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 xarray io_buffers
;
350 struct xarray personalities
;
354 unsigned cached_cq_tail
;
357 atomic_t cq_timeouts
;
358 unsigned cq_last_tm_flush
;
359 unsigned long cq_check_overflow
;
360 struct wait_queue_head cq_wait
;
361 struct fasync_struct
*cq_fasync
;
362 struct eventfd_ctx
*cq_ev_fd
;
363 } ____cacheline_aligned_in_smp
;
366 struct mutex uring_lock
;
367 wait_queue_head_t wait
;
368 } ____cacheline_aligned_in_smp
;
371 spinlock_t completion_lock
;
374 * ->iopoll_list is protected by the ctx->uring_lock for
375 * io_uring instances that don't use IORING_SETUP_SQPOLL.
376 * For SQPOLL, only the single threaded io_sq_thread() will
377 * manipulate the list, hence no extra locking is needed there.
379 struct list_head iopoll_list
;
380 struct hlist_head
*cancel_hash
;
381 unsigned cancel_hash_bits
;
382 bool poll_multi_file
;
384 spinlock_t inflight_lock
;
385 struct list_head inflight_list
;
386 } ____cacheline_aligned_in_smp
;
388 struct delayed_work file_put_work
;
389 struct llist_head file_put_llist
;
391 struct work_struct exit_work
;
392 struct io_restriction restrictions
;
396 * First field must be the file pointer in all the
397 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
399 struct io_poll_iocb
{
401 struct wait_queue_head
*head
;
405 struct wait_queue_entry wait
;
408 struct io_poll_remove
{
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
|
863 [IORING_OP_RECVMSG
] = {
865 .unbound_nonreg_file
= 1,
868 .needs_async_data
= 1,
869 .async_size
= sizeof(struct io_async_msghdr
),
870 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
873 [IORING_OP_TIMEOUT
] = {
874 .needs_async_data
= 1,
875 .async_size
= sizeof(struct io_timeout_data
),
876 .work_flags
= IO_WQ_WORK_MM
,
878 [IORING_OP_TIMEOUT_REMOVE
] = {
879 /* used by timeout updates' prep() */
880 .work_flags
= IO_WQ_WORK_MM
,
882 [IORING_OP_ACCEPT
] = {
884 .unbound_nonreg_file
= 1,
886 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
,
888 [IORING_OP_ASYNC_CANCEL
] = {},
889 [IORING_OP_LINK_TIMEOUT
] = {
890 .needs_async_data
= 1,
891 .async_size
= sizeof(struct io_timeout_data
),
892 .work_flags
= IO_WQ_WORK_MM
,
894 [IORING_OP_CONNECT
] = {
896 .unbound_nonreg_file
= 1,
898 .needs_async_data
= 1,
899 .async_size
= sizeof(struct io_async_connect
),
900 .work_flags
= IO_WQ_WORK_MM
,
902 [IORING_OP_FALLOCATE
] = {
904 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
906 [IORING_OP_OPENAT
] = {
907 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
|
908 IO_WQ_WORK_FS
| IO_WQ_WORK_MM
,
910 [IORING_OP_CLOSE
] = {
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
] = {
932 .unbound_nonreg_file
= 1,
935 .async_size
= sizeof(struct io_async_rw
),
936 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
939 [IORING_OP_FADVISE
] = {
941 .work_flags
= IO_WQ_WORK_BLKCG
,
943 [IORING_OP_MADVISE
] = {
944 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
948 .unbound_nonreg_file
= 1,
950 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
954 .unbound_nonreg_file
= 1,
957 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
959 [IORING_OP_OPENAT2
] = {
960 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_FS
|
961 IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
963 [IORING_OP_EPOLL_CTL
] = {
964 .unbound_nonreg_file
= 1,
965 .work_flags
= IO_WQ_WORK_FILES
,
967 [IORING_OP_SPLICE
] = {
970 .unbound_nonreg_file
= 1,
971 .work_flags
= IO_WQ_WORK_BLKCG
,
973 [IORING_OP_PROVIDE_BUFFERS
] = {},
974 [IORING_OP_REMOVE_BUFFERS
] = {},
978 .unbound_nonreg_file
= 1,
980 [IORING_OP_SHUTDOWN
] = {
983 [IORING_OP_RENAMEAT
] = {
984 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
985 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
987 [IORING_OP_UNLINKAT
] = {
988 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
989 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
993 enum io_mem_account
{
998 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
999 struct task_struct
*task
,
1000 struct files_struct
*files
);
1001 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
);
1002 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
1003 struct io_ring_ctx
*ctx
);
1005 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
1006 struct io_comp_state
*cs
);
1007 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
1008 static void io_put_req(struct io_kiocb
*req
);
1009 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1010 static void io_double_put_req(struct io_kiocb
*req
);
1011 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1012 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
1013 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1014 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
1015 struct io_uring_files_update
*ip
,
1017 static void __io_clean_op(struct io_kiocb
*req
);
1018 static struct file
*io_file_get(struct io_submit_state
*state
,
1019 struct io_kiocb
*req
, int fd
, bool fixed
);
1020 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
);
1021 static void io_file_put_work(struct work_struct
*work
);
1023 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
1024 struct iovec
**iovec
, struct iov_iter
*iter
,
1026 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
1027 const struct iovec
*fast_iov
,
1028 struct iov_iter
*iter
, bool force
);
1029 static void io_req_drop_files(struct io_kiocb
*req
);
1030 static void io_req_task_queue(struct io_kiocb
*req
);
1032 static struct kmem_cache
*req_cachep
;
1034 static const struct file_operations io_uring_fops
;
1036 struct sock
*io_uring_get_socket(struct file
*file
)
1038 #if defined(CONFIG_UNIX)
1039 if (file
->f_op
== &io_uring_fops
) {
1040 struct io_ring_ctx
*ctx
= file
->private_data
;
1042 return ctx
->ring_sock
->sk
;
1047 EXPORT_SYMBOL(io_uring_get_socket
);
1049 #define io_for_each_link(pos, head) \
1050 for (pos = (head); pos; pos = pos->link)
1052 static inline void io_clean_op(struct io_kiocb
*req
)
1054 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
))
1058 static inline void io_set_resource_node(struct io_kiocb
*req
)
1060 struct io_ring_ctx
*ctx
= req
->ctx
;
1062 if (!req
->fixed_file_refs
) {
1063 req
->fixed_file_refs
= &ctx
->file_data
->node
->refs
;
1064 percpu_ref_get(req
->fixed_file_refs
);
1068 static bool io_match_task(struct io_kiocb
*head
,
1069 struct task_struct
*task
,
1070 struct files_struct
*files
)
1072 struct io_kiocb
*req
;
1074 if (task
&& head
->task
!= task
) {
1075 /* in terms of cancelation, always match if req task is dead */
1076 if (head
->task
->flags
& PF_EXITING
)
1083 io_for_each_link(req
, head
) {
1084 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1086 if (req
->file
&& req
->file
->f_op
== &io_uring_fops
)
1088 if ((req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1089 req
->work
.identity
->files
== files
)
1095 static void io_sq_thread_drop_mm_files(void)
1097 struct files_struct
*files
= current
->files
;
1098 struct mm_struct
*mm
= current
->mm
;
1101 kthread_unuse_mm(mm
);
1106 struct nsproxy
*nsproxy
= current
->nsproxy
;
1109 current
->files
= NULL
;
1110 current
->nsproxy
= NULL
;
1111 task_unlock(current
);
1112 put_files_struct(files
);
1113 put_nsproxy(nsproxy
);
1117 static int __io_sq_thread_acquire_files(struct io_ring_ctx
*ctx
)
1119 if (current
->flags
& PF_EXITING
)
1122 if (!current
->files
) {
1123 struct files_struct
*files
;
1124 struct nsproxy
*nsproxy
;
1126 task_lock(ctx
->sqo_task
);
1127 files
= ctx
->sqo_task
->files
;
1129 task_unlock(ctx
->sqo_task
);
1132 atomic_inc(&files
->count
);
1133 get_nsproxy(ctx
->sqo_task
->nsproxy
);
1134 nsproxy
= ctx
->sqo_task
->nsproxy
;
1135 task_unlock(ctx
->sqo_task
);
1138 current
->files
= files
;
1139 current
->nsproxy
= nsproxy
;
1140 task_unlock(current
);
1145 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1147 struct mm_struct
*mm
;
1149 if (current
->flags
& PF_EXITING
)
1154 /* Should never happen */
1155 if (unlikely(!(ctx
->flags
& IORING_SETUP_SQPOLL
)))
1158 task_lock(ctx
->sqo_task
);
1159 mm
= ctx
->sqo_task
->mm
;
1160 if (unlikely(!mm
|| !mmget_not_zero(mm
)))
1162 task_unlock(ctx
->sqo_task
);
1172 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx
*ctx
,
1173 struct io_kiocb
*req
)
1175 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1178 if (def
->work_flags
& IO_WQ_WORK_MM
) {
1179 ret
= __io_sq_thread_acquire_mm(ctx
);
1184 if (def
->needs_file
|| (def
->work_flags
& IO_WQ_WORK_FILES
)) {
1185 ret
= __io_sq_thread_acquire_files(ctx
);
1193 static void io_sq_thread_associate_blkcg(struct io_ring_ctx
*ctx
,
1194 struct cgroup_subsys_state
**cur_css
)
1197 #ifdef CONFIG_BLK_CGROUP
1198 /* puts the old one when swapping */
1199 if (*cur_css
!= ctx
->sqo_blkcg_css
) {
1200 kthread_associate_blkcg(ctx
->sqo_blkcg_css
);
1201 *cur_css
= ctx
->sqo_blkcg_css
;
1206 static void io_sq_thread_unassociate_blkcg(void)
1208 #ifdef CONFIG_BLK_CGROUP
1209 kthread_associate_blkcg(NULL
);
1213 static inline void req_set_fail_links(struct io_kiocb
*req
)
1215 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1216 req
->flags
|= REQ_F_FAIL_LINK
;
1220 * None of these are dereferenced, they are simply used to check if any of
1221 * them have changed. If we're under current and check they are still the
1222 * same, we're fine to grab references to them for actual out-of-line use.
1224 static void io_init_identity(struct io_identity
*id
)
1226 id
->files
= current
->files
;
1227 id
->mm
= current
->mm
;
1228 #ifdef CONFIG_BLK_CGROUP
1230 id
->blkcg_css
= blkcg_css();
1233 id
->creds
= current_cred();
1234 id
->nsproxy
= current
->nsproxy
;
1235 id
->fs
= current
->fs
;
1236 id
->fsize
= rlimit(RLIMIT_FSIZE
);
1238 id
->loginuid
= current
->loginuid
;
1239 id
->sessionid
= current
->sessionid
;
1241 refcount_set(&id
->count
, 1);
1244 static inline void __io_req_init_async(struct io_kiocb
*req
)
1246 memset(&req
->work
, 0, sizeof(req
->work
));
1247 req
->flags
|= REQ_F_WORK_INITIALIZED
;
1251 * Note: must call io_req_init_async() for the first time you
1252 * touch any members of io_wq_work.
1254 static inline void io_req_init_async(struct io_kiocb
*req
)
1256 struct io_uring_task
*tctx
= current
->io_uring
;
1258 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
1261 __io_req_init_async(req
);
1263 /* Grab a ref if this isn't our static identity */
1264 req
->work
.identity
= tctx
->identity
;
1265 if (tctx
->identity
!= &tctx
->__identity
)
1266 refcount_inc(&req
->work
.identity
->count
);
1269 static inline bool io_async_submit(struct io_ring_ctx
*ctx
)
1271 return ctx
->flags
& IORING_SETUP_SQPOLL
;
1274 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1276 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1278 complete(&ctx
->ref_comp
);
1281 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1283 return !req
->timeout
.off
;
1286 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1288 struct io_ring_ctx
*ctx
;
1291 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1295 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
1296 if (!ctx
->fallback_req
)
1300 * Use 5 bits less than the max cq entries, that should give us around
1301 * 32 entries per hash list if totally full and uniformly spread.
1303 hash_bits
= ilog2(p
->cq_entries
);
1307 ctx
->cancel_hash_bits
= hash_bits
;
1308 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1310 if (!ctx
->cancel_hash
)
1312 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1314 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1315 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1318 ctx
->flags
= p
->flags
;
1319 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1320 INIT_LIST_HEAD(&ctx
->sqd_list
);
1321 init_waitqueue_head(&ctx
->cq_wait
);
1322 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1323 init_completion(&ctx
->ref_comp
);
1324 init_completion(&ctx
->sq_thread_comp
);
1325 xa_init_flags(&ctx
->io_buffers
, XA_FLAGS_ALLOC1
);
1326 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1327 mutex_init(&ctx
->uring_lock
);
1328 init_waitqueue_head(&ctx
->wait
);
1329 spin_lock_init(&ctx
->completion_lock
);
1330 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1331 INIT_LIST_HEAD(&ctx
->defer_list
);
1332 INIT_LIST_HEAD(&ctx
->timeout_list
);
1333 spin_lock_init(&ctx
->inflight_lock
);
1334 INIT_LIST_HEAD(&ctx
->inflight_list
);
1335 INIT_DELAYED_WORK(&ctx
->file_put_work
, io_file_put_work
);
1336 init_llist_head(&ctx
->file_put_llist
);
1339 if (ctx
->fallback_req
)
1340 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1341 kfree(ctx
->cancel_hash
);
1346 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1348 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1349 struct io_ring_ctx
*ctx
= req
->ctx
;
1351 return seq
!= ctx
->cached_cq_tail
1352 + READ_ONCE(ctx
->cached_cq_overflow
);
1358 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1360 struct io_rings
*rings
= ctx
->rings
;
1362 /* order cqe stores with ring update */
1363 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1366 static void io_put_identity(struct io_uring_task
*tctx
, struct io_kiocb
*req
)
1368 if (req
->work
.identity
== &tctx
->__identity
)
1370 if (refcount_dec_and_test(&req
->work
.identity
->count
))
1371 kfree(req
->work
.identity
);
1374 static void io_req_clean_work(struct io_kiocb
*req
)
1376 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1379 req
->flags
&= ~REQ_F_WORK_INITIALIZED
;
1381 if (req
->work
.flags
& IO_WQ_WORK_MM
) {
1382 mmdrop(req
->work
.identity
->mm
);
1383 req
->work
.flags
&= ~IO_WQ_WORK_MM
;
1385 #ifdef CONFIG_BLK_CGROUP
1386 if (req
->work
.flags
& IO_WQ_WORK_BLKCG
) {
1387 css_put(req
->work
.identity
->blkcg_css
);
1388 req
->work
.flags
&= ~IO_WQ_WORK_BLKCG
;
1391 if (req
->work
.flags
& IO_WQ_WORK_CREDS
) {
1392 put_cred(req
->work
.identity
->creds
);
1393 req
->work
.flags
&= ~IO_WQ_WORK_CREDS
;
1395 if (req
->work
.flags
& IO_WQ_WORK_FS
) {
1396 struct fs_struct
*fs
= req
->work
.identity
->fs
;
1398 spin_lock(&req
->work
.identity
->fs
->lock
);
1401 spin_unlock(&req
->work
.identity
->fs
->lock
);
1404 req
->work
.flags
&= ~IO_WQ_WORK_FS
;
1406 if (req
->flags
& REQ_F_INFLIGHT
)
1407 io_req_drop_files(req
);
1409 io_put_identity(req
->task
->io_uring
, req
);
1413 * Create a private copy of io_identity, since some fields don't match
1414 * the current context.
1416 static bool io_identity_cow(struct io_kiocb
*req
)
1418 struct io_uring_task
*tctx
= current
->io_uring
;
1419 const struct cred
*creds
= NULL
;
1420 struct io_identity
*id
;
1422 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1423 creds
= req
->work
.identity
->creds
;
1425 id
= kmemdup(req
->work
.identity
, sizeof(*id
), GFP_KERNEL
);
1426 if (unlikely(!id
)) {
1427 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1432 * We can safely just re-init the creds we copied Either the field
1433 * matches the current one, or we haven't grabbed it yet. The only
1434 * exception is ->creds, through registered personalities, so handle
1435 * that one separately.
1437 io_init_identity(id
);
1441 /* add one for this request */
1442 refcount_inc(&id
->count
);
1444 /* drop tctx and req identity references, if needed */
1445 if (tctx
->identity
!= &tctx
->__identity
&&
1446 refcount_dec_and_test(&tctx
->identity
->count
))
1447 kfree(tctx
->identity
);
1448 if (req
->work
.identity
!= &tctx
->__identity
&&
1449 refcount_dec_and_test(&req
->work
.identity
->count
))
1450 kfree(req
->work
.identity
);
1452 req
->work
.identity
= id
;
1453 tctx
->identity
= id
;
1457 static bool io_grab_identity(struct io_kiocb
*req
)
1459 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1460 struct io_identity
*id
= req
->work
.identity
;
1461 struct io_ring_ctx
*ctx
= req
->ctx
;
1463 if (def
->work_flags
& IO_WQ_WORK_FSIZE
) {
1464 if (id
->fsize
!= rlimit(RLIMIT_FSIZE
))
1466 req
->work
.flags
|= IO_WQ_WORK_FSIZE
;
1468 #ifdef CONFIG_BLK_CGROUP
1469 if (!(req
->work
.flags
& IO_WQ_WORK_BLKCG
) &&
1470 (def
->work_flags
& IO_WQ_WORK_BLKCG
)) {
1472 if (id
->blkcg_css
!= blkcg_css()) {
1477 * This should be rare, either the cgroup is dying or the task
1478 * is moving cgroups. Just punt to root for the handful of ios.
1480 if (css_tryget_online(id
->blkcg_css
))
1481 req
->work
.flags
|= IO_WQ_WORK_BLKCG
;
1485 if (!(req
->work
.flags
& IO_WQ_WORK_CREDS
)) {
1486 if (id
->creds
!= current_cred())
1488 get_cred(id
->creds
);
1489 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
1492 if (!uid_eq(current
->loginuid
, id
->loginuid
) ||
1493 current
->sessionid
!= id
->sessionid
)
1496 if (!(req
->work
.flags
& IO_WQ_WORK_FS
) &&
1497 (def
->work_flags
& IO_WQ_WORK_FS
)) {
1498 if (current
->fs
!= id
->fs
)
1500 spin_lock(&id
->fs
->lock
);
1501 if (!id
->fs
->in_exec
) {
1503 req
->work
.flags
|= IO_WQ_WORK_FS
;
1505 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1507 spin_unlock(¤t
->fs
->lock
);
1509 if (!(req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1510 (def
->work_flags
& IO_WQ_WORK_FILES
) &&
1511 !(req
->flags
& REQ_F_NO_FILE_TABLE
)) {
1512 if (id
->files
!= current
->files
||
1513 id
->nsproxy
!= current
->nsproxy
)
1515 atomic_inc(&id
->files
->count
);
1516 get_nsproxy(id
->nsproxy
);
1518 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1519 req
->flags
|= REQ_F_INFLIGHT
;
1521 spin_lock_irq(&ctx
->inflight_lock
);
1522 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1523 spin_unlock_irq(&ctx
->inflight_lock
);
1525 req
->work
.flags
|= IO_WQ_WORK_FILES
;
1527 if (!(req
->work
.flags
& IO_WQ_WORK_MM
) &&
1528 (def
->work_flags
& IO_WQ_WORK_MM
)) {
1529 if (id
->mm
!= current
->mm
)
1532 req
->work
.flags
|= IO_WQ_WORK_MM
;
1538 static void io_prep_async_work(struct io_kiocb
*req
)
1540 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1541 struct io_ring_ctx
*ctx
= req
->ctx
;
1543 io_req_init_async(req
);
1545 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1546 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1548 if (req
->flags
& REQ_F_ISREG
) {
1549 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1550 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1551 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
1552 if (def
->unbound_nonreg_file
)
1553 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1556 /* if we fail grabbing identity, we must COW, regrab, and retry */
1557 if (io_grab_identity(req
))
1560 if (!io_identity_cow(req
))
1563 /* can't fail at this point */
1564 if (!io_grab_identity(req
))
1568 static void io_prep_async_link(struct io_kiocb
*req
)
1570 struct io_kiocb
*cur
;
1572 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
1573 struct io_ring_ctx
*ctx
= req
->ctx
;
1575 spin_lock_irq(&ctx
->completion_lock
);
1576 io_for_each_link(cur
, req
)
1577 io_prep_async_work(cur
);
1578 spin_unlock_irq(&ctx
->completion_lock
);
1580 io_for_each_link(cur
, req
)
1581 io_prep_async_work(cur
);
1585 static struct io_kiocb
*__io_queue_async_work(struct io_kiocb
*req
)
1587 struct io_ring_ctx
*ctx
= req
->ctx
;
1588 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1590 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1591 &req
->work
, req
->flags
);
1592 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1596 static void io_queue_async_work(struct io_kiocb
*req
)
1598 struct io_kiocb
*link
;
1600 /* init ->work of the whole link before punting */
1601 io_prep_async_link(req
);
1602 link
= __io_queue_async_work(req
);
1605 io_queue_linked_timeout(link
);
1608 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1610 struct io_timeout_data
*io
= req
->async_data
;
1613 ret
= hrtimer_try_to_cancel(&io
->timer
);
1616 req_set_fail_links(req
);
1617 atomic_set(&req
->ctx
->cq_timeouts
,
1618 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1619 list_del_init(&req
->timeout
.list
);
1620 io_cqring_fill_event(req
, status
);
1621 io_put_req_deferred(req
, 1);
1626 * Returns true if we found and killed one or more timeouts
1628 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
1629 struct files_struct
*files
)
1631 struct io_kiocb
*req
, *tmp
;
1634 spin_lock_irq(&ctx
->completion_lock
);
1635 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1636 if (io_match_task(req
, tsk
, files
)) {
1637 io_kill_timeout(req
, -ECANCELED
);
1641 spin_unlock_irq(&ctx
->completion_lock
);
1642 return canceled
!= 0;
1645 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1648 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1649 struct io_defer_entry
, list
);
1651 if (req_need_defer(de
->req
, de
->seq
))
1653 list_del_init(&de
->list
);
1654 io_req_task_queue(de
->req
);
1656 } while (!list_empty(&ctx
->defer_list
));
1659 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1663 if (list_empty(&ctx
->timeout_list
))
1666 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1669 u32 events_needed
, events_got
;
1670 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1671 struct io_kiocb
, timeout
.list
);
1673 if (io_is_timeout_noseq(req
))
1677 * Since seq can easily wrap around over time, subtract
1678 * the last seq at which timeouts were flushed before comparing.
1679 * Assuming not more than 2^31-1 events have happened since,
1680 * these subtractions won't have wrapped, so we can check if
1681 * target is in [last_seq, current_seq] by comparing the two.
1683 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1684 events_got
= seq
- ctx
->cq_last_tm_flush
;
1685 if (events_got
< events_needed
)
1688 list_del_init(&req
->timeout
.list
);
1689 io_kill_timeout(req
, 0);
1690 } while (!list_empty(&ctx
->timeout_list
));
1692 ctx
->cq_last_tm_flush
= seq
;
1695 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1697 io_flush_timeouts(ctx
);
1698 __io_commit_cqring(ctx
);
1700 if (unlikely(!list_empty(&ctx
->defer_list
)))
1701 __io_queue_deferred(ctx
);
1704 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1706 struct io_rings
*r
= ctx
->rings
;
1708 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1711 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1713 struct io_rings
*rings
= ctx
->rings
;
1716 tail
= ctx
->cached_cq_tail
;
1718 * writes to the cq entry need to come after reading head; the
1719 * control dependency is enough as we're using WRITE_ONCE to
1722 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1725 ctx
->cached_cq_tail
++;
1726 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1729 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1733 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1735 if (!ctx
->eventfd_async
)
1737 return io_wq_current_is_worker();
1740 static inline unsigned __io_cqring_events(struct io_ring_ctx
*ctx
)
1742 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1745 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1747 /* see waitqueue_active() comment */
1750 if (waitqueue_active(&ctx
->wait
))
1751 wake_up(&ctx
->wait
);
1752 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1753 wake_up(&ctx
->sq_data
->wait
);
1754 if (io_should_trigger_evfd(ctx
))
1755 eventfd_signal(ctx
->cq_ev_fd
, 1);
1756 if (waitqueue_active(&ctx
->cq_wait
)) {
1757 wake_up_interruptible(&ctx
->cq_wait
);
1758 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1762 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1764 /* see waitqueue_active() comment */
1767 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1768 if (waitqueue_active(&ctx
->wait
))
1769 wake_up(&ctx
->wait
);
1771 if (io_should_trigger_evfd(ctx
))
1772 eventfd_signal(ctx
->cq_ev_fd
, 1);
1773 if (waitqueue_active(&ctx
->cq_wait
)) {
1774 wake_up_interruptible(&ctx
->cq_wait
);
1775 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1779 /* Returns true if there are no backlogged entries after the flush */
1780 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1781 struct task_struct
*tsk
,
1782 struct files_struct
*files
)
1784 struct io_rings
*rings
= ctx
->rings
;
1785 struct io_kiocb
*req
, *tmp
;
1786 struct io_uring_cqe
*cqe
;
1787 unsigned long flags
;
1788 bool all_flushed
, posted
;
1791 if (!force
&& __io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1795 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1796 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1797 if (!io_match_task(req
, tsk
, files
))
1800 cqe
= io_get_cqring(ctx
);
1804 list_move(&req
->compl.list
, &list
);
1806 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1807 WRITE_ONCE(cqe
->res
, req
->result
);
1808 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1810 ctx
->cached_cq_overflow
++;
1811 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1812 ctx
->cached_cq_overflow
);
1817 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1819 clear_bit(0, &ctx
->sq_check_overflow
);
1820 clear_bit(0, &ctx
->cq_check_overflow
);
1821 WRITE_ONCE(ctx
->rings
->sq_flags
,
1822 ctx
->rings
->sq_flags
& ~IORING_SQ_CQ_OVERFLOW
);
1826 io_commit_cqring(ctx
);
1827 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1829 io_cqring_ev_posted(ctx
);
1831 while (!list_empty(&list
)) {
1832 req
= list_first_entry(&list
, struct io_kiocb
, compl.list
);
1833 list_del(&req
->compl.list
);
1840 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1841 struct task_struct
*tsk
,
1842 struct files_struct
*files
)
1846 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1847 /* iopoll syncs against uring_lock, not completion_lock */
1848 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1849 mutex_lock(&ctx
->uring_lock
);
1850 ret
= __io_cqring_overflow_flush(ctx
, force
, tsk
, files
);
1851 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1852 mutex_unlock(&ctx
->uring_lock
);
1858 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
,
1859 unsigned int cflags
)
1861 struct io_ring_ctx
*ctx
= req
->ctx
;
1862 struct io_uring_cqe
*cqe
;
1864 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1867 * If we can't get a cq entry, userspace overflowed the
1868 * submission (by quite a lot). Increment the overflow count in
1871 cqe
= io_get_cqring(ctx
);
1873 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1874 WRITE_ONCE(cqe
->res
, res
);
1875 WRITE_ONCE(cqe
->flags
, cflags
);
1876 } else if (ctx
->cq_overflow_flushed
||
1877 atomic_read(&req
->task
->io_uring
->in_idle
)) {
1879 * If we're in ring overflow flush mode, or in task cancel mode,
1880 * then we cannot store the request for later flushing, we need
1881 * to drop it on the floor.
