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
;
201 struct list_head list
;
208 struct fixed_rsrc_table
{
212 struct fixed_rsrc_ref_node
{
213 struct percpu_ref refs
;
214 struct list_head node
;
215 struct list_head rsrc_list
;
216 struct fixed_rsrc_data
*rsrc_data
;
217 void (*rsrc_put
)(struct io_ring_ctx
*ctx
,
218 struct io_rsrc_put
*prsrc
);
219 struct llist_node llist
;
223 struct fixed_rsrc_data
{
224 struct fixed_rsrc_table
*table
;
225 struct io_ring_ctx
*ctx
;
227 struct fixed_rsrc_ref_node
*node
;
228 struct percpu_ref refs
;
229 struct completion done
;
233 struct list_head list
;
239 struct io_restriction
{
240 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
241 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
242 u8 sqe_flags_allowed
;
243 u8 sqe_flags_required
;
251 /* ctx's that are using this sqd */
252 struct list_head ctx_list
;
253 struct list_head ctx_new_list
;
254 struct mutex ctx_lock
;
256 struct task_struct
*thread
;
257 struct wait_queue_head wait
;
259 unsigned sq_thread_idle
;
264 struct percpu_ref refs
;
265 } ____cacheline_aligned_in_smp
;
269 unsigned int compat
: 1;
270 unsigned int limit_mem
: 1;
271 unsigned int cq_overflow_flushed
: 1;
272 unsigned int drain_next
: 1;
273 unsigned int eventfd_async
: 1;
274 unsigned int restricted
: 1;
275 unsigned int sqo_dead
: 1;
278 * Ring buffer of indices into array of io_uring_sqe, which is
279 * mmapped by the application using the IORING_OFF_SQES offset.
281 * This indirection could e.g. be used to assign fixed
282 * io_uring_sqe entries to operations and only submit them to
283 * the queue when needed.
285 * The kernel modifies neither the indices array nor the entries
289 unsigned cached_sq_head
;
292 unsigned sq_thread_idle
;
293 unsigned cached_sq_dropped
;
294 unsigned cached_cq_overflow
;
295 unsigned long sq_check_overflow
;
297 struct list_head defer_list
;
298 struct list_head timeout_list
;
299 struct list_head cq_overflow_list
;
301 struct io_uring_sqe
*sq_sqes
;
302 } ____cacheline_aligned_in_smp
;
304 struct io_rings
*rings
;
310 * For SQPOLL usage - we hold a reference to the parent task, so we
311 * have access to the ->files
313 struct task_struct
*sqo_task
;
315 /* Only used for accounting purposes */
316 struct mm_struct
*mm_account
;
318 #ifdef CONFIG_BLK_CGROUP
319 struct cgroup_subsys_state
*sqo_blkcg_css
;
322 struct io_sq_data
*sq_data
; /* if using sq thread polling */
324 struct wait_queue_head sqo_sq_wait
;
325 struct list_head sqd_list
;
328 * If used, fixed file set. Writers must ensure that ->refs is dead,
329 * readers must ensure that ->refs is alive as long as the file* is
330 * used. Only updated through io_uring_register(2).
332 struct fixed_rsrc_data
*file_data
;
333 unsigned nr_user_files
;
335 /* if used, fixed mapped user buffers */
336 unsigned nr_user_bufs
;
337 struct io_mapped_ubuf
*user_bufs
;
339 struct user_struct
*user
;
341 const struct cred
*creds
;
345 unsigned int sessionid
;
348 struct completion ref_comp
;
349 struct completion sq_thread_comp
;
351 /* if all else fails... */
352 struct io_kiocb
*fallback_req
;
354 #if defined(CONFIG_UNIX)
355 struct socket
*ring_sock
;
358 struct idr io_buffer_idr
;
360 struct idr personality_idr
;
363 unsigned cached_cq_tail
;
366 atomic_t cq_timeouts
;
367 unsigned cq_last_tm_flush
;
368 unsigned long cq_check_overflow
;
369 struct wait_queue_head cq_wait
;
370 struct fasync_struct
*cq_fasync
;
371 struct eventfd_ctx
*cq_ev_fd
;
372 } ____cacheline_aligned_in_smp
;
375 struct mutex uring_lock
;
376 wait_queue_head_t wait
;
377 } ____cacheline_aligned_in_smp
;
380 spinlock_t completion_lock
;
383 * ->iopoll_list is protected by the ctx->uring_lock for
384 * io_uring instances that don't use IORING_SETUP_SQPOLL.
385 * For SQPOLL, only the single threaded io_sq_thread() will
386 * manipulate the list, hence no extra locking is needed there.
388 struct list_head iopoll_list
;
389 struct hlist_head
*cancel_hash
;
390 unsigned cancel_hash_bits
;
391 bool poll_multi_file
;
393 spinlock_t inflight_lock
;
394 struct list_head inflight_list
;
395 } ____cacheline_aligned_in_smp
;
397 struct delayed_work rsrc_put_work
;
398 struct llist_head rsrc_put_llist
;
399 struct list_head rsrc_ref_list
;
400 spinlock_t rsrc_ref_lock
;
402 struct work_struct exit_work
;
403 struct io_restriction restrictions
;
407 * First field must be the file pointer in all the
408 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
410 struct io_poll_iocb
{
412 struct wait_queue_head
*head
;
416 struct wait_queue_entry wait
;
419 struct io_poll_remove
{
429 struct io_timeout_data
{
430 struct io_kiocb
*req
;
431 struct hrtimer timer
;
432 struct timespec64 ts
;
433 enum hrtimer_mode mode
;
438 struct sockaddr __user
*addr
;
439 int __user
*addr_len
;
441 unsigned long nofile
;
461 struct list_head list
;
462 /* head of the link, used by linked timeouts only */
463 struct io_kiocb
*head
;
466 struct io_timeout_rem
{
471 struct timespec64 ts
;
476 /* NOTE: kiocb has the file as the first member, so don't do it here */
484 struct sockaddr __user
*addr
;
491 struct user_msghdr __user
*umsg
;
497 struct io_buffer
*kbuf
;
503 struct filename
*filename
;
505 unsigned long nofile
;
508 struct io_rsrc_update
{
534 struct epoll_event event
;
538 struct file
*file_out
;
539 struct file
*file_in
;
546 struct io_provide_buf
{
560 const char __user
*filename
;
561 struct statx __user
*buffer
;
573 struct filename
*oldpath
;
574 struct filename
*newpath
;
582 struct filename
*filename
;
585 struct io_completion
{
587 struct list_head list
;
591 struct io_async_connect
{
592 struct sockaddr_storage address
;
595 struct io_async_msghdr
{
596 struct iovec fast_iov
[UIO_FASTIOV
];
598 struct sockaddr __user
*uaddr
;
600 struct sockaddr_storage addr
;
604 struct iovec fast_iov
[UIO_FASTIOV
];
605 const struct iovec
*free_iovec
;
606 struct iov_iter iter
;
608 struct wait_page_queue wpq
;
612 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
613 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
614 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
615 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
616 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
617 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
623 REQ_F_LINK_TIMEOUT_BIT
,
625 REQ_F_NEED_CLEANUP_BIT
,
627 REQ_F_BUFFER_SELECTED_BIT
,
628 REQ_F_NO_FILE_TABLE_BIT
,
629 REQ_F_WORK_INITIALIZED_BIT
,
630 REQ_F_LTIMEOUT_ACTIVE_BIT
,
631 REQ_F_COMPLETE_INLINE_BIT
,
633 /* not a real bit, just to check we're not overflowing the space */
639 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
640 /* drain existing IO first */
641 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
643 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
644 /* doesn't sever on completion < 0 */
645 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
647 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
648 /* IOSQE_BUFFER_SELECT */
649 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
651 /* fail rest of links */
652 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
653 /* on inflight list */
654 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
655 /* read/write uses file position */
656 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
657 /* must not punt to workers */
658 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
659 /* has or had linked timeout */
660 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
662 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
664 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
665 /* already went through poll handler */
666 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
667 /* buffer already selected */
668 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
669 /* doesn't need file table for this request */
670 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
671 /* io_wq_work is initialized */
672 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
673 /* linked timeout is active, i.e. prepared by link's head */
674 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
675 /* completion is deferred through io_comp_state */
676 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
680 struct io_poll_iocb poll
;
681 struct io_poll_iocb
*double_poll
;
685 * NOTE! Each of the iocb union members has the file pointer
686 * as the first entry in their struct definition. So you can
687 * access the file pointer through any of the sub-structs,
688 * or directly as just 'ki_filp' in this struct.
694 struct io_poll_iocb poll
;
695 struct io_poll_remove poll_remove
;
696 struct io_accept accept
;
698 struct io_cancel cancel
;
699 struct io_timeout timeout
;
700 struct io_timeout_rem timeout_rem
;
701 struct io_connect connect
;
702 struct io_sr_msg sr_msg
;
704 struct io_close close
;
705 struct io_rsrc_update rsrc_update
;
706 struct io_fadvise fadvise
;
707 struct io_madvise madvise
;
708 struct io_epoll epoll
;
709 struct io_splice splice
;
710 struct io_provide_buf pbuf
;
711 struct io_statx statx
;
712 struct io_shutdown shutdown
;
713 struct io_rename rename
;
714 struct io_unlink unlink
;
715 /* use only after cleaning per-op data, see io_clean_op() */
716 struct io_completion
compl;
719 /* opcode allocated if it needs to store data for async defer */
722 /* polled IO has completed */
728 struct io_ring_ctx
*ctx
;
731 struct task_struct
*task
;
734 struct io_kiocb
*link
;
735 struct percpu_ref
*fixed_rsrc_refs
;
738 * 1. used with ctx->iopoll_list with reads/writes
739 * 2. to track reqs with ->files (see io_op_def::file_table)
741 struct list_head inflight_entry
;
742 struct callback_head task_work
;
743 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
744 struct hlist_node hash_node
;
745 struct async_poll
*apoll
;
746 struct io_wq_work work
;
749 struct io_defer_entry
{
750 struct list_head list
;
751 struct io_kiocb
*req
;
755 #define IO_IOPOLL_BATCH 8
757 struct io_comp_state
{
759 struct list_head list
;
760 struct io_ring_ctx
*ctx
;
763 struct io_submit_state
{
764 struct blk_plug plug
;
767 * io_kiocb alloc cache
769 void *reqs
[IO_IOPOLL_BATCH
];
770 unsigned int free_reqs
;
775 * Batch completion logic
777 struct io_comp_state comp
;
780 * File reference cache
784 unsigned int file_refs
;
785 unsigned int ios_left
;
789 /* needs req->file assigned */
790 unsigned needs_file
: 1;
791 /* hash wq insertion if file is a regular file */
792 unsigned hash_reg_file
: 1;
793 /* unbound wq insertion if file is a non-regular file */
794 unsigned unbound_nonreg_file
: 1;
795 /* opcode is not supported by this kernel */
796 unsigned not_supported
: 1;
797 /* set if opcode supports polled "wait" */
799 unsigned pollout
: 1;
800 /* op supports buffer selection */
801 unsigned buffer_select
: 1;
802 /* must always have async data allocated */
803 unsigned needs_async_data
: 1;
804 /* should block plug */
806 /* size of async data needed, if any */
807 unsigned short async_size
;
811 static const struct io_op_def io_op_defs
[] = {
812 [IORING_OP_NOP
] = {},
813 [IORING_OP_READV
] = {
815 .unbound_nonreg_file
= 1,
818 .needs_async_data
= 1,
820 .async_size
= sizeof(struct io_async_rw
),
821 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
823 [IORING_OP_WRITEV
] = {
826 .unbound_nonreg_file
= 1,
828 .needs_async_data
= 1,
830 .async_size
= sizeof(struct io_async_rw
),
831 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
834 [IORING_OP_FSYNC
] = {
836 .work_flags
= IO_WQ_WORK_BLKCG
,
838 [IORING_OP_READ_FIXED
] = {
840 .unbound_nonreg_file
= 1,
843 .async_size
= sizeof(struct io_async_rw
),
844 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
846 [IORING_OP_WRITE_FIXED
] = {
849 .unbound_nonreg_file
= 1,
852 .async_size
= sizeof(struct io_async_rw
),
853 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
|
856 [IORING_OP_POLL_ADD
] = {
858 .unbound_nonreg_file
= 1,
860 [IORING_OP_POLL_REMOVE
] = {},
861 [IORING_OP_SYNC_FILE_RANGE
] = {
863 .work_flags
= IO_WQ_WORK_BLKCG
,
865 [IORING_OP_SENDMSG
] = {
867 .unbound_nonreg_file
= 1,
869 .needs_async_data
= 1,
870 .async_size
= sizeof(struct io_async_msghdr
),
871 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
873 [IORING_OP_RECVMSG
] = {
875 .unbound_nonreg_file
= 1,
878 .needs_async_data
= 1,
879 .async_size
= sizeof(struct io_async_msghdr
),
880 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
882 [IORING_OP_TIMEOUT
] = {
883 .needs_async_data
= 1,
884 .async_size
= sizeof(struct io_timeout_data
),
885 .work_flags
= IO_WQ_WORK_MM
,
887 [IORING_OP_TIMEOUT_REMOVE
] = {
888 /* used by timeout updates' prep() */
889 .work_flags
= IO_WQ_WORK_MM
,
891 [IORING_OP_ACCEPT
] = {
893 .unbound_nonreg_file
= 1,
895 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
,
897 [IORING_OP_ASYNC_CANCEL
] = {},
898 [IORING_OP_LINK_TIMEOUT
] = {
899 .needs_async_data
= 1,
900 .async_size
= sizeof(struct io_timeout_data
