we finish setup
9747 ret
= io_uring_install_fd(ctx
, file
);
9749 io_disable_sqo_submit(ctx
);
9750 /* fput will clean it up */
9755 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9758 io_disable_sqo_submit(ctx
);
9759 io_ring_ctx_wait_and_kill(ctx
);
9764 * Sets up an aio uring context, and returns the fd. Applications asks for a
9765 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9766 * params structure passed in.
9768 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9770 struct io_uring_params p
;
9773 if (copy_from_user(&p
, params
, sizeof(p
)))
9775 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9780 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9781 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9782 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9783 IORING_SETUP_R_DISABLED
))
9786 return io_uring_create(entries
, &p
, params
);
9789 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9790 struct io_uring_params __user
*, params
)
9792 return io_uring_setup(entries
, params
);
9795 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9797 struct io_uring_probe
*p
;
9801 size
= struct_size(p
, ops
, nr_args
);
9802 if (size
== SIZE_MAX
)
9804 p
= kzalloc(size
, GFP_KERNEL
);
9809 if (copy_from_user(p
, arg
, size
))
9812 if (memchr_inv(p
, 0, size
))
9815 p
->last_op
= IORING_OP_LAST
- 1;
9816 if (nr_args
> IORING_OP_LAST
)
9817 nr_args
= IORING_OP_LAST
;
9819 for (i
= 0; i
< nr_args
; i
++) {
9821 if (!io_op_defs
[i
].not_supported
)
9822 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9827 if (copy_to_user(arg
, p
, size
))
9834 static int io_register_personality(struct io_ring_ctx
*ctx
)
9836 struct io_identity
*id
;
9839 id
= kmalloc(sizeof(*id
), GFP_KERNEL
);
9843 io_init_identity(id
);
9844 id
->creds
= get_current_cred();
9846 ret
= idr_alloc_cyclic(&ctx
->personality_idr
, id
, 1, USHRT_MAX
, GFP_KERNEL
);
9848 put_cred(id
->creds
);
9854 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9856 struct io_identity
*iod
;
9858 iod
= idr_remove(&ctx
->personality_idr
, id
);
9860 put_cred(iod
->creds
);
9861 if (refcount_dec_and_test(&iod
->count
))
9869 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9870 unsigned int nr_args
)
9872 struct io_uring_restriction
*res
;
9876 /* Restrictions allowed only if rings started disabled */
9877 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9880 /* We allow only a single restrictions registration */
9881 if (ctx
->restrictions
.registered
)
9884 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9887 size
= array_size(nr_args
, sizeof(*res
));
9888 if (size
== SIZE_MAX
)
9891 res
= memdup_user(arg
, size
);
9893 return PTR_ERR(res
);
9897 for (i
= 0; i
< nr_args
; i
++) {
9898 switch (res
[i
].opcode
) {
9899 case IORING_RESTRICTION_REGISTER_OP
:
9900 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9905 __set_bit(res
[i
].register_op
,
9906 ctx
->restrictions
.register_op
);
9908 case IORING_RESTRICTION_SQE_OP
:
9909 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9914 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9916 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9917 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9919 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9920 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9929 /* Reset all restrictions if an error happened */
9931 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9933 ctx
->restrictions
.registered
= true;
9939 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
9941 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9944 if (ctx
->restrictions
.registered
)
9945 ctx
->restricted
= 1;
9947 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
9949 io_sq_offload_start(ctx
);
9954 static bool io_register_op_must_quiesce(int op
)
9957 case IORING_UNREGISTER_FILES
:
9958 case IORING_REGISTER_FILES_UPDATE
:
9959 case IORING_REGISTER_PROBE
:
9960 case IORING_REGISTER_PERSONALITY
:
9961 case IORING_UNREGISTER_PERSONALITY
:
9968 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
9969 void __user
*arg
, unsigned nr_args
)
9970 __releases(ctx
->uring_lock
)
9971 __acquires(ctx
->uring_lock
)
9976 * We're inside the ring mutex, if the ref is already dying, then
9977 * someone else killed the ctx or is already going through
9978 * io_uring_register().