1883 ctx
->cached_cq_overflow
++;
1884 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1886 if (list_empty(&ctx
->cq_overflow_list
)) {
1887 set_bit(0, &ctx
->sq_check_overflow
);
1888 set_bit(0, &ctx
->cq_check_overflow
);
1889 WRITE_ONCE(ctx
->rings
->sq_flags
,
1890 ctx
->rings
->sq_flags
| IORING_SQ_CQ_OVERFLOW
);
1895 req
->compl.cflags
= cflags
;
1896 refcount_inc(&req
->refs
);
1897 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1901 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1903 __io_cqring_fill_event(req
, res
, 0);
1906 static void io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1908 struct io_ring_ctx
*ctx
= req
->ctx
;
1909 unsigned long flags
;
1911 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1912 __io_cqring_fill_event(req
, res
, cflags
);
1913 io_commit_cqring(ctx
);
1914 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1916 io_cqring_ev_posted(ctx
);
1919 static void io_submit_flush_completions(struct io_comp_state
*cs
)
1921 struct io_ring_ctx
*ctx
= cs
->ctx
;
1923 spin_lock_irq(&ctx
->completion_lock
);
1924 while (!list_empty(&cs
->list
)) {
1925 struct io_kiocb
*req
;
1927 req
= list_first_entry(&cs
->list
, struct io_kiocb
, compl.list
);
1928 list_del(&req
->compl.list
);
1929 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
1932 * io_free_req() doesn't care about completion_lock unless one
1933 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1934 * because of a potential deadlock with req->work.fs->lock
1936 if (req
->flags
& (REQ_F_FAIL_LINK
|REQ_F_LINK_TIMEOUT
1937 |REQ_F_WORK_INITIALIZED
)) {
1938 spin_unlock_irq(&ctx
->completion_lock
);
1940 spin_lock_irq(&ctx
->completion_lock
);
1945 io_commit_cqring(ctx
);
1946 spin_unlock_irq(&ctx
->completion_lock
);
1948 io_cqring_ev_posted(ctx
);
1952 static void __io_req_complete(struct io_kiocb
*req
, long res
, unsigned cflags
,
1953 struct io_comp_state
*cs
)
1956 io_cqring_add_event(req
, res
, cflags
);
1961 req
->compl.cflags
= cflags
;
1962 list_add_tail(&req
->compl.list
, &cs
->list
);
1964 io_submit_flush_completions(cs
);
1968 static void io_req_complete(struct io_kiocb
*req
, long res
)
1970 __io_req_complete(req
, res
, 0, NULL
);
1973 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1975 return req
== (struct io_kiocb
*)
1976 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1979 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1981 struct io_kiocb
*req
;
1983 req
= ctx
->fallback_req
;
1984 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1990 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1991 struct io_submit_state
*state
)
1993 if (!state
->free_reqs
) {
1994 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1998 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1999 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
2002 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2003 * retry single alloc to be on the safe side.
2005 if (unlikely(ret
<= 0)) {
2006 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
2007 if (!state
->reqs
[0])
2011 state
->free_reqs
= ret
;
2015 return state
->reqs
[state
->free_reqs
];
2017 return io_get_fallback_req(ctx
);
2020 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
2027 static void io_dismantle_req(struct io_kiocb
*req
)
2031 if (req
->async_data
)
2032 kfree(req
->async_data
);
2034 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
2035 if (req
->fixed_file_refs
)
2036 percpu_ref_put(req
->fixed_file_refs
);
2037 io_req_clean_work(req
);
2040 static void __io_free_req(struct io_kiocb
*req
)
2042 struct io_uring_task
*tctx
= req
->task
->io_uring
;
2043 struct io_ring_ctx
*ctx
= req
->ctx
;
2045 io_dismantle_req(req
);
2047 percpu_counter_dec(&tctx
->inflight
);
2048 if (atomic_read(&tctx
->in_idle
))
2049 wake_up(&tctx
->wait
);
2050 put_task_struct(req
->task
);
2052 if (likely(!io_is_fallback_req(req
)))
2053 kmem_cache_free(req_cachep
, req
);
2055 clear_bit_unlock(0, (unsigned long *) &ctx
->fallback_req
);
2056 percpu_ref_put(&ctx
->refs
);
2059 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2061 struct io_kiocb
*nxt
= req
->link
;
2063 req
->link
= nxt
->link
;
2067 static void io_kill_linked_timeout(struct io_kiocb
*req
)
2069 struct io_ring_ctx
*ctx
= req
->ctx
;
2070 struct io_kiocb
*link
;
2071 bool cancelled
= false;
2072 unsigned long flags
;
2074 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2078 * Can happen if a linked timeout fired and link had been like
2079 * req -> link t-out -> link t-out [-> ...]
2081 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
2082 struct io_timeout_data
*io
= link
->async_data
;
2085 io_remove_next_linked(req
);
2086 link
->timeout
.head
= NULL
;
2087 ret
= hrtimer_try_to_cancel(&io
->timer
);
2089 io_cqring_fill_event(link
, -ECANCELED
);
2090 io_commit_cqring(ctx
);
2094 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
2095 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2098 io_cqring_ev_posted(ctx
);
2104 static void io_fail_links(struct io_kiocb
*req
)
2106 struct io_kiocb
*link
, *nxt
;
2107 struct io_ring_ctx
*ctx
= req
->ctx
;
2108 unsigned long flags
;
2110 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2118 trace_io_uring_fail_link(req
, link
);
2119 io_cqring_fill_event(link
, -ECANCELED
);
2122 * It's ok to free under spinlock as they're not linked anymore,
2123 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2126 if (link
->flags
& REQ_F_WORK_INITIALIZED
)
2127 io_put_req_deferred(link
, 2);
2129 io_double_put_req(link
);
2132 io_commit_cqring(ctx
);
2133 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2135 io_cqring_ev_posted(ctx
);
2138 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2140 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
2141 io_kill_linked_timeout(req
);
2144 * If LINK is set, we have dependent requests in this chain. If we
2145 * didn't fail this request, queue the first one up, moving any other
2146 * dependencies to the next request. In case of failure, fail the rest
2149 if (likely(!(req
->flags
& REQ_F_FAIL_LINK
))) {
2150 struct io_kiocb
*nxt
= req
->link
;
2159 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2161 if (likely(!(req
->link
) && !(req
->flags
& REQ_F_LINK_TIMEOUT
)))
2163 return __io_req_find_next(req
);
2166 static int io_req_task_work_add(struct io_kiocb
*req
)
2168 struct task_struct
*tsk
= req
->task
;
2169 struct io_ring_ctx
*ctx
= req
->ctx
;
2170 enum task_work_notify_mode notify
;
2173 if (tsk
->flags
& PF_EXITING
)
2177 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2178 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2179 * processing task_work. There's no reliable way to tell if TWA_RESUME
2183 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
))
2184 notify
= TWA_SIGNAL
;
2186 ret
= task_work_add(tsk
, &req
->task_work
, notify
);
2188 wake_up_process(tsk
);
2193 static void io_req_task_work_add_fallback(struct io_kiocb
*req
,
2194 void (*cb
)(struct callback_head
*))
2196 struct task_struct
*tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2198 init_task_work(&req
->task_work
, cb
);
2199 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2200 wake_up_process(tsk
);
2203 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
2205 struct io_ring_ctx
*ctx
= req
->ctx
;
2207 spin_lock_irq(&ctx
->completion_lock
);
2208 io_cqring_fill_event(req
, error
);
2209 io_commit_cqring(ctx
);
2210 spin_unlock_irq(&ctx
->completion_lock
);
2212 io_cqring_ev_posted(ctx
);
2213 req_set_fail_links(req
);
2214 io_double_put_req(req
);
2217 static void io_req_task_cancel(struct callback_head
*cb
)
2219 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2220 struct io_ring_ctx
*ctx
= req
->ctx
;
2222 mutex_lock(&ctx
->uring_lock
);
2223 __io_req_task_cancel(req
, -ECANCELED
);
2224 mutex_unlock(&ctx
->uring_lock
);
2225 percpu_ref_put(&ctx
->refs
);
2228 static void __io_req_task_submit(struct io_kiocb
*req
)
2230 struct io_ring_ctx
*ctx
= req
->ctx
;
2232 mutex_lock(&ctx
->uring_lock
);
2233 if (!ctx
->sqo_dead
&&
2234 !__io_sq_thread_acquire_mm(ctx
) &&
2235 !__io_sq_thread_acquire_files(ctx
))
2236 __io_queue_sqe(req
, NULL
);
2238 __io_req_task_cancel(req
, -EFAULT
);
2239 mutex_unlock(&ctx
->uring_lock
);
2241 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
2242 io_sq_thread_drop_mm_files();
2245 static void io_req_task_submit(struct callback_head
*cb
)
2247 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2248 struct io_ring_ctx
*ctx
= req
->ctx
;
2250 __io_req_task_submit(req
);
2251 percpu_ref_put(&ctx
->refs
);
2254 static void io_req_task_queue(struct io_kiocb
*req
)
2258 init_task_work(&req
->task_work
, io_req_task_submit
);
2259 percpu_ref_get(&req
->ctx
->refs
);
2261 ret
= io_req_task_work_add(req
);
2263 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2266 static inline void io_queue_next(struct io_kiocb
*req
)
2268 struct io_kiocb
*nxt
= io_req_find_next(req
);
2271 io_req_task_queue(nxt
);
2274 static void io_free_req(struct io_kiocb
*req
)
2281 void *reqs
[IO_IOPOLL_BATCH
];
2284 struct task_struct
*task
;
2288 static inline void io_init_req_batch(struct req_batch
*rb
)
2295 static void __io_req_free_batch_flush(struct io_ring_ctx
*ctx
,
2296 struct req_batch
*rb
)
2298 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
2299 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
2303 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2304 struct req_batch
*rb
)
2307 __io_req_free_batch_flush(ctx
, rb
);
2309 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2311 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2312 if (atomic_read(&tctx
->in_idle
))
2313 wake_up(&tctx
->wait
);
2314 put_task_struct_many(rb
->task
, rb
->task_refs
);
2319 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
)
2321 if (unlikely(io_is_fallback_req(req
))) {
2327 if (req
->task
!= rb
->task
) {
2329 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2331 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2332 if (atomic_read(&tctx
->in_idle
))
2333 wake_up(&tctx
->wait
);
2334 put_task_struct_many(rb
->task
, rb
->task_refs
);
2336 rb
->task
= req
->task
;
2341 io_dismantle_req(req
);
2342 rb
->reqs
[rb
->to_free
++] = req
;
2343 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
2344 __io_req_free_batch_flush(req
->ctx
, rb
);
2348 * Drop reference to request, return next in chain (if there is one) if this
2349 * was the last reference to this request.
2351 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2353 struct io_kiocb
*nxt
= NULL
;
2355 if (refcount_dec_and_test(&req
->refs
)) {
2356 nxt
= io_req_find_next(req
);
2362 static void io_put_req(struct io_kiocb
*req
)
2364 if (refcount_dec_and_test(&req
->refs
))
2368 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2370 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2375 static void io_free_req_deferred(struct io_kiocb
*req
)
2379 init_task_work(&req
->task_work
, io_put_req_deferred_cb
);
2380 ret
= io_req_task_work_add(req
);
2382 io_req_task_work_add_fallback(req
, io_put_req_deferred_cb
);
2385 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2387 if (refcount_sub_and_test(refs
, &req
->refs
))
2388 io_free_req_deferred(req
);
2391 static void io_double_put_req(struct io_kiocb
*req
)
2393 /* drop both submit and complete references */
2394 if (refcount_sub_and_test(2, &req
->refs
))
2398 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2400 /* See comment at the top of this file */
2402 return __io_cqring_events(ctx
);
2405 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2407 struct io_rings
*rings
= ctx
->rings
;
2409 /* make sure SQ entry isn't read before tail */
2410 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2413 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2415 unsigned int cflags
;
2417 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2418 cflags
|= IORING_CQE_F_BUFFER
;
2419 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2424 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2426 struct io_buffer
*kbuf
;
2428 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2429 return io_put_kbuf(req
, kbuf
);
2432 static inline bool io_run_task_work(void)
2435 * Not safe to run on exiting task, and the task_work handling will
2436 * not add work to such a task.
2438 if (unlikely(current
->flags
& PF_EXITING
))
2440 if (current
->task_works
) {
2441 __set_current_state(TASK_RUNNING
);
2449 static void io_iopoll_queue(struct list_head
*again
)
2451 struct io_kiocb
*req
;
2454 req
= list_first_entry(again
, struct io_kiocb
, inflight_entry
);
2455 list_del(&req
->inflight_entry
);
2456 __io_complete_rw(req
, -EAGAIN
, 0, NULL
);
2457 } while (!list_empty(again
));
2461 * Find and free completed poll iocbs
2463 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2464 struct list_head
*done
)
2466 struct req_batch rb
;
2467 struct io_kiocb
*req
;
2470 /* order with ->result store in io_complete_rw_iopoll() */
2473 io_init_req_batch(&rb
);
2474 while (!list_empty(done
)) {
2477 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2478 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2480 req
->iopoll_completed
= 0;
2481 list_move_tail(&req
->inflight_entry
, &again
);
2484 list_del(&req
->inflight_entry
);
2486 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2487 cflags
= io_put_rw_kbuf(req
);
2489 __io_cqring_fill_event(req
, req
->result
, cflags
);
2492 if (refcount_dec_and_test(&req
->refs
))
2493 io_req_free_batch(&rb
, req
);
2496 io_commit_cqring(ctx
);
2497 io_cqring_ev_posted_iopoll(ctx
);
2498 io_req_free_batch_finish(ctx
, &rb
);
2500 if (!list_empty(&again
))
2501 io_iopoll_queue(&again
);
2504 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2507 struct io_kiocb
*req
, *tmp
;
2513 * Only spin for completions if we don't have multiple devices hanging
2514 * off our complete list, and we're under the requested amount.
2516 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2519 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2520 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2523 * Move completed and retryable entries to our local lists.
2524 * If we find a request that requires polling, break out
2525 * and complete those lists first, if we have entries there.
2527 if (READ_ONCE(req
->iopoll_completed
)) {
2528 list_move_tail(&req
->inflight_entry
, &done
);
2531 if (!list_empty(&done
))
2534 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2538 /* iopoll may have completed current req */
2539 if (READ_ONCE(req
->iopoll_completed
))
2540 list_move_tail(&req
->inflight_entry
, &done
);
2547 if (!list_empty(&done
))
2548 io_iopoll_complete(ctx
, nr_events
, &done
);
2554 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2555 * non-spinning poll check - we'll still enter the driver poll loop, but only
2556 * as a non-spinning completion check.
2558 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2561 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2564 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2567 if (*nr_events
>= min
)
2575 * We can't just wait for polled events to come to us, we have to actively
2576 * find and complete them.
2578 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2580 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2583 mutex_lock(&ctx
->uring_lock
);
2584 while (!list_empty(&ctx
->iopoll_list
)) {
2585 unsigned int nr_events
= 0;
2587 io_do_iopoll(ctx
, &nr_events
, 0);
2589 /* let it sleep and repeat later if can't complete a request */
2593 * Ensure we allow local-to-the-cpu processing to take place,
2594 * in this case we need to ensure that we reap all events.
2595 * Also let task_work, etc. to progress by releasing the mutex
2597 if (need_resched()) {
2598 mutex_unlock(&ctx
->uring_lock
);
2600 mutex_lock(&ctx
->uring_lock
);
2603 mutex_unlock(&ctx
->uring_lock
);
2606 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2608 unsigned int nr_events
= 0;
2609 int iters
= 0, ret
= 0;
2612 * We disallow the app entering submit/complete with polling, but we
2613 * still need to lock the ring to prevent racing with polled issue
2614 * that got punted to a workqueue.
2616 mutex_lock(&ctx
->uring_lock
);
2619 * Don't enter poll loop if we already have events pending.
2620 * If we do, we can potentially be spinning for commands that
2621 * already triggered a CQE (eg in error).
2623 if (test_bit(0, &ctx
->cq_check_overflow
))
2624 __io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2625 if (io_cqring_events(ctx
))
2629 * If a submit got punted to a workqueue, we can have the
2630 * application entering polling for a command before it gets
2631 * issued. That app will hold the uring_lock for the duration
2632 * of the poll right here, so we need to take a breather every
2633 * now and then to ensure that the issue has a chance to add
2634 * the poll to the issued list. Otherwise we can spin here
2635 * forever, while the workqueue is stuck trying to acquire the
2638 if (!(++iters
& 7)) {
2639 mutex_unlock(&ctx
->uring_lock
);
2641 mutex_lock(&ctx
->uring_lock
);
2644 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2648 } while (min
&& !nr_events
&& !need_resched());
2650 mutex_unlock(&ctx
->uring_lock
);
2654 static void kiocb_end_write(struct io_kiocb
*req
)
2657 * Tell lockdep we inherited freeze protection from submission
2660 if (req
->flags
& REQ_F_ISREG
) {
2661 struct inode
*inode
= file_inode(req
->file
);
2663 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2665 file_end_write(req
->file
);
2668 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
,
2669 struct io_comp_state
*cs
)
2671 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2674 if (kiocb
->ki_flags
& IOCB_WRITE
)
2675 kiocb_end_write(req
);
2677 if (res
!= req
->result
)
2678 req_set_fail_links(req
);
2679 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2680 cflags
= io_put_rw_kbuf(req
);
2681 __io_req_complete(req
, res
, cflags
, cs
);
2685 static bool io_resubmit_prep(struct io_kiocb
*req
, int error
)
2687 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2688 ssize_t ret
= -ECANCELED
;
2689 struct iov_iter iter
;
2697 switch (req
->opcode
) {
2698 case IORING_OP_READV
:
2699 case IORING_OP_READ_FIXED
:
2700 case IORING_OP_READ
:
2703 case IORING_OP_WRITEV
:
2704 case IORING_OP_WRITE_FIXED
:
2705 case IORING_OP_WRITE
:
2709 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2714 if (!req
->async_data
) {
2715 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2718 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2726 req_set_fail_links(req
);
2731 static bool io_rw_reissue(struct io_kiocb
*req
, long res
)
2734 umode_t mode
= file_inode(req
->file
)->i_mode
;
2737 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2739 if ((res
!= -EAGAIN
&& res
!= -EOPNOTSUPP
) || io_wq_current_is_worker())
2742 * If ref is dying, we might be running poll reap from the exit work.
2743 * Don't attempt to reissue from that path, just let it fail with
2746 if (percpu_ref_is_dying(&req
->ctx
->refs
))
2749 lockdep_assert_held(&req
->ctx
->uring_lock
);
2751 ret
= io_sq_thread_acquire_mm_files(req
->ctx
, req
);
2753 if (io_resubmit_prep(req
, ret
)) {
2754 refcount_inc(&req
->refs
);
2755 io_queue_async_work(req
);
2763 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2764 struct io_comp_state
*cs
)
2766 if (!io_rw_reissue(req
, res
))
2767 io_complete_rw_common(&req
->rw
.kiocb
, res
, cs
);
2770 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2772 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2774 __io_complete_rw(req
, res
, res2
, NULL
);
2777 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2779 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2781 if (kiocb
->ki_flags
& IOCB_WRITE
)
2782 kiocb_end_write(req
);
2784 if (res
!= -EAGAIN
&& res
!= req
->result
)
2785 req_set_fail_links(req
);
2787 WRITE_ONCE(req
->result
, res
);
2788 /* order with io_poll_complete() checking ->result */
2790 WRITE_ONCE(req
->iopoll_completed
, 1);
2794 * After the iocb has been issued, it's safe to be found on the poll list.
2795 * Adding the kiocb to the list AFTER submission ensures that we don't
2796 * find it from a io_iopoll_getevents() thread before the issuer is done
2797 * accessing the kiocb cookie.
2799 static void io_iopoll_req_issued(struct io_kiocb
*req
, bool in_async
)
2801 struct io_ring_ctx
*ctx
= req
->ctx
;
2804 * Track whether we have multiple files in our lists. This will impact
2805 * how we do polling eventually, not spinning if we're on potentially
2806 * different devices.
2808 if (list_empty(&ctx
->iopoll_list
)) {
2809 ctx
->poll_multi_file
= false;
2810 } else if (!ctx
->poll_multi_file
) {
2811 struct io_kiocb
*list_req
;
2813 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2815 if (list_req
->file
!= req
->file
)
2816 ctx
->poll_multi_file
= true;
2820 * For fast devices, IO may have already completed. If it has, add
2821 * it to the front so we find it first.
2823 if (READ_ONCE(req
->iopoll_completed
))
2824 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2826 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2829 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2830 * task context or in io worker task context. If current task context is
2831 * sq thread, we don't need to check whether should wake up sq thread.
2833 if (in_async
&& (ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2834 wq_has_sleeper(&ctx
->sq_data
->wait
))
2835 wake_up(&ctx
->sq_data
->wait
);
2838 static inline void __io_state_file_put(struct io_submit_state
*state
)
2840 fput_many(state
->file
, state
->file_refs
);
2841 state
->file_refs
= 0;
2844 static inline void io_state_file_put(struct io_submit_state
*state
)
2846 if (state
->file_refs
)
2847 __io_state_file_put(state
);
2851 * Get as many references to a file as we have IOs left in this submission,
2852 * assuming most submissions are for one file, or at least that each file
2853 * has more than one submission.
2855 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2860 if (state
->file_refs
) {
2861 if (state
->fd
== fd
) {
2865 __io_state_file_put(state
);
2867 state
->file
= fget_many(fd
, state
->ios_left
);
2868 if (unlikely(!state
->file
))
2872 state
->file_refs
= state
->ios_left
- 1;
2876 static bool io_bdev_nowait(struct block_device
*bdev
)
2878 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2882 * If we tracked the file through the SCM inflight mechanism, we could support
2883 * any file. For now, just ensure that anything potentially problematic is done
2886 static bool io_file_supports_async(struct file
*file
, int rw
)
2888 umode_t mode
= file_inode(file
)->i_mode
;
2890 if (S_ISBLK(mode
)) {
2891 if (IS_ENABLED(CONFIG_BLOCK
) &&
2892 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2898 if (S_ISREG(mode
)) {
2899 if (IS_ENABLED(CONFIG_BLOCK
) &&
2900 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2901 file
->f_op
!= &io_uring_fops
)
2906 /* any ->read/write should understand O_NONBLOCK */
2907 if (file
->f_flags
& O_NONBLOCK
)
2910 if (!(file
->f_mode
& FMODE_NOWAIT
))
2914 return file
->f_op
->read_iter
!= NULL
;
2916 return file
->f_op
->write_iter
!= NULL
;
2919 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2921 struct io_ring_ctx
*ctx
= req
->ctx
;
2922 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2926 if (S_ISREG(file_inode(req
->file
)->i_mode
))
2927 req
->flags
|= REQ_F_ISREG
;
2929 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2930 if (kiocb
->ki_pos
== -1 && !(req
->file
->f_mode
& FMODE_STREAM
)) {
2931 req
->flags
|= REQ_F_CUR_POS
;
2932 kiocb
->ki_pos
= req
->file
->f_pos
;
2934 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2935 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2936 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2940 ioprio
= READ_ONCE(sqe
->ioprio
);
2942 ret
= ioprio_check_cap(ioprio
);
2946 kiocb
->ki_ioprio
= ioprio
;
2948 kiocb
->ki_ioprio
= get_current_ioprio();
2950 /* don't allow async punt if RWF_NOWAIT was requested */
2951 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2952 req
->flags
|= REQ_F_NOWAIT
;
2954 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2955 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2956 !kiocb
->ki_filp
->f_op
->iopoll
)
2959 kiocb
->ki_flags
|= IOCB_HIPRI
;
2960 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2961 req
->iopoll_completed
= 0;
2963 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2965 kiocb
->ki_complete
= io_complete_rw
;
2968 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2969 req
->rw
.len
= READ_ONCE(sqe
->len
);
2970 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2974 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2980 case -ERESTARTNOINTR
:
2981 case -ERESTARTNOHAND
:
2982 case -ERESTART_RESTARTBLOCK
:
2984 * We can't just restart the syscall, since previously
2985 * submitted sqes may already be in progress. Just fail this
2991 kiocb
->ki_complete(kiocb
, ret
, 0);
2995 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2996 struct io_comp_state
*cs
)
2998 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2999 struct io_async_rw
*io
= req
->async_data
;
3001 /* add previously done IO, if any */
3002 if (io
&& io
->bytes_done
> 0) {
3004 ret
= io
->bytes_done
;
3006 ret
+= io
->bytes_done
;
3009 if (req
->flags
& REQ_F_CUR_POS
)
3010 req
->file
->f_pos
= kiocb
->ki_pos
;
3011 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
3012 __io_complete_rw(req
, ret
, 0, cs
);
3014 io_rw_done(kiocb
, ret
);
3017 static ssize_t
io_import_fixed(struct io_kiocb
*req
, int rw
,
3018 struct iov_iter
*iter
)
3020 struct io_ring_ctx
*ctx
= req
->ctx
;
3021 size_t len
= req
->rw
.len
;
3022 struct io_mapped_ubuf
*imu
;
3023 u16 index
, buf_index
= req
->buf_index
;
3027 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3029 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3030 imu
= &ctx
->user_bufs
[index
];
3031 buf_addr
= req
->rw
.addr
;
3034 if (buf_addr
+ len
< buf_addr
)
3036 /* not inside the mapped region */
3037 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
3041 * May not be a start of buffer, set size appropriately
3042 * and advance us to the beginning.