),
901 .work_flags
= IO_WQ_WORK_MM
,
903 [IORING_OP_CONNECT
] = {
905 .unbound_nonreg_file
= 1,
907 .needs_async_data
= 1,
908 .async_size
= sizeof(struct io_async_connect
),
909 .work_flags
= IO_WQ_WORK_MM
,
911 [IORING_OP_FALLOCATE
] = {
913 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
915 [IORING_OP_OPENAT
] = {
916 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
|
917 IO_WQ_WORK_FS
| IO_WQ_WORK_MM
,
919 [IORING_OP_CLOSE
] = {
920 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
,
922 [IORING_OP_FILES_UPDATE
] = {
923 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
,
925 [IORING_OP_STATX
] = {
926 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
|
927 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
931 .unbound_nonreg_file
= 1,
935 .async_size
= sizeof(struct io_async_rw
),
936 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
938 [IORING_OP_WRITE
] = {
940 .unbound_nonreg_file
= 1,
943 .async_size
= sizeof(struct io_async_rw
),
944 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
947 [IORING_OP_FADVISE
] = {
949 .work_flags
= IO_WQ_WORK_BLKCG
,
951 [IORING_OP_MADVISE
] = {
952 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
956 .unbound_nonreg_file
= 1,
958 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
962 .unbound_nonreg_file
= 1,
965 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
967 [IORING_OP_OPENAT2
] = {
968 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_FS
|
969 IO_WQ_WORK_BLKCG
| IO_WQ_WORK_MM
,
971 [IORING_OP_EPOLL_CTL
] = {
972 .unbound_nonreg_file
= 1,
973 .work_flags
= IO_WQ_WORK_FILES
,
975 [IORING_OP_SPLICE
] = {
978 .unbound_nonreg_file
= 1,
979 .work_flags
= IO_WQ_WORK_BLKCG
,
981 [IORING_OP_PROVIDE_BUFFERS
] = {},
982 [IORING_OP_REMOVE_BUFFERS
] = {},
986 .unbound_nonreg_file
= 1,
988 [IORING_OP_SHUTDOWN
] = {
991 [IORING_OP_RENAMEAT
] = {
992 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
993 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
995 [IORING_OP_UNLINKAT
] = {
996 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
|
997 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
1001 enum io_mem_account
{
1006 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1007 struct task_struct
*task
,
1008 struct files_struct
*files
);
1009 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node
*ref_node
);
1010 static struct fixed_rsrc_ref_node
*alloc_fixed_rsrc_ref_node(
1011 struct io_ring_ctx
*ctx
);
1012 static void init_fixed_file_ref_node(struct io_ring_ctx
*ctx
,
1013 struct fixed_rsrc_ref_node
*ref_node
);
1015 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
1016 struct io_comp_state
*cs
);
1017 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
1018 static void io_put_req(struct io_kiocb
*req
);
1019 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1020 static void io_double_put_req(struct io_kiocb
*req
);
1021 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1022 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
1023 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1024 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
1025 struct io_uring_rsrc_update
*ip
,
1027 static void __io_clean_op(struct io_kiocb
*req
);
1028 static struct file
*io_file_get(struct io_submit_state
*state
,
1029 struct io_kiocb
*req
, int fd
, bool fixed
);
1030 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
);
1031 static void io_rsrc_put_work(struct work_struct
*work
);
1033 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
1034 struct iov_iter
*iter
, bool needs_lock
);
1035 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
1036 const struct iovec
*fast_iov
,
1037 struct iov_iter
*iter
, bool force
);
1038 static void io_req_task_queue(struct io_kiocb
*req
);
1040 static struct kmem_cache
*req_cachep
;
1042 static const struct file_operations io_uring_fops
;
1044 struct sock
*io_uring_get_socket(struct file
*file
)
1046 #if defined(CONFIG_UNIX)
1047 if (file
->f_op
== &io_uring_fops
) {
1048 struct io_ring_ctx
*ctx
= file
->private_data
;
1050 return ctx
->ring_sock
->sk
;
1055 EXPORT_SYMBOL(io_uring_get_socket
);
1057 #define io_for_each_link(pos, head) \
1058 for (pos = (head); pos; pos = pos->link)
1060 static inline void io_clean_op(struct io_kiocb
*req
)
1062 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
))
1066 static inline void io_set_resource_node(struct io_kiocb
*req
)
1068 struct io_ring_ctx
*ctx
= req
->ctx
;
1070 if (!req
->fixed_rsrc_refs
) {
1071 req
->fixed_rsrc_refs
= &ctx
->file_data
->node
->refs
;
1072 percpu_ref_get(req
->fixed_rsrc_refs
);
1076 static bool io_match_task(struct io_kiocb
*head
,
1077 struct task_struct
*task
,
1078 struct files_struct
*files
)
1080 struct io_kiocb
*req
;
1082 if (task
&& head
->task
!= task
) {
1083 /* in terms of cancelation, always match if req task is dead */
1084 if (head
->task
->flags
& PF_EXITING
)
1091 io_for_each_link(req
, head
) {
1092 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1094 if (req
->file
&& req
->file
->f_op
== &io_uring_fops
)
1096 if ((req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1097 req
->work
.identity
->files
== files
)
1103 static void io_sq_thread_drop_mm_files(void)
1105 struct files_struct
*files
= current
->files
;
1106 struct mm_struct
*mm
= current
->mm
;
1109 kthread_unuse_mm(mm
);
1114 struct nsproxy
*nsproxy
= current
->nsproxy
;
1117 current
->files
= NULL
;
1118 current
->nsproxy
= NULL
;
1119 task_unlock(current
);
1120 put_files_struct(files
);
1121 put_nsproxy(nsproxy
);
1125 static int __io_sq_thread_acquire_files(struct io_ring_ctx
*ctx
)
1127 if (current
->flags
& PF_EXITING
)
1130 if (!current
->files
) {
1131 struct files_struct
*files
;
1132 struct nsproxy
*nsproxy
;
1134 task_lock(ctx
->sqo_task
);
1135 files
= ctx
->sqo_task
->files
;
1137 task_unlock(ctx
->sqo_task
);
1140 atomic_inc(&files
->count
);
1141 get_nsproxy(ctx
->sqo_task
->nsproxy
);
1142 nsproxy
= ctx
->sqo_task
->nsproxy
;
1143 task_unlock(ctx
->sqo_task
);
1146 current
->files
= files
;
1147 current
->nsproxy
= nsproxy
;
1148 task_unlock(current
);
1153 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1155 struct mm_struct
*mm
;
1157 if (current
->flags
& PF_EXITING
)
1162 /* Should never happen */
1163 if (unlikely(!(ctx
->flags
& IORING_SETUP_SQPOLL
)))
1166 task_lock(ctx
->sqo_task
);
1167 mm
= ctx
->sqo_task
->mm
;
1168 if (unlikely(!mm
|| !mmget_not_zero(mm
)))
1170 task_unlock(ctx
->sqo_task
);
1180 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx
*ctx
,
1181 struct io_kiocb
*req
)
1183 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1186 if (def
->work_flags
& IO_WQ_WORK_MM
) {
1187 ret
= __io_sq_thread_acquire_mm(ctx
);
1192 if (def
->needs_file
|| (def
->work_flags
& IO_WQ_WORK_FILES
)) {
1193 ret
= __io_sq_thread_acquire_files(ctx
);
1201 static void io_sq_thread_associate_blkcg(struct io_ring_ctx
*ctx
,
1202 struct cgroup_subsys_state
**cur_css
)
1205 #ifdef CONFIG_BLK_CGROUP
1206 /* puts the old one when swapping */
1207 if (*cur_css
!= ctx
->sqo_blkcg_css
) {
1208 kthread_associate_blkcg(ctx
->sqo_blkcg_css
);
1209 *cur_css
= ctx
->sqo_blkcg_css
;
1214 static void io_sq_thread_unassociate_blkcg(void)
1216 #ifdef CONFIG_BLK_CGROUP
1217 kthread_associate_blkcg(NULL
);
1221 static inline void req_set_fail_links(struct io_kiocb
*req
)
1223 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1224 req
->flags
|= REQ_F_FAIL_LINK
;
1228 * None of these are dereferenced, they are simply used to check if any of
1229 * them have changed. If we're under current and check they are still the
1230 * same, we're fine to grab references to them for actual out-of-line use.
1232 static void io_init_identity(struct io_identity
*id
)
1234 id
->files
= current
->files
;
1235 id
->mm
= current
->mm
;
1236 #ifdef CONFIG_BLK_CGROUP
1238 id
->blkcg_css
= blkcg_css();
1241 id
->creds
= current_cred();
1242 id
->nsproxy
= current
->nsproxy
;
1243 id
->fs
= current
->fs
;
1244 id
->fsize
= rlimit(RLIMIT_FSIZE
);
1246 id
->loginuid
= current
->loginuid
;
1247 id
->sessionid
= current
->sessionid
;
1249 refcount_set(&id
->count
, 1);
1252 static inline void __io_req_init_async(struct io_kiocb
*req
)
1254 memset(&req
->work
, 0, sizeof(req
->work
));
1255 req
->flags
|= REQ_F_WORK_INITIALIZED
;
1259 * Note: must call io_req_init_async() for the first time you
1260 * touch any members of io_wq_work.
1262 static inline void io_req_init_async(struct io_kiocb
*req
)
1264 struct io_uring_task
*tctx
= current
->io_uring
;
1266 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
1269 __io_req_init_async(req
);
1271 /* Grab a ref if this isn't our static identity */
1272 req
->work
.identity
= tctx
->identity
;
1273 if (tctx
->identity
!= &tctx
->__identity
)
1274 refcount_inc(&req
->work
.identity
->count
);
1277 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1279 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1281 complete(&ctx
->ref_comp
);
1284 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1286 return !req
->timeout
.off
;
1289 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1291 struct io_ring_ctx
*ctx
;
1294 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1298 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
1299 if (!ctx
->fallback_req
)
1303 * Use 5 bits less than the max cq entries, that should give us around
1304 * 32 entries per hash list if totally full and uniformly spread.
1306 hash_bits
= ilog2(p
->cq_entries
);
1310 ctx
->cancel_hash_bits
= hash_bits
;
1311 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1313 if (!ctx
->cancel_hash
)
1315 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1317 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1318 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1321 ctx
->flags
= p
->flags
;
1322 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1323 INIT_LIST_HEAD(&ctx
->sqd_list
);
1324 init_waitqueue_head(&ctx
->cq_wait
);
1325 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1326 init_completion(&ctx
->ref_comp
);
1327 init_completion(&ctx
->sq_thread_comp
);
1328 idr_init(&ctx
->io_buffer_idr
);
1329 idr_init(&ctx
->personality_idr
);
1330 mutex_init(&ctx
->uring_lock
);
1331 init_waitqueue_head(&ctx
->wait
);
1332 spin_lock_init(&ctx
->completion_lock
);
1333 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1334 INIT_LIST_HEAD(&ctx
->defer_list
);
1335 INIT_LIST_HEAD(&ctx
->timeout_list
);
1336 spin_lock_init(&ctx
->inflight_lock
);
1337 INIT_LIST_HEAD(&ctx
->inflight_list
);
1338 spin_lock_init(&ctx
->rsrc_ref_lock
);
1339 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1340 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1341 init_llist_head(&ctx
->rsrc_put_llist
);
1344 if (ctx
->fallback_req
)
1345 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1346 kfree(ctx
->cancel_hash
);
1351 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1353 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1354 struct io_ring_ctx
*ctx
= req
->ctx
;
1356 return seq
!= ctx
->cached_cq_tail
1357 + READ_ONCE(ctx
->cached_cq_overflow
);
1363 static void io_put_identity(struct io_uring_task
*tctx
, struct io_kiocb
*req
)
1365 if (req
->work
.identity
== &tctx
->__identity
)
1367 if (refcount_dec_and_test(&req
->work
.identity
->count
))
1368 kfree(req
->work
.identity
);
1371 static void io_req_clean_work(struct io_kiocb
*req
)
1373 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1376 if (req
->work
.flags
& IO_WQ_WORK_MM
)
1377 mmdrop(req
->work
.identity
->mm
);
1378 #ifdef CONFIG_BLK_CGROUP
1379 if (req
->work
.flags
& IO_WQ_WORK_BLKCG
)
1380 css_put(req
->work
.identity
->blkcg_css
);
1382 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1383 put_cred(req
->work
.identity
->creds
);
1384 if (req
->work
.flags
& IO_WQ_WORK_FS
) {
1385 struct fs_struct
*fs
= req
->work
.identity
->fs
;
1387 spin_lock(&req
->work
.identity
->fs
->lock
);
1390 spin_unlock(&req
->work
.identity
->fs
->lock
);
1394 if (req
->work
.flags
& IO_WQ_WORK_FILES
) {
1395 put_files_struct(req
->work
.identity
->files
);
1396 put_nsproxy(req
->work
.identity
->nsproxy
);
1398 if (req
->flags
& REQ_F_INFLIGHT
) {
1399 struct io_ring_ctx
*ctx
= req
->ctx
;
1400 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1401 unsigned long flags
;
1403 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1404 list_del(&req
->inflight_entry
);
1405 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1406 req
->flags
&= ~REQ_F_INFLIGHT
;
1407 if (atomic_read(&tctx
->in_idle
))
1408 wake_up(&tctx
->wait
);
1411 req
->flags
&= ~REQ_F_WORK_INITIALIZED
;
1412 req
->work
.flags
&= ~(IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FS
|
1413 IO_WQ_WORK_CREDS
| IO_WQ_WORK_FILES
);
1414 io_put_identity(req
->task
->io_uring
, req
);
1418 * Create a private copy of io_identity, since some fields don't match
1419 * the current context.