9980 if (percpu_ref_is_dying(&ctx
->refs
))
9983 if (io_register_op_must_quiesce(opcode
)) {
9984 percpu_ref_kill(&ctx
->refs
);
9987 * Drop uring mutex before waiting for references to exit. If
9988 * another thread is currently inside io_uring_enter() it might
9989 * need to grab the uring_lock to make progress. If we hold it
9990 * here across the drain wait, then we can deadlock. It's safe
9991 * to drop the mutex here, since no new references will come in
9992 * after we've killed the percpu ref.
9994 mutex_unlock(&ctx
->uring_lock
);
9996 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
9999 ret
= io_run_task_work_sig();
10004 mutex_lock(&ctx
->uring_lock
);
10007 percpu_ref_resurrect(&ctx
->refs
);
10012 if (ctx
->restricted
) {
10013 if (opcode
>= IORING_REGISTER_LAST
) {
10018 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
10025 case IORING_REGISTER_BUFFERS
:
10026 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
10028 case IORING_UNREGISTER_BUFFERS
:
10030 if (arg
|| nr_args
)
10032 ret
= io_sqe_buffer_unregister(ctx
);
10034 case IORING_REGISTER_FILES
:
10035 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
10037 case IORING_UNREGISTER_FILES
:
10039 if (arg
|| nr_args
)
10041 ret
= io_sqe_files_unregister(ctx
);
10043 case IORING_REGISTER_FILES_UPDATE
:
10044 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
10046 case IORING_REGISTER_EVENTFD
:
10047 case IORING_REGISTER_EVENTFD_ASYNC
:
10051 ret
= io_eventfd_register(ctx
, arg
);
10054 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10055 ctx
->eventfd_async
= 1;
10057 ctx
->eventfd_async
= 0;
10059 case IORING_UNREGISTER_EVENTFD
:
10061 if (arg
|| nr_args
)
10063 ret
= io_eventfd_unregister(ctx
);
10065 case IORING_REGISTER_PROBE
:
10067 if (!arg
|| nr_args
> 256)
10069 ret
= io_probe(ctx
, arg
, nr_args
);
10071 case IORING_REGISTER_PERSONALITY
:
10073 if (arg
|| nr_args
)
10075 ret
= io_register_personality(ctx
);
10077 case IORING_UNREGISTER_PERSONALITY
:
10081 ret
= io_unregister_personality(ctx
, nr_args
);
10083 case IORING_REGISTER_ENABLE_RINGS
:
10085 if (arg
|| nr_args
)
10087 ret
= io_register_enable_rings(ctx
);
10089 case IORING_REGISTER_RESTRICTIONS
:
10090 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10098 if (io_register_op_must_quiesce(opcode
)) {
10099 /* bring the ctx back to life */
10100 percpu_ref_reinit(&ctx
->refs
);
10102 reinit_completion(&ctx
->ref_comp
);
10107 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10108 void __user
*, arg
, unsigned int, nr_args
)
10110 struct io_ring_ctx
*ctx
;
10119 if (f
.file
->f_op
!= &io_uring_fops
)
10122 ctx
= f
.file
->private_data
;
10124 mutex_lock(&ctx
->uring_lock
);
10125 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10126 mutex_unlock(&ctx
->uring_lock
);
10127 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10128 ctx
->cq_ev_fd
!= NULL
, ret
);
10134 static int __init
io_uring_init(void)
10136 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10137 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10138 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10141 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10142 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10143 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10144 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10145 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10146 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10147 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10148 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10149 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10150 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10151 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10152 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10153 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10154 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10155 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10156 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10157 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10158 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10159 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10160 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10161 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10162 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10163 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10164 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10165 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10166 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10167 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10168 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10169 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10170 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10171 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10173 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10174 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
10175 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
10178 __initcall(io_uring_init
);