3044 offset
= buf_addr
- imu
->ubuf
;
3045 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3049 * Don't use iov_iter_advance() here, as it's really slow for
3050 * using the latter parts of a big fixed buffer - it iterates
3051 * over each segment manually. We can cheat a bit here, because
3054 * 1) it's a BVEC iter, we set it up
3055 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3056 * first and last bvec
3058 * So just find our index, and adjust the iterator afterwards.
3059 * If the offset is within the first bvec (or the whole first
3060 * bvec, just use iov_iter_advance(). This makes it easier
3061 * since we can just skip the first segment, which may not
3062 * be PAGE_SIZE aligned.
3064 const struct bio_vec
*bvec
= imu
->bvec
;
3066 if (offset
<= bvec
->bv_len
) {
3067 iov_iter_advance(iter
, offset
);
3069 unsigned long seg_skip
;
3071 /* skip first vec */
3072 offset
-= bvec
->bv_len
;
3073 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3075 iter
->bvec
= bvec
+ seg_skip
;
3076 iter
->nr_segs
-= seg_skip
;
3077 iter
->count
-= bvec
->bv_len
+ offset
;
3078 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3085 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3088 mutex_unlock(&ctx
->uring_lock
);
3091 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3094 * "Normal" inline submissions always hold the uring_lock, since we
3095 * grab it from the system call. Same is true for the SQPOLL offload.
3096 * The only exception is when we've detached the request and issue it
3097 * from an async worker thread, grab the lock for that case.
3100 mutex_lock(&ctx
->uring_lock
);
3103 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3104 int bgid
, struct io_buffer
*kbuf
,
3107 struct io_buffer
*head
;
3109 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3112 io_ring_submit_lock(req
->ctx
, needs_lock
);
3114 lockdep_assert_held(&req
->ctx
->uring_lock
);
3116 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
3118 if (!list_empty(&head
->list
)) {
3119 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3121 list_del(&kbuf
->list
);
3124 xa_erase(&req
->ctx
->io_buffers
, bgid
);
3126 if (*len
> kbuf
->len
)
3129 kbuf
= ERR_PTR(-ENOBUFS
);
3132 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3137 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3140 struct io_buffer
*kbuf
;
3143 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3144 bgid
= req
->buf_index
;
3145 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3148 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3149 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3150 return u64_to_user_ptr(kbuf
->addr
);
3153 #ifdef CONFIG_COMPAT
3154 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3157 struct compat_iovec __user
*uiov
;
3158 compat_ssize_t clen
;
3162 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3163 if (!access_ok(uiov
, sizeof(*uiov
)))
3165 if (__get_user(clen
, &uiov
->iov_len
))
3171 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3173 return PTR_ERR(buf
);
3174 iov
[0].iov_base
= buf
;
3175 iov
[0].iov_len
= (compat_size_t
) len
;
3180 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3183 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3187 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3190 len
= iov
[0].iov_len
;
3193 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3195 return PTR_ERR(buf
);
3196 iov
[0].iov_base
= buf
;
3197 iov
[0].iov_len
= len
;
3201 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3204 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3205 struct io_buffer
*kbuf
;
3207 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3208 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3209 iov
[0].iov_len
= kbuf
->len
;
3212 if (req
->rw
.len
!= 1)
3215 #ifdef CONFIG_COMPAT
3216 if (req
->ctx
->compat
)
3217 return io_compat_import(req
, iov
, needs_lock
);
3220 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3223 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
3224 struct iovec
**iovec
, struct iov_iter
*iter
,
3227 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3228 size_t sqe_len
= req
->rw
.len
;
3232 opcode
= req
->opcode
;
3233 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3235 return io_import_fixed(req
, rw
, iter
);
3238 /* buffer index only valid with fixed read/write, or buffer select */
3239 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3242 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3243 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3244 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3246 return PTR_ERR(buf
);
3247 req
->rw
.len
= sqe_len
;
3250 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3255 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3256 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3258 ret
= (*iovec
)->iov_len
;
3259 iov_iter_init(iter
, rw
, *iovec
, 1, ret
);
3265 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3269 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3271 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3275 * For files that don't have ->read_iter() and ->write_iter(), handle them
3276 * by looping over ->read() or ->write() manually.
3278 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3280 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3281 struct file
*file
= req
->file
;
3285 * Don't support polled IO through this interface, and we can't
3286 * support non-blocking either. For the latter, this just causes
3287 * the kiocb to be handled from an async context.
3289 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3291 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3294 while (iov_iter_count(iter
)) {
3298 if (!iov_iter_is_bvec(iter
)) {
3299 iovec
= iov_iter_iovec(iter
);
3301 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3302 iovec
.iov_len
= req
->rw
.len
;
3306 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3307 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3309 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3310 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3318 if (!iov_iter_is_bvec(iter
)) {
3319 iov_iter_advance(iter
, nr
);
3325 if (nr
!= iovec
.iov_len
)
3332 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3333 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3335 struct io_async_rw
*rw
= req
->async_data
;
3337 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3338 rw
->free_iovec
= iovec
;
3340 /* can only be fixed buffers, no need to do anything */
3341 if (iov_iter_is_bvec(iter
))
3344 unsigned iov_off
= 0;
3346 rw
->iter
.iov
= rw
->fast_iov
;
3347 if (iter
->iov
!= fast_iov
) {
3348 iov_off
= iter
->iov
- fast_iov
;
3349 rw
->iter
.iov
+= iov_off
;
3351 if (rw
->fast_iov
!= fast_iov
)
3352 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3353 sizeof(struct iovec
) * iter
->nr_segs
);
3355 req
->flags
|= REQ_F_NEED_CLEANUP
;
3359 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3361 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3362 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3363 return req
->async_data
== NULL
;
3366 static int io_alloc_async_data(struct io_kiocb
*req
)
3368 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3371 return __io_alloc_async_data(req
);
3374 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3375 const struct iovec
*fast_iov
,
3376 struct iov_iter
*iter
, bool force
)
3378 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3380 if (!req
->async_data
) {
3381 if (__io_alloc_async_data(req
))
3384 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3389 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3391 struct io_async_rw
*iorw
= req
->async_data
;
3392 struct iovec
*iov
= iorw
->fast_iov
;
3395 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3396 if (unlikely(ret
< 0))
3399 iorw
->bytes_done
= 0;
3400 iorw
->free_iovec
= iov
;
3402 req
->flags
|= REQ_F_NEED_CLEANUP
;
3406 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3410 ret
= io_prep_rw(req
, sqe
);
3414 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3417 /* either don't need iovec imported or already have it */
3418 if (!req
->async_data
)
3420 return io_rw_prep_async(req
, READ
);
3424 * This is our waitqueue callback handler, registered through lock_page_async()
3425 * when we initially tried to do the IO with the iocb armed our waitqueue.
3426 * This gets called when the page is unlocked, and we generally expect that to
3427 * happen when the page IO is completed and the page is now uptodate. This will
3428 * queue a task_work based retry of the operation, attempting to copy the data
3429 * again. If the latter fails because the page was NOT uptodate, then we will
3430 * do a thread based blocking retry of the operation. That's the unexpected
3433 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3434 int sync
, void *arg
)
3436 struct wait_page_queue
*wpq
;
3437 struct io_kiocb
*req
= wait
->private;
3438 struct wait_page_key
*key
= arg
;
3441 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3443 if (!wake_page_match(wpq
, key
))
3446 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3447 list_del_init(&wait
->entry
);
3449 init_task_work(&req
->task_work
, io_req_task_submit
);
3450 percpu_ref_get(&req
->ctx
->refs
);
3452 /* submit ref gets dropped, acquire a new one */
3453 refcount_inc(&req
->refs
);
3454 ret
= io_req_task_work_add(req
);
3456 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
3461 * This controls whether a given IO request should be armed for async page
3462 * based retry. If we return false here, the request is handed to the async
3463 * worker threads for retry. If we're doing buffered reads on a regular file,
3464 * we prepare a private wait_page_queue entry and retry the operation. This
3465 * will either succeed because the page is now uptodate and unlocked, or it
3466 * will register a callback when the page is unlocked at IO completion. Through
3467 * that callback, io_uring uses task_work to setup a retry of the operation.
3468 * That retry will attempt the buffered read again. The retry will generally
3469 * succeed, or in rare cases where it fails, we then fall back to using the
3470 * async worker threads for a blocking retry.
3472 static bool io_rw_should_retry(struct io_kiocb
*req
)
3474 struct io_async_rw
*rw
= req
->async_data
;
3475 struct wait_page_queue
*wait
= &rw
->wpq
;
3476 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3478 /* never retry for NOWAIT, we just complete with -EAGAIN */
3479 if (req
->flags
& REQ_F_NOWAIT
)
3482 /* Only for buffered IO */
3483 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3487 * just use poll if we can, and don't attempt if the fs doesn't
3488 * support callback based unlocks
3490 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3493 wait
->wait
.func
= io_async_buf_func
;
3494 wait
->wait
.private = req
;
3495 wait
->wait
.flags
= 0;
3496 INIT_LIST_HEAD(&wait
->wait
.entry
);
3497 kiocb
->ki_flags
|= IOCB_WAITQ
;
3498 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3499 kiocb
->ki_waitq
= wait
;
3503 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3505 if (req
->file
->f_op
->read_iter
)
3506 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3507 else if (req
->file
->f_op
->read
)
3508 return loop_rw_iter(READ
, req
, iter
);
3513 static int io_read(struct io_kiocb
*req
, bool force_nonblock
,
3514 struct io_comp_state
*cs
)
3516 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3517 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3518 struct iov_iter __iter
, *iter
= &__iter
;
3519 struct io_async_rw
*rw
= req
->async_data
;
3520 ssize_t io_size
, ret
, ret2
;
3527 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3531 io_size
= iov_iter_count(iter
);
3532 req
->result
= io_size
;
3535 /* Ensure we clear previously set non-block flag */
3536 if (!force_nonblock
)
3537 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3539 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3541 /* If the file doesn't support async, just async punt */
3542 no_async
= force_nonblock
&& !io_file_supports_async(req
->file
, READ
);
3546 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3550 ret
= io_iter_do_read(req
, iter
);
3552 if (ret
== -EIOCBQUEUED
) {
3555 } else if (ret
== -EAGAIN
) {
3556 /* IOPOLL retry should happen for io-wq threads */
3557 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3559 /* no retry on NONBLOCK marked file */
3560 if (req
->file
->f_flags
& O_NONBLOCK
)
3562 /* some cases will consume bytes even on error returns */
3563 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3566 } else if (ret
<= 0) {
3567 /* make sure -ERESTARTSYS -> -EINTR is done */
3571 /* read it all, or we did blocking attempt. no retry. */
3572 if (!iov_iter_count(iter
) || !force_nonblock
||
3573 (req
->file
->f_flags
& O_NONBLOCK
) || !(req
->flags
& REQ_F_ISREG
))
3578 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3585 rw
= req
->async_data
;
3586 /* it's copied and will be cleaned with ->io */
3588 /* now use our persistent iterator, if we aren't already */
3591 rw
->bytes_done
+= ret
;
3592 /* if we can retry, do so with the callbacks armed */
3593 if (!io_rw_should_retry(req
)) {
3594 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3599 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3600 * get -EIOCBQUEUED, then we'll get a notification when the desired
3601 * page gets unlocked. We can also get a partial read here, and if we
3602 * do, then just retry at the new offset.
3604 ret
= io_iter_do_read(req
, iter
);
3605 if (ret
== -EIOCBQUEUED
) {
3608 } else if (ret
> 0 && ret
< io_size
) {
3609 /* we got some bytes, but not all. retry. */
3610 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3614 kiocb_done(kiocb
, ret
, cs
);
3617 /* it's reportedly faster than delegating the null check to kfree() */
3623 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3627 ret
= io_prep_rw(req
, sqe
);
3631 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3634 /* either don't need iovec imported or already have it */
3635 if (!req
->async_data
)
3637 return io_rw_prep_async(req
, WRITE
);
3640 static int io_write(struct io_kiocb
*req
, bool force_nonblock
,
3641 struct io_comp_state
*cs
)
3643 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3644 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3645 struct iov_iter __iter
, *iter
= &__iter
;
3646 struct io_async_rw
*rw
= req
->async_data
;
3647 ssize_t ret
, ret2
, io_size
;
3653 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3657 io_size
= iov_iter_count(iter
);
3658 req
->result
= io_size
;
3660 /* Ensure we clear previously set non-block flag */
3661 if (!force_nonblock
)
3662 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3664 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3666 /* If the file doesn't support async, just async punt */
3667 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3670 /* file path doesn't support NOWAIT for non-direct_IO */
3671 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3672 (req
->flags
& REQ_F_ISREG
))
3675 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3680 * Open-code file_start_write here to grab freeze protection,
3681 * which will be released by another thread in
3682 * io_complete_rw(). Fool lockdep by telling it the lock got
3683 * released so that it doesn't complain about the held lock when
3684 * we return to userspace.
3686 if (req
->flags
& REQ_F_ISREG
) {
3687 sb_start_write(file_inode(req
->file
)->i_sb
);
3688 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3691 kiocb
->ki_flags
|= IOCB_WRITE
;
3693 if (req
->file
->f_op
->write_iter
)
3694 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3695 else if (req
->file
->f_op
->write
)
3696 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3701 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3702 * retry them without IOCB_NOWAIT.
3704 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3706 /* no retry on NONBLOCK marked file */
3707 if (ret2
== -EAGAIN
&& (req
->file
->f_flags
& O_NONBLOCK
))
3709 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3710 /* IOPOLL retry should happen for io-wq threads */
3711 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3714 kiocb_done(kiocb
, ret2
, cs
);
3717 /* some cases will consume bytes even on error returns */
3718 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3719 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3724 /* it's reportedly faster than delegating the null check to kfree() */
3730 static int io_renameat_prep(struct io_kiocb
*req
,
3731 const struct io_uring_sqe
*sqe
)
3733 struct io_rename
*ren
= &req
->rename
;
3734 const char __user
*oldf
, *newf
;
3736 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3739 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3740 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3741 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3742 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3743 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3745 ren
->oldpath
= getname(oldf
);
3746 if (IS_ERR(ren
->oldpath
))
3747 return PTR_ERR(ren
->oldpath
);
3749 ren
->newpath
= getname(newf
);
3750 if (IS_ERR(ren
->newpath
)) {
3751 putname(ren
->oldpath
);
3752 return PTR_ERR(ren
->newpath
);
3755 req
->flags
|= REQ_F_NEED_CLEANUP
;
3759 static int io_renameat(struct io_kiocb
*req
, bool force_nonblock
)
3761 struct io_rename
*ren
= &req
->rename
;
3767 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3768 ren
->newpath
, ren
->flags
);
3770 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3772 req_set_fail_links(req
);
3773 io_req_complete(req
, ret
);
3777 static int io_unlinkat_prep(struct io_kiocb
*req
,
3778 const struct io_uring_sqe
*sqe
)
3780 struct io_unlink
*un
= &req
->unlink
;
3781 const char __user
*fname
;
3783 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3786 un
->dfd
= READ_ONCE(sqe
->fd
);
3788 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3789 if (un
->flags
& ~AT_REMOVEDIR
)
3792 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3793 un
->filename
= getname(fname
);
3794 if (IS_ERR(un
->filename
))
3795 return PTR_ERR(un
->filename
);
3797 req
->flags
|= REQ_F_NEED_CLEANUP
;
3801 static int io_unlinkat(struct io_kiocb
*req
, bool force_nonblock
)
3803 struct io_unlink
*un
= &req
->unlink
;
3809 if (un
->flags
& AT_REMOVEDIR
)
3810 ret
= do_rmdir(un
->dfd
, un
->filename
);
3812 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3814 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3816 req_set_fail_links(req
);
3817 io_req_complete(req
, ret
);
3821 static int io_shutdown_prep(struct io_kiocb
*req
,
3822 const struct io_uring_sqe
*sqe
)
3824 #if defined(CONFIG_NET)
3825 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3827 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3831 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3838 static int io_shutdown(struct io_kiocb
*req
, bool force_nonblock
)
3840 #if defined(CONFIG_NET)
3841 struct socket
*sock
;
3847 sock
= sock_from_file(req
->file
);
3848 if (unlikely(!sock
))
3851 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3853 req_set_fail_links(req
);
3854 io_req_complete(req
, ret
);
3861 static int __io_splice_prep(struct io_kiocb
*req
,
3862 const struct io_uring_sqe
*sqe
)
3864 struct io_splice
*sp
= &req
->splice
;
3865 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3867 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3871 sp
->len
= READ_ONCE(sqe
->len
);
3872 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3874 if (unlikely(sp
->flags
& ~valid_flags
))
3877 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3878 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3881 req
->flags
|= REQ_F_NEED_CLEANUP
;
3883 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3885 * Splice operation will be punted aync, and here need to
3886 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3888 io_req_init_async(req
);
3889 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3895 static int io_tee_prep(struct io_kiocb
*req
,
3896 const struct io_uring_sqe
*sqe
)
3898 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3900 return __io_splice_prep(req
, sqe
);
3903 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3905 struct io_splice
*sp
= &req
->splice
;
3906 struct file
*in
= sp
->file_in
;
3907 struct file
*out
= sp
->file_out
;
3908 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3914 ret
= do_tee(in
, out
, sp
->len
, flags
);
3916 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3917 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3920 req_set_fail_links(req
);
3921 io_req_complete(req
, ret
);
3925 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3927 struct io_splice
*sp
= &req
->splice
;
3929 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3930 sp
->off_out
= READ_ONCE(sqe
->off
);
3931 return __io_splice_prep(req
, sqe
);
3934 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3936 struct io_splice
*sp
= &req
->splice
;
3937 struct file
*in
= sp
->file_in
;
3938 struct file
*out
= sp
->file_out
;
3939 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3940 loff_t
*poff_in
, *poff_out
;
3946 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3947 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3950 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3952 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3953 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3956 req_set_fail_links(req
);
3957 io_req_complete(req
, ret
);
3962 * IORING_OP_NOP just posts a completion event, nothing else.
3964 static int io_nop(struct io_kiocb
*req
, struct io_comp_state
*cs
)
3966 struct io_ring_ctx
*ctx
= req
->ctx
;
3968 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3971 __io_req_complete(req
, 0, 0, cs
);
3975 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3977 struct io_ring_ctx
*ctx
= req
->ctx
;
3982 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3984 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
3988 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3989 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3992 req
->sync
.off
= READ_ONCE(sqe
->off
);
3993 req
->sync
.len
= READ_ONCE(sqe
->len
);
3997 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3999 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
4002 /* fsync always requires a blocking context */
4006 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
4007 end
> 0 ? end
: LLONG_MAX
,
4008 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
4010 req_set_fail_links(req
);
4011 io_req_complete(req
, ret
);
4015 static int io_fallocate_prep(struct io_kiocb
*req
,
4016 const struct io_uring_sqe
*sqe
)
4018 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
||
4021 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4024 req
->sync
.off
= READ_ONCE(sqe
->off
);
4025 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4026 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4030 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
4034 /* fallocate always requiring blocking context */
4037 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4040 req_set_fail_links(req
);
4041 io_req_complete(req
, ret
);
4045 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4047 const char __user
*fname
;
4050 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
))
4052 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4055 /* open.how should be already initialised */
4056 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4057 req
->open
.how
.flags
|= O_LARGEFILE
;
4059 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4060 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4061 req
->open
.filename
= getname(fname
);
4062 if (IS_ERR(req
->open
.filename
)) {
4063 ret
= PTR_ERR(req
->open
.filename
);
4064 req
->open
.filename
= NULL
;
4067 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4068 req
->open
.ignore_nonblock
= false;
4069 req
->flags
|= REQ_F_NEED_CLEANUP
;
4073 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4077 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4079 mode
= READ_ONCE(sqe
->len
);
4080 flags
= READ_ONCE(sqe
->open_flags
);
4081 req
->open
.how
= build_open_how(flags
, mode
);
4082 return __io_openat_prep(req
, sqe
);
4085 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4087 struct open_how __user
*how
;
4091 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4093 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4094 len
= READ_ONCE(sqe
->len
);
4095 if (len
< OPEN_HOW_SIZE_VER0
)
4098 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4103 return __io_openat_prep(req
, sqe
);
4106 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
4108 struct open_flags op
;
4112 if (force_nonblock
&& !req
->open
.ignore_nonblock
)
4115 ret
= build_open_flags(&req
->open
.how
, &op
);
4119 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4123 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4126 ret
= PTR_ERR(file
);
4128 * A work-around to ensure that /proc/self works that way
4129 * that it should - if we get -EOPNOTSUPP back, then assume
4130 * that proc_self_get_link() failed us because we're in async
4131 * context. We should be safe to retry this from the task
4132 * itself with force_nonblock == false set, as it should not
4133 * block on lookup. Would be nice to know this upfront and
4134 * avoid the async dance, but doesn't seem feasible.