1421 static bool io_identity_cow(struct io_kiocb
*req
)
1423 struct io_uring_task
*tctx
= current
->io_uring
;
1424 const struct cred
*creds
= NULL
;
1425 struct io_identity
*id
;
1427 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1428 creds
= req
->work
.identity
->creds
;
1430 id
= kmemdup(req
->work
.identity
, sizeof(*id
), GFP_KERNEL
);
1431 if (unlikely(!id
)) {
1432 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1437 * We can safely just re-init the creds we copied Either the field
1438 * matches the current one, or we haven't grabbed it yet. The only
1439 * exception is ->creds, through registered personalities, so handle
1440 * that one separately.
1442 io_init_identity(id
);
1446 /* add one for this request */
1447 refcount_inc(&id
->count
);
1449 /* drop tctx and req identity references, if needed */
1450 if (tctx
->identity
!= &tctx
->__identity
&&
1451 refcount_dec_and_test(&tctx
->identity
->count
))
1452 kfree(tctx
->identity
);
1453 if (req
->work
.identity
!= &tctx
->__identity
&&
1454 refcount_dec_and_test(&req
->work
.identity
->count
))
1455 kfree(req
->work
.identity
);
1457 req
->work
.identity
= id
;
1458 tctx
->identity
= id
;
1462 static void io_req_track_inflight(struct io_kiocb
*req
)
1464 struct io_ring_ctx
*ctx
= req
->ctx
;
1466 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1467 io_req_init_async(req
);
1468 req
->flags
|= REQ_F_INFLIGHT
;
1470 spin_lock_irq(&ctx
->inflight_lock
);
1471 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1472 spin_unlock_irq(&ctx
->inflight_lock
);
1476 static bool io_grab_identity(struct io_kiocb
*req
)
1478 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1479 struct io_identity
*id
= req
->work
.identity
;
1481 if (def
->work_flags
& IO_WQ_WORK_FSIZE
) {
1482 if (id
->fsize
!= rlimit(RLIMIT_FSIZE
))
1484 req
->work
.flags
|= IO_WQ_WORK_FSIZE
;
1486 #ifdef CONFIG_BLK_CGROUP
1487 if (!(req
->work
.flags
& IO_WQ_WORK_BLKCG
) &&
1488 (def
->work_flags
& IO_WQ_WORK_BLKCG
)) {
1490 if (id
->blkcg_css
!= blkcg_css()) {
1495 * This should be rare, either the cgroup is dying or the task
1496 * is moving cgroups. Just punt to root for the handful of ios.
1498 if (css_tryget_online(id
->blkcg_css
))
1499 req
->work
.flags
|= IO_WQ_WORK_BLKCG
;
1503 if (!(req
->work
.flags
& IO_WQ_WORK_CREDS
)) {
1504 if (id
->creds
!= current_cred())
1506 get_cred(id
->creds
);
1507 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
1510 if (!uid_eq(current
->loginuid
, id
->loginuid
) ||
1511 current
->sessionid
!= id
->sessionid
)
1514 if (!(req
->work
.flags
& IO_WQ_WORK_FS
) &&
1515 (def
->work_flags
& IO_WQ_WORK_FS
)) {
1516 if (current
->fs
!= id
->fs
)
1518 spin_lock(&id
->fs
->lock
);
1519 if (!id
->fs
->in_exec
) {
1521 req
->work
.flags
|= IO_WQ_WORK_FS
;
1523 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1525 spin_unlock(¤t
->fs
->lock
);
1527 if (!(req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1528 (def
->work_flags
& IO_WQ_WORK_FILES
) &&
1529 !(req
->flags
& REQ_F_NO_FILE_TABLE
)) {
1530 if (id
->files
!= current
->files
||
1531 id
->nsproxy
!= current
->nsproxy
)
1533 atomic_inc(&id
->files
->count
);
1534 get_nsproxy(id
->nsproxy
);
1535 req
->work
.flags
|= IO_WQ_WORK_FILES
;
1536 io_req_track_inflight(req
);
1538 if (!(req
->work
.flags
& IO_WQ_WORK_MM
) &&
1539 (def
->work_flags
& IO_WQ_WORK_MM
)) {
1540 if (id
->mm
!= current
->mm
)
1543 req
->work
.flags
|= IO_WQ_WORK_MM
;
1549 static void io_prep_async_work(struct io_kiocb
*req
)
1551 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1552 struct io_ring_ctx
*ctx
= req
->ctx
;
1554 io_req_init_async(req
);
1556 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1557 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1559 if (req
->flags
& REQ_F_ISREG
) {
1560 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1561 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1563 if (def
->unbound_nonreg_file
)
1564 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1567 /* if we fail grabbing identity, we must COW, regrab, and retry */
1568 if (io_grab_identity(req
))
1571 if (!io_identity_cow(req
))
1574 /* can't fail at this point */
1575 if (!io_grab_identity(req
))
1579 static void io_prep_async_link(struct io_kiocb
*req
)
1581 struct io_kiocb
*cur
;
1583 io_for_each_link(cur
, req
)
1584 io_prep_async_work(cur
);
1587 static struct io_kiocb
*__io_queue_async_work(struct io_kiocb
*req
)
1589 struct io_ring_ctx
*ctx
= req
->ctx
;
1590 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1592 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1593 &req
->work
, req
->flags
);
1594 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1598 static void io_queue_async_work(struct io_kiocb
*req
)
1600 struct io_kiocb
*link
;
1602 /* init ->work of the whole link before punting */
1603 io_prep_async_link(req
);
1604 link
= __io_queue_async_work(req
);
1607 io_queue_linked_timeout(link
);
1610 static void io_kill_timeout(struct io_kiocb
*req
)
1612 struct io_timeout_data
*io
= req
->async_data
;
1615 ret
= hrtimer_try_to_cancel(&io
->timer
);
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
, 0);
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
);
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
);
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
);
1699 /* order cqe stores with ring update */
1700 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1702 if (unlikely(!list_empty(&ctx
->defer_list
)))
1703 __io_queue_deferred(ctx
);
1706 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1708 struct io_rings
*r
= ctx
->rings
;
1710 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1713 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1715 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1718 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1720 struct io_rings
*rings
= ctx
->rings
;
1724 * writes to the cq entry need to come after reading head; the
1725 * control dependency is enough as we're using WRITE_ONCE to
1728 if (__io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1731 tail
= ctx
->cached_cq_tail
++;
1732 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1735 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1739 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1741 if (!ctx
->eventfd_async
)
1743 return io_wq_current_is_worker();
1746 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1748 /* see waitqueue_active() comment */
1751 if (waitqueue_active(&ctx
->wait
))
1752 wake_up(&ctx
->wait
);
1753 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1754 wake_up(&ctx
->sq_data
->wait
);
1755 if (io_should_trigger_evfd(ctx
))
1756 eventfd_signal(ctx
->cq_ev_fd
, 1);
1757 if (waitqueue_active(&ctx
->cq_wait
)) {
1758 wake_up_interruptible(&ctx
->cq_wait
);
1759 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1763 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1765 /* see waitqueue_active() comment */
1768 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1769 if (waitqueue_active(&ctx
->wait
))
1770 wake_up(&ctx
->wait
);
1772 if (io_should_trigger_evfd(ctx
))
1773 eventfd_signal(ctx
->cq_ev_fd
, 1);
1774 if (waitqueue_active(&ctx
->cq_wait
)) {
1775 wake_up_interruptible(&ctx
->cq_wait
);
1776 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1780 /* Returns true if there are no backlogged entries after the flush */
1781 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1782 struct task_struct
*tsk
,
1783 struct files_struct
*files
)
1785 struct io_rings
*rings
= ctx
->rings
;
1786 struct io_kiocb
*req
, *tmp
;
1787 struct io_uring_cqe
*cqe
;
1788 unsigned long flags
;
1789 bool all_flushed
, posted
;
1792 if (!force
&& __io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1796 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1797 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1798 if (!io_match_task(req
, tsk
, files
))
1801 cqe
= io_get_cqring(ctx
);
1805 list_move(&req
->compl.list
, &list
);
1807 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1808 WRITE_ONCE(cqe
->res
, req
->result
);
1809 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1811 ctx
->cached_cq_overflow
++;
1812 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1813 ctx
->cached_cq_overflow
);
1818 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1820 clear_bit(0, &ctx
->sq_check_overflow
);
1821 clear_bit(0, &ctx
->cq_check_overflow
);
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 void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1841 struct task_struct
*tsk
,
1842 struct files_struct
*files
)
1844 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1845 /* iopoll syncs against uring_lock, not completion_lock */
1846 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1847 mutex_lock(&ctx
->uring_lock
);
1848 __io_cqring_overflow_flush(ctx
, force
, tsk
, files
);
1849 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1850 mutex_unlock(&ctx
->uring_lock
);
1854 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1856 struct io_ring_ctx
*ctx
= req
->ctx
;
1857 struct io_uring_cqe
*cqe
;
1859 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1862 * If we can't get a cq entry, userspace overflowed the
1863 * submission (by quite a lot). Increment the overflow count in
1866 cqe
= io_get_cqring(ctx
);
1868 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1869 WRITE_ONCE(cqe
->res
, res
);
1870 WRITE_ONCE(cqe
->flags
, cflags
);
1871 } else if (ctx
->cq_overflow_flushed
||
1872 atomic_read(&req
->task
->io_uring
->in_idle
)) {
1874 * If we're in ring overflow flush mode, or in task cancel mode,
1875 * then we cannot store the request for later flushing, we need
1876 * to drop it on the floor.
1878 ctx
->cached_cq_overflow
++;
1879 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1881 if (list_empty(&ctx
->cq_overflow_list
)) {
1882 set_bit(0, &ctx
->sq_check_overflow
);
1883 set_bit(0, &ctx
->cq_check_overflow
);
1884 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1888 req
->compl.cflags
= cflags
;
1889 refcount_inc(&req
->refs
);
1890 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1894 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1896 __io_cqring_fill_event(req
, res
, 0);
1899 static void io_req_complete_post(struct io_kiocb
*req
, long res
,
1900 unsigned int cflags
)
1902 struct io_ring_ctx
*ctx
= req
->ctx
;
1903 unsigned long flags
;
1905 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1906 __io_cqring_fill_event(req
, res
, cflags
);
1907 io_commit_cqring(ctx
);
1908 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1910 io_cqring_ev_posted(ctx
);
1913 static inline void io_req_complete_nostate(struct io_kiocb
*req
, long res
,
1914 unsigned int cflags
)
1916 io_req_complete_post(req
, res
, cflags
);
1920 static void io_submit_flush_completions(struct io_comp_state
*cs
)
1922 struct io_ring_ctx
*ctx
= cs
->ctx
;
1924 spin_lock_irq(&ctx
->completion_lock
);
1925 while (!list_empty(&cs
->list
)) {
1926 struct io_kiocb
*req
;
1928 req
= list_first_entry(&cs
->list
, struct io_kiocb
, compl.list
);
1929 list_del(&req
->compl.list
);
1930 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
1933 * io_free_req() doesn't care about completion_lock unless one
1934 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1935 * because of a potential deadlock with req->work.fs->lock
1936 * We defer both, completion and submission refs.
1938 if (req
->flags
& (REQ_F_FAIL_LINK
|REQ_F_LINK_TIMEOUT
1939 |REQ_F_WORK_INITIALIZED
)) {
1940 spin_unlock_irq(&ctx
->completion_lock
);
1941 io_double_put_req(req
);
1942 spin_lock_irq(&ctx
->completion_lock
);
1944 io_double_put_req(req
);
1947 io_commit_cqring(ctx
);
1948 spin_unlock_irq(&ctx
->completion_lock
);
1950 io_cqring_ev_posted(ctx
);
1954 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1955 unsigned int cflags
, struct io_comp_state
*cs
)
1959 req
->compl.cflags
= cflags
;
1960 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1963 static inline void __io_req_complete(struct io_kiocb
*req
, long res
,
1964 unsigned cflags
, struct io_comp_state
*cs
)
1967 io_req_complete_nostate(req
, res
, cflags
);
1969 io_req_complete_state(req
, res
, cflags
, cs
);
1972 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1974 __io_req_complete(req
, res
, 0, NULL
);
1977 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1979 return req
== (struct io_kiocb
*)
1980 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1983 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1985 struct io_kiocb
*req
;
1987 req
= ctx
->fallback_req
;
1988 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1994 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1995 struct io_submit_state
*state
)
1997 if (!state
->free_reqs
) {
1998 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
2002 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
2003 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
2006 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2007 * retry single alloc to be on the safe side.