4136 if (ret
== -EOPNOTSUPP
&& io_wq_current_is_worker()) {
4137 req
->open
.ignore_nonblock
= true;
4138 refcount_inc(&req
->refs
);
4139 io_req_task_queue(req
);
4143 fsnotify_open(file
);
4144 fd_install(ret
, file
);
4147 putname(req
->open
.filename
);
4148 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4150 req_set_fail_links(req
);
4151 io_req_complete(req
, ret
);
4155 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
4157 return io_openat2(req
, force_nonblock
);
4160 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4161 const struct io_uring_sqe
*sqe
)
4163 struct io_provide_buf
*p
= &req
->pbuf
;
4166 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
4170 tmp
= READ_ONCE(sqe
->fd
);
4171 if (!tmp
|| tmp
> USHRT_MAX
)
4174 memset(p
, 0, sizeof(*p
));
4176 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4180 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4181 int bgid
, unsigned nbufs
)
4185 /* shouldn't happen */
4189 /* the head kbuf is the list itself */
4190 while (!list_empty(&buf
->list
)) {
4191 struct io_buffer
*nxt
;
4193 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4194 list_del(&nxt
->list
);
4201 xa_erase(&ctx
->io_buffers
, bgid
);
4206 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4207 struct io_comp_state
*cs
)
4209 struct io_provide_buf
*p
= &req
->pbuf
;
4210 struct io_ring_ctx
*ctx
= req
->ctx
;
4211 struct io_buffer
*head
;
4214 io_ring_submit_lock(ctx
, !force_nonblock
);
4216 lockdep_assert_held(&ctx
->uring_lock
);
4219 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4221 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4223 req_set_fail_links(req
);
4225 /* need to hold the lock to complete IOPOLL requests */
4226 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4227 __io_req_complete(req
, ret
, 0, cs
);
4228 io_ring_submit_unlock(ctx
, !force_nonblock
);
4230 io_ring_submit_unlock(ctx
, !force_nonblock
);
4231 __io_req_complete(req
, ret
, 0, cs
);
4236 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4237 const struct io_uring_sqe
*sqe
)
4239 unsigned long size
, tmp_check
;
4240 struct io_provide_buf
*p
= &req
->pbuf
;
4243 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
4246 tmp
= READ_ONCE(sqe
->fd
);
4247 if (!tmp
|| tmp
> USHRT_MAX
)
4250 p
->addr
= READ_ONCE(sqe
->addr
);
4251 p
->len
= READ_ONCE(sqe
->len
);
4253 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
4256 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
4259 size
= (unsigned long)p
->len
* p
->nbufs
;
4260 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4263 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4264 tmp
= READ_ONCE(sqe
->off
);
4265 if (tmp
> USHRT_MAX
)
4271 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4273 struct io_buffer
*buf
;
4274 u64 addr
= pbuf
->addr
;
4275 int i
, bid
= pbuf
->bid
;
4277 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4278 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
4283 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
4288 INIT_LIST_HEAD(&buf
->list
);
4291 list_add_tail(&buf
->list
, &(*head
)->list
);
4295 return i
? i
: -ENOMEM
;
4298 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4299 struct io_comp_state
*cs
)
4301 struct io_provide_buf
*p
= &req
->pbuf
;
4302 struct io_ring_ctx
*ctx
= req
->ctx
;
4303 struct io_buffer
*head
, *list
;
4306 io_ring_submit_lock(ctx
, !force_nonblock
);
4308 lockdep_assert_held(&ctx
->uring_lock
);
4310 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4312 ret
= io_add_buffers(p
, &head
);
4313 if (ret
>= 0 && !list
) {
4314 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4316 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4319 req_set_fail_links(req
);
4321 /* need to hold the lock to complete IOPOLL requests */
4322 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4323 __io_req_complete(req
, ret
, 0, cs
);
4324 io_ring_submit_unlock(ctx
, !force_nonblock
);
4326 io_ring_submit_unlock(ctx
, !force_nonblock
);
4327 __io_req_complete(req
, ret
, 0, cs
);
4332 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4333 const struct io_uring_sqe
*sqe
)
4335 #if defined(CONFIG_EPOLL)
4336 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4338 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4341 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4342 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4343 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4345 if (ep_op_has_event(req
->epoll
.op
)) {
4346 struct epoll_event __user
*ev
;
4348 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4349 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4359 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
,
4360 struct io_comp_state
*cs
)
4362 #if defined(CONFIG_EPOLL)
4363 struct io_epoll
*ie
= &req
->epoll
;
4366 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4367 if (force_nonblock
&& ret
== -EAGAIN
)
4371 req_set_fail_links(req
);
4372 __io_req_complete(req
, ret
, 0, cs
);
4379 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4381 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4382 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->splice_fd_in
)
4384 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4387 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4388 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4389 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4396 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
4398 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4399 struct io_madvise
*ma
= &req
->madvise
;
4405 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4407 req_set_fail_links(req
);
4408 io_req_complete(req
, ret
);
4415 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4417 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
|| sqe
->splice_fd_in
)
4419 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4422 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4423 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4424 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4428 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
4430 struct io_fadvise
*fa
= &req
->fadvise
;
4433 if (force_nonblock
) {
4434 switch (fa
->advice
) {
4435 case POSIX_FADV_NORMAL
:
4436 case POSIX_FADV_RANDOM
:
4437 case POSIX_FADV_SEQUENTIAL
:
4444 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4446 req_set_fail_links(req
);
4447 io_req_complete(req
, ret
);
4451 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4453 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4455 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4457 if (req
->flags
& REQ_F_FIXED_FILE
)
4460 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4461 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4462 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4463 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4464 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4469 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
4471 struct io_statx
*ctx
= &req
->statx
;
4474 if (force_nonblock
) {
4475 /* only need file table for an actual valid fd */
4476 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4477 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4481 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4485 req_set_fail_links(req
);
4486 io_req_complete(req
, ret
);
4490 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4492 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4494 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4495 sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4497 if (req
->flags
& REQ_F_FIXED_FILE
)
4500 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4504 static int io_close(struct io_kiocb
*req
, bool force_nonblock
,
4505 struct io_comp_state
*cs
)
4507 struct files_struct
*files
= current
->files
;
4508 struct io_close
*close
= &req
->close
;
4509 struct fdtable
*fdt
;
4515 spin_lock(&files
->file_lock
);
4516 fdt
= files_fdtable(files
);
4517 if (close
->fd
>= fdt
->max_fds
) {
4518 spin_unlock(&files
->file_lock
);
4521 file
= fdt
->fd
[close
->fd
];
4523 spin_unlock(&files
->file_lock
);
4527 if (file
->f_op
== &io_uring_fops
) {
4528 spin_unlock(&files
->file_lock
);
4533 /* if the file has a flush method, be safe and punt to async */
4534 if (file
->f_op
->flush
&& force_nonblock
) {
4535 spin_unlock(&files
->file_lock
);
4539 ret
= __close_fd_get_file(close
->fd
, &file
);
4540 spin_unlock(&files
->file_lock
);
4547 /* No ->flush() or already async, safely close from here */
4548 ret
= filp_close(file
, current
->files
);
4551 req_set_fail_links(req
);
4554 __io_req_complete(req
, ret
, 0, cs
);
4558 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4560 struct io_ring_ctx
*ctx
= req
->ctx
;
4565 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4567 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4571 req
->sync
.off
= READ_ONCE(sqe
->off
);
4572 req
->sync
.len
= READ_ONCE(sqe
->len
);
4573 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4577 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
4581 /* sync_file_range always requires a blocking context */
4585 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4588 req_set_fail_links(req
);
4589 io_req_complete(req
, ret
);
4593 #if defined(CONFIG_NET)
4594 static int io_setup_async_msg(struct io_kiocb
*req
,
4595 struct io_async_msghdr
*kmsg
)
4597 struct io_async_msghdr
*async_msg
= req
->async_data
;
4601 if (io_alloc_async_data(req
)) {
4602 if (kmsg
->iov
!= kmsg
->fast_iov
)
4606 async_msg
= req
->async_data
;
4607 req
->flags
|= REQ_F_NEED_CLEANUP
;
4608 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4612 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4613 struct io_async_msghdr
*iomsg
)
4615 iomsg
->iov
= iomsg
->fast_iov
;
4616 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4617 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4618 req
->sr_msg
.msg_flags
, &iomsg
->iov
);
4621 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4623 struct io_async_msghdr
*async_msg
= req
->async_data
;
4624 struct io_sr_msg
*sr
= &req
->sr_msg
;
4627 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4630 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4631 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4632 sr
->len
= READ_ONCE(sqe
->len
);
4634 #ifdef CONFIG_COMPAT
4635 if (req
->ctx
->compat
)
4636 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4639 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4641 ret
= io_sendmsg_copy_hdr(req
, async_msg
);
4643 req
->flags
|= REQ_F_NEED_CLEANUP
;
4647 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4648 struct io_comp_state
*cs
)
4650 struct io_async_msghdr iomsg
, *kmsg
;
4651 struct socket
*sock
;
4656 sock
= sock_from_file(req
->file
);
4657 if (unlikely(!sock
))
4660 if (req
->async_data
) {
4661 kmsg
= req
->async_data
;
4662 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4663 /* if iov is set, it's allocated already */
4665 kmsg
->iov
= kmsg
->fast_iov
;
4666 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4668 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4674 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4675 if (flags
& MSG_DONTWAIT
)
4676 req
->flags
|= REQ_F_NOWAIT
;
4677 else if (force_nonblock
)
4678 flags
|= MSG_DONTWAIT
;
4680 if (flags
& MSG_WAITALL
)
4681 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4683 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4684 if (force_nonblock
&& ret
== -EAGAIN
)
4685 return io_setup_async_msg(req
, kmsg
);
4686 if (ret
== -ERESTARTSYS
)
4689 if (kmsg
->iov
!= kmsg
->fast_iov
)
4691 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4693 req_set_fail_links(req
);
4694 __io_req_complete(req
, ret
, 0, cs
);
4698 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4699 struct io_comp_state
*cs
)
4701 struct io_sr_msg
*sr
= &req
->sr_msg
;
4704 struct socket
*sock
;
4709 sock
= sock_from_file(req
->file
);
4710 if (unlikely(!sock
))
4713 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4717 msg
.msg_name
= NULL
;
4718 msg
.msg_control
= NULL
;
4719 msg
.msg_controllen
= 0;
4720 msg
.msg_namelen
= 0;
4722 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4723 if (flags
& MSG_DONTWAIT
)
4724 req
->flags
|= REQ_F_NOWAIT
;
4725 else if (force_nonblock
)
4726 flags
|= MSG_DONTWAIT
;
4728 if (flags
& MSG_WAITALL
)
4729 min_ret
= iov_iter_count(&msg
.msg_iter
);
4731 msg
.msg_flags
= flags
;
4732 ret
= sock_sendmsg(sock
, &msg
);
4733 if (force_nonblock
&& ret
== -EAGAIN
)
4735 if (ret
== -ERESTARTSYS
)
4739 req_set_fail_links(req
);
4740 __io_req_complete(req
, ret
, 0, cs
);
4744 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4745 struct io_async_msghdr
*iomsg
)
4747 struct io_sr_msg
*sr
= &req
->sr_msg
;
4748 struct iovec __user
*uiov
;
4752 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4753 &iomsg
->uaddr
, &uiov
, &iov_len
);
4757 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4760 if (copy_from_user(iomsg
->iov
, uiov
, sizeof(*uiov
)))
4762 sr
->len
= iomsg
->iov
[0].iov_len
;
4763 iov_iter_init(&iomsg
->msg
.msg_iter
, READ
, iomsg
->iov
, 1,
4767 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4768 &iomsg
->iov
, &iomsg
->msg
.msg_iter
,
4777 #ifdef CONFIG_COMPAT
4778 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4779 struct io_async_msghdr
*iomsg
)
4781 struct compat_msghdr __user
*msg_compat
;
4782 struct io_sr_msg
*sr
= &req
->sr_msg
;
4783 struct compat_iovec __user
*uiov
;
4788 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4789 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4794 uiov
= compat_ptr(ptr
);
4795 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4796 compat_ssize_t clen
;
4800 if (!access_ok(uiov
, sizeof(*uiov
)))
4802 if (__get_user(clen
, &uiov
->iov_len
))
4807 iomsg
->iov
[0].iov_len
= clen
;
4810 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4811 UIO_FASTIOV
, &iomsg
->iov
,
4812 &iomsg
->msg
.msg_iter
, true);
4821 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4822 struct io_async_msghdr
*iomsg
)
4824 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4825 iomsg
->iov
= iomsg
->fast_iov
;
4827 #ifdef CONFIG_COMPAT
4828 if (req
->ctx
->compat
)
4829 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4832 return __io_recvmsg_copy_hdr(req
, iomsg
);
4835 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4838 struct io_sr_msg
*sr
= &req
->sr_msg
;
4839 struct io_buffer
*kbuf
;
4841 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4846 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4850 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4852 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4855 static int io_recvmsg_prep(struct io_kiocb
*req
,
4856 const struct io_uring_sqe
*sqe
)
4858 struct io_async_msghdr
*async_msg
= req
->async_data
;
4859 struct io_sr_msg
*sr
= &req
->sr_msg
;
4862 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4865 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4866 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4867 sr
->len
= READ_ONCE(sqe
->len
);
4868 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4870 #ifdef CONFIG_COMPAT
4871 if (req
->ctx
->compat
)
4872 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4875 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4877 ret
= io_recvmsg_copy_hdr(req
, async_msg
);
4879 req
->flags
|= REQ_F_NEED_CLEANUP
;
4883 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
4884 struct io_comp_state
*cs
)
4886 struct io_async_msghdr iomsg
, *kmsg
;
4887 struct socket
*sock
;
4888 struct io_buffer
*kbuf
;
4891 int ret
, cflags
= 0;
4893 sock
= sock_from_file(req
->file
);
4894 if (unlikely(!sock
))
4897 if (req
->async_data
) {
4898 kmsg
= req
->async_data
;
4899 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4900 /* if iov is set, it's allocated already */
4902 kmsg
->iov
= kmsg
->fast_iov
;
4903 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4905 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4911 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4912 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4914 return PTR_ERR(kbuf
);
4915 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4916 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4917 1, req
->sr_msg
.len
);
4920 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4921 if (flags
& MSG_DONTWAIT
)
4922 req
->flags
|= REQ_F_NOWAIT
;
4923 else if (force_nonblock
)
4924 flags
|= MSG_DONTWAIT
;
4926 if (flags
& MSG_WAITALL
)
4927 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4929 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4930 kmsg
->uaddr
, flags
);
4931 if (force_nonblock
&& ret
== -EAGAIN
)
4932 return io_setup_async_msg(req
, kmsg
);
4933 if (ret
== -ERESTARTSYS
)
4936 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4937 cflags
= io_put_recv_kbuf(req
);
4938 if (kmsg
->iov
!= kmsg
->fast_iov
)
4940 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4941 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4942 req_set_fail_links(req
);
4943 __io_req_complete(req
, ret
, cflags
, cs
);
4947 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
4948 struct io_comp_state
*cs
)
4950 struct io_buffer
*kbuf
;
4951 struct io_sr_msg
*sr
= &req
->sr_msg
;
4953 void __user
*buf
= sr
->buf
;
4954 struct socket
*sock
;
4958 int ret
, cflags
= 0;
4960 sock
= sock_from_file(req
->file
);
4961 if (unlikely(!sock
))
4964 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4965 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4967 return PTR_ERR(kbuf
);
4968 buf
= u64_to_user_ptr(kbuf
->addr
);
4971 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4975 msg
.msg_name
= NULL
;
4976 msg
.msg_control
= NULL
;
4977 msg
.msg_controllen
= 0;
4978 msg
.msg_namelen
= 0;
4979 msg
.msg_iocb
= NULL
;
4982 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4983 if (flags
& MSG_DONTWAIT
)
4984 req
->flags
|= REQ_F_NOWAIT
;
4985 else if (force_nonblock
)
4986 flags
|= MSG_DONTWAIT
;
4988 if (flags
& MSG_WAITALL
)
4989 min_ret
= iov_iter_count(&msg
.msg_iter
);
4991 ret
= sock_recvmsg(sock
, &msg
, flags
);
4992 if (force_nonblock
&& ret
== -EAGAIN
)
4994 if (ret
== -ERESTARTSYS
)
4997 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4998 cflags
= io_put_recv_kbuf(req
);
4999 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5000 req_set_fail_links(req
);
5001 __io_req_complete(req
, ret
, cflags
, cs
);
5005 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5007 struct io_accept
*accept
= &req
->accept
;
5009 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5011 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
5014 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5015 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5016 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
5017 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
5021 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
5022 struct io_comp_state
*cs
)
5024 struct io_accept
*accept
= &req
->accept
;
5025 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5028 if (req
->file
->f_flags
& O_NONBLOCK
)
5029 req
->flags
|= REQ_F_NOWAIT
;
5031 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
5032 accept
->addr_len
, accept
->flags
,
5034 if (ret
== -EAGAIN
&& force_nonblock
)
5037 if (ret
== -ERESTARTSYS
)
5039 req_set_fail_links(req
);
5041 __io_req_complete(req
, ret
, 0, cs
);
5045 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5047 struct io_connect
*conn
= &req
->connect
;
5048 struct io_async_connect
*io
= req
->async_data
;
5050 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5052 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
||
5056 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5057 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5062 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
5066 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5067 struct io_comp_state
*cs
)
5069 struct io_async_connect __io
, *io
;
5070 unsigned file_flags
;
5073 if (req
->async_data
) {
5074 io
= req
->async_data
;
5076 ret
= move_addr_to_kernel(req
->connect
.addr
,
5077 req
->connect
.addr_len
,
5084 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5086 ret
= __sys_connect_file(req
->file
, &io
->address
,
5087 req
->connect
.addr_len
, file_flags
);
5088 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5089 if (req
->async_data
)
5091 if (io_alloc_async_data(req
)) {
5095 io
= req
->async_data
;
5096 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5099 if (ret
== -ERESTARTSYS
)
5103 req_set_fail_links(req
);
5104 __io_req_complete(req
, ret
, 0, cs
);
5107 #else /* !CONFIG_NET */
5108 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5113 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
5114 struct io_comp_state
*cs
)
5119 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
5120 struct io_comp_state
*cs
)
5125 static int io_recvmsg_prep(struct io_kiocb
*req
,
5126 const struct io_uring_sqe
*sqe
)
5131 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
5132 struct io_comp_state
*cs
)
5137 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
5138 struct io_comp_state
*cs
)
5143 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5148 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
5149 struct io_comp_state
*cs
)
5154 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5159 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5160 struct io_comp_state
*cs
)
5164 #endif /* CONFIG_NET */
5166 struct io_poll_table
{
5167 struct poll_table_struct pt
;
5168 struct io_kiocb
*req
;
5173 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5174 __poll_t mask
, task_work_func_t func
)
5178 /* for instances that support it check for an event match first: */
5179 if (mask
&& !(mask
& poll
->events
))
5182 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5184 list_del_init(&poll
->wait
.entry
);
5187 init_task_work(&req
->task_work
, func
);
5188 percpu_ref_get(&req
->ctx
->refs
);
5191 * If this fails, then the task is exiting. When a task exits, the
5192 * work gets canceled, so just cancel this request as well instead
5193 * of executing it. We can't safely execute it anyway, as we may not
5194 * have the needed state needed for it anyway.
5196 ret
= io_req_task_work_add(req
);
5197 if (unlikely(ret
)) {
5198 WRITE_ONCE(poll
->canceled
, true);
5199 io_req_task_work_add_fallback(req
, func
);
5204 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5205 __acquires(&req
->ctx
->completion_lock
)
5207 struct io_ring_ctx
*ctx
= req
->ctx
;
5209 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5210 struct poll_table_struct pt
= { ._key
= poll
->events
};
5212 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5215 spin_lock_irq(&ctx
->completion_lock
);
5216 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5217 add_wait_queue(poll
->head
, &poll
->wait
);
5224 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5226 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5227 if (req
->opcode
== IORING_OP_POLL_ADD
)
5228 return req
->async_data
;
5229 return req
->apoll
->double_poll
;
5232 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5234 if (req
->opcode
== IORING_OP_POLL_ADD
)
5236 return &req
->apoll
->poll
;
5239 static void io_poll_remove_double(struct io_kiocb
*req
)
5241 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5243 lockdep_assert_held(&req
->ctx
->completion_lock
);
5245 if (poll
&& poll
->head
) {
5246 struct wait_queue_head
*head
= poll
->head
;
5248 spin_lock(&head
->lock
);
5249 list_del_init(&poll
->wait
.entry
);
5250 if (poll
->wait
.private)
5251 refcount_dec(&req
->refs
);
5253 spin_unlock(&head
->lock
);
5257 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
5259 struct io_ring_ctx
*ctx
= req
->ctx
;
5261 io_poll_remove_double(req
);
5262 req
->poll
.done
= true;
5263 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
5264 io_commit_cqring(ctx
);
5267 static void io_poll_task_func(struct callback_head
*cb
)
5269 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5270 struct io_ring_ctx
*ctx
= req
->ctx
;
5271 struct io_kiocb
*nxt
;
5273 if (io_poll_rewait(req
, &req
->poll
)) {
5274 spin_unlock_irq(&ctx
->completion_lock
);
5276 hash_del(&req
->hash_node
);
5277 io_poll_complete(req
, req
->result
, 0);
5278 spin_unlock_irq(&ctx
->completion_lock
);
5280 nxt
= io_put_req_find_next(req
);
5281 io_cqring_ev_posted(ctx
);
5283 __io_req_task_submit(nxt
);
5286 percpu_ref_put(&ctx
->refs
);
5289 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5290 int sync
, void *key
)
5292 struct io_kiocb
*req
= wait
->private;
5293 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5294 __poll_t mask
= key_to_poll(key
);
5296 /* for instances that support it check for an event match first: */
5297 if (mask
&& !(mask
& poll
->events
))
5300 list_del_init(&wait
->entry
);
5302 if (poll
&& poll
->head
) {
5305 spin_lock(&poll
->head
->lock
);
5306 done
= list_empty(&poll
->wait
.entry
);
5308 list_del_init(&poll
->wait
.entry
);
5309 /* make sure double remove sees this as being gone */
5310 wait
->private = NULL
;
5311 spin_unlock(&poll
->head
->lock
);
5313 /* use wait func handler, so it matches the rq type */
5314 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5317 refcount_dec(&req
->refs
);
5321 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5322 wait_queue_func_t wake_func
)
5326 poll
->canceled
= false;
5327 poll
->events
= events
;
5328 INIT_LIST_HEAD(&poll
->wait
.entry
);
5329 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5332 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5333 struct wait_queue_head
*head
,
5334 struct io_poll_iocb
**poll_ptr
)
5336 struct io_kiocb
*req
= pt
->req
;
5339 * The file being polled uses multiple waitqueues for poll handling
5340 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5343 if (unlikely(pt
->nr_entries
)) {
5344 struct io_poll_iocb
*poll_one
= poll
;
5346 /* already have a 2nd entry, fail a third attempt */
5348 pt
->error
= -EINVAL
;
5351 /* double add on the same waitqueue head, ignore */
5352 if (poll
->head
== head
)
5354 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5356 pt
->error
= -ENOMEM
;
5359 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5360 refcount_inc(&req
->refs
);
5361 poll
->wait
.private = req
;
5368 if (poll
->events
& EPOLLEXCLUSIVE
)
5369 add_wait_queue_exclusive(head
, &poll
->wait
);
5371 add_wait_queue(head
, &poll
->wait
);
5374 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5375 struct poll_table_struct
*p
)
5377 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5378 struct async_poll
*apoll
= pt
->req
->apoll
;
5380 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5383 static void io_async_task_func(struct callback_head
*cb
)
5385 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5386 struct async_poll
*apoll
= req
->apoll
;
5387 struct io_ring_ctx
*ctx
= req
->ctx
;
5389 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5391 if (io_poll_rewait(req
, &apoll
->poll
)) {
5392 spin_unlock_irq(&ctx
->completion_lock
);
5393 percpu_ref_put(&ctx
->refs
);
5397 /* If req is still hashed, it cannot have been canceled. Don't check. */
5398 if (hash_hashed(&req
->hash_node
))
5399 hash_del(&req
->hash_node
);
5401 io_poll_remove_double(req
);
5402 spin_unlock_irq(&ctx
->completion_lock
);
5404 if (!READ_ONCE(apoll
->poll
.canceled
))
5405 __io_req_task_submit(req
);
5407 __io_req_task_cancel(req
, -ECANCELED
);
5409 percpu_ref_put(&ctx
->refs
);
5410 kfree(apoll
->double_poll
);
5414 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5417 struct io_kiocb
*req
= wait
->private;
5418 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5420 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5423 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5426 static void io_poll_req_insert(struct io_kiocb
*req
)
5428 struct io_ring_ctx
*ctx
= req
->ctx
;
5429 struct hlist_head
*list
;
5431 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5432 hlist_add_head(&req
->hash_node
, list
);
5435 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5436 struct io_poll_iocb
*poll
,
5437 struct io_poll_table
*ipt
, __poll_t mask
,
5438 wait_queue_func_t wake_func
)
5439 __acquires(&ctx
->completion_lock
)
5441 struct io_ring_ctx
*ctx
= req
->ctx
;
5442 bool cancel
= false;
5444 INIT_HLIST_NODE(&req
->hash_node
);
5445 io_init_poll_iocb(poll
, mask
, wake_func
);
5446 poll
->file
= req
->file
;
5447 poll
->wait
.private = req
;
5449 ipt
->pt
._key
= mask
;
5452 ipt
->nr_entries
= 0;
5454 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5455 if (unlikely(!ipt
->nr_entries
) && !ipt
->error
)
5456 ipt
->error
= -EINVAL
;
5458 spin_lock_irq(&ctx
->completion_lock
);
5460 io_poll_remove_double(req
);
5461 if (likely(poll
->head
)) {
5462 spin_lock(&poll
->head
->lock
);
5463 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5469 if (mask
|| ipt
->error
)
5470 list_del_init(&poll
->wait
.entry
);
5472 WRITE_ONCE(poll
->canceled
, true);
5473 else if (!poll
->done
) /* actually waiting for an event */
5474 io_poll_req_insert(req
);
5475 spin_unlock(&poll
->head
->lock
);
5481 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5483 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5484 struct io_ring_ctx
*ctx
= req
->ctx
;
5485 struct async_poll
*apoll
;
5486 struct io_poll_table ipt
;
5490 if (!req
->file
|| !file_can_poll(req
->file
))
5492 if (req
->flags
& REQ_F_POLLED
)
5496 else if (def
->pollout
)
5500 /* if we can't nonblock try, then no point in arming a poll handler */
5501 if (!io_file_supports_async(req
->file
, rw
))
5504 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5505 if (unlikely(!apoll
))
5507 apoll
->double_poll
= NULL
;
5509 req
->flags
|= REQ_F_POLLED
;
5514 mask
|= POLLIN
| POLLRDNORM
;
5516 mask
|= POLLOUT
| POLLWRNORM
;
5518 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5519 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5520 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5523 mask
|= POLLERR
| POLLPRI
;
5525 ipt
.pt
._qproc
= io_async_queue_proc
;
5527 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5529 if (ret
|| ipt
.error
) {
5530 io_poll_remove_double(req
);
5531 spin_unlock_irq(&ctx
->completion_lock
);
5532 kfree(apoll
->double_poll
);
5536 spin_unlock_irq(&ctx
->completion_lock
);
5537 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5538 apoll
->poll
.events
);
5542 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5543 struct io_poll_iocb
*poll
)
5545 bool do_complete
= false;
5547 spin_lock(&poll
->head
->lock
);
5548 WRITE_ONCE(poll
->canceled
, true);
5549 if (!list_empty(&poll
->wait
.entry
)) {
5550 list_del_init(&poll
->wait
.entry
);
5553 spin_unlock(&poll
->head
->lock
);
5554 hash_del(&req
->hash_node
);
5558 static bool io_poll_remove_one(struct io_kiocb
*req
)
5562 io_poll_remove_double(req
);
5564 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5565 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5567 struct async_poll
*apoll
= req
->apoll
;
5569 /* non-poll requests have submit ref still */
5570 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5573 kfree(apoll
->double_poll
);
5579 io_cqring_fill_event(req
, -ECANCELED
);
5580 io_commit_cqring(req
->ctx
);
5581 req_set_fail_links(req
);
5582 io_put_req_deferred(req
, 1);
5589 * Returns true if we found and killed one or more poll requests
5591 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5592 struct files_struct
*files
)
5594 struct hlist_node
*tmp
;
5595 struct io_kiocb
*req
;
5598 spin_lock_irq(&ctx
->completion_lock
);
5599 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5600 struct hlist_head
*list
;
5602 list
= &ctx
->cancel_hash
[i
];
5603 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5604 if (io_match_task(req
, tsk
, files
))
5605 posted
+= io_poll_remove_one(req
);
5608 spin_unlock_irq(&ctx
->completion_lock
);
5611 io_cqring_ev_posted(ctx
);
5616 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5618 struct hlist_head
*list
;
5619 struct io_kiocb
*req
;
5621 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5622 hlist_for_each_entry(req
, list
, hash_node
) {
5623 if (sqe_addr
!= req
->user_data
)
5625 if (io_poll_remove_one(req
))
5633 static int io_poll_remove_prep(struct io_kiocb
*req
,
5634 const struct io_uring_sqe
*sqe
)
5636 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5638 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5642 req
->poll_remove
.addr
= READ_ONCE(sqe
->addr
);
5647 * Find a running poll command that matches one specified in sqe->addr,
5648 * and remove it if found.
5650 static int io_poll_remove(struct io_kiocb
*req
)
5652 struct io_ring_ctx
*ctx
= req
->ctx
;
5655 spin_lock_irq(&ctx
->completion_lock
);
5656 ret
= io_poll_cancel(ctx
, req
->poll_remove
.addr
);
5657 spin_unlock_irq(&ctx
->completion_lock
);
5660 req_set_fail_links(req
);
5661 io_req_complete(req
, ret
);
5665 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5668 struct io_kiocb
*req
= wait
->private;
5669 struct io_poll_iocb
*poll
= &req
->poll
;
5671 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5674 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5675 struct poll_table_struct
*p
)
5677 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5679 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5682 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5684 struct io_poll_iocb
*poll
= &req
->poll
;
5687 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5689 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5692 events
= READ_ONCE(sqe
->poll32_events
);
5694 events
= swahw32(events
);
5696 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5697 (events
& EPOLLEXCLUSIVE
);
5701 static int io_poll_add(struct io_kiocb
*req
)
5703 struct io_poll_iocb
*poll
= &req
->poll
;
5704 struct io_ring_ctx
*ctx
= req
->ctx
;
5705 struct io_poll_table ipt
;
5708 ipt
.pt
._qproc
= io_poll_queue_proc
;
5710 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5713 if (mask
) { /* no async, we'd stolen it */
5715 io_poll_complete(req
, mask
, 0);
5717 spin_unlock_irq(&ctx
->completion_lock
);
5720 io_cqring_ev_posted(ctx
);
5726 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5728 struct io_timeout_data
*data
= container_of(timer
,
5729 struct io_timeout_data
, timer
);
5730 struct io_kiocb
*req
= data
->req
;
5731 struct io_ring_ctx
*ctx
= req
->ctx
;
5732 unsigned long flags
;
5734 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5735 list_del_init(&req
->timeout
.list
);
5736 atomic_set(&req
->ctx
->cq_timeouts
,
5737 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5739 io_cqring_fill_event(req
, -ETIME
);
5740 io_commit_cqring(ctx
);
5741 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5743 io_cqring_ev_posted(ctx
);
5744 req_set_fail_links(req
);
5746 return HRTIMER_NORESTART
;
5749 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5752 struct io_timeout_data
*io
;
5753 struct io_kiocb
*req
;
5756 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5757 if (user_data
== req
->user_data
) {
5764 return ERR_PTR(ret
);
5766 io
= req
->async_data
;
5767 ret
= hrtimer_try_to_cancel(&io
->timer
);
5769 return ERR_PTR(-EALREADY
);
5770 list_del_init(&req
->timeout
.list
);
5774 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5776 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5779 return PTR_ERR(req
);
5781 req_set_fail_links(req
);
5782 io_cqring_fill_event(req
, -ECANCELED
);
5783 io_put_req_deferred(req
, 1);
5787 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5788 struct timespec64
*ts
, enum hrtimer_mode mode
)
5790 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5791 struct io_timeout_data
*data
;
5794 return PTR_ERR(req
);
5796 req
->timeout
.off
= 0; /* noseq */
5797 data
= req
->async_data
;
5798 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5799 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5800 data
->timer
.function
= io_timeout_fn
;
5801 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5805 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5806 const struct io_uring_sqe
*sqe
)
5808 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5810 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5812 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5814 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->splice_fd_in
)
5817 tr
->addr
= READ_ONCE(sqe
->addr
);
5818 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5819 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5820 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5822 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5824 } else if (tr
->flags
) {
5825 /* timeout removal doesn't support flags */
5833 * Remove or update an existing timeout command
5835 static int io_timeout_remove(struct io_kiocb
*req
)
5837 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5838 struct io_ring_ctx
*ctx
= req
->ctx
;
5841 spin_lock_irq(&ctx
->completion_lock
);
5842 if (req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
) {
5843 enum hrtimer_mode mode
= (tr
->flags
& IORING_TIMEOUT_ABS
)
5844 ? HRTIMER_MODE_ABS
: HRTIMER_MODE_REL
;
5846 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
5848 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5851 io_cqring_fill_event(req
, ret
);
5852 io_commit_cqring(ctx
);
5853 spin_unlock_irq(&ctx
->completion_lock
);
5854 io_cqring_ev_posted(ctx
);
5856 req_set_fail_links(req
);
5861 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5862 bool is_timeout_link
)
5864 struct io_timeout_data
*data
;
5866 u32 off
= READ_ONCE(sqe
->off
);
5868 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5870 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1 ||
5873 if (off
&& is_timeout_link
)
5875 flags
= READ_ONCE(sqe
->timeout_flags
);
5876 if (flags
& ~IORING_TIMEOUT_ABS
)
5879 req
->timeout
.off
= off
;
5881 if (!req
->async_data
&& io_alloc_async_data(req
))
5884 data
= req
->async_data
;
5887 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5890 if (flags
& IORING_TIMEOUT_ABS
)
5891 data
->mode
= HRTIMER_MODE_ABS
;
5893 data
->mode
= HRTIMER_MODE_REL
;
5895 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5899 static int io_timeout(struct io_kiocb
*req
)
5901 struct io_ring_ctx
*ctx
= req
->ctx
;
5902 struct io_timeout_data
*data
= req
->async_data
;
5903 struct list_head
*entry
;
5904 u32 tail
, off
= req
->timeout
.off
;
5906 spin_lock_irq(&ctx
->completion_lock
);
5909 * sqe->off holds how many events that need to occur for this
5910 * timeout event to be satisfied. If it isn't set, then this is
5911 * a pure timeout request, sequence isn't used.
5913 if (io_is_timeout_noseq(req
)) {
5914 entry
= ctx
->timeout_list
.prev
;
5918 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5919 req
->timeout
.target_seq
= tail
+ off
;
5921 /* Update the last seq here in case io_flush_timeouts() hasn't.
5922 * This is safe because ->completion_lock is held, and submissions
5923 * and completions are never mixed in the same ->completion_lock section.
5925 ctx
->cq_last_tm_flush
= tail
;
5928 * Insertion sort, ensuring the first entry in the list is always
5929 * the one we need first.
5931 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5932 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5935 if (io_is_timeout_noseq(nxt
))
5937 /* nxt.seq is behind @tail, otherwise would've been completed */
5938 if (off
>= nxt
->timeout
.target_seq
- tail
)
5942 list_add(&req
->timeout
.list
, entry
);
5943 data
->timer
.function
= io_timeout_fn
;
5944 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5945 spin_unlock_irq(&ctx
->completion_lock
);
5949 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5951 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5953 return req
->user_data
== (unsigned long) data
;
5956 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5958 enum io_wq_cancel cancel_ret
;
5961 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5962 switch (cancel_ret
) {
5963 case IO_WQ_CANCEL_OK
:
5966 case IO_WQ_CANCEL_RUNNING
:
5969 case IO_WQ_CANCEL_NOTFOUND
:
5977 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5978 struct io_kiocb
*req
, __u64 sqe_addr
,
5981 unsigned long flags
;
5984 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5985 if (ret
!= -ENOENT
) {
5986 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5990 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5991 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5994 ret
= io_poll_cancel(ctx
, sqe_addr
);
5998 io_cqring_fill_event(req
, ret
);
5999 io_commit_cqring(ctx
);
6000 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6001 io_cqring_ev_posted(ctx
);
6004 req_set_fail_links(req
);
6008 static int io_async_cancel_prep(struct io_kiocb
*req
,
6009 const struct io_uring_sqe
*sqe
)
6011 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6013 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6015 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
||
6019 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
6023 static int io_async_cancel(struct io_kiocb
*req
)
6025 struct io_ring_ctx
*ctx
= req
->ctx
;
6027 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
6031 static int io_files_update_prep(struct io_kiocb
*req
,
6032 const struct io_uring_sqe
*sqe
)
6034 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
6036 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6038 if (sqe
->ioprio
|| sqe
->rw_flags
)
6041 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
6042 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
6043 if (!req
->files_update
.nr_args
)
6045 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
6049 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
,
6050 struct io_comp_state
*cs
)
6052 struct io_ring_ctx
*ctx
= req
->ctx
;
6053 struct io_uring_files_update up
;
6059 up
.offset
= req
->files_update
.offset
;
6060 up
.fds
= req
->files_update
.arg
;
6062 mutex_lock(&ctx
->uring_lock
);
6063 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
6064 mutex_unlock(&ctx
->uring_lock
);
6067 req_set_fail_links(req
);
6068 __io_req_complete(req
, ret
, 0, cs
);
6072 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6074 switch (req
->opcode
) {
6077 case IORING_OP_READV
:
6078 case IORING_OP_READ_FIXED
:
6079 case IORING_OP_READ
:
6080 return io_read_prep(req
, sqe
);
6081 case IORING_OP_WRITEV
:
6082 case IORING_OP_WRITE_FIXED
:
6083 case IORING_OP_WRITE
:
6084 return io_write_prep(req
, sqe
);
6085 case IORING_OP_POLL_ADD
:
6086 return io_poll_add_prep(req
, sqe
);
6087 case IORING_OP_POLL_REMOVE
:
6088 return io_poll_remove_prep(req
, sqe
);
6089 case IORING_OP_FSYNC
:
6090 return io_prep_fsync(req
, sqe
);
6091 case IORING_OP_SYNC_FILE_RANGE
:
6092 return io_prep_sfr(req
, sqe
);
6093 case IORING_OP_SENDMSG
:
6094 case IORING_OP_SEND
:
6095 return io_sendmsg_prep(req
, sqe
);
6096 case IORING_OP_RECVMSG
:
6097 case IORING_OP_RECV
:
6098 return io_recvmsg_prep(req
, sqe
);
6099 case IORING_OP_CONNECT
:
6100 return io_connect_prep(req
, sqe
);
6101 case IORING_OP_TIMEOUT
:
6102 return io_timeout_prep(req
, sqe
, false);
6103 case IORING_OP_TIMEOUT_REMOVE
:
6104 return io_timeout_remove_prep(req
, sqe
);
6105 case IORING_OP_ASYNC_CANCEL
:
6106 return io_async_cancel_prep(req
, sqe
);
6107 case IORING_OP_LINK_TIMEOUT
:
6108 return io_timeout_prep(req
, sqe
, true);
6109 case IORING_OP_ACCEPT
:
6110 return io_accept_prep(req
, sqe
);
6111 case IORING_OP_FALLOCATE
:
6112 return io_fallocate_prep(req
, sqe
);
6113 case IORING_OP_OPENAT
:
6114 return io_openat_prep(req
, sqe
);
6115 case IORING_OP_CLOSE
:
6116 return io_close_prep(req
, sqe
);
6117 case IORING_OP_FILES_UPDATE
:
6118 return io_files_update_prep(req
, sqe
);
6119 case IORING_OP_STATX
:
6120 return io_statx_prep(req
, sqe
);
6121 case IORING_OP_FADVISE
:
6122 return io_fadvise_prep(req
, sqe
);
6123 case IORING_OP_MADVISE
:
6124 return io_madvise_prep(req
, sqe
);
6125 case IORING_OP_OPENAT2
:
6126 return io_openat2_prep(req
, sqe
);
6127 case IORING_OP_EPOLL_CTL
:
6128 return io_epoll_ctl_prep(req
, sqe
);
6129 case IORING_OP_SPLICE
:
6130 return io_splice_prep(req
, sqe
);
6131 case IORING_OP_PROVIDE_BUFFERS
:
6132 return io_provide_buffers_prep(req
, sqe
);
6133 case IORING_OP_REMOVE_BUFFERS
:
6134 return io_remove_buffers_prep(req
, sqe
);
6136 return io_tee_prep(req
, sqe
);
6137 case IORING_OP_SHUTDOWN
:
6138 return io_shutdown_prep(req
, sqe
);
6139 case IORING_OP_RENAMEAT
:
6140 return io_renameat_prep(req
, sqe
);
6141 case IORING_OP_UNLINKAT
:
6142 return io_unlinkat_prep(req
, sqe
);
6145 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6150 static int io_req_defer_prep(struct io_kiocb
*req
,
6151 const struct io_uring_sqe
*sqe
)
6155 if (io_alloc_async_data(req
))
6157 return io_req_prep(req
, sqe
);
6160 static u32
io_get_sequence(struct io_kiocb
*req
)
6162 struct io_kiocb
*pos
;
6163 struct io_ring_ctx
*ctx
= req
->ctx
;
6164 u32 total_submitted
, nr_reqs
= 0;
6166 io_for_each_link(pos
, req
)
6169 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
6170 return total_submitted
- nr_reqs
;
6173 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6175 struct io_ring_ctx
*ctx
= req
->ctx
;
6176 struct io_defer_entry
*de
;
6180 /* Still need defer if there is pending req in defer list. */
6181 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6182 !(req
->flags
& REQ_F_IO_DRAIN
)))
6185 seq
= io_get_sequence(req
);
6186 /* Still a chance to pass the sequence check */
6187 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6190 if (!req
->async_data
) {
6191 ret
= io_req_defer_prep(req
, sqe
);
6195 io_prep_async_link(req
);
6196 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6200 spin_lock_irq(&ctx
->completion_lock
);
6201 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6202 spin_unlock_irq(&ctx
->completion_lock
);
6204 io_queue_async_work(req
);
6205 return -EIOCBQUEUED
;
6208 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6211 list_add_tail(&de
->list
, &ctx
->defer_list
);
6212 spin_unlock_irq(&ctx
->completion_lock
);
6213 return -EIOCBQUEUED
;
6216 static void io_req_drop_files(struct io_kiocb
*req
)
6218 struct io_ring_ctx
*ctx
= req
->ctx
;
6219 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6220 unsigned long flags
;
6222 if (req
->work
.flags
& IO_WQ_WORK_FILES
) {
6223 put_files_struct(req
->work
.identity
->files
);
6224 put_nsproxy(req
->work
.identity
->nsproxy
);
6226 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
6227 list_del(&req
->inflight_entry
);
6228 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
6229 req
->flags
&= ~REQ_F_INFLIGHT
;
6230 req
->work
.flags
&= ~IO_WQ_WORK_FILES
;
6231 if (atomic_read(&tctx
->in_idle
))
6232 wake_up(&tctx
->wait
);
6235 static void __io_clean_op(struct io_kiocb
*req
)
6237 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6238 switch (req
->opcode
) {
6239 case IORING_OP_READV
:
6240 case IORING_OP_READ_FIXED
:
6241 case IORING_OP_READ
:
6242 kfree((void *)(unsigned long)req
->rw
.addr
);
6244 case IORING_OP_RECVMSG
:
6245 case IORING_OP_RECV
:
6246 kfree(req
->sr_msg
.kbuf
);
6249 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
6252 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6253 switch (req
->opcode
) {
6254 case IORING_OP_READV
:
6255 case IORING_OP_READ_FIXED
:
6256 case IORING_OP_READ
:
6257 case IORING_OP_WRITEV
:
6258 case IORING_OP_WRITE_FIXED
:
6259 case IORING_OP_WRITE
: {
6260 struct io_async_rw
*io
= req
->async_data
;
6262 kfree(io
->free_iovec
);
6265 case IORING_OP_RECVMSG
:
6266 case IORING_OP_SENDMSG
: {
6267 struct io_async_msghdr
*io
= req
->async_data
;
6268 if (io
->iov
!= io
->fast_iov
)
6272 case IORING_OP_SPLICE
:
6274 io_put_file(req
, req
->splice
.file_in
,
6275 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
6277 case IORING_OP_OPENAT
:
6278 case IORING_OP_OPENAT2
:
6279 if (req
->open
.filename
)
6280 putname(req
->open
.filename
);
6282 case IORING_OP_RENAMEAT
:
6283 putname(req
->rename
.oldpath
);
6284 putname(req
->rename
.newpath
);
6286 case IORING_OP_UNLINKAT
:
6287 putname(req
->unlink
.filename
);
6290 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6294 static int io_issue_sqe(struct io_kiocb
*req
, bool force_nonblock
,
6295 struct io_comp_state
*cs
)
6297 struct io_ring_ctx
*ctx
= req
->ctx
;
6300 switch (req
->opcode
) {
6302 ret
= io_nop(req
, cs
);
6304 case IORING_OP_READV
:
6305 case IORING_OP_READ_FIXED
:
6306 case IORING_OP_READ
:
6307 ret
= io_read(req
, force_nonblock
, cs
);
6309 case IORING_OP_WRITEV
:
6310 case IORING_OP_WRITE_FIXED
:
6311 case IORING_OP_WRITE
:
6312 ret
= io_write(req
, force_nonblock
, cs
);
6314 case IORING_OP_FSYNC
:
6315 ret
= io_fsync(req
, force_nonblock
);
6317 case IORING_OP_POLL_ADD
:
6318 ret
= io_poll_add(req
);
6320 case IORING_OP_POLL_REMOVE
:
6321 ret
= io_poll_remove(req
);
6323 case IORING_OP_SYNC_FILE_RANGE
:
6324 ret
= io_sync_file_range(req
, force_nonblock
);
6326 case IORING_OP_SENDMSG
:
6327 ret
= io_sendmsg(req
, force_nonblock
, cs
);
6329 case IORING_OP_SEND
:
6330 ret
= io_send(req
, force_nonblock
, cs
);
6332 case IORING_OP_RECVMSG
:
6333 ret
= io_recvmsg(req
, force_nonblock
, cs
);
6335 case IORING_OP_RECV
:
6336 ret
= io_recv(req
, force_nonblock
, cs
);
6338 case IORING_OP_TIMEOUT
:
6339 ret
= io_timeout(req
);
6341 case IORING_OP_TIMEOUT_REMOVE
:
6342 ret
= io_timeout_remove(req
);
6344 case IORING_OP_ACCEPT
:
6345 ret
= io_accept(req
, force_nonblock
, cs
);
6347 case IORING_OP_CONNECT
:
6348 ret
= io_connect(req
, force_nonblock
, cs
);
6350 case IORING_OP_ASYNC_CANCEL
:
6351 ret
= io_async_cancel(req
);
6353 case IORING_OP_FALLOCATE
:
6354 ret
= io_fallocate(req
, force_nonblock
);
6356 case IORING_OP_OPENAT
:
6357 ret
= io_openat(req
, force_nonblock
);
6359 case IORING_OP_CLOSE
:
6360 ret
= io_close(req
, force_nonblock
, cs
);
6362 case IORING_OP_FILES_UPDATE
:
6363 ret
= io_files_update(req
, force_nonblock
, cs
);
6365 case IORING_OP_STATX
:
6366 ret
= io_statx(req
, force_nonblock
);
6368 case IORING_OP_FADVISE
:
6369 ret
= io_fadvise(req
, force_nonblock
);
6371 case IORING_OP_MADVISE
:
6372 ret
= io_madvise(req
, force_nonblock
);
6374 case IORING_OP_OPENAT2
:
6375 ret
= io_openat2(req
, force_nonblock
);
6377 case IORING_OP_EPOLL_CTL
:
6378 ret
= io_epoll_ctl(req
, force_nonblock
, cs
);
6380 case IORING_OP_SPLICE
:
6381 ret
= io_splice(req
, force_nonblock
);
6383 case IORING_OP_PROVIDE_BUFFERS
:
6384 ret
= io_provide_buffers(req
, force_nonblock
, cs
);
6386 case IORING_OP_REMOVE_BUFFERS
:
6387 ret
= io_remove_buffers(req
, force_nonblock
, cs
);
6390 ret
= io_tee(req
, force_nonblock
);
6392 case IORING_OP_SHUTDOWN
:
6393 ret
= io_shutdown(req
, force_nonblock
);
6395 case IORING_OP_RENAMEAT
:
6396 ret
= io_renameat(req
, force_nonblock
);
6398 case IORING_OP_UNLINKAT
:
6399 ret
= io_unlinkat(req
, force_nonblock
);
6409 /* If the op doesn't have a file, we're not polling for it */
6410 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6411 const bool in_async
= io_wq_current_is_worker();
6413 /* workqueue context doesn't hold uring_lock, grab it now */
6415 mutex_lock(&ctx
->uring_lock
);
6417 io_iopoll_req_issued(req
, in_async
);
6420 mutex_unlock(&ctx
->uring_lock
);
6426 static void io_wq_submit_work(struct io_wq_work
*work
)
6428 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6429 struct io_kiocb
*timeout
;
6432 timeout
= io_prep_linked_timeout(req
);
6434 io_queue_linked_timeout(timeout
);
6436 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
6437 /* io-wq is going to take down one */
6438 refcount_inc(&req
->refs
);
6439 percpu_ref_get(&req
->ctx
->refs
);
6440 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
6446 ret
= io_issue_sqe(req
, false, NULL
);
6448 * We can get EAGAIN for polled IO even though we're
6449 * forcing a sync submission from here, since we can't
6450 * wait for request slots on the block side.
6459 struct io_ring_ctx
*lock_ctx
= NULL
;
6461 if (req
->ctx
->flags
& IORING_SETUP_IOPOLL
)
6462 lock_ctx
= req
->ctx
;
6465 * io_iopoll_complete() does not hold completion_lock to
6466 * complete polled io, so here for polled io, we can not call
6467 * io_req_complete() directly, otherwise there maybe concurrent
6468 * access to cqring, defer_list, etc, which is not safe. Given
6469 * that io_iopoll_complete() is always called under uring_lock,
6470 * so here for polled io, we also get uring_lock to complete
6474 mutex_lock(&lock_ctx
->uring_lock
);
6476 req_set_fail_links(req
);
6477 io_req_complete(req
, ret
);
6480 mutex_unlock(&lock_ctx
->uring_lock
);
6484 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6487 struct fixed_file_table
*table
;
6489 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6490 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6493 static struct file
*io_file_get(struct io_submit_state
*state
,
6494 struct io_kiocb
*req
, int fd
, bool fixed
)
6496 struct io_ring_ctx
*ctx
= req
->ctx
;
6500 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6502 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6503 file
= io_file_from_index(ctx
, fd
);
6504 io_set_resource_node(req
);
6506 trace_io_uring_file_get(ctx
, fd
);
6507 file
= __io_file_get(state
, fd
);
6510 if (file
&& file
->f_op
== &io_uring_fops
&&
6511 !(req
->flags
& REQ_F_INFLIGHT
)) {
6512 io_req_init_async(req
);
6513 req
->flags
|= REQ_F_INFLIGHT
;
6515 spin_lock_irq(&ctx
->inflight_lock
);
6516 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
6517 spin_unlock_irq(&ctx
->inflight_lock
);
6523 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6525 struct io_timeout_data
*data
= container_of(timer
,
6526 struct io_timeout_data
, timer
);
6527 struct io_kiocb
*prev
, *req
= data
->req
;
6528 struct io_ring_ctx
*ctx
= req
->ctx
;
6529 unsigned long flags
;
6531 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6532 prev
= req
->timeout
.head
;
6533 req
->timeout
.head
= NULL
;
6536 * We don't expect the list to be empty, that will only happen if we
6537 * race with the completion of the linked work.
6540 io_remove_next_linked(prev
);
6541 if (!refcount_inc_not_zero(&prev
->refs
))
6544 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6547 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6548 io_put_req_deferred(prev
, 1);
6550 io_cqring_add_event(req
, -ETIME
, 0);
6551 io_put_req_deferred(req
, 1);
6553 return HRTIMER_NORESTART
;
6556 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6559 * If the back reference is NULL, then our linked request finished
6560 * before we got a chance to setup the timer
6562 if (req
->timeout
.head
) {
6563 struct io_timeout_data
*data
= req
->async_data
;
6565 data
->timer
.function
= io_link_timeout_fn
;
6566 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6571 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6573 struct io_ring_ctx
*ctx
= req
->ctx
;
6575 spin_lock_irq(&ctx
->completion_lock
);
6576 __io_queue_linked_timeout(req
);
6577 spin_unlock_irq(&ctx
->completion_lock
);
6579 /* drop submission reference */
6583 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6585 struct io_kiocb
*nxt
= req
->link
;
6587 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6588 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6591 nxt
->timeout
.head
= req
;
6592 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6593 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6597 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
)
6599 struct io_kiocb
*linked_timeout
;
6600 const struct cred
*old_creds
= NULL
;
6604 linked_timeout
= io_prep_linked_timeout(req
);
6606 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
6607 (req
->work
.flags
& IO_WQ_WORK_CREDS
) &&
6608 req
->work
.identity
->creds
!= current_cred()) {
6610 revert_creds(old_creds
);
6611 if (old_creds
== req
->work
.identity
->creds
)
6612 old_creds
= NULL
; /* restored original creds */
6614 old_creds
= override_creds(req
->work
.identity
->creds
);
6617 ret
= io_issue_sqe(req
, true, cs
);
6620 * We async punt it if the file wasn't marked NOWAIT, or if the file
6621 * doesn't support non-blocking read/write attempts
6623 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6624 if (!io_arm_poll_handler(req
)) {
6626 * Queued up for async execution, worker will release
6627 * submit reference when the iocb is actually submitted.