2009 if (unlikely(ret
<= 0)) {
2010 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
2011 if (!state
->reqs
[0])
2012 return io_get_fallback_req(ctx
);
2015 state
->free_reqs
= ret
;
2019 return state
->reqs
[state
->free_reqs
];
2022 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
2029 static void io_dismantle_req(struct io_kiocb
*req
)
2033 if (req
->async_data
)
2034 kfree(req
->async_data
);
2036 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
2037 if (req
->fixed_rsrc_refs
)
2038 percpu_ref_put(req
->fixed_rsrc_refs
);
2039 io_req_clean_work(req
);
2042 static inline void io_put_task(struct task_struct
*task
, int nr
)
2044 struct io_uring_task
*tctx
= task
->io_uring
;
2046 percpu_counter_sub(&tctx
->inflight
, nr
);
2047 if (unlikely(atomic_read(&tctx
->in_idle
)))
2048 wake_up(&tctx
->wait
);
2049 put_task_struct_many(task
, nr
);
2052 static void __io_free_req(struct io_kiocb
*req
)
2054 struct io_ring_ctx
*ctx
= req
->ctx
;
2056 io_dismantle_req(req
);
2057 io_put_task(req
->task
, 1);
2059 if (likely(!io_is_fallback_req(req
)))
2060 kmem_cache_free(req_cachep
, req
);
2062 clear_bit_unlock(0, (unsigned long *) &ctx
->fallback_req
);
2063 percpu_ref_put(&ctx
->refs
);
2066 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2068 struct io_kiocb
*nxt
= req
->link
;
2070 req
->link
= nxt
->link
;
2074 static void io_kill_linked_timeout(struct io_kiocb
*req
)
2076 struct io_ring_ctx
*ctx
= req
->ctx
;
2077 struct io_kiocb
*link
;
2078 bool cancelled
= false;
2079 unsigned long flags
;
2081 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2085 * Can happen if a linked timeout fired and link had been like
2086 * req -> link t-out -> link t-out [-> ...]
2088 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
2089 struct io_timeout_data
*io
= link
->async_data
;
2092 io_remove_next_linked(req
);
2093 link
->timeout
.head
= NULL
;
2094 ret
= hrtimer_try_to_cancel(&io
->timer
);
2096 io_cqring_fill_event(link
, -ECANCELED
);
2097 io_commit_cqring(ctx
);
2101 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
2102 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2105 io_cqring_ev_posted(ctx
);
2111 static void io_fail_links(struct io_kiocb
*req
)
2113 struct io_kiocb
*link
, *nxt
;
2114 struct io_ring_ctx
*ctx
= req
->ctx
;
2115 unsigned long flags
;
2117 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
2125 trace_io_uring_fail_link(req
, link
);
2126 io_cqring_fill_event(link
, -ECANCELED
);
2129 * It's ok to free under spinlock as they're not linked anymore,
2130 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2133 if (link
->flags
& REQ_F_WORK_INITIALIZED
)
2134 io_put_req_deferred(link
, 2);
2136 io_double_put_req(link
);
2139 io_commit_cqring(ctx
);
2140 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
2142 io_cqring_ev_posted(ctx
);
2145 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2147 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
2148 io_kill_linked_timeout(req
);
2151 * If LINK is set, we have dependent requests in this chain. If we
2152 * didn't fail this request, queue the first one up, moving any other
2153 * dependencies to the next request. In case of failure, fail the rest
2156 if (likely(!(req
->flags
& REQ_F_FAIL_LINK
))) {
2157 struct io_kiocb
*nxt
= req
->link
;
2166 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2168 if (likely(!(req
->link
) && !(req
->flags
& REQ_F_LINK_TIMEOUT
)))
2170 return __io_req_find_next(req
);
2173 static int io_req_task_work_add(struct io_kiocb
*req
)
2175 struct task_struct
*tsk
= req
->task
;
2176 struct io_ring_ctx
*ctx
= req
->ctx
;
2177 enum task_work_notify_mode notify
;
2180 if (tsk
->flags
& PF_EXITING
)
2184 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2185 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2186 * processing task_work. There's no reliable way to tell if TWA_RESUME
2190 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
))
2191 notify
= TWA_SIGNAL
;
2193 ret
= task_work_add(tsk
, &req
->task_work
, notify
);
2195 wake_up_process(tsk
);
2200 static void io_req_task_work_add_fallback(struct io_kiocb
*req
,
2201 void (*cb
)(struct callback_head
*))
2203 struct task_struct
*tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2205 init_task_work(&req
->task_work
, cb
);
2206 task_work_add(tsk
, &req
->task_work
, TWA_NONE
);
2207 wake_up_process(tsk
);
2210 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
2212 struct io_ring_ctx
*ctx
= req
->ctx
;
2214 spin_lock_irq(&ctx
->completion_lock
);
2215 io_cqring_fill_event(req
, error
);
2216 io_commit_cqring(ctx
);
2217 spin_unlock_irq(&ctx
->completion_lock
);
2219 io_cqring_ev_posted(ctx
);
2220 req_set_fail_links(req
);
2221 io_double_put_req(req
);
2224 static void io_req_task_cancel(struct callback_head
*cb
)
2226 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2227 struct io_ring_ctx
*ctx
= req
->ctx
;
2229 __io_req_task_cancel(req
, -ECANCELED
);
2230 percpu_ref_put(&ctx
->refs
);
2233 static void __io_req_task_submit(struct io_kiocb
*req
)
2235 struct io_ring_ctx
*ctx
= req
->ctx
;
2237 mutex_lock(&ctx
->uring_lock
);
2238 if (!ctx
->sqo_dead
&&
2239 !__io_sq_thread_acquire_mm(ctx
) &&
2240 !__io_sq_thread_acquire_files(ctx
))
2241 __io_queue_sqe(req
, NULL
);
2243 __io_req_task_cancel(req
, -EFAULT
);
2244 mutex_unlock(&ctx
->uring_lock
);
2247 static void io_req_task_submit(struct callback_head
*cb
)
2249 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2250 struct io_ring_ctx
*ctx
= req
->ctx
;
2252 __io_req_task_submit(req
);
2253 percpu_ref_put(&ctx
->refs
);
2256 static void io_req_task_queue(struct io_kiocb
*req
)
2260 init_task_work(&req
->task_work
, io_req_task_submit
);
2261 percpu_ref_get(&req
->ctx
->refs
);
2263 ret
= io_req_task_work_add(req
);
2265 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2268 static inline void io_queue_next(struct io_kiocb
*req
)
2270 struct io_kiocb
*nxt
= io_req_find_next(req
);
2273 io_req_task_queue(nxt
);
2276 static void io_free_req(struct io_kiocb
*req
)
2283 void *reqs
[IO_IOPOLL_BATCH
];
2286 struct task_struct
*task
;
2290 static inline void io_init_req_batch(struct req_batch
*rb
)
2297 static void __io_req_free_batch_flush(struct io_ring_ctx
*ctx
,
2298 struct req_batch
*rb
)
2300 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
2301 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
2305 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2306 struct req_batch
*rb
)
2309 __io_req_free_batch_flush(ctx
, rb
);
2311 io_put_task(rb
->task
, rb
->task_refs
);
2316 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
)
2318 if (unlikely(io_is_fallback_req(req
))) {
2324 if (req
->task
!= rb
->task
) {
2326 io_put_task(rb
->task
, rb
->task_refs
);
2327 rb
->task
= req
->task
;
2332 io_dismantle_req(req
);
2333 rb
->reqs
[rb
->to_free
++] = req
;
2334 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
2335 __io_req_free_batch_flush(req
->ctx
, rb
);
2339 * Drop reference to request, return next in chain (if there is one) if this
2340 * was the last reference to this request.
2342 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2344 struct io_kiocb
*nxt
= NULL
;
2346 if (refcount_dec_and_test(&req
->refs
)) {
2347 nxt
= io_req_find_next(req
);
2353 static void io_put_req(struct io_kiocb
*req
)
2355 if (refcount_dec_and_test(&req
->refs
))
2359 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2361 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2366 static void io_free_req_deferred(struct io_kiocb
*req
)
2370 init_task_work(&req
->task_work
, io_put_req_deferred_cb
);
2371 ret
= io_req_task_work_add(req
);
2373 io_req_task_work_add_fallback(req
, io_put_req_deferred_cb
);
2376 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2378 if (refcount_sub_and_test(refs
, &req
->refs
))
2379 io_free_req_deferred(req
);
2382 static void io_double_put_req(struct io_kiocb
*req
)
2384 /* drop both submit and complete references */
2385 if (refcount_sub_and_test(2, &req
->refs
))
2389 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2391 /* See comment at the top of this file */
2393 return __io_cqring_events(ctx
);
2396 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2398 struct io_rings
*rings
= ctx
->rings
;
2400 /* make sure SQ entry isn't read before tail */
2401 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2404 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2406 unsigned int cflags
;
2408 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2409 cflags
|= IORING_CQE_F_BUFFER
;
2410 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2415 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2417 struct io_buffer
*kbuf
;
2419 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2420 return io_put_kbuf(req
, kbuf
);
2423 static inline bool io_run_task_work(void)
2426 * Not safe to run on exiting task, and the task_work handling will
2427 * not add work to such a task.
2429 if (unlikely(current
->flags
& PF_EXITING
))
2431 if (current
->task_works
) {
2432 __set_current_state(TASK_RUNNING
);
2440 static void io_iopoll_queue(struct list_head
*again
)
2442 struct io_kiocb
*req
;
2445 req
= list_first_entry(again
, struct io_kiocb
, inflight_entry
);
2446 list_del(&req
->inflight_entry
);
2447 __io_complete_rw(req
, -EAGAIN
, 0, NULL
);
2448 } while (!list_empty(again
));
2452 * Find and free completed poll iocbs
2454 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2455 struct list_head
*done
)
2457 struct req_batch rb
;
2458 struct io_kiocb
*req
;
2461 /* order with ->result store in io_complete_rw_iopoll() */
2464 io_init_req_batch(&rb
);
2465 while (!list_empty(done
)) {
2468 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2469 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2471 req
->iopoll_completed
= 0;
2472 list_move_tail(&req
->inflight_entry
, &again
);
2475 list_del(&req
->inflight_entry
);
2477 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2478 cflags
= io_put_rw_kbuf(req
);
2480 __io_cqring_fill_event(req
, req
->result
, cflags
);
2483 if (refcount_dec_and_test(&req
->refs
))
2484 io_req_free_batch(&rb
, req
);
2487 io_commit_cqring(ctx
);
2488 io_cqring_ev_posted_iopoll(ctx
);
2489 io_req_free_batch_finish(ctx
, &rb
);
2491 if (!list_empty(&again
))
2492 io_iopoll_queue(&again
);
2495 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2498 struct io_kiocb
*req
, *tmp
;
2504 * Only spin for completions if we don't have multiple devices hanging
2505 * off our complete list, and we're under the requested amount.
2507 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2510 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2511 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2514 * Move completed and retryable entries to our local lists.
2515 * If we find a request that requires polling, break out
2516 * and complete those lists first, if we have entries there.
2518 if (READ_ONCE(req
->iopoll_completed
)) {
2519 list_move_tail(&req
->inflight_entry
, &done
);
2522 if (!list_empty(&done
))
2525 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2529 /* iopoll may have completed current req */
2530 if (READ_ONCE(req
->iopoll_completed
))
2531 list_move_tail(&req
->inflight_entry
, &done
);
2538 if (!list_empty(&done
))
2539 io_iopoll_complete(ctx
, nr_events
, &done
);
2545 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2546 * non-spinning poll check - we'll still enter the driver poll loop, but only
2547 * as a non-spinning completion check.
2549 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2552 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2555 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2558 if (*nr_events
>= min
)
2566 * We can't just wait for polled events to come to us, we have to actively
2567 * find and complete them.
2569 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2571 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2574 mutex_lock(&ctx
->uring_lock
);
2575 while (!list_empty(&ctx
->iopoll_list
)) {
2576 unsigned int nr_events
= 0;
2578 io_do_iopoll(ctx
, &nr_events
, 0);
2580 /* let it sleep and repeat later if can't complete a request */
2584 * Ensure we allow local-to-the-cpu processing to take place,
2585 * in this case we need to ensure that we reap all events.
2586 * Also let task_work, etc. to progress by releasing the mutex
2588 if (need_resched()) {
2589 mutex_unlock(&ctx
->uring_lock
);
2591 mutex_lock(&ctx
->uring_lock
);
2594 mutex_unlock(&ctx
->uring_lock
);
2597 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2599 unsigned int nr_events
= 0;
2600 int iters
= 0, ret
= 0;
2603 * We disallow the app entering submit/complete with polling, but we
2604 * still need to lock the ring to prevent racing with polled issue
2605 * that got punted to a workqueue.
2607 mutex_lock(&ctx
->uring_lock
);
2610 * Don't enter poll loop if we already have events pending.
2611 * If we do, we can potentially be spinning for commands that
2612 * already triggered a CQE (eg in error).
2614 if (test_bit(0, &ctx
->cq_check_overflow
))
2615 __io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2616 if (io_cqring_events(ctx
))
2620 * If a submit got punted to a workqueue, we can have the
2621 * application entering polling for a command before it gets
2622 * issued. That app will hold the uring_lock for the duration
2623 * of the poll right here, so we need to take a breather every
2624 * now and then to ensure that the issue has a chance to add
2625 * the poll to the issued list. Otherwise we can spin here
2626 * forever, while the workqueue is stuck trying to acquire the
2629 if (!(++iters
& 7)) {
2630 mutex_unlock(&ctx
->uring_lock
);
2632 mutex_lock(&ctx
->uring_lock
);
2635 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2639 } while (min
&& !nr_events
&& !need_resched());
2641 mutex_unlock(&ctx
->uring_lock
);
2645 static void kiocb_end_write(struct io_kiocb
*req
)
2648 * Tell lockdep we inherited freeze protection from submission
2651 if (req
->flags
& REQ_F_ISREG
) {
2652 struct inode
*inode
= file_inode(req
->file
);
2654 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2656 file_end_write(req
->file
);
2659 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
,
2660 struct io_comp_state
*cs
)
2662 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2665 if (kiocb
->ki_flags
& IOCB_WRITE
)
2666 kiocb_end_write(req
);
2668 if (res
!= req
->result
)
2669 req_set_fail_links(req
);
2670 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2671 cflags
= io_put_rw_kbuf(req
);
2672 __io_req_complete(req
, res
, cflags
, cs
);
2676 static bool io_resubmit_prep(struct io_kiocb
*req
)
2678 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2679 int rw
, ret
= -ECANCELED
;
2680 struct iov_iter iter
;
2682 /* already prepared */
2683 if (req
->async_data
)
2686 switch (req
->opcode
) {
2687 case IORING_OP_READV
:
2688 case IORING_OP_READ_FIXED
:
2689 case IORING_OP_READ
:
2692 case IORING_OP_WRITEV
:
2693 case IORING_OP_WRITE_FIXED
:
2694 case IORING_OP_WRITE
:
2698 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2703 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2706 return !io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2710 static bool io_rw_reissue(struct io_kiocb
*req
, long res
)
2716 if (res
!= -EAGAIN
&& res
!= -EOPNOTSUPP
)
2718 mode
= file_inode(req
->file
)->i_mode
;
2719 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2721 if ((req
->flags
& REQ_F_NOWAIT
) || io_wq_current_is_worker())
2724 lockdep_assert_held(&req
->ctx
->uring_lock
);
2726 ret
= io_sq_thread_acquire_mm_files(req
->ctx
, req
);
2728 if (!ret
&& io_resubmit_prep(req
)) {
2729 refcount_inc(&req
->refs
);
2730 io_queue_async_work(req
);
2733 req_set_fail_links(req
);
2738 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2739 struct io_comp_state
*cs
)
2741 if (!io_rw_reissue(req
, res
))
2742 io_complete_rw_common(&req
->rw
.kiocb
, res
, cs
);
2745 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2747 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2749 __io_complete_rw(req
, res
, res2
, NULL
);
2752 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2754 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2756 if (kiocb
->ki_flags
& IOCB_WRITE
)
2757 kiocb_end_write(req
);
2759 if (res
!= -EAGAIN
&& res
!= req
->result
)
2760 req_set_fail_links(req
);
2762 WRITE_ONCE(req
->result
, res
);
2763 /* order with io_poll_complete() checking ->result */
2765 WRITE_ONCE(req
->iopoll_completed
, 1);
2769 * After the iocb has been issued, it's safe to be found on the poll list.
2770 * Adding the kiocb to the list AFTER submission ensures that we don't
2771 * find it from a io_iopoll_getevents() thread before the issuer is done
2772 * accessing the kiocb cookie.
2774 static void io_iopoll_req_issued(struct io_kiocb
*req
, bool in_async
)
2776 struct io_ring_ctx
*ctx
= req
->ctx
;
2779 * Track whether we have multiple files in our lists. This will impact
2780 * how we do polling eventually, not spinning if we're on potentially
2781 * different devices.
2783 if (list_empty(&ctx
->iopoll_list
)) {
2784 ctx
->poll_multi_file
= false;
2785 } else if (!ctx
->poll_multi_file
) {
2786 struct io_kiocb
*list_req
;
2788 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2790 if (list_req
->file
!= req
->file
)
2791 ctx
->poll_multi_file
= true;
2795 * For fast devices, IO may have already completed. If it has, add
2796 * it to the front so we find it first.
2798 if (READ_ONCE(req
->iopoll_completed
))
2799 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2801 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2804 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2805 * task context or in io worker task context. If current task context is
2806 * sq thread, we don't need to check whether should wake up sq thread.
2808 if (in_async
&& (ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2809 wq_has_sleeper(&ctx
->sq_data
->wait
))
2810 wake_up(&ctx
->sq_data
->wait
);
2813 static inline void io_state_file_put(struct io_submit_state
*state
)
2815 if (state
->file_refs
) {
2816 fput_many(state
->file
, state
->file_refs
);
2817 state
->file_refs
= 0;
2822 * Get as many references to a file as we have IOs left in this submission,
2823 * assuming most submissions are for one file, or at least that each file
2824 * has more than one submission.
2826 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2831 if (state
->file_refs
) {
2832 if (state
->fd
== fd
) {
2836 io_state_file_put(state
);
2838 state
->file
= fget_many(fd
, state
->ios_left
);
2839 if (unlikely(!state
->file
))
2843 state
->file_refs
= state
->ios_left
- 1;
2847 static bool io_bdev_nowait(struct block_device
*bdev
)
2849 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2853 * If we tracked the file through the SCM inflight mechanism, we could support
2854 * any file. For now, just ensure that anything potentially problematic is done
2857 static bool io_file_supports_async(struct file
*file
, int rw
)
2859 umode_t mode
= file_inode(file
)->i_mode
;
2861 if (S_ISBLK(mode
)) {
2862 if (IS_ENABLED(CONFIG_BLOCK
) &&
2863 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2867 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2869 if (S_ISREG(mode
)) {
2870 if (IS_ENABLED(CONFIG_BLOCK
) &&
2871 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2872 file
->f_op
!= &io_uring_fops
)
2877 /* any ->read/write should understand O_NONBLOCK */
2878 if (file
->f_flags
& O_NONBLOCK
)
2881 if (!(file
->f_mode
& FMODE_NOWAIT
))
2885 return file
->f_op
->read_iter
!= NULL
;
2887 return file
->f_op
->write_iter
!= NULL
;
2890 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2892 struct io_ring_ctx
*ctx
= req
->ctx
;
2893 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2894 struct file
*file
= req
->file
;
2898 if (S_ISREG(file_inode(file
)->i_mode
))
2899 req
->flags
|= REQ_F_ISREG
;
2901 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2902 if (kiocb
->ki_pos
== -1 && !(file
->f_mode
& FMODE_STREAM
)) {
2903 req
->flags
|= REQ_F_CUR_POS
;
2904 kiocb
->ki_pos
= file
->f_pos
;
2906 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2907 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2908 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2912 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2913 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) || (file
->f_flags
& O_NONBLOCK
))
2914 req
->flags
|= REQ_F_NOWAIT
;
2916 ioprio
= READ_ONCE(sqe
->ioprio
);
2918 ret
= ioprio_check_cap(ioprio
);
2922 kiocb
->ki_ioprio
= ioprio
;
2924 kiocb
->ki_ioprio
= get_current_ioprio();
2926 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2927 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2928 !kiocb
->ki_filp
->f_op
->iopoll
)
2931 kiocb
->ki_flags
|= IOCB_HIPRI
;
2932 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2933 req
->iopoll_completed
= 0;
2935 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2937 kiocb
->ki_complete
= io_complete_rw
;
2940 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2941 req
->rw
.len
= READ_ONCE(sqe
->len
);
2942 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2946 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2952 case -ERESTARTNOINTR
:
2953 case -ERESTARTNOHAND
:
2954 case -ERESTART_RESTARTBLOCK
:
2956 * We can't just restart the syscall, since previously
2957 * submitted sqes may already be in progress. Just fail this
2963 kiocb
->ki_complete(kiocb
, ret
, 0);
2967 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2968 struct io_comp_state
*cs
)
2970 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2971 struct io_async_rw
*io
= req
->async_data
;
2973 /* add previously done IO, if any */
2974 if (io
&& io
->bytes_done
> 0) {
2976 ret
= io
->bytes_done
;
2978 ret
+= io
->bytes_done
;
2981 if (req
->flags
& REQ_F_CUR_POS
)
2982 req
->file
->f_pos
= kiocb
->ki_pos
;
2983 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2984 __io_complete_rw(req
, ret
, 0, cs
);
2986 io_rw_done(kiocb
, ret
);
2989 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
2991 struct io_ring_ctx
*ctx
= req
->ctx
;
2992 size_t len
= req
->rw
.len
;
2993 struct io_mapped_ubuf
*imu
;
2994 u16 index
, buf_index
= req
->buf_index
;
2998 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3000 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3001 imu
= &ctx
->user_bufs
[index
];
3002 buf_addr
= req
->rw
.addr
;
3005 if (buf_addr
+ len
< buf_addr
)
3007 /* not inside the mapped region */
3008 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
3012 * May not be a start of buffer, set size appropriately
3013 * and advance us to the beginning.
3015 offset
= buf_addr
- imu
->ubuf
;
3016 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3020 * Don't use iov_iter_advance() here, as it's really slow for
3021 * using the latter parts of a big fixed buffer - it iterates
3022 * over each segment manually. We can cheat a bit here, because
3025 * 1) it's a BVEC iter, we set it up
3026 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3027 * first and last bvec
3029 * So just find our index, and adjust the iterator afterwards.
3030 * If the offset is within the first bvec (or the whole first
3031 * bvec, just use iov_iter_advance(). This makes it easier
3032 * since we can just skip the first segment, which may not
3033 * be PAGE_SIZE aligned.
3035 const struct bio_vec
*bvec
= imu
->bvec
;
3037 if (offset
<= bvec
->bv_len
) {
3038 iov_iter_advance(iter
, offset
);
3040 unsigned long seg_skip
;
3042 /* skip first vec */
3043 offset
-= bvec
->bv_len
;
3044 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3046 iter
->bvec
= bvec
+ seg_skip
;
3047 iter
->nr_segs
-= seg_skip
;
3048 iter
->count
-= bvec
->bv_len
+ offset
;
3049 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3056 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3059 mutex_unlock(&ctx
->uring_lock
);
3062 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3065 * "Normal" inline submissions always hold the uring_lock, since we
3066 * grab it from the system call. Same is true for the SQPOLL offload.
3067 * The only exception is when we've detached the request and issue it
3068 * from an async worker thread, grab the lock for that case.
3071 mutex_lock(&ctx
->uring_lock
);
3074 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3075 int bgid
, struct io_buffer
*kbuf
,
3078 struct io_buffer
*head
;
3080 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3083 io_ring_submit_lock(req
->ctx
, needs_lock
);
3085 lockdep_assert_held(&req
->ctx
->uring_lock
);
3087 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
3089 if (!list_empty(&head
->list
)) {
3090 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3092 list_del(&kbuf
->list
);
3095 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
3097 if (*len
> kbuf
->len
)
3100 kbuf
= ERR_PTR(-ENOBUFS
);
3103 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3108 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3111 struct io_buffer
*kbuf
;
3114 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3115 bgid
= req
->buf_index
;
3116 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3119 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3120 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3121 return u64_to_user_ptr(kbuf
->addr
);
3124 #ifdef CONFIG_COMPAT
3125 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3128 struct compat_iovec __user
*uiov
;
3129 compat_ssize_t clen
;
3133 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3134 if (!access_ok(uiov
, sizeof(*uiov
)))
3136 if (__get_user(clen
, &uiov
->iov_len
))
3142 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3144 return PTR_ERR(buf
);
3145 iov
[0].iov_base
= buf
;
3146 iov
[0].iov_len
= (compat_size_t
) len
;
3151 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3154 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3158 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3161 len
= iov
[0].iov_len
;
3164 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3166 return PTR_ERR(buf
);
3167 iov
[0].iov_base
= buf
;
3168 iov
[0].iov_len
= len
;
3172 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3175 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3176 struct io_buffer
*kbuf
;
3178 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3179 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3180 iov
[0].iov_len
= kbuf
->len
;
3183 if (req
->rw
.len
!= 1)
3186 #ifdef CONFIG_COMPAT
3187 if (req
->ctx
->compat
)
3188 return io_compat_import(req
, iov
, needs_lock
);
3191 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3194 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
3195 struct iov_iter
*iter
, bool needs_lock
)
3197 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3198 size_t sqe_len
= req
->rw
.len
;
3199 u8 opcode
= req
->opcode
;
3202 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3204 return io_import_fixed(req
, rw
, iter
);
3207 /* buffer index only valid with fixed read/write, or buffer select */
3208 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3211 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3212 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3213 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3215 return PTR_ERR(buf
);
3216 req
->rw
.len
= sqe_len
;
3219 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3224 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3225 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3227 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3232 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3236 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3238 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3242 * For files that don't have ->read_iter() and ->write_iter(), handle them
3243 * by looping over ->read() or ->write() manually.
3245 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3247 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3248 struct file
*file
= req
->file
;
3252 * Don't support polled IO through this interface, and we can't
3253 * support non-blocking either. For the latter, this just causes
3254 * the kiocb to be handled from an async context.