6629 io_queue_async_work(req
);
6633 io_queue_linked_timeout(linked_timeout
);
6634 } else if (likely(!ret
)) {
6635 /* drop submission reference */
6636 req
= io_put_req_find_next(req
);
6638 io_queue_linked_timeout(linked_timeout
);
6641 if (!(req
->flags
& REQ_F_FORCE_ASYNC
))
6643 io_queue_async_work(req
);
6646 /* un-prep timeout, so it'll be killed as any other linked */
6647 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
6648 req_set_fail_links(req
);
6650 io_req_complete(req
, ret
);
6654 revert_creds(old_creds
);
6657 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6658 struct io_comp_state
*cs
)
6662 ret
= io_req_defer(req
, sqe
);
6664 if (ret
!= -EIOCBQUEUED
) {
6666 req_set_fail_links(req
);
6668 io_req_complete(req
, ret
);
6670 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6671 if (!req
->async_data
) {
6672 ret
= io_req_defer_prep(req
, sqe
);
6676 io_queue_async_work(req
);
6679 ret
= io_req_prep(req
, sqe
);
6683 __io_queue_sqe(req
, cs
);
6687 static inline void io_queue_link_head(struct io_kiocb
*req
,
6688 struct io_comp_state
*cs
)
6690 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
6692 io_req_complete(req
, -ECANCELED
);
6694 io_queue_sqe(req
, NULL
, cs
);
6697 struct io_submit_link
{
6698 struct io_kiocb
*head
;
6699 struct io_kiocb
*last
;
6702 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6703 struct io_submit_link
*link
, struct io_comp_state
*cs
)
6705 struct io_ring_ctx
*ctx
= req
->ctx
;
6709 * If we already have a head request, queue this one for async
6710 * submittal once the head completes. If we don't have a head but
6711 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6712 * submitted sync once the chain is complete. If none of those
6713 * conditions are true (normal request), then just queue it.
6716 struct io_kiocb
*head
= link
->head
;
6719 * Taking sequential execution of a link, draining both sides
6720 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6721 * requests in the link. So, it drains the head and the
6722 * next after the link request. The last one is done via
6723 * drain_next flag to persist the effect across calls.
6725 if (req
->flags
& REQ_F_IO_DRAIN
) {
6726 head
->flags
|= REQ_F_IO_DRAIN
;
6727 ctx
->drain_next
= 1;
6729 ret
= io_req_defer_prep(req
, sqe
);
6730 if (unlikely(ret
)) {
6731 /* fail even hard links since we don't submit */
6732 head
->flags
|= REQ_F_FAIL_LINK
;
6735 trace_io_uring_link(ctx
, req
, head
);
6736 link
->last
->link
= req
;
6739 /* last request of a link, enqueue the link */
6740 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6741 io_queue_link_head(head
, cs
);
6745 if (unlikely(ctx
->drain_next
)) {
6746 req
->flags
|= REQ_F_IO_DRAIN
;
6747 ctx
->drain_next
= 0;
6749 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6750 ret
= io_req_defer_prep(req
, sqe
);
6752 req
->flags
|= REQ_F_FAIL_LINK
;
6756 io_queue_sqe(req
, sqe
, cs
);
6764 * Batched submission is done, ensure local IO is flushed out.
6766 static void io_submit_state_end(struct io_submit_state
*state
)
6768 if (!list_empty(&state
->comp
.list
))
6769 io_submit_flush_completions(&state
->comp
);
6770 if (state
->plug_started
)
6771 blk_finish_plug(&state
->plug
);
6772 io_state_file_put(state
);
6773 if (state
->free_reqs
)
6774 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6778 * Start submission side cache.
6780 static void io_submit_state_start(struct io_submit_state
*state
,
6781 struct io_ring_ctx
*ctx
, unsigned int max_ios
)
6783 state
->plug_started
= false;
6785 INIT_LIST_HEAD(&state
->comp
.list
);
6786 state
->comp
.ctx
= ctx
;
6787 state
->free_reqs
= 0;
6788 state
->file_refs
= 0;
6789 state
->ios_left
= max_ios
;
6792 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6794 struct io_rings
*rings
= ctx
->rings
;
6797 * Ensure any loads from the SQEs are done at this point,
6798 * since once we write the new head, the application could
6799 * write new data to them.
6801 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6805 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6806 * that is mapped by userspace. This means that care needs to be taken to
6807 * ensure that reads are stable, as we cannot rely on userspace always
6808 * being a good citizen. If members of the sqe are validated and then later
6809 * used, it's important that those reads are done through READ_ONCE() to
6810 * prevent a re-load down the line.
6812 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6814 u32
*sq_array
= ctx
->sq_array
;
6818 * The cached sq head (or cq tail) serves two purposes:
6820 * 1) allows us to batch the cost of updating the user visible
6822 * 2) allows the kernel side to track the head on its own, even
6823 * though the application is the one updating it.
6825 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6826 if (likely(head
< ctx
->sq_entries
))
6827 return &ctx
->sq_sqes
[head
];
6829 /* drop invalid entries */
6830 ctx
->cached_sq_dropped
++;
6831 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6835 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6837 ctx
->cached_sq_head
++;
6841 * Check SQE restrictions (opcode and flags).
6843 * Returns 'true' if SQE is allowed, 'false' otherwise.
6845 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6846 struct io_kiocb
*req
,
6847 unsigned int sqe_flags
)
6849 if (!ctx
->restricted
)
6852 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6855 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6856 ctx
->restrictions
.sqe_flags_required
)
6859 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6860 ctx
->restrictions
.sqe_flags_required
))
6866 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6867 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6868 IOSQE_BUFFER_SELECT)
6870 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6871 const struct io_uring_sqe
*sqe
,
6872 struct io_submit_state
*state
)
6874 unsigned int sqe_flags
;
6877 req
->opcode
= READ_ONCE(sqe
->opcode
);
6878 req
->user_data
= READ_ONCE(sqe
->user_data
);
6879 req
->async_data
= NULL
;
6884 req
->fixed_file_refs
= NULL
;
6885 /* one is dropped after submission, the other at completion */
6886 refcount_set(&req
->refs
, 2);
6887 req
->task
= current
;
6890 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6893 if (unlikely(io_sq_thread_acquire_mm_files(ctx
, req
)))
6896 sqe_flags
= READ_ONCE(sqe
->flags
);
6897 /* enforce forwards compatibility on users */
6898 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6901 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6904 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6905 !io_op_defs
[req
->opcode
].buffer_select
)
6908 id
= READ_ONCE(sqe
->personality
);
6910 struct io_identity
*iod
;
6912 iod
= xa_load(&ctx
->personalities
, id
);
6915 refcount_inc(&iod
->count
);
6917 __io_req_init_async(req
);
6918 get_cred(iod
->creds
);
6919 req
->work
.identity
= iod
;
6920 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
6923 /* same numerical values with corresponding REQ_F_*, safe to copy */
6924 req
->flags
|= sqe_flags
;
6927 * Plug now if we have more than 1 IO left after this, and the target
6928 * is potentially a read/write to block based storage.
6930 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6931 io_op_defs
[req
->opcode
].plug
) {
6932 blk_start_plug(&state
->plug
);
6933 state
->plug_started
= true;
6937 if (io_op_defs
[req
->opcode
].needs_file
) {
6938 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6940 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6941 if (unlikely(!req
->file
&&
6942 !io_op_defs
[req
->opcode
].needs_file_no_error
))
6950 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6952 struct io_submit_state state
;
6953 struct io_submit_link link
;
6954 int i
, submitted
= 0;
6956 /* if we have a backlog and couldn't flush it all, return BUSY */
6957 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6958 if (!__io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6962 /* make sure SQ entry isn't read before tail */
6963 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6965 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6968 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6969 refcount_add(nr
, ¤t
->usage
);
6971 io_submit_state_start(&state
, ctx
, nr
);
6974 for (i
= 0; i
< nr
; i
++) {
6975 const struct io_uring_sqe
*sqe
;
6976 struct io_kiocb
*req
;
6979 sqe
= io_get_sqe(ctx
);
6980 if (unlikely(!sqe
)) {
6981 io_consume_sqe(ctx
);
6984 req
= io_alloc_req(ctx
, &state
);
6985 if (unlikely(!req
)) {
6987 submitted
= -EAGAIN
;
6990 io_consume_sqe(ctx
);
6991 /* will complete beyond this point, count as submitted */
6994 err
= io_init_req(ctx
, req
, sqe
, &state
);
6995 if (unlikely(err
)) {
6998 io_req_complete(req
, err
);
7002 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
7003 true, io_async_submit(ctx
));
7004 err
= io_submit_sqe(req
, sqe
, &link
, &state
.comp
);
7009 if (unlikely(submitted
!= nr
)) {
7010 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
7011 struct io_uring_task
*tctx
= current
->io_uring
;
7012 int unused
= nr
- ref_used
;
7014 percpu_ref_put_many(&ctx
->refs
, unused
);
7015 percpu_counter_sub(&tctx
->inflight
, unused
);
7016 put_task_struct_many(current
, unused
);
7019 io_queue_link_head(link
.head
, &state
.comp
);
7020 io_submit_state_end(&state
);
7022 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7023 io_commit_sqring(ctx
);
7028 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
7030 /* Tell userspace we may need a wakeup call */
7031 spin_lock_irq(&ctx
->completion_lock
);
7032 WRITE_ONCE(ctx
->rings
->sq_flags
,
7033 ctx
->rings
->sq_flags
| IORING_SQ_NEED_WAKEUP
);
7034 spin_unlock_irq(&ctx
->completion_lock
);
7037 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
7039 spin_lock_irq(&ctx
->completion_lock
);
7040 WRITE_ONCE(ctx
->rings
->sq_flags
,
7041 ctx
->rings
->sq_flags
& ~IORING_SQ_NEED_WAKEUP
);
7042 spin_unlock_irq(&ctx
->completion_lock
);
7045 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
7047 unsigned int to_submit
;
7050 to_submit
= io_sqring_entries(ctx
);
7051 /* if we're handling multiple rings, cap submit size for fairness */
7052 if (cap_entries
&& to_submit
> 8)
7055 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
7056 unsigned nr_events
= 0;
7058 mutex_lock(&ctx
->uring_lock
);
7059 if (!list_empty(&ctx
->iopoll_list
))
7060 io_do_iopoll(ctx
, &nr_events
, 0);
7062 if (to_submit
&& !ctx
->sqo_dead
&&
7063 likely(!percpu_ref_is_dying(&ctx
->refs
)))
7064 ret
= io_submit_sqes(ctx
, to_submit
);
7065 mutex_unlock(&ctx
->uring_lock
);
7068 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
7069 wake_up(&ctx
->sqo_sq_wait
);
7074 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7076 struct io_ring_ctx
*ctx
;
7077 unsigned sq_thread_idle
= 0;
7079 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7080 if (sq_thread_idle
< ctx
->sq_thread_idle
)
7081 sq_thread_idle
= ctx
->sq_thread_idle
;
7084 sqd
->sq_thread_idle
= sq_thread_idle
;
7087 static void io_sqd_init_new(struct io_sq_data
*sqd
)
7089 struct io_ring_ctx
*ctx
;
7091 while (!list_empty(&sqd
->ctx_new_list
)) {
7092 ctx
= list_first_entry(&sqd
->ctx_new_list
, struct io_ring_ctx
, sqd_list
);
7093 list_move_tail(&ctx
->sqd_list
, &sqd
->ctx_list
);
7094 complete(&ctx
->sq_thread_comp
);
7097 io_sqd_update_thread_idle(sqd
);
7100 static int io_sq_thread(void *data
)
7102 struct cgroup_subsys_state
*cur_css
= NULL
;
7103 struct files_struct
*old_files
= current
->files
;
7104 struct nsproxy
*old_nsproxy
= current
->nsproxy
;
7105 const struct cred
*old_cred
= NULL
;
7106 struct io_sq_data
*sqd
= data
;
7107 struct io_ring_ctx
*ctx
;
7108 unsigned long timeout
= 0;
7112 current
->files
= NULL
;
7113 current
->nsproxy
= NULL
;
7114 task_unlock(current
);
7116 while (!kthread_should_stop()) {
7118 bool cap_entries
, sqt_spin
, needs_sched
;
7121 * Any changes to the sqd lists are synchronized through the
7122 * kthread parking. This synchronizes the thread vs users,
7123 * the users are synchronized on the sqd->ctx_lock.
7125 if (kthread_should_park()) {
7128 * When sq thread is unparked, in case the previous park operation
7129 * comes from io_put_sq_data(), which means that sq thread is going
7130 * to be stopped, so here needs to have a check.
7132 if (kthread_should_stop())
7136 if (unlikely(!list_empty(&sqd
->ctx_new_list
))) {
7137 io_sqd_init_new(sqd
);
7138 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7142 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7143 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7144 if (current
->cred
!= ctx
->creds
) {
7146 revert_creds(old_cred
);
7147 old_cred
= override_creds(ctx
->creds
);
7149 io_sq_thread_associate_blkcg(ctx
, &cur_css
);
7151 current
->loginuid
= ctx
->loginuid
;
7152 current
->sessionid
= ctx
->sessionid
;
7155 ret
= __io_sq_thread(ctx
, cap_entries
);
7156 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7159 io_sq_thread_drop_mm_files();
7162 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7164 io_sq_thread_drop_mm_files();
7167 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7171 if (kthread_should_park())
7175 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7176 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7177 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7178 !list_empty_careful(&ctx
->iopoll_list
)) {
7179 needs_sched
= false;
7182 if (io_sqring_entries(ctx
)) {
7183 needs_sched
= false;
7189 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7190 io_ring_set_wakeup_flag(ctx
);
7193 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7194 io_ring_clear_wakeup_flag(ctx
);
7197 finish_wait(&sqd
->wait
, &wait
);
7198 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7202 io_sq_thread_drop_mm_files();
7205 io_sq_thread_unassociate_blkcg();
7207 revert_creds(old_cred
);
7210 current
->files
= old_files
;
7211 current
->nsproxy
= old_nsproxy
;
7212 task_unlock(current
);
7219 struct io_wait_queue
{
7220 struct wait_queue_entry wq
;
7221 struct io_ring_ctx
*ctx
;
7223 unsigned nr_timeouts
;
7226 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7228 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7231 * Wake up if we have enough events, or if a timeout occurred since we
7232 * started waiting. For timeouts, we always want to return to userspace,
7233 * regardless of event count.
7235 return io_cqring_events(ctx
) >= iowq
->to_wait
||
7236 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7239 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7240 int wake_flags
, void *key
)
7242 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7246 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7247 * the task, and the next invocation will do it.
7249 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->cq_check_overflow
))
7250 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7254 static int io_run_task_work_sig(void)
7256 if (io_run_task_work())
7258 if (!signal_pending(current
))
7260 if (test_tsk_thread_flag(current
, TIF_NOTIFY_SIGNAL
))
7261 return -ERESTARTSYS
;
7265 /* when returns >0, the caller should retry */
7266 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7267 struct io_wait_queue
*iowq
,
7268 signed long *timeout
)
7272 /* make sure we run task_work before checking for signals */
7273 ret
= io_run_task_work_sig();
7274 if (ret
|| io_should_wake(iowq
))
7276 /* let the caller flush overflows, retry */
7277 if (test_bit(0, &ctx
->cq_check_overflow
))
7280 *timeout
= schedule_timeout(*timeout
);
7281 return !*timeout
? -ETIME
: 1;
7285 * Wait until events become available, if we don't already have some. The
7286 * application must reap them itself, as they reside on the shared cq ring.
7288 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7289 const sigset_t __user
*sig
, size_t sigsz
,
7290 struct __kernel_timespec __user
*uts
)
7292 struct io_wait_queue iowq
= {
7295 .func
= io_wake_function
,
7296 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
7299 .to_wait
= min_events
,
7301 struct io_rings
*rings
= ctx
->rings
;
7302 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
7306 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7307 if (io_cqring_events(ctx
) >= min_events
)
7309 if (!io_run_task_work())
7314 #ifdef CONFIG_COMPAT
7315 if (in_compat_syscall())
7316 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7320 ret
= set_user_sigmask(sig
, sigsz
);
7327 struct timespec64 ts
;
7329 if (get_timespec64(&ts
, uts
))
7331 timeout
= timespec64_to_jiffies(&ts
);
7334 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7335 trace_io_uring_cqring_wait(ctx
, min_events
);
7337 /* if we can't even flush overflow, don't wait for more */
7338 if (!io_cqring_overflow_flush(ctx
, false, NULL
, NULL
)) {
7342 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
7343 TASK_INTERRUPTIBLE
);
7344 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
7345 finish_wait(&ctx
->wait
, &iowq
.wq
);
7349 restore_saved_sigmask_unless(ret
== -EINTR
);
7351 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7354 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7356 #if defined(CONFIG_UNIX)
7357 if (ctx
->ring_sock
) {
7358 struct sock
*sock
= ctx
->ring_sock
->sk
;
7359 struct sk_buff
*skb
;
7361 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7367 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7370 file
= io_file_from_index(ctx
, i
);
7377 static void io_file_ref_kill(struct percpu_ref
*ref
)
7379 struct fixed_file_data
*data
;
7381 data
= container_of(ref
, struct fixed_file_data
, refs
);
7382 complete(&data
->done
);
7385 static void io_sqe_files_set_node(struct fixed_file_data
*file_data
,
7386 struct fixed_file_ref_node
*ref_node
)
7388 spin_lock_bh(&file_data
->lock
);
7389 file_data
->node
= ref_node
;
7390 list_add_tail(&ref_node
->node
, &file_data
->ref_list
);
7391 spin_unlock_bh(&file_data
->lock
);
7392 percpu_ref_get(&file_data
->refs
);
7395 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7397 struct fixed_file_data
*data
= ctx
->file_data
;
7398 struct fixed_file_ref_node
*backup_node
, *ref_node
= NULL
;
7399 unsigned nr_tables
, i
;
7404 backup_node
= alloc_fixed_file_ref_node(ctx
);
7408 spin_lock_bh(&data
->lock
);
7409 ref_node
= data
->node
;
7410 spin_unlock_bh(&data
->lock
);
7412 percpu_ref_kill(&ref_node
->refs
);
7414 percpu_ref_kill(&data
->refs
);
7416 /* wait for all refs nodes to complete */
7417 flush_delayed_work(&ctx
->file_put_work
);
7419 ret
= wait_for_completion_interruptible(&data
->done
);
7422 ret
= io_run_task_work_sig();
7424 percpu_ref_resurrect(&data
->refs
);
7425 reinit_completion(&data
->done
);
7426 io_sqe_files_set_node(data
, backup_node
);
7431 __io_sqe_files_unregister(ctx
);
7432 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
7433 for (i
= 0; i
< nr_tables
; i
++)
7434 kfree(data
->table
[i
].files
);
7436 percpu_ref_exit(&data
->refs
);
7438 ctx
->file_data
= NULL
;
7439 ctx
->nr_user_files
= 0;
7440 destroy_fixed_file_ref_node(backup_node
);
7444 static void io_put_sq_data(struct io_sq_data
*sqd
)
7446 if (refcount_dec_and_test(&sqd
->refs
)) {
7448 * The park is a bit of a work-around, without it we get
7449 * warning spews on shutdown with SQPOLL set and affinity
7450 * set to a single CPU.
7453 kthread_park(sqd
->thread
);
7454 kthread_stop(sqd
->thread
);
7461 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7463 struct io_ring_ctx
*ctx_attach
;
7464 struct io_sq_data
*sqd
;
7467 f
= fdget(p
->wq_fd
);
7469 return ERR_PTR(-ENXIO
);
7470 if (f
.file
->f_op
!= &io_uring_fops
) {
7472 return ERR_PTR(-EINVAL
);
7475 ctx_attach
= f
.file
->private_data
;
7476 sqd
= ctx_attach
->sq_data
;
7479 return ERR_PTR(-EINVAL
);
7482 refcount_inc(&sqd
->refs
);
7487 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
)
7489 struct io_sq_data
*sqd
;
7491 if (p
->flags
& IORING_SETUP_ATTACH_WQ
)
7492 return io_attach_sq_data(p
);
7494 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7496 return ERR_PTR(-ENOMEM
);
7498 refcount_set(&sqd
->refs
, 1);
7499 INIT_LIST_HEAD(&sqd
->ctx_list
);
7500 INIT_LIST_HEAD(&sqd
->ctx_new_list
);
7501 mutex_init(&sqd
->ctx_lock
);
7502 mutex_init(&sqd
->lock
);
7503 init_waitqueue_head(&sqd
->wait
);
7507 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7508 __releases(&sqd
->lock
)
7512 kthread_unpark(sqd
->thread
);
7513 mutex_unlock(&sqd
->lock
);
7516 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7517 __acquires(&sqd
->lock
)
7521 mutex_lock(&sqd
->lock
);
7522 kthread_park(sqd
->thread
);
7525 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
7527 struct io_sq_data
*sqd
= ctx
->sq_data
;
7532 * We may arrive here from the error branch in
7533 * io_sq_offload_create() where the kthread is created
7534 * without being waked up, thus wake it up now to make
7535 * sure the wait will complete.
7537 wake_up_process(sqd
->thread
);
7538 wait_for_completion(&ctx
->sq_thread_comp
);
7540 io_sq_thread_park(sqd
);
7543 mutex_lock(&sqd
->ctx_lock
);
7544 list_del(&ctx
->sqd_list
);
7545 io_sqd_update_thread_idle(sqd
);
7546 mutex_unlock(&sqd
->ctx_lock
);
7549 io_sq_thread_unpark(sqd
);
7551 io_put_sq_data(sqd
);
7552 ctx
->sq_data
= NULL
;
7556 static void io_finish_async(struct io_ring_ctx
*ctx
)
7558 io_sq_thread_stop(ctx
);
7561 io_wq_destroy(ctx
->io_wq
);
7566 #if defined(CONFIG_UNIX)
7568 * Ensure the UNIX gc is aware of our file set, so we are certain that
7569 * the io_uring can be safely unregistered on process exit, even if we have
7570 * loops in the file referencing.
7572 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7574 struct sock
*sk
= ctx
->ring_sock
->sk
;
7575 struct scm_fp_list
*fpl
;
7576 struct sk_buff
*skb
;
7579 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7583 skb
= alloc_skb(0, GFP_KERNEL
);
7592 fpl
->user
= get_uid(ctx
->user
);
7593 for (i
= 0; i
< nr
; i
++) {
7594 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7598 fpl
->fp
[nr_files
] = get_file(file
);
7599 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7604 fpl
->max
= SCM_MAX_FD
;
7605 fpl
->count
= nr_files
;
7606 UNIXCB(skb
).fp
= fpl
;
7607 skb
->destructor
= unix_destruct_scm
;
7608 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7609 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7611 for (i
= 0; i
< nr_files
; i
++)
7622 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7623 * causes regular reference counting to break down. We rely on the UNIX
7624 * garbage collection to take care of this problem for us.
7626 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7628 unsigned left
, total
;
7632 left
= ctx
->nr_user_files
;
7634 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7636 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7640 total
+= this_files
;
7646 while (total
< ctx
->nr_user_files
) {
7647 struct file
*file
= io_file_from_index(ctx
, total
);
7657 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7663 static int io_sqe_alloc_file_tables(struct fixed_file_data
*file_data
,
7664 unsigned nr_tables
, unsigned nr_files
)
7668 for (i
= 0; i
< nr_tables
; i
++) {
7669 struct fixed_file_table
*table
= &file_data
->table
[i
];
7670 unsigned this_files
;
7672 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7673 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7674 GFP_KERNEL_ACCOUNT
);
7677 nr_files
-= this_files
;
7683 for (i
= 0; i
< nr_tables
; i
++) {
7684 struct fixed_file_table
*table
= &file_data
->table
[i
];
7685 kfree(table
->files
);
7690 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
7692 #if defined(CONFIG_UNIX)
7693 struct sock
*sock
= ctx
->ring_sock
->sk
;
7694 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7695 struct sk_buff
*skb
;
7698 __skb_queue_head_init(&list
);
7701 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7702 * remove this entry and rearrange the file array.