3256 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3258 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3261 while (iov_iter_count(iter
)) {
3265 if (!iov_iter_is_bvec(iter
)) {
3266 iovec
= iov_iter_iovec(iter
);
3268 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3269 iovec
.iov_len
= req
->rw
.len
;
3273 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3274 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3276 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3277 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3286 if (nr
!= iovec
.iov_len
)
3290 iov_iter_advance(iter
, nr
);
3296 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3297 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3299 struct io_async_rw
*rw
= req
->async_data
;
3301 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3302 rw
->free_iovec
= iovec
;
3304 /* can only be fixed buffers, no need to do anything */
3305 if (iov_iter_is_bvec(iter
))
3308 unsigned iov_off
= 0;
3310 rw
->iter
.iov
= rw
->fast_iov
;
3311 if (iter
->iov
!= fast_iov
) {
3312 iov_off
= iter
->iov
- fast_iov
;
3313 rw
->iter
.iov
+= iov_off
;
3315 if (rw
->fast_iov
!= fast_iov
)
3316 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3317 sizeof(struct iovec
) * iter
->nr_segs
);
3319 req
->flags
|= REQ_F_NEED_CLEANUP
;
3323 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3325 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3326 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3327 return req
->async_data
== NULL
;
3330 static int io_alloc_async_data(struct io_kiocb
*req
)
3332 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3335 return __io_alloc_async_data(req
);
3338 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3339 const struct iovec
*fast_iov
,
3340 struct iov_iter
*iter
, bool force
)
3342 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3344 if (!req
->async_data
) {
3345 if (__io_alloc_async_data(req
)) {
3350 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3355 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3357 struct io_async_rw
*iorw
= req
->async_data
;
3358 struct iovec
*iov
= iorw
->fast_iov
;
3361 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3362 if (unlikely(ret
< 0))
3365 iorw
->bytes_done
= 0;
3366 iorw
->free_iovec
= iov
;
3368 req
->flags
|= REQ_F_NEED_CLEANUP
;
3372 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3376 ret
= io_prep_rw(req
, sqe
);
3380 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3383 /* either don't need iovec imported or already have it */
3384 if (!req
->async_data
)
3386 return io_rw_prep_async(req
, READ
);
3390 * This is our waitqueue callback handler, registered through lock_page_async()
3391 * when we initially tried to do the IO with the iocb armed our waitqueue.
3392 * This gets called when the page is unlocked, and we generally expect that to
3393 * happen when the page IO is completed and the page is now uptodate. This will
3394 * queue a task_work based retry of the operation, attempting to copy the data
3395 * again. If the latter fails because the page was NOT uptodate, then we will
3396 * do a thread based blocking retry of the operation. That's the unexpected
3399 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3400 int sync
, void *arg
)
3402 struct wait_page_queue
*wpq
;
3403 struct io_kiocb
*req
= wait
->private;
3404 struct wait_page_key
*key
= arg
;
3407 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3409 if (!wake_page_match(wpq
, key
))
3412 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3413 list_del_init(&wait
->entry
);
3415 init_task_work(&req
->task_work
, io_req_task_submit
);
3416 percpu_ref_get(&req
->ctx
->refs
);
3418 /* submit ref gets dropped, acquire a new one */
3419 refcount_inc(&req
->refs
);
3420 ret
= io_req_task_work_add(req
);
3422 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
3427 * This controls whether a given IO request should be armed for async page
3428 * based retry. If we return false here, the request is handed to the async
3429 * worker threads for retry. If we're doing buffered reads on a regular file,
3430 * we prepare a private wait_page_queue entry and retry the operation. This
3431 * will either succeed because the page is now uptodate and unlocked, or it
3432 * will register a callback when the page is unlocked at IO completion. Through
3433 * that callback, io_uring uses task_work to setup a retry of the operation.
3434 * That retry will attempt the buffered read again. The retry will generally
3435 * succeed, or in rare cases where it fails, we then fall back to using the
3436 * async worker threads for a blocking retry.
3438 static bool io_rw_should_retry(struct io_kiocb
*req
)
3440 struct io_async_rw
*rw
= req
->async_data
;
3441 struct wait_page_queue
*wait
= &rw
->wpq
;
3442 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3444 /* never retry for NOWAIT, we just complete with -EAGAIN */
3445 if (req
->flags
& REQ_F_NOWAIT
)
3448 /* Only for buffered IO */
3449 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3453 * just use poll if we can, and don't attempt if the fs doesn't
3454 * support callback based unlocks
3456 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3459 wait
->wait
.func
= io_async_buf_func
;
3460 wait
->wait
.private = req
;
3461 wait
->wait
.flags
= 0;
3462 INIT_LIST_HEAD(&wait
->wait
.entry
);
3463 kiocb
->ki_flags
|= IOCB_WAITQ
;
3464 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3465 kiocb
->ki_waitq
= wait
;
3469 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3471 if (req
->file
->f_op
->read_iter
)
3472 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3473 else if (req
->file
->f_op
->read
)
3474 return loop_rw_iter(READ
, req
, iter
);
3479 static int io_read(struct io_kiocb
*req
, bool force_nonblock
,
3480 struct io_comp_state
*cs
)
3482 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3483 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3484 struct iov_iter __iter
, *iter
= &__iter
;
3485 struct io_async_rw
*rw
= req
->async_data
;
3486 ssize_t io_size
, ret
, ret2
;
3492 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3496 io_size
= iov_iter_count(iter
);
3497 req
->result
= io_size
;
3499 /* Ensure we clear previously set non-block flag */
3500 if (!force_nonblock
)
3501 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3503 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3505 /* If the file doesn't support async, just async punt */
3506 if (force_nonblock
&& !io_file_supports_async(req
->file
, READ
)) {
3507 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3508 return ret
?: -EAGAIN
;
3511 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3512 if (unlikely(ret
)) {
3517 ret
= io_iter_do_read(req
, iter
);
3519 if (ret
== -EIOCBQUEUED
) {
3520 /* it's faster to check here then delegate to kfree */
3524 } else if (ret
== -EAGAIN
) {
3525 /* IOPOLL retry should happen for io-wq threads */
3526 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3528 /* no retry on NONBLOCK nor RWF_NOWAIT */
3529 if (req
->flags
& REQ_F_NOWAIT
)
3531 /* some cases will consume bytes even on error returns */
3532 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3534 } else if (ret
<= 0 || ret
== io_size
|| !force_nonblock
||
3535 (req
->flags
& REQ_F_NOWAIT
) || !(req
->flags
& REQ_F_ISREG
)) {
3536 /* read all, failed, already did sync or don't want to retry */
3540 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3544 rw
= req
->async_data
;
3545 /* now use our persistent iterator, if we aren't already */
3550 rw
->bytes_done
+= ret
;
3551 /* if we can retry, do so with the callbacks armed */
3552 if (!io_rw_should_retry(req
)) {
3553 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3558 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3559 * we get -EIOCBQUEUED, then we'll get a notification when the
3560 * desired page gets unlocked. We can also get a partial read
3561 * here, and if we do, then just retry at the new offset.
3563 ret
= io_iter_do_read(req
, iter
);
3564 if (ret
== -EIOCBQUEUED
)
3566 /* we got some bytes, but not all. retry. */
3567 } while (ret
> 0 && ret
< io_size
);
3569 kiocb_done(kiocb
, ret
, cs
);
3573 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3577 ret
= io_prep_rw(req
, sqe
);
3581 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3584 /* either don't need iovec imported or already have it */
3585 if (!req
->async_data
)
3587 return io_rw_prep_async(req
, WRITE
);
3590 static int io_write(struct io_kiocb
*req
, bool force_nonblock
,
3591 struct io_comp_state
*cs
)
3593 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3594 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3595 struct iov_iter __iter
, *iter
= &__iter
;
3596 struct io_async_rw
*rw
= req
->async_data
;
3597 ssize_t ret
, ret2
, io_size
;
3603 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3607 io_size
= iov_iter_count(iter
);
3608 req
->result
= io_size
;
3610 /* Ensure we clear previously set non-block flag */
3611 if (!force_nonblock
)
3612 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3614 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3616 /* If the file doesn't support async, just async punt */
3617 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3620 /* file path doesn't support NOWAIT for non-direct_IO */
3621 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3622 (req
->flags
& REQ_F_ISREG
))
3625 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3630 * Open-code file_start_write here to grab freeze protection,
3631 * which will be released by another thread in
3632 * io_complete_rw(). Fool lockdep by telling it the lock got
3633 * released so that it doesn't complain about the held lock when
3634 * we return to userspace.
3636 if (req
->flags
& REQ_F_ISREG
) {
3637 sb_start_write(file_inode(req
->file
)->i_sb
);
3638 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3641 kiocb
->ki_flags
|= IOCB_WRITE
;
3643 if (req
->file
->f_op
->write_iter
)
3644 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3645 else if (req
->file
->f_op
->write
)
3646 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3651 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3652 * retry them without IOCB_NOWAIT.
3654 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3656 /* no retry on NONBLOCK nor RWF_NOWAIT */
3657 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3659 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3660 /* IOPOLL retry should happen for io-wq threads */
3661 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3664 kiocb_done(kiocb
, ret2
, cs
);
3667 /* some cases will consume bytes even on error returns */
3668 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3669 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3670 return ret
?: -EAGAIN
;
3673 /* it's reportedly faster than delegating the null check to kfree() */
3679 static int io_renameat_prep(struct io_kiocb
*req
,
3680 const struct io_uring_sqe
*sqe
)
3682 struct io_rename
*ren
= &req
->rename
;
3683 const char __user
*oldf
, *newf
;
3685 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3688 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3689 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3690 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3691 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3692 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3694 ren
->oldpath
= getname(oldf
);
3695 if (IS_ERR(ren
->oldpath
))
3696 return PTR_ERR(ren
->oldpath
);
3698 ren
->newpath
= getname(newf
);
3699 if (IS_ERR(ren
->newpath
)) {
3700 putname(ren
->oldpath
);
3701 return PTR_ERR(ren
->newpath
);
3704 req
->flags
|= REQ_F_NEED_CLEANUP
;
3708 static int io_renameat(struct io_kiocb
*req
, bool force_nonblock
)
3710 struct io_rename
*ren
= &req
->rename
;
3716 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3717 ren
->newpath
, ren
->flags
);
3719 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3721 req_set_fail_links(req
);
3722 io_req_complete(req
, ret
);
3726 static int io_unlinkat_prep(struct io_kiocb
*req
,
3727 const struct io_uring_sqe
*sqe
)
3729 struct io_unlink
*un
= &req
->unlink
;
3730 const char __user
*fname
;
3732 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3735 un
->dfd
= READ_ONCE(sqe
->fd
);
3737 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3738 if (un
->flags
& ~AT_REMOVEDIR
)
3741 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3742 un
->filename
= getname(fname
);
3743 if (IS_ERR(un
->filename
))
3744 return PTR_ERR(un
->filename
);
3746 req
->flags
|= REQ_F_NEED_CLEANUP
;
3750 static int io_unlinkat(struct io_kiocb
*req
, bool force_nonblock
)
3752 struct io_unlink
*un
= &req
->unlink
;
3758 if (un
->flags
& AT_REMOVEDIR
)
3759 ret
= do_rmdir(un
->dfd
, un
->filename
);
3761 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3763 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3765 req_set_fail_links(req
);
3766 io_req_complete(req
, ret
);
3770 static int io_shutdown_prep(struct io_kiocb
*req
,
3771 const struct io_uring_sqe
*sqe
)
3773 #if defined(CONFIG_NET)
3774 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3776 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3780 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3787 static int io_shutdown(struct io_kiocb
*req
, bool force_nonblock
)
3789 #if defined(CONFIG_NET)
3790 struct socket
*sock
;
3796 sock
= sock_from_file(req
->file
);
3797 if (unlikely(!sock
))
3800 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3802 req_set_fail_links(req
);
3803 io_req_complete(req
, ret
);
3810 static int __io_splice_prep(struct io_kiocb
*req
,
3811 const struct io_uring_sqe
*sqe
)
3813 struct io_splice
* sp
= &req
->splice
;
3814 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3816 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3820 sp
->len
= READ_ONCE(sqe
->len
);
3821 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3823 if (unlikely(sp
->flags
& ~valid_flags
))
3826 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3827 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3830 req
->flags
|= REQ_F_NEED_CLEANUP
;
3832 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3834 * Splice operation will be punted aync, and here need to
3835 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3837 io_req_init_async(req
);
3838 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3844 static int io_tee_prep(struct io_kiocb
*req
,
3845 const struct io_uring_sqe
*sqe
)
3847 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3849 return __io_splice_prep(req
, sqe
);
3852 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3854 struct io_splice
*sp
= &req
->splice
;
3855 struct file
*in
= sp
->file_in
;
3856 struct file
*out
= sp
->file_out
;
3857 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3863 ret
= do_tee(in
, out
, sp
->len
, flags
);
3865 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3866 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3869 req_set_fail_links(req
);
3870 io_req_complete(req
, ret
);
3874 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3876 struct io_splice
* sp
= &req
->splice
;
3878 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3879 sp
->off_out
= READ_ONCE(sqe
->off
);
3880 return __io_splice_prep(req
, sqe
);
3883 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3885 struct io_splice
*sp
= &req
->splice
;
3886 struct file
*in
= sp
->file_in
;
3887 struct file
*out
= sp
->file_out
;
3888 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3889 loff_t
*poff_in
, *poff_out
;
3895 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3896 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3899 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3901 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3902 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3905 req_set_fail_links(req
);
3906 io_req_complete(req
, ret
);
3911 * IORING_OP_NOP just posts a completion event, nothing else.