7704 skb
= skb_dequeue(head
);
7706 struct scm_fp_list
*fp
;
7708 fp
= UNIXCB(skb
).fp
;
7709 for (i
= 0; i
< fp
->count
; i
++) {
7712 if (fp
->fp
[i
] != file
)
7715 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7716 left
= fp
->count
- 1 - i
;
7718 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7719 left
* sizeof(struct file
*));
7726 __skb_queue_tail(&list
, skb
);
7736 __skb_queue_tail(&list
, skb
);
7738 skb
= skb_dequeue(head
);
7741 if (skb_peek(&list
)) {
7742 spin_lock_irq(&head
->lock
);
7743 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7744 __skb_queue_tail(head
, skb
);
7745 spin_unlock_irq(&head
->lock
);
7752 struct io_file_put
{
7753 struct list_head list
;
7757 static void __io_file_put_work(struct fixed_file_ref_node
*ref_node
)
7759 struct fixed_file_data
*file_data
= ref_node
->file_data
;
7760 struct io_ring_ctx
*ctx
= file_data
->ctx
;
7761 struct io_file_put
*pfile
, *tmp
;
7763 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
7764 list_del(&pfile
->list
);
7765 io_ring_file_put(ctx
, pfile
->file
);
7769 percpu_ref_exit(&ref_node
->refs
);
7771 percpu_ref_put(&file_data
->refs
);
7774 static void io_file_put_work(struct work_struct
*work
)
7776 struct io_ring_ctx
*ctx
;
7777 struct llist_node
*node
;
7779 ctx
= container_of(work
, struct io_ring_ctx
, file_put_work
.work
);
7780 node
= llist_del_all(&ctx
->file_put_llist
);
7783 struct fixed_file_ref_node
*ref_node
;
7784 struct llist_node
*next
= node
->next
;
7786 ref_node
= llist_entry(node
, struct fixed_file_ref_node
, llist
);
7787 __io_file_put_work(ref_node
);
7792 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
7794 struct fixed_file_ref_node
*ref_node
;
7795 struct fixed_file_data
*data
;
7796 struct io_ring_ctx
*ctx
;
7797 bool first_add
= false;
7800 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
7801 data
= ref_node
->file_data
;
7804 spin_lock_bh(&data
->lock
);
7805 ref_node
->done
= true;
7807 while (!list_empty(&data
->ref_list
)) {
7808 ref_node
= list_first_entry(&data
->ref_list
,
7809 struct fixed_file_ref_node
, node
);
7810 /* recycle ref nodes in order */
7811 if (!ref_node
->done
)
7813 list_del(&ref_node
->node
);
7814 first_add
|= llist_add(&ref_node
->llist
, &ctx
->file_put_llist
);
7816 spin_unlock_bh(&data
->lock
);
7818 if (percpu_ref_is_dying(&data
->refs
))
7822 mod_delayed_work(system_wq
, &ctx
->file_put_work
, 0);
7824 queue_delayed_work(system_wq
, &ctx
->file_put_work
, delay
);
7827 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
7828 struct io_ring_ctx
*ctx
)
7830 struct fixed_file_ref_node
*ref_node
;
7832 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7836 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
7841 INIT_LIST_HEAD(&ref_node
->node
);
7842 INIT_LIST_HEAD(&ref_node
->file_list
);
7843 ref_node
->file_data
= ctx
->file_data
;
7844 ref_node
->done
= false;
7848 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
7850 percpu_ref_exit(&ref_node
->refs
);
7854 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7857 __s32 __user
*fds
= (__s32 __user
*) arg
;
7858 unsigned nr_tables
, i
;
7860 int fd
, ret
= -ENOMEM
;
7861 struct fixed_file_ref_node
*ref_node
;
7862 struct fixed_file_data
*file_data
;
7868 if (nr_args
> IORING_MAX_FIXED_FILES
)
7870 if (nr_args
> rlimit(RLIMIT_NOFILE
))
7873 file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL_ACCOUNT
);
7876 file_data
->ctx
= ctx
;
7877 init_completion(&file_data
->done
);
7878 INIT_LIST_HEAD(&file_data
->ref_list
);
7879 spin_lock_init(&file_data
->lock
);
7881 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7882 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7883 GFP_KERNEL_ACCOUNT
);
7884 if (!file_data
->table
)
7887 if (percpu_ref_init(&file_data
->refs
, io_file_ref_kill
,
7888 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
7891 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7893 ctx
->file_data
= file_data
;
7895 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7896 struct fixed_file_table
*table
;
7899 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7903 /* allow sparse sets */
7913 * Don't allow io_uring instances to be registered. If UNIX
7914 * isn't enabled, then this causes a reference cycle and this
7915 * instance can never get freed. If UNIX is enabled we'll
7916 * handle it just fine, but there's still no point in allowing
7917 * a ring fd as it doesn't support regular read/write anyway.
7919 if (file
->f_op
== &io_uring_fops
) {
7923 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7924 index
= i
& IORING_FILE_TABLE_MASK
;
7925 table
->files
[index
] = file
;
7928 ret
= io_sqe_files_scm(ctx
);
7930 io_sqe_files_unregister(ctx
);
7934 ref_node
= alloc_fixed_file_ref_node(ctx
);
7936 io_sqe_files_unregister(ctx
);
7940 io_sqe_files_set_node(file_data
, ref_node
);
7943 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7944 file
= io_file_from_index(ctx
, i
);
7948 for (i
= 0; i
< nr_tables
; i
++)
7949 kfree(file_data
->table
[i
].files
);
7950 ctx
->nr_user_files
= 0;
7952 percpu_ref_exit(&file_data
->refs
);
7954 kfree(file_data
->table
);
7956 ctx
->file_data
= NULL
;
7960 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7963 #if defined(CONFIG_UNIX)
7964 struct sock
*sock
= ctx
->ring_sock
->sk
;
7965 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7966 struct sk_buff
*skb
;
7969 * See if we can merge this file into an existing skb SCM_RIGHTS
7970 * file set. If there's no room, fall back to allocating a new skb
7971 * and filling it in.
7973 spin_lock_irq(&head
->lock
);
7974 skb
= skb_peek(head
);
7976 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7978 if (fpl
->count
< SCM_MAX_FD
) {
7979 __skb_unlink(skb
, head
);
7980 spin_unlock_irq(&head
->lock
);
7981 fpl
->fp
[fpl
->count
] = get_file(file
);
7982 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7984 spin_lock_irq(&head
->lock
);
7985 __skb_queue_head(head
, skb
);
7990 spin_unlock_irq(&head
->lock
);
7997 return __io_sqe_files_scm(ctx
, 1, index
);
8003 static int io_queue_file_removal(struct fixed_file_data
*data
,
8006 struct io_file_put
*pfile
;
8007 struct fixed_file_ref_node
*ref_node
= data
->node
;
8009 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
8014 list_add(&pfile
->list
, &ref_node
->file_list
);
8019 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
8020 struct io_uring_files_update
*up
,
8023 struct fixed_file_data
*data
= ctx
->file_data
;
8024 struct fixed_file_ref_node
*ref_node
;
8029 bool needs_switch
= false;
8031 if (check_add_overflow(up
->offset
, nr_args
, &done
))
8033 if (done
> ctx
->nr_user_files
)
8036 ref_node
= alloc_fixed_file_ref_node(ctx
);
8041 fds
= u64_to_user_ptr(up
->fds
);
8043 struct fixed_file_table
*table
;
8047 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8051 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
8052 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
8053 index
= i
& IORING_FILE_TABLE_MASK
;
8054 if (table
->files
[index
]) {
8055 file
= table
->files
[index
];
8056 err
= io_queue_file_removal(data
, file
);
8059 table
->files
[index
] = NULL
;
8060 needs_switch
= true;
8069 * Don't allow io_uring instances to be registered. If
8070 * UNIX isn't enabled, then this causes a reference
8071 * cycle and this instance can never get freed. If UNIX
8072 * is enabled we'll handle it just fine, but there's
8073 * still no point in allowing a ring fd as it doesn't
8074 * support regular read/write anyway.
8076 if (file
->f_op
== &io_uring_fops
) {
8081 table
->files
[index
] = file
;
8082 err
= io_sqe_file_register(ctx
, file
, i
);
8084 table
->files
[index
] = NULL
;
8095 percpu_ref_kill(&data
->node
->refs
);
8096 io_sqe_files_set_node(data
, ref_node
);
8098 destroy_fixed_file_ref_node(ref_node
);
8100 return done
? done
: err
;
8103 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
8106 struct io_uring_files_update up
;
8108 if (!ctx
->file_data
)
8112 if (copy_from_user(&up
, arg
, sizeof(up
)))
8117 return __io_sqe_files_update(ctx
, &up
, nr_args
);
8120 static struct io_wq_work
*io_free_work(struct io_wq_work
*work
)
8122 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8124 req
= io_put_req_find_next(req
);
8125 return req
? &req
->work
: NULL
;
8128 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
8129 struct io_uring_params
*p
)
8131 struct io_wq_data data
;
8133 struct io_ring_ctx
*ctx_attach
;
8134 unsigned int concurrency
;
8137 data
.user
= ctx
->user
;
8138 data
.free_work
= io_free_work
;
8139 data
.do_work
= io_wq_submit_work
;
8141 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
8142 /* Do QD, or 4 * CPUS, whatever is smallest */
8143 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8145 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
8146 if (IS_ERR(ctx
->io_wq
)) {
8147 ret
= PTR_ERR(ctx
->io_wq
);
8153 f
= fdget(p
->wq_fd
);
8157 if (f
.file
->f_op
!= &io_uring_fops
) {
8162 ctx_attach
= f
.file
->private_data
;
8163 /* @io_wq is protected by holding the fd */
8164 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
8169 ctx
->io_wq
= ctx_attach
->io_wq
;
8175 static int io_uring_alloc_task_context(struct task_struct
*task
)
8177 struct io_uring_task
*tctx
;
8180 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
8181 if (unlikely(!tctx
))
8184 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8185 if (unlikely(ret
)) {
8191 init_waitqueue_head(&tctx
->wait
);
8193 atomic_set(&tctx
->in_idle
, 0);
8194 tctx
->sqpoll
= false;
8195 io_init_identity(&tctx
->__identity
);
8196 tctx
->identity
= &tctx
->__identity
;
8197 task
->io_uring
= tctx
;
8201 void __io_uring_free(struct task_struct
*tsk
)
8203 struct io_uring_task
*tctx
= tsk
->io_uring
;
8205 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8206 WARN_ON_ONCE(refcount_read(&tctx
->identity
->count
) != 1);
8207 if (tctx
->identity
!= &tctx
->__identity
)
8208 kfree(tctx
->identity
);
8209 percpu_counter_destroy(&tctx
->inflight
);
8211 tsk
->io_uring
= NULL
;
8214 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8215 struct io_uring_params
*p
)
8219 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8220 struct io_sq_data
*sqd
;
8223 if (!capable(CAP_SYS_ADMIN
) && !capable(CAP_SYS_NICE
))
8226 sqd
= io_get_sq_data(p
);
8233 io_sq_thread_park(sqd
);
8234 mutex_lock(&sqd
->ctx_lock
);
8235 list_add(&ctx
->sqd_list
, &sqd
->ctx_new_list
);
8236 mutex_unlock(&sqd
->ctx_lock
);
8237 io_sq_thread_unpark(sqd
);
8239 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8240 if (!ctx
->sq_thread_idle
)
8241 ctx
->sq_thread_idle
= HZ
;
8246 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8247 int cpu
= p
->sq_thread_cpu
;
8250 if (cpu
>= nr_cpu_ids
)
8252 if (!cpu_online(cpu
))
8255 sqd
->thread
= kthread_create_on_cpu(io_sq_thread
, sqd
,
8256 cpu
, "io_uring-sq");
8258 sqd
->thread
= kthread_create(io_sq_thread
, sqd
,
8261 if (IS_ERR(sqd
->thread
)) {
8262 ret
= PTR_ERR(sqd
->thread
);
8266 ret
= io_uring_alloc_task_context(sqd
->thread
);
8269 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8270 /* Can't have SQ_AFF without SQPOLL */
8276 ret
= io_init_wq_offload(ctx
, p
);
8282 io_finish_async(ctx
);
8286 static void io_sq_offload_start(struct io_ring_ctx
*ctx
)
8288 struct io_sq_data
*sqd
= ctx
->sq_data
;
8290 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
8291 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && sqd
&& sqd
->thread
)
8292 wake_up_process(sqd
->thread
);
8295 static inline void __io_unaccount_mem(struct user_struct
*user
,
8296 unsigned long nr_pages
)
8298 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8301 static inline int __io_account_mem(struct user_struct
*user
,
8302 unsigned long nr_pages
)
8304 unsigned long page_limit
, cur_pages
, new_pages
;
8306 /* Don't allow more pages than we can safely lock */
8307 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8310 cur_pages
= atomic_long_read(&user
->locked_vm
);
8311 new_pages
= cur_pages
+ nr_pages
;
8312 if (new_pages
> page_limit
)
8314 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8315 new_pages
) != cur_pages
);
8320 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8321 enum io_mem_account acct
)
8324 __io_unaccount_mem(ctx
->user
, nr_pages
);
8326 if (ctx
->mm_account
) {
8327 if (acct
== ACCT_LOCKED
) {
8328 mmap_write_lock(ctx
->mm_account
);
8329 ctx
->mm_account
->locked_vm
-= nr_pages
;
8330 mmap_write_unlock(ctx
->mm_account
);
8331 }else if (acct
== ACCT_PINNED
) {
8332 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8337 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8338 enum io_mem_account acct
)
8342 if (ctx
->limit_mem
) {
8343 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8348 if (ctx
->mm_account
) {
8349 if (acct
== ACCT_LOCKED
) {
8350 mmap_write_lock(ctx
->mm_account
);
8351 ctx
->mm_account
->locked_vm
+= nr_pages
;
8352 mmap_write_unlock(ctx
->mm_account
);
8353 } else if (acct
== ACCT_PINNED
) {
8354 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8361 static void io_mem_free(void *ptr
)
8368 page
= virt_to_head_page(ptr
);
8369 if (put_page_testzero(page
))
8370 free_compound_page(page
);
8373 static void *io_mem_alloc(size_t size
)
8375 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8378 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8381 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8384 struct io_rings
*rings
;
8385 size_t off
, sq_array_size
;
8387 off
= struct_size(rings
, cqes
, cq_entries
);
8388 if (off
== SIZE_MAX
)
8392 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8400 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8401 if (sq_array_size
== SIZE_MAX
)
8404 if (check_add_overflow(off
, sq_array_size
, &off
))
8410 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
8414 pages
= (size_t)1 << get_order(
8415 rings_size(sq_entries
, cq_entries
, NULL
));
8416 pages
+= (size_t)1 << get_order(
8417 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
8422 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
8426 if (!ctx
->user_bufs
)
8429 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8430 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8432 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
8433 unpin_user_page(imu
->bvec
[j
].bv_page
);
8435 if (imu
->acct_pages
)
8436 io_unaccount_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8441 kfree(ctx
->user_bufs
);
8442 ctx
->user_bufs
= NULL
;
8443 ctx
->nr_user_bufs
= 0;
8447 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8448 void __user
*arg
, unsigned index
)
8450 struct iovec __user
*src
;
8452 #ifdef CONFIG_COMPAT
8454 struct compat_iovec __user
*ciovs
;
8455 struct compat_iovec ciov
;
8457 ciovs
= (struct compat_iovec __user
*) arg
;
8458 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8461 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8462 dst
->iov_len
= ciov
.iov_len
;
8466 src
= (struct iovec __user
*) arg
;
8467 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8473 * Not super efficient, but this is just a registration time. And we do cache
8474 * the last compound head, so generally we'll only do a full search if we don't
8477 * We check if the given compound head page has already been accounted, to
8478 * avoid double accounting it. This allows us to account the full size of the
8479 * page, not just the constituent pages of a huge page.
8481 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8482 int nr_pages
, struct page
*hpage
)
8486 /* check current page array */
8487 for (i
= 0; i
< nr_pages
; i
++) {
8488 if (!PageCompound(pages
[i
]))
8490 if (compound_head(pages
[i
]) == hpage
)
8494 /* check previously registered pages */
8495 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8496 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8498 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8499 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8501 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8509 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8510 int nr_pages
, struct io_mapped_ubuf
*imu
,
8511 struct page
**last_hpage
)
8515 for (i
= 0; i
< nr_pages
; i
++) {
8516 if (!PageCompound(pages
[i
])) {
8521 hpage
= compound_head(pages
[i
]);
8522 if (hpage
== *last_hpage
)
8524 *last_hpage
= hpage
;
8525 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8527 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8531 if (!imu
->acct_pages
)
8534 ret
= io_account_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8536 imu
->acct_pages
= 0;
8540 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8543 struct vm_area_struct
**vmas
= NULL
;
8544 struct page
**pages
= NULL
;
8545 struct page
*last_hpage
= NULL
;
8546 int i
, j
, got_pages
= 0;
8551 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8554 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8556 if (!ctx
->user_bufs
)
8559 for (i
= 0; i
< nr_args
; i
++) {
8560 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8561 unsigned long off
, start
, end
, ubuf
;
8566 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8571 * Don't impose further limits on the size and buffer
8572 * constraints here, we'll -EINVAL later when IO is
8573 * submitted if they are wrong.
8576 if (!iov
.iov_base
|| !iov
.iov_len
)
8579 /* arbitrary limit, but we need something */
8580 if (iov
.iov_len
> SZ_1G
)
8583 ubuf
= (unsigned long) iov
.iov_base
;
8584 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8585 start
= ubuf
>> PAGE_SHIFT
;
8586 nr_pages
= end
- start
;
8589 if (!pages
|| nr_pages
> got_pages
) {
8592 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
8594 vmas
= kvmalloc_array(nr_pages
,
8595 sizeof(struct vm_area_struct
*),
8597 if (!pages
|| !vmas
) {
8601 got_pages
= nr_pages
;
8604 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8611 mmap_read_lock(current
->mm
);
8612 pret
= pin_user_pages(ubuf
, nr_pages
,
8613 FOLL_WRITE
| FOLL_LONGTERM
,
8615 if (pret
== nr_pages
) {
8616 /* don't support file backed memory */
8617 for (j
= 0; j
< nr_pages
; j
++) {
8618 struct vm_area_struct
*vma
= vmas
[j
];
8621 !is_file_hugepages(vma
->vm_file
)) {
8627 ret
= pret
< 0 ? pret
: -EFAULT
;
8629 mmap_read_unlock(current
->mm
);
8632 * if we did partial map, or found file backed vmas,
8633 * release any pages we did get
8636 unpin_user_pages(pages
, pret
);
8641 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, &last_hpage
);
8643 unpin_user_pages(pages
, pret
);
8648 off
= ubuf
& ~PAGE_MASK
;
8650 for (j
= 0; j
< nr_pages
; j
++) {
8653 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8654 imu
->bvec
[j
].bv_page
= pages
[j
];
8655 imu
->bvec
[j
].bv_len
= vec_len
;
8656 imu
->bvec
[j
].bv_offset
= off
;
8660 /* store original address for later verification */
8662 imu
->len
= iov
.iov_len
;
8663 imu
->nr_bvecs
= nr_pages
;
8665 ctx
->nr_user_bufs
++;
8673 io_sqe_buffer_unregister(ctx
);
8677 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8679 __s32 __user
*fds
= arg
;
8685 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8688 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8689 if (IS_ERR(ctx
->cq_ev_fd
)) {
8690 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8692 ctx
->cq_ev_fd
= NULL
;
8699 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8701 if (ctx
->cq_ev_fd
) {
8702 eventfd_ctx_put(ctx
->cq_ev_fd
);
8703 ctx
->cq_ev_fd
= NULL
;
8710 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8712 struct io_buffer
*buf
;
8713 unsigned long index
;
8715 xa_for_each(&ctx
->io_buffers
, index
, buf
)
8716 __io_remove_buffers(ctx
, buf
, index
, -1U);
8719 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8721 io_finish_async(ctx
);
8722 io_sqe_buffer_unregister(ctx
);
8724 if (ctx
->sqo_task
) {
8725 put_task_struct(ctx
->sqo_task
);
8726 ctx
->sqo_task
= NULL
;
8727 mmdrop(ctx
->mm_account
);
8728 ctx
->mm_account
= NULL
;
8731 #ifdef CONFIG_BLK_CGROUP
8732 if (ctx
->sqo_blkcg_css
)
8733 css_put(ctx
->sqo_blkcg_css
);
8736 io_sqe_files_unregister(ctx
);
8737 io_eventfd_unregister(ctx
);
8738 io_destroy_buffers(ctx
);
8740 #if defined(CONFIG_UNIX)
8741 if (ctx
->ring_sock
) {
8742 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8743 sock_release(ctx
->ring_sock
);
8747 io_mem_free(ctx
->rings
);
8748 io_mem_free(ctx
->sq_sqes
);
8750 percpu_ref_exit(&ctx
->refs
);
8751 free_uid(ctx
->user
);
8752 put_cred(ctx
->creds
);
8753 kfree(ctx
->cancel_hash
);
8754 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
8758 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8760 struct io_ring_ctx
*ctx
= file
->private_data
;
8763 poll_wait(file
, &ctx
->cq_wait
, wait
);
8765 * synchronizes with barrier from wq_has_sleeper call in
8769 if (!io_sqring_full(ctx
))
8770 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8772 /* prevent SQPOLL from submitting new requests */
8774 io_sq_thread_park(ctx
->sq_data
);
8775 list_del_init(&ctx
->sqd_list
);
8776 io_sqd_update_thread_idle(ctx
->sq_data
);
8777 io_sq_thread_unpark(ctx
->sq_data
);
8781 * Don't flush cqring overflow list here, just do a simple check.
8782 * Otherwise there could possible be ABBA deadlock:
8785 * lock(&ctx->uring_lock);
8787 * lock(&ctx->uring_lock);
8790 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8791 * pushs them to do the flush.
8793 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->cq_check_overflow
))
8794 mask
|= EPOLLIN
| EPOLLRDNORM
;
8799 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8801 struct io_ring_ctx
*ctx
= file
->private_data
;
8803 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8806 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8808 struct io_identity
*iod
;
8810 iod
= xa_erase(&ctx
->personalities
, id
);
8812 put_cred(iod
->creds
);
8813 if (refcount_dec_and_test(&iod
->count
))
8821 static void io_ring_exit_work(struct work_struct
*work
)
8823 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
8827 * If we're doing polled IO and end up having requests being
8828 * submitted async (out-of-line), then completions can come in while
8829 * we're waiting for refs to drop. We need to reap these manually,
8830 * as nobody else will be looking for them.
8833 io_uring_try_cancel_requests(ctx
, NULL
, NULL
);
8834 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8835 io_ring_ctx_free(ctx
);
8838 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8840 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8842 return req
->ctx
== data
;
8845 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8847 unsigned long index
;
8848 struct io_identify
*iod
;
8850 mutex_lock(&ctx
->uring_lock
);
8851 percpu_ref_kill(&ctx
->refs
);
8853 if (WARN_ON_ONCE((ctx
->flags
& IORING_SETUP_SQPOLL
) && !ctx
->sqo_dead
))
8856 /* if force is set, the ring is going away. always drop after that */
8857 ctx
->cq_overflow_flushed
= 1;
8859 __io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8860 mutex_unlock(&ctx
->uring_lock
);
8862 io_kill_timeouts(ctx
, NULL
, NULL
);
8863 io_poll_remove_all(ctx
, NULL
, NULL
);
8866 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8868 /* if we failed setting up the ctx, we might not have any rings */
8869 io_iopoll_try_reap_events(ctx
);
8870 xa_for_each(&ctx
->personalities
, index
, iod
)
8871 io_unregister_personality(ctx
, index
);
8874 * Do this upfront, so we won't have a grace period where the ring
8875 * is closed but resources aren't reaped yet. This can cause
8876 * spurious failure in setting up a new ring.
8878 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
8881 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8883 * Use system_unbound_wq to avoid spawning tons of event kworkers
8884 * if we're exiting a ton of rings at the same time. It just adds
8885 * noise and overhead, there's no discernable change in runtime
8886 * over using system_wq.
8888 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8891 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8893 struct io_ring_ctx
*ctx
= file
->private_data
;
8895 file
->private_data
= NULL
;
8896 io_ring_ctx_wait_and_kill(ctx
);
8900 struct io_task_cancel
{
8901 struct task_struct
*task
;
8902 struct files_struct
*files
;
8905 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8907 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8908 struct io_task_cancel
*cancel
= data
;
8911 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8912 unsigned long flags
;
8913 struct io_ring_ctx
*ctx
= req
->ctx
;
8915 /* protect against races with linked timeouts */
8916 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8917 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8918 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8920 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8925 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8926 struct task_struct
*task
,
8927 struct files_struct
*files
)
8929 struct io_defer_entry
*de
;
8932 spin_lock_irq(&ctx
->completion_lock
);
8933 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8934 if (io_match_task(de
->req
, task
, files
)) {
8935 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8939 spin_unlock_irq(&ctx
->completion_lock
);
8940 if (list_empty(&list
))
8943 while (!list_empty(&list
)) {
8944 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8945 list_del_init(&de
->list
);
8946 req_set_fail_links(de
->req
);
8947 io_put_req(de
->req
);
8948 io_req_complete(de
->req
, -ECANCELED
);
8954 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
8955 struct task_struct
*task
,
8956 struct files_struct
*files
)
8958 struct io_task_cancel cancel
= { .task
= task
, .files
= files
, };
8961 enum io_wq_cancel cret
;
8965 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
,
8967 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8970 /* SQPOLL thread does its own polling */
8971 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && !files
) ||
8972 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
8973 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8974 io_iopoll_try_reap_events(ctx
);
8979 ret
|= io_cancel_defer_files(ctx
, task
, files
);
8980 ret
|= io_poll_remove_all(ctx
, task
, files
);
8981 ret
|= io_kill_timeouts(ctx
, task
, files
);
8982 ret
|= io_run_task_work();
8983 io_cqring_overflow_flush(ctx
, true, task
, files
);
8990 static int io_uring_count_inflight(struct io_ring_ctx
*ctx
,
8991 struct task_struct
*task
,
8992 struct files_struct
*files
)
8994 struct io_kiocb
*req
;
8997 spin_lock_irq(&ctx
->inflight_lock
);
8998 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
)
8999 cnt
+= io_match_task(req
, task
, files
);
9000 spin_unlock_irq(&ctx
->inflight_lock
);
9004 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
9005 struct task_struct
*task
,
9006 struct files_struct
*files
)
9008 while (!list_empty_careful(&ctx
->inflight_list
)) {
9012 inflight
= io_uring_count_inflight(ctx
, task
, files
);
9016 io_uring_try_cancel_requests(ctx
, task
, files
);
9019 io_sq_thread_unpark(ctx
->sq_data
);
9020 prepare_to_wait(&task
->io_uring
->wait
, &wait
,
9021 TASK_UNINTERRUPTIBLE
);
9022 if (inflight
== io_uring_count_inflight(ctx
, task
, files
))
9024 finish_wait(&task
->io_uring
->wait
, &wait
);
9026 io_sq_thread_park(ctx
->sq_data
);
9030 static void io_disable_sqo_submit(struct io_ring_ctx
*ctx
)
9032 mutex_lock(&ctx
->uring_lock
);
9034 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9035 io_sq_offload_start(ctx
);
9036 mutex_unlock(&ctx
->uring_lock
);
9038 /* make sure callers enter the ring to get error */
9040 io_ring_set_wakeup_flag(ctx
);
9044 * We need to iteratively cancel requests, in case a request has dependent
9045 * hard links. These persist even for failure of cancelations, hence keep
9046 * looping until none are found.
9048 static void io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
9049 struct files_struct
*files
)
9051 struct task_struct
*task
= current
;
9053 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
9054 io_disable_sqo_submit(ctx
);
9055 task
= ctx
->sq_data
->thread
;
9056 atomic_inc(&task
->io_uring
->in_idle
);
9057 io_sq_thread_park(ctx
->sq_data
);
9060 io_uring_cancel_files(ctx
, task
, files
);
9062 io_uring_try_cancel_requests(ctx
, task
, NULL
);
9064 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
9065 atomic_dec(&task
->io_uring
->in_idle
);
9066 io_sq_thread_unpark(ctx
->sq_data
);
9071 * Note that this task has used io_uring. We use it for cancelation purposes.
9073 static int io_uring_add_task_file(struct io_ring_ctx
*ctx
, struct file
*file
)
9075 struct io_uring_task
*tctx
= current
->io_uring
;
9078 if (unlikely(!tctx
)) {
9079 ret
= io_uring_alloc_task_context(current
);
9082 tctx
= current
->io_uring
;
9084 if (tctx
->last
!= file
) {
9085 void *old
= xa_load(&tctx
->xa
, (unsigned long)file
);
9089 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)file
,
9100 * This is race safe in that the task itself is doing this, hence it
9101 * cannot be going through the exit/cancel paths at the same time.
9102 * This cannot be modified while exit/cancel is running.
9104 if (!tctx
->sqpoll
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9105 tctx
->sqpoll
= true;
9111 * Remove this io_uring_file -> task mapping.
9113 static void io_uring_del_task_file(struct file
*file
)
9115 struct io_uring_task
*tctx
= current
->io_uring
;
9117 if (tctx
->last
== file
)
9119 file
= xa_erase(&tctx
->xa
, (unsigned long)file
);
9124 static void io_uring_remove_task_files(struct io_uring_task
*tctx
)
9127 unsigned long index
;
9129 xa_for_each(&tctx
->xa
, index
, file
)
9130 io_uring_del_task_file(file
);
9133 void __io_uring_files_cancel(struct files_struct
*files
)
9135 struct io_uring_task
*tctx
= current
->io_uring
;
9137 unsigned long index
;
9139 /* make sure overflow events are dropped */
9140 atomic_inc(&tctx
->in_idle
);
9141 xa_for_each(&tctx
->xa
, index
, file
)
9142 io_uring_cancel_task_requests(file
->private_data
, files
);
9143 atomic_dec(&tctx
->in_idle
);
9146 io_uring_remove_task_files(tctx
);
9149 static s64
tctx_inflight(struct io_uring_task
*tctx
)
9151 unsigned long index
;
9155 inflight
= percpu_counter_sum(&tctx
->inflight
);
9160 * If we have SQPOLL rings, then we need to iterate and find them, and
9161 * add the pending count for those.
9163 xa_for_each(&tctx
->xa
, index
, file
) {
9164 struct io_ring_ctx
*ctx
= file
->private_data
;
9166 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9167 struct io_uring_task
*__tctx
= ctx
->sqo_task
->io_uring
;
9169 inflight
+= percpu_counter_sum(&__tctx
->inflight
);
9177 * Find any io_uring fd that this task has registered or done IO on, and cancel
9180 void __io_uring_task_cancel(void)
9182 struct io_uring_task
*tctx
= current
->io_uring
;
9186 /* make sure overflow events are dropped */
9187 atomic_inc(&tctx
->in_idle
);
9189 /* trigger io_disable_sqo_submit() */
9191 __io_uring_files_cancel(NULL
);
9194 /* read completions before cancelations */
9195 inflight
= tctx_inflight(tctx
);
9198 __io_uring_files_cancel(NULL
);
9200 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9203 * If we've seen completions, retry without waiting. This
9204 * avoids a race where a completion comes in before we did
9205 * prepare_to_wait().
9207 if (inflight
== tctx_inflight(tctx
))
9209 finish_wait(&tctx
->wait
, &wait
);
9212 atomic_dec(&tctx
->in_idle
);
9214 io_uring_remove_task_files(tctx
);
9217 static int io_uring_flush(struct file
*file
, void *data
)
9219 struct io_uring_task
*tctx
= current
->io_uring
;
9220 struct io_ring_ctx
*ctx
= file
->private_data
;
9222 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
))
9223 io_uring_cancel_task_requests(ctx
, NULL
);
9228 /* we should have cancelled and erased it before PF_EXITING */
9229 WARN_ON_ONCE((current
->flags
& PF_EXITING
) &&
9230 xa_load(&tctx
->xa
, (unsigned long)file
));
9233 * fput() is pending, will be 2 if the only other ref is our potential
9234 * task file note. If the task is exiting, drop regardless of count.
9236 if (atomic_long_read(&file
->f_count
) != 2)
9239 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9240 /* there is only one file note, which is owned by sqo_task */
9241 WARN_ON_ONCE(ctx
->sqo_task
!= current
&&
9242 xa_load(&tctx
->xa
, (unsigned long)file
));
9243 /* sqo_dead check is for when this happens after cancellation */
9244 WARN_ON_ONCE(ctx
->sqo_task
== current
&& !ctx
->sqo_dead
&&
9245 !xa_load(&tctx
->xa
, (unsigned long)file
));
9247 io_disable_sqo_submit(ctx
);
9250 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) || ctx
->sqo_task
== current
)
9251 io_uring_del_task_file(file
);
9255 static void *io_uring_validate_mmap_request(struct file
*file
,
9256 loff_t pgoff
, size_t sz
)
9258 struct io_ring_ctx
*ctx
= file
->private_data
;
9259 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9264 case IORING_OFF_SQ_RING
:
9265 case IORING_OFF_CQ_RING
:
9268 case IORING_OFF_SQES
:
9272 return ERR_PTR(-EINVAL
);
9275 page
= virt_to_head_page(ptr
);
9276 if (sz
> page_size(page
))
9277 return ERR_PTR(-EINVAL
);
9284 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9286 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9290 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9292 return PTR_ERR(ptr
);
9294 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9295 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9298 #else /* !CONFIG_MMU */
9300 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9302 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9305 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9307 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9310 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9311 unsigned long addr
, unsigned long len
,
9312 unsigned long pgoff
, unsigned long flags
)
9316 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9318 return PTR_ERR(ptr
);
9320 return (unsigned long) ptr
;
9323 #endif /* !CONFIG_MMU */
9325 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9331 if (!io_sqring_full(ctx
))
9334 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9336 if (unlikely(ctx
->sqo_dead
)) {
9341 if (!io_sqring_full(ctx
))
9345 } while (!signal_pending(current
));
9347 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9352 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9353 struct __kernel_timespec __user
**ts
,
9354 const sigset_t __user
**sig
)
9356 struct io_uring_getevents_arg arg
;
9359 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9360 * is just a pointer to the sigset_t.
9362 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9363 *sig
= (const sigset_t __user
*) argp
;
9369 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9370 * timespec and sigset_t pointers if good.
9372 if (*argsz
!= sizeof(arg
))
9374 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9376 *sig
= u64_to_user_ptr(arg
.sigmask
);
9377 *argsz
= arg
.sigmask_sz
;
9378 *ts
= u64_to_user_ptr(arg
.ts
);
9382 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9383 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9386 struct io_ring_ctx
*ctx
;
9393 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9394 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
))
9402 if (f
.file
->f_op
!= &io_uring_fops
)
9406 ctx
= f
.file
->private_data
;
9407 if (!percpu_ref_tryget(&ctx
->refs
))
9411 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9415 * For SQ polling, the thread will do all submissions and completions.
9416 * Just return the requested submit count, and wake the thread if
9420 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9421 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
9423 if (unlikely(ctx
->sqo_dead
)) {
9427 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9428 wake_up(&ctx
->sq_data
->wait
);
9429 if (flags
& IORING_ENTER_SQ_WAIT
) {
9430 ret
= io_sqpoll_wait_sq(ctx
);
9434 submitted
= to_submit
;
9435 } else if (to_submit
) {
9436 ret
= io_uring_add_task_file(ctx
, f
.file
);
9439 mutex_lock(&ctx
->uring_lock
);
9440 submitted
= io_submit_sqes(ctx
, to_submit
);
9441 mutex_unlock(&ctx
->uring_lock
);
9443 if (submitted
!= to_submit
)
9446 if (flags
& IORING_ENTER_GETEVENTS
) {
9447 const sigset_t __user
*sig
;
9448 struct __kernel_timespec __user
*ts
;
9450 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9454 min_complete
= min(min_complete
, ctx
->cq_entries
);
9457 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9458 * space applications don't need to do io completion events
9459 * polling again, they can rely on io_sq_thread to do polling
9460 * work, which can reduce cpu usage and uring_lock contention.
9462 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9463 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9464 ret
= io_iopoll_check(ctx
, min_complete
);
9466 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9471 percpu_ref_put(&ctx
->refs
);
9474 return submitted
? submitted
: ret
;
9477 #ifdef CONFIG_PROC_FS
9478 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
9479 const struct io_identity
*iod
)
9481 const struct cred
*cred
= iod
->creds
;
9482 struct user_namespace
*uns
= seq_user_ns(m
);
9483 struct group_info
*gi
;
9488 seq_printf(m
, "%5d\n", id
);
9489 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9490 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9491 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9492 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9493 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9494 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9495 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9496 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9497 seq_puts(m
, "\n\tGroups:\t");
9498 gi
= cred
->group_info
;
9499 for (g
= 0; g
< gi
->ngroups
; g
++) {
9500 seq_put_decimal_ull(m
, g
? " " : "",
9501 from_kgid_munged(uns
, gi
->gid
[g
]));
9503 seq_puts(m
, "\n\tCapEff:\t");
9504 cap
= cred
->cap_effective
;
9505 CAP_FOR_EACH_U32(__capi
)
9506 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9511 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9513 struct io_sq_data
*sq
= NULL
;
9518 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9519 * since fdinfo case grabs it in the opposite direction of normal use
9520 * cases. If we fail to get the lock, we just don't iterate any
9521 * structures that could be going away outside the io_uring mutex.
9523 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9525 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9528 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9529 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9530 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9531 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9532 struct fixed_file_table
*table
;
9535 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
9536 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
9538 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9540 seq_printf(m
, "%5u: <none>\n", i
);
9542 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9543 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9544 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9546 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9547 (unsigned int) buf
->len
);
9549 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
9550 unsigned long index
;
9551 const struct io_identity
*iod
;
9553 seq_printf(m
, "Personalities:\n");
9554 xa_for_each(&ctx
->personalities
, index
, iod
)
9555 io_uring_show_cred(m
, index
, iod
);
9557 seq_printf(m
, "PollList:\n");
9558 spin_lock_irq(&ctx
->completion_lock
);
9559 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9560 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9561 struct io_kiocb
*req
;
9563 hlist_for_each_entry(req
, list
, hash_node
)
9564 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9565 req
->task
->task_works
!= NULL
);
9567 spin_unlock_irq(&ctx
->completion_lock
);
9569 mutex_unlock(&ctx
->uring_lock
);
9572 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9574 struct io_ring_ctx
*ctx
= f
->private_data
;
9576 if (percpu_ref_tryget(&ctx
->refs
)) {
9577 __io_uring_show_fdinfo(ctx
, m
);
9578 percpu_ref_put(&ctx
->refs
);
9583 static const struct file_operations io_uring_fops
= {
9584 .release
= io_uring_release
,
9585 .flush
= io_uring_flush
,
9586 .mmap
= io_uring_mmap
,
9588 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9589 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9591 .poll
= io_uring_poll
,
9592 .fasync
= io_uring_fasync
,
9593 #ifdef CONFIG_PROC_FS
9594 .show_fdinfo
= io_uring_show_fdinfo
,
9598 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9599 struct io_uring_params
*p
)
9601 struct io_rings
*rings
;
9602 size_t size
, sq_array_offset
;
9604 /* make sure these are sane, as we already accounted them */
9605 ctx
->sq_entries
= p
->sq_entries
;
9606 ctx
->cq_entries
= p
->cq_entries
;
9608 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9609 if (size
== SIZE_MAX
)
9612 rings
= io_mem_alloc(size
);
9617 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9618 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9619 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9620 rings
->sq_ring_entries
= p
->sq_entries
;
9621 rings
->cq_ring_entries
= p
->cq_entries
;
9622 ctx
->sq_mask
= rings
->sq_ring_mask
;
9623 ctx
->cq_mask
= rings
->cq_ring_mask
;
9625 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9626 if (size
== SIZE_MAX
) {
9627 io_mem_free(ctx
->rings
);
9632 ctx
->sq_sqes
= io_mem_alloc(size
);
9633 if (!ctx
->sq_sqes
) {
9634 io_mem_free(ctx
->rings
);
9642 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9646 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9650 ret
= io_uring_add_task_file(ctx
, file
);
9655 fd_install(fd
, file
);
9660 * Allocate an anonymous fd, this is what constitutes the application
9661 * visible backing of an io_uring instance. The application mmaps this
9662 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9663 * we have to tie this fd to a socket for file garbage collection purposes.
9665 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9668 #if defined(CONFIG_UNIX)
9671 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9674 return ERR_PTR(ret
);
9677 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9678 O_RDWR
| O_CLOEXEC
);
9679 #if defined(CONFIG_UNIX)
9681 sock_release(ctx
->ring_sock
);
9682 ctx
->ring_sock
= NULL
;
9684 ctx
->ring_sock
->file
= file
;
9690 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9691 struct io_uring_params __user
*params
)
9693 struct user_struct
*user
= NULL
;
9694 struct io_ring_ctx
*ctx
;
9701 if (entries
> IORING_MAX_ENTRIES
) {
9702 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9704 entries
= IORING_MAX_ENTRIES
;
9708 * Use twice as many entries for the CQ ring. It's possible for the
9709 * application to drive a higher depth than the size of the SQ ring,
9710 * since the sqes are only used at submission time. This allows for
9711 * some flexibility in overcommitting a bit. If the application has
9712 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9713 * of CQ ring entries manually.
9715 p
->sq_entries
= roundup_pow_of_two(entries
);
9716 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9718 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9719 * to a power-of-two, if it isn't already. We do NOT impose
9720 * any cq vs sq ring sizing.
9724 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9725 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9727 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9729 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9730 if (p
->cq_entries
< p
->sq_entries
)
9733 p
->cq_entries
= 2 * p
->sq_entries
;
9736 user
= get_uid(current_user());
9737 limit_mem
= !capable(CAP_IPC_LOCK
);
9740 ret
= __io_account_mem(user
,
9741 ring_pages(p
->sq_entries
, p
->cq_entries
));
9748 ctx
= io_ring_ctx_alloc(p
);
9751 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
9756 ctx
->compat
= in_compat_syscall();
9758 ctx
->creds
= get_current_cred();
9760 ctx
->loginuid
= current
->loginuid
;
9761 ctx
->sessionid
= current
->sessionid
;
9763 ctx
->sqo_task
= get_task_struct(current
);
9766 * This is just grabbed for accounting purposes. When a process exits,
9767 * the mm is exited and dropped before the files, hence we need to hang
9768 * on to this mm purely for the purposes of being able to unaccount
9769 * memory (locked/pinned vm). It's not used for anything else.
9771 mmgrab(current
->mm
);
9772 ctx
->mm_account
= current
->mm
;
9774 #ifdef CONFIG_BLK_CGROUP
9776 * The sq thread will belong to the original cgroup it was inited in.
9777 * If the cgroup goes offline (e.g. disabling the io controller), then
9778 * issued bios will be associated with the closest cgroup later in the
9782 ctx
->sqo_blkcg_css
= blkcg_css();
9783 ret
= css_tryget_online(ctx
->sqo_blkcg_css
);
9786 /* don't init against a dying cgroup, have the user try again */
9787 ctx
->sqo_blkcg_css
= NULL
;
9794 * Account memory _before_ installing the file descriptor. Once
9795 * the descriptor is installed, it can get closed at any time. Also
9796 * do this before hitting the general error path, as ring freeing
9797 * will un-account as well.
9799 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
9801 ctx
->limit_mem
= limit_mem
;
9803 ret
= io_allocate_scq_urings(ctx
, p
);
9807 ret
= io_sq_offload_create(ctx
, p
);
9811 if (!(p
->flags
& IORING_SETUP_R_DISABLED
))
9812 io_sq_offload_start(ctx
);
9814 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9815 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9816 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9817 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9818 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9819 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9820 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9821 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9823 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9824 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9825 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9826 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9827 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9828 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9829 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9830 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9832 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9833 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9834 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9835 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9836 IORING_FEAT_EXT_ARG
;
9838 if (copy_to_user(params
, p
, sizeof(*p
))) {
9843 file
= io_uring_get_file(ctx
);
9845 ret
= PTR_ERR(file
);
9850 * Install ring fd as the very last thing, so we don't risk someone
9851 * having closed it before we finish setup
9853 ret
= io_uring_install_fd(ctx
, file
);
9855 io_disable_sqo_submit(ctx
);
9856 /* fput will clean it up */
9861 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9864 io_disable_sqo_submit(ctx
);
9865 io_ring_ctx_wait_and_kill(ctx
);
9870 * Sets up an aio uring context, and returns the fd. Applications asks for a
9871 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9872 * params structure passed in.
9874 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9876 struct io_uring_params p
;
9879 if (copy_from_user(&p
, params
, sizeof(p
)))
9881 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9886 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9887 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9888 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9889 IORING_SETUP_R_DISABLED
))
9892 return io_uring_create(entries
, &p
, params
);
9895 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9896 struct io_uring_params __user
*, params
)
9898 return io_uring_setup(entries
, params
);
9901 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9903 struct io_uring_probe
*p
;
9907 size
= struct_size(p
, ops
, nr_args
);
9908 if (size
== SIZE_MAX
)
9910 p
= kzalloc(size
, GFP_KERNEL
);
9915 if (copy_from_user(p
, arg
, size
))
9918 if (memchr_inv(p
, 0, size
))
9921 p
->last_op
= IORING_OP_LAST
- 1;
9922 if (nr_args
> IORING_OP_LAST
)
9923 nr_args
= IORING_OP_LAST
;
9925 for (i
= 0; i
< nr_args
; i
++) {
9927 if (!io_op_defs
[i
].not_supported
)
9928 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9933 if (copy_to_user(arg
, p
, size
))
9940 static int io_register_personality(struct io_ring_ctx
*ctx
)
9942 struct io_identity
*iod
;
9946 iod
= kmalloc(sizeof(*iod
), GFP_KERNEL
);
9950 io_init_identity(iod
);
9951 iod
->creds
= get_current_cred();
9953 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)iod
,
9954 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
9956 put_cred(iod
->creds
);
9963 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9964 unsigned int nr_args
)
9966 struct io_uring_restriction
*res
;
9970 /* Restrictions allowed only if rings started disabled */
9971 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9974 /* We allow only a single restrictions registration */
9975 if (ctx
->restrictions
.registered
)
9978 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9981 size
= array_size(nr_args
, sizeof(*res
));
9982 if (size
== SIZE_MAX
)
9985 res
= memdup_user(arg
, size
);
9987 return PTR_ERR(res
);
9991 for (i
= 0; i
< nr_args
; i
++) {
9992 switch (res
[i
].opcode
) {
9993 case IORING_RESTRICTION_REGISTER_OP
:
9994 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9999 __set_bit(res
[i
].register_op
,
10000 ctx
->restrictions
.register_op
);
10002 case IORING_RESTRICTION_SQE_OP
:
10003 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
10008 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
10010 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
10011 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
10013 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
10014 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
10023 /* Reset all restrictions if an error happened */
10025 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
10027 ctx
->restrictions
.registered
= true;
10033 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10035 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10038 if (ctx
->restrictions
.registered
)
10039 ctx
->restricted
= 1;
10041 io_sq_offload_start(ctx
);
10045 static bool io_register_op_must_quiesce(int op
)
10048 case IORING_UNREGISTER_FILES
:
10049 case IORING_REGISTER_FILES_UPDATE
:
10050 case IORING_REGISTER_PROBE
:
10051 case IORING_REGISTER_PERSONALITY
:
10052 case IORING_UNREGISTER_PERSONALITY
:
10059 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10060 void __user
*arg
, unsigned nr_args
)
10061 __releases(ctx
->uring_lock
)
10062 __acquires(ctx
->uring_lock
)
10067 * We're inside the ring mutex, if the ref is already dying, then
10068 * someone else killed the ctx or is already going through
10069 * io_uring_register().
10071 if (percpu_ref_is_dying(&ctx
->refs
))
10074 if (io_register_op_must_quiesce(opcode
)) {
10075 percpu_ref_kill(&ctx
->refs
);
10078 * Drop uring mutex before waiting for references to exit. If
10079 * another thread is currently inside io_uring_enter() it might
10080 * need to grab the uring_lock to make progress. If we hold it
10081 * here across the drain wait, then we can deadlock. It's safe
10082 * to drop the mutex here, since no new references will come in
10083 * after we've killed the percpu ref.
10085 mutex_unlock(&ctx
->uring_lock
);
10087 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10090 ret
= io_run_task_work_sig();
10095 mutex_lock(&ctx
->uring_lock
);
10098 percpu_ref_resurrect(&ctx
->refs
);
10103 if (ctx
->restricted
) {
10104 if (opcode
>= IORING_REGISTER_LAST
) {
10109 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
10116 case IORING_REGISTER_BUFFERS
:
10117 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
10119 case IORING_UNREGISTER_BUFFERS
:
10121 if (arg
|| nr_args
)
10123 ret
= io_sqe_buffer_unregister(ctx
);
10125 case IORING_REGISTER_FILES
:
10126 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
10128 case IORING_UNREGISTER_FILES
:
10130 if (arg
|| nr_args
)
10132 ret
= io_sqe_files_unregister(ctx
);
10134 case IORING_REGISTER_FILES_UPDATE
:
10135 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
10137 case IORING_REGISTER_EVENTFD
:
10138 case IORING_REGISTER_EVENTFD_ASYNC
:
10142 ret
= io_eventfd_register(ctx
, arg
);
10145 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10146 ctx
->eventfd_async
= 1;
10148 ctx
->eventfd_async
= 0;
10150 case IORING_UNREGISTER_EVENTFD
:
10152 if (arg
|| nr_args
)
10154 ret
= io_eventfd_unregister(ctx
);
10156 case IORING_REGISTER_PROBE
:
10158 if (!arg
|| nr_args
> 256)
10160 ret
= io_probe(ctx
, arg
, nr_args
);
10162 case IORING_REGISTER_PERSONALITY
:
10164 if (arg
|| nr_args
)
10166 ret
= io_register_personality(ctx
);
10168 case IORING_UNREGISTER_PERSONALITY
:
10172 ret
= io_unregister_personality(ctx
, nr_args
);
10174 case IORING_REGISTER_ENABLE_RINGS
:
10176 if (arg
|| nr_args
)
10178 ret
= io_register_enable_rings(ctx
);
10180 case IORING_REGISTER_RESTRICTIONS
:
10181 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10189 if (io_register_op_must_quiesce(opcode
)) {
10190 /* bring the ctx back to life */
10191 percpu_ref_reinit(&ctx
->refs
);
10193 reinit_completion(&ctx
->ref_comp
);
10198 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10199 void __user
*, arg
, unsigned int, nr_args
)
10201 struct io_ring_ctx
*ctx
;
10210 if (f
.file
->f_op
!= &io_uring_fops
)
10213 ctx
= f
.file
->private_data
;
10215 mutex_lock(&ctx
->uring_lock
);
10216 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10217 mutex_unlock(&ctx
->uring_lock
);
10218 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10219 ctx
->cq_ev_fd
!= NULL
, ret
);
10225 static int __init
io_uring_init(void)
10227 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10228 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10229 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10232 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10233 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10234 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10235 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10236 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10237 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10238 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10239 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10240 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10241 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10242 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10243 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10244 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10245 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10246 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10247 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10248 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10249 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10250 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10251 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10252 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10253 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10254 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10255 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10256 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10257 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10258 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10259 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10260 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10261 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10262 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10264 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10265 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
10266 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
10269 __initcall(io_uring_init
);