3913 static int io_nop(struct io_kiocb
*req
, struct io_comp_state
*cs
)
3915 struct io_ring_ctx
*ctx
= req
->ctx
;
3917 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3920 __io_req_complete(req
, 0, 0, cs
);
3924 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3926 struct io_ring_ctx
*ctx
= req
->ctx
;
3931 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3933 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3936 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3937 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3940 req
->sync
.off
= READ_ONCE(sqe
->off
);
3941 req
->sync
.len
= READ_ONCE(sqe
->len
);
3945 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3947 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3950 /* fsync always requires a blocking context */
3954 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3955 end
> 0 ? end
: LLONG_MAX
,
3956 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3958 req_set_fail_links(req
);
3959 io_req_complete(req
, ret
);
3963 static int io_fallocate_prep(struct io_kiocb
*req
,
3964 const struct io_uring_sqe
*sqe
)
3966 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3968 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3971 req
->sync
.off
= READ_ONCE(sqe
->off
);
3972 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3973 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3977 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
3981 /* fallocate always requiring blocking context */
3984 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3987 req_set_fail_links(req
);
3988 io_req_complete(req
, ret
);
3992 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3994 const char __user
*fname
;
3997 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3999 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4002 /* open.how should be already initialised */
4003 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4004 req
->open
.how
.flags
|= O_LARGEFILE
;
4006 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4007 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4008 req
->open
.filename
= getname(fname
);
4009 if (IS_ERR(req
->open
.filename
)) {
4010 ret
= PTR_ERR(req
->open
.filename
);
4011 req
->open
.filename
= NULL
;
4014 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4015 req
->flags
|= REQ_F_NEED_CLEANUP
;
4019 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4023 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4025 mode
= READ_ONCE(sqe
->len
);
4026 flags
= READ_ONCE(sqe
->open_flags
);
4027 req
->open
.how
= build_open_how(flags
, mode
);
4028 return __io_openat_prep(req
, sqe
);
4031 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4033 struct open_how __user
*how
;
4037 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4039 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4040 len
= READ_ONCE(sqe
->len
);
4041 if (len
< OPEN_HOW_SIZE_VER0
)
4044 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4049 return __io_openat_prep(req
, sqe
);
4052 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
4054 struct open_flags op
;
4057 bool resolve_nonblock
;
4060 ret
= build_open_flags(&req
->open
.how
, &op
);
4063 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
4064 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
4065 if (force_nonblock
) {
4067 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4068 * it'll always -EAGAIN
4070 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
4072 op
.lookup_flags
|= LOOKUP_CACHED
;
4073 op
.open_flag
|= O_NONBLOCK
;
4076 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4080 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4081 /* only retry if RESOLVE_CACHED wasn't already set by application */
4082 if ((!resolve_nonblock
&& force_nonblock
) && file
== ERR_PTR(-EAGAIN
)) {
4084 * We could hang on to this 'fd', but seems like marginal
4085 * gain for something that is now known to be a slower path.
4086 * So just put it, and we'll get a new one when we retry.
4094 ret
= PTR_ERR(file
);
4096 if (force_nonblock
&& !nonblock_set
)
4097 file
->f_flags
&= ~O_NONBLOCK
;
4098 fsnotify_open(file
);
4099 fd_install(ret
, file
);
4102 putname(req
->open
.filename
);
4103 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4105 req_set_fail_links(req
);
4106 io_req_complete(req
, ret
);
4110 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
4112 return io_openat2(req
, force_nonblock
);
4115 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4116 const struct io_uring_sqe
*sqe
)
4118 struct io_provide_buf
*p
= &req
->pbuf
;
4121 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
4124 tmp
= READ_ONCE(sqe
->fd
);
4125 if (!tmp
|| tmp
> USHRT_MAX
)
4128 memset(p
, 0, sizeof(*p
));
4130 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4134 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4135 int bgid
, unsigned nbufs
)
4139 /* shouldn't happen */
4143 /* the head kbuf is the list itself */
4144 while (!list_empty(&buf
->list
)) {
4145 struct io_buffer
*nxt
;
4147 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4148 list_del(&nxt
->list
);
4155 idr_remove(&ctx
->io_buffer_idr
, bgid
);
4160 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4161 struct io_comp_state
*cs
)
4163 struct io_provide_buf
*p
= &req
->pbuf
;
4164 struct io_ring_ctx
*ctx
= req
->ctx
;
4165 struct io_buffer
*head
;
4168 io_ring_submit_lock(ctx
, !force_nonblock
);
4170 lockdep_assert_held(&ctx
->uring_lock
);
4173 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4175 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4177 req_set_fail_links(req
);
4179 /* need to hold the lock to complete IOPOLL requests */
4180 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4181 __io_req_complete(req
, ret
, 0, cs
);
4182 io_ring_submit_unlock(ctx
, !force_nonblock
);
4184 io_ring_submit_unlock(ctx
, !force_nonblock
);
4185 __io_req_complete(req
, ret
, 0, cs
);
4190 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4191 const struct io_uring_sqe
*sqe
)
4193 struct io_provide_buf
*p
= &req
->pbuf
;
4196 if (sqe
->ioprio
|| sqe
->rw_flags
)
4199 tmp
= READ_ONCE(sqe
->fd
);
4200 if (!tmp
|| tmp
> USHRT_MAX
)
4203 p
->addr
= READ_ONCE(sqe
->addr
);
4204 p
->len
= READ_ONCE(sqe
->len
);
4206 if (!access_ok(u64_to_user_ptr(p
->addr
), (p
->len
* p
->nbufs
)))
4209 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4210 tmp
= READ_ONCE(sqe
->off
);
4211 if (tmp
> USHRT_MAX
)
4217 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4219 struct io_buffer
*buf
;
4220 u64 addr
= pbuf
->addr
;
4221 int i
, bid
= pbuf
->bid
;
4223 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4224 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
4229 buf
->len
= pbuf
->len
;
4234 INIT_LIST_HEAD(&buf
->list
);
4237 list_add_tail(&buf
->list
, &(*head
)->list
);
4241 return i
? i
: -ENOMEM
;
4244 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
,
4245 struct io_comp_state
*cs
)
4247 struct io_provide_buf
*p
= &req
->pbuf
;
4248 struct io_ring_ctx
*ctx
= req
->ctx
;
4249 struct io_buffer
*head
, *list
;
4252 io_ring_submit_lock(ctx
, !force_nonblock
);
4254 lockdep_assert_held(&ctx
->uring_lock
);
4256 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
4258 ret
= io_add_buffers(p
, &head
);
4263 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
4266 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4272 req_set_fail_links(req
);
4274 /* need to hold the lock to complete IOPOLL requests */
4275 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4276 __io_req_complete(req
, ret
, 0, cs
);
4277 io_ring_submit_unlock(ctx
, !force_nonblock
);
4279 io_ring_submit_unlock(ctx
, !force_nonblock
);
4280 __io_req_complete(req
, ret
, 0, cs
);
4285 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4286 const struct io_uring_sqe
*sqe
)
4288 #if defined(CONFIG_EPOLL)
4289 if (sqe
->ioprio
|| sqe
->buf_index
)
4291 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4294 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4295 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4296 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4298 if (ep_op_has_event(req
->epoll
.op
)) {
4299 struct epoll_event __user
*ev
;
4301 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4302 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4312 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
,
4313 struct io_comp_state
*cs
)
4315 #if defined(CONFIG_EPOLL)
4316 struct io_epoll
*ie
= &req
->epoll
;
4319 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4320 if (force_nonblock
&& ret
== -EAGAIN
)
4324 req_set_fail_links(req
);
4325 __io_req_complete(req
, ret
, 0, cs
);
4332 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4334 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4335 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4337 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4340 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4341 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4342 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4349 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
4351 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4352 struct io_madvise
*ma
= &req
->madvise
;
4358 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4360 req_set_fail_links(req
);
4361 io_req_complete(req
, ret
);
4368 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4370 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4372 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4375 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4376 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4377 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4381 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
4383 struct io_fadvise
*fa
= &req
->fadvise
;
4386 if (force_nonblock
) {
4387 switch (fa
->advice
) {
4388 case POSIX_FADV_NORMAL
:
4389 case POSIX_FADV_RANDOM
:
4390 case POSIX_FADV_SEQUENTIAL
:
4397 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4399 req_set_fail_links(req
);
4400 io_req_complete(req
, ret
);
4404 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4406 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4408 if (sqe
->ioprio
|| sqe
->buf_index
)
4410 if (req
->flags
& REQ_F_FIXED_FILE
)
4413 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4414 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4415 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4416 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4417 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4422 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
4424 struct io_statx
*ctx
= &req
->statx
;
4427 if (force_nonblock
) {
4428 /* only need file table for an actual valid fd */
4429 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4430 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4434 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4438 req_set_fail_links(req
);
4439 io_req_complete(req
, ret
);
4443 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4445 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4447 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4448 sqe
->rw_flags
|| sqe
->buf_index
)
4450 if (req
->flags
& REQ_F_FIXED_FILE
)
4453 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4457 static int io_close(struct io_kiocb
*req
, bool force_nonblock
,
4458 struct io_comp_state
*cs
)
4460 struct files_struct
*files
= current
->files
;
4461 struct io_close
*close
= &req
->close
;
4462 struct fdtable
*fdt
;
4468 spin_lock(&files
->file_lock
);
4469 fdt
= files_fdtable(files
);
4470 if (close
->fd
>= fdt
->max_fds
) {
4471 spin_unlock(&files
->file_lock
);
4474 file
= fdt
->fd
[close
->fd
];
4476 spin_unlock(&files
->file_lock
);
4480 if (file
->f_op
== &io_uring_fops
) {
4481 spin_unlock(&files
->file_lock
);
4486 /* if the file has a flush method, be safe and punt to async */
4487 if (file
->f_op
->flush
&& force_nonblock
) {
4488 spin_unlock(&files
->file_lock
);
4492 ret
= __close_fd_get_file(close
->fd
, &file
);
4493 spin_unlock(&files
->file_lock
);
4500 /* No ->flush() or already async, safely close from here */
4501 ret
= filp_close(file
, current
->files
);
4504 req_set_fail_links(req
);
4507 __io_req_complete(req
, ret
, 0, cs
);
4511 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4513 struct io_ring_ctx
*ctx
= req
->ctx
;
4518 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4520 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4523 req
->sync
.off
= READ_ONCE(sqe
->off
);
4524 req
->sync
.len
= READ_ONCE(sqe
->len
);
4525 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4529 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
4533 /* sync_file_range always requires a blocking context */
4537 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4540 req_set_fail_links(req
);
4541 io_req_complete(req
, ret
);
4545 #if defined(CONFIG_NET)
4546 static int io_setup_async_msg(struct io_kiocb
*req
,
4547 struct io_async_msghdr
*kmsg
)
4549 struct io_async_msghdr
*async_msg
= req
->async_data
;
4553 if (io_alloc_async_data(req
)) {
4554 if (kmsg
->iov
!= kmsg
->fast_iov
)
4558 async_msg
= req
->async_data
;
4559 req
->flags
|= REQ_F_NEED_CLEANUP
;
4560 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4564 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4565 struct io_async_msghdr
*iomsg
)
4567 iomsg
->iov
= iomsg
->fast_iov
;
4568 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4569 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4570 req
->sr_msg
.msg_flags
, &iomsg
->iov
);
4573 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4575 struct io_async_msghdr
*async_msg
= req
->async_data
;
4576 struct io_sr_msg
*sr
= &req
->sr_msg
;
4579 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4582 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4583 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4584 sr
->len
= READ_ONCE(sqe
->len
);
4586 #ifdef CONFIG_COMPAT
4587 if (req
->ctx
->compat
)
4588 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4591 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4593 ret
= io_sendmsg_copy_hdr(req
, async_msg
);
4595 req
->flags
|= REQ_F_NEED_CLEANUP
;
4599 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4600 struct io_comp_state
*cs
)
4602 struct io_async_msghdr iomsg
, *kmsg
;
4603 struct socket
*sock
;
4607 sock
= sock_from_file(req
->file
);
4608 if (unlikely(!sock
))
4611 if (req
->async_data
) {
4612 kmsg
= req
->async_data
;
4613 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4614 /* if iov is set, it's allocated already */
4616 kmsg
->iov
= kmsg
->fast_iov
;
4617 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4619 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4625 flags
= req
->sr_msg
.msg_flags
;
4626 if (flags
& MSG_DONTWAIT
)
4627 req
->flags
|= REQ_F_NOWAIT
;
4628 else if (force_nonblock
)
4629 flags
|= MSG_DONTWAIT
;
4631 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4632 if (force_nonblock
&& ret
== -EAGAIN
)
4633 return io_setup_async_msg(req
, kmsg
);
4634 if (ret
== -ERESTARTSYS
)
4637 if (kmsg
->iov
!= kmsg
->fast_iov
)
4639 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4641 req_set_fail_links(req
);
4642 __io_req_complete(req
, ret
, 0, cs
);
4646 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4647 struct io_comp_state
*cs
)
4649 struct io_sr_msg
*sr
= &req
->sr_msg
;
4652 struct socket
*sock
;
4656 sock
= sock_from_file(req
->file
);
4657 if (unlikely(!sock
))
4660 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4664 msg
.msg_name
= NULL
;
4665 msg
.msg_control
= NULL
;
4666 msg
.msg_controllen
= 0;
4667 msg
.msg_namelen
= 0;
4669 flags
= req
->sr_msg
.msg_flags
;
4670 if (flags
& MSG_DONTWAIT
)
4671 req
->flags
|= REQ_F_NOWAIT
;
4672 else if (force_nonblock
)
4673 flags
|= MSG_DONTWAIT
;
4675 msg
.msg_flags
= flags
;
4676 ret
= sock_sendmsg(sock
, &msg
);
4677 if (force_nonblock
&& ret
== -EAGAIN
)
4679 if (ret
== -ERESTARTSYS
)
4683 req_set_fail_links(req
);
4684 __io_req_complete(req
, ret
, 0, cs
);
4688 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4689 struct io_async_msghdr
*iomsg
)
4691 struct io_sr_msg
*sr
= &req
->sr_msg
;
4692 struct iovec __user
*uiov
;
4696 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4697 &iomsg
->uaddr
, &uiov
, &iov_len
);
4701 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4704 if (copy_from_user(iomsg
->iov
, uiov
, sizeof(*uiov
)))
4706 sr
->len
= iomsg
->iov
[0].iov_len
;
4707 iov_iter_init(&iomsg
->msg
.msg_iter
, READ
, iomsg
->iov
, 1,
4711 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4712 &iomsg
->iov
, &iomsg
->msg
.msg_iter
,
4721 #ifdef CONFIG_COMPAT
4722 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4723 struct io_async_msghdr
*iomsg
)
4725 struct compat_msghdr __user
*msg_compat
;
4726 struct io_sr_msg
*sr
= &req
->sr_msg
;
4727 struct compat_iovec __user
*uiov
;
4732 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4733 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4738 uiov
= compat_ptr(ptr
);
4739 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4740 compat_ssize_t clen
;
4744 if (!access_ok(uiov
, sizeof(*uiov
)))
4746 if (__get_user(clen
, &uiov
->iov_len
))
4751 iomsg
->iov
[0].iov_len
= clen
;
4754 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4755 UIO_FASTIOV
, &iomsg
->iov
,
4756 &iomsg
->msg
.msg_iter
, true);
4765 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4766 struct io_async_msghdr
*iomsg
)
4768 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4769 iomsg
->iov
= iomsg
->fast_iov
;
4771 #ifdef CONFIG_COMPAT
4772 if (req
->ctx
->compat
)
4773 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4776 return __io_recvmsg_copy_hdr(req
, iomsg
);
4779 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4782 struct io_sr_msg
*sr
= &req
->sr_msg
;
4783 struct io_buffer
*kbuf
;
4785 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4790 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4794 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4796 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4799 static int io_recvmsg_prep(struct io_kiocb
*req
,
4800 const struct io_uring_sqe
*sqe
)
4802 struct io_async_msghdr
*async_msg
= req
->async_data
;
4803 struct io_sr_msg
*sr
= &req
->sr_msg
;
4806 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4809 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4810 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4811 sr
->len
= READ_ONCE(sqe
->len
);
4812 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4814 #ifdef CONFIG_COMPAT
4815 if (req
->ctx
->compat
)
4816 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4819 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4821 ret
= io_recvmsg_copy_hdr(req
, async_msg
);
4823 req
->flags
|= REQ_F_NEED_CLEANUP
;
4827 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
4828 struct io_comp_state
*cs
)
4830 struct io_async_msghdr iomsg
, *kmsg
;
4831 struct socket
*sock
;
4832 struct io_buffer
*kbuf
;
4834 int ret
, cflags
= 0;
4836 sock
= sock_from_file(req
->file
);
4837 if (unlikely(!sock
))
4840 if (req
->async_data
) {
4841 kmsg
= req
->async_data
;
4842 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4843 /* if iov is set, it's allocated already */
4845 kmsg
->iov
= kmsg
->fast_iov
;
4846 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4848 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4854 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4855 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4857 return PTR_ERR(kbuf
);
4858 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4859 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4860 1, req
->sr_msg
.len
);
4863 flags
= req
->sr_msg
.msg_flags
;
4864 if (flags
& MSG_DONTWAIT
)
4865 req
->flags
|= REQ_F_NOWAIT
;
4866 else if (force_nonblock
)
4867 flags
|= MSG_DONTWAIT
;
4869 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4870 kmsg
->uaddr
, flags
);
4871 if (force_nonblock
&& ret
== -EAGAIN
)
4872 return io_setup_async_msg(req
, kmsg
);
4873 if (ret
== -ERESTARTSYS
)
4876 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4877 cflags
= io_put_recv_kbuf(req
);
4878 if (kmsg
->iov
!= kmsg
->fast_iov
)
4880 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4882 req_set_fail_links(req
);
4883 __io_req_complete(req
, ret
, cflags
, cs
);
4887 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
4888 struct io_comp_state
*cs
)
4890 struct io_buffer
*kbuf
;
4891 struct io_sr_msg
*sr
= &req
->sr_msg
;
4893 void __user
*buf
= sr
->buf
;
4894 struct socket
*sock
;
4897 int ret
, cflags
= 0;
4899 sock
= sock_from_file(req
->file
);
4900 if (unlikely(!sock
))
4903 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4904 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4906 return PTR_ERR(kbuf
);
4907 buf
= u64_to_user_ptr(kbuf
->addr
);
4910 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4914 msg
.msg_name
= NULL
;
4915 msg
.msg_control
= NULL
;
4916 msg
.msg_controllen
= 0;
4917 msg
.msg_namelen
= 0;
4918 msg
.msg_iocb
= NULL
;
4921 flags
= req
->sr_msg
.msg_flags
;
4922 if (flags
& MSG_DONTWAIT
)
4923 req
->flags
|= REQ_F_NOWAIT
;
4924 else if (force_nonblock
)
4925 flags
|= MSG_DONTWAIT
;
4927 ret
= sock_recvmsg(sock
, &msg
, flags
);
4928 if (force_nonblock
&& ret
== -EAGAIN
)
4930 if (ret
== -ERESTARTSYS
)
4933 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4934 cflags
= io_put_recv_kbuf(req
);
4936 req_set_fail_links(req
);
4937 __io_req_complete(req
, ret
, cflags
, cs
);
4941 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4943 struct io_accept
*accept
= &req
->accept
;
4945 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4947 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4950 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4951 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4952 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4953 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4957 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
4958 struct io_comp_state
*cs
)
4960 struct io_accept
*accept
= &req
->accept
;
4961 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4964 if (req
->file
->f_flags
& O_NONBLOCK
)
4965 req
->flags
|= REQ_F_NOWAIT
;
4967 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4968 accept
->addr_len
, accept
->flags
,
4970 if (ret
== -EAGAIN
&& force_nonblock
)
4973 if (ret
== -ERESTARTSYS
)
4975 req_set_fail_links(req
);
4977 __io_req_complete(req
, ret
, 0, cs
);
4981 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4983 struct io_connect
*conn
= &req
->connect
;
4984 struct io_async_connect
*io
= req
->async_data
;
4986 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4988 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4991 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4992 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4997 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
5001 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5002 struct io_comp_state
*cs
)
5004 struct io_async_connect __io
, *io
;
5005 unsigned file_flags
;
5008 if (req
->async_data
) {
5009 io
= req
->async_data
;
5011 ret
= move_addr_to_kernel(req
->connect
.addr
,
5012 req
->connect
.addr_len
,
5019 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5021 ret
= __sys_connect_file(req
->file
, &io
->address
,
5022 req
->connect
.addr_len
, file_flags
);
5023 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5024 if (req
->async_data
)
5026 if (io_alloc_async_data(req
)) {
5030 io
= req
->async_data
;
5031 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5034 if (ret
== -ERESTARTSYS
)
5038 req_set_fail_links(req
);
5039 __io_req_complete(req
, ret
, 0, cs
);
5042 #else /* !CONFIG_NET */
5043 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5048 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
5049 struct io_comp_state
*cs
)
5054 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
5055 struct io_comp_state
*cs
)
5060 static int io_recvmsg_prep(struct io_kiocb
*req
,
5061 const struct io_uring_sqe
*sqe
)
5066 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
5067 struct io_comp_state
*cs
)
5072 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
5073 struct io_comp_state
*cs
)
5078 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5083 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
5084 struct io_comp_state
*cs
)
5089 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5094 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
5095 struct io_comp_state
*cs
)
5099 #endif /* CONFIG_NET */
5101 struct io_poll_table
{
5102 struct poll_table_struct pt
;
5103 struct io_kiocb
*req
;
5107 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5108 __poll_t mask
, task_work_func_t func
)
5112 /* for instances that support it check for an event match first: */
5113 if (mask
&& !(mask
& poll
->events
))
5116 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5118 list_del_init(&poll
->wait
.entry
);
5121 init_task_work(&req
->task_work
, func
);
5122 percpu_ref_get(&req
->ctx
->refs
);
5125 * If this fails, then the task is exiting. When a task exits, the
5126 * work gets canceled, so just cancel this request as well instead
5127 * of executing it. We can't safely execute it anyway, as we may not
5128 * have the needed state needed for it anyway.
5130 ret
= io_req_task_work_add(req
);
5131 if (unlikely(ret
)) {
5132 WRITE_ONCE(poll
->canceled
, true);
5133 io_req_task_work_add_fallback(req
, func
);
5138 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5139 __acquires(&req
->ctx
->completion_lock
)
5141 struct io_ring_ctx
*ctx
= req
->ctx
;
5143 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5144 struct poll_table_struct pt
= { ._key
= poll
->events
};
5146 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5149 spin_lock_irq(&ctx
->completion_lock
);
5150 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5151 add_wait_queue(poll
->head
, &poll
->wait
);
5158 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5160 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5161 if (req
->opcode
== IORING_OP_POLL_ADD
)
5162 return req
->async_data
;
5163 return req
->apoll
->double_poll
;
5166 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5168 if (req
->opcode
== IORING_OP_POLL_ADD
)
5170 return &req
->apoll
->poll
;
5173 static void io_poll_remove_double(struct io_kiocb
*req
)
5175 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5177 lockdep_assert_held(&req
->ctx
->completion_lock
);
5179 if (poll
&& poll
->head
) {
5180 struct wait_queue_head
*head
= poll
->head
;
5182 spin_lock(&head
->lock
);
5183 list_del_init(&poll
->wait
.entry
);
5184 if (poll
->wait
.private)
5185 refcount_dec(&req
->refs
);
5187 spin_unlock(&head
->lock
);
5191 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
5193 struct io_ring_ctx
*ctx
= req
->ctx
;
5195 io_poll_remove_double(req
);
5196 req
->poll
.done
= true;
5197 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
5198 io_commit_cqring(ctx
);
5201 static void io_poll_task_func(struct callback_head
*cb
)
5203 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5204 struct io_ring_ctx
*ctx
= req
->ctx
;
5205 struct io_kiocb
*nxt
;
5207 if (io_poll_rewait(req
, &req
->poll
)) {
5208 spin_unlock_irq(&ctx
->completion_lock
);
5210 hash_del(&req
->hash_node
);
5211 io_poll_complete(req
, req
->result
, 0);
5212 spin_unlock_irq(&ctx
->completion_lock
);
5214 nxt
= io_put_req_find_next(req
);
5215 io_cqring_ev_posted(ctx
);
5217 __io_req_task_submit(nxt
);
5220 percpu_ref_put(&ctx
->refs
);
5223 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5224 int sync
, void *key
)
5226 struct io_kiocb
*req
= wait
->private;
5227 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5228 __poll_t mask
= key_to_poll(key
);
5230 /* for instances that support it check for an event match first: */
5231 if (mask
&& !(mask
& poll
->events
))
5234 list_del_init(&wait
->entry
);
5236 if (poll
&& poll
->head
) {
5239 spin_lock(&poll
->head
->lock
);
5240 done
= list_empty(&poll
->wait
.entry
);
5242 list_del_init(&poll
->wait
.entry
);
5243 /* make sure double remove sees this as being gone */
5244 wait
->private = NULL
;
5245 spin_unlock(&poll
->head
->lock
);
5247 /* use wait func handler, so it matches the rq type */
5248 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5251 refcount_dec(&req
->refs
);
5255 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5256 wait_queue_func_t wake_func
)
5260 poll
->canceled
= false;
5261 poll
->events
= events
;
5262 INIT_LIST_HEAD(&poll
->wait
.entry
);
5263 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5266 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5267 struct wait_queue_head
*head
,
5268 struct io_poll_iocb
**poll_ptr
)
5270 struct io_kiocb
*req
= pt
->req
;
5273 * If poll->head is already set, it's because the file being polled
5274 * uses multiple waitqueues for poll handling (eg one for read, one
5275 * for write). Setup a separate io_poll_iocb if this happens.
5277 if (unlikely(poll
->head
)) {
5278 struct io_poll_iocb
*poll_one
= poll
;
5280 /* already have a 2nd entry, fail a third attempt */
5282 pt
->error
= -EINVAL
;
5285 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5287 pt
->error
= -ENOMEM
;
5290 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5291 refcount_inc(&req
->refs
);
5292 poll
->wait
.private = req
;
5299 if (poll
->events
& EPOLLEXCLUSIVE
)
5300 add_wait_queue_exclusive(head
, &poll
->wait
);
5302 add_wait_queue(head
, &poll
->wait
);
5305 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5306 struct poll_table_struct
*p
)
5308 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5309 struct async_poll
*apoll
= pt
->req
->apoll
;
5311 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5314 static void io_async_task_func(struct callback_head
*cb
)
5316 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5317 struct async_poll
*apoll
= req
->apoll
;
5318 struct io_ring_ctx
*ctx
= req
->ctx
;
5320 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5322 if (io_poll_rewait(req
, &apoll
->poll
)) {
5323 spin_unlock_irq(&ctx
->completion_lock
);
5324 percpu_ref_put(&ctx
->refs
);
5328 /* If req is still hashed, it cannot have been canceled. Don't check. */
5329 if (hash_hashed(&req
->hash_node
))
5330 hash_del(&req
->hash_node
);
5332 io_poll_remove_double(req
);
5333 spin_unlock_irq(&ctx
->completion_lock
);
5335 if (!READ_ONCE(apoll
->poll
.canceled
))
5336 __io_req_task_submit(req
);
5338 __io_req_task_cancel(req
, -ECANCELED
);
5340 percpu_ref_put(&ctx
->refs
);
5341 kfree(apoll
->double_poll
);
5345 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5348 struct io_kiocb
*req
= wait
->private;
5349 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5351 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5354 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5357 static void io_poll_req_insert(struct io_kiocb
*req
)
5359 struct io_ring_ctx
*ctx
= req
->ctx
;
5360 struct hlist_head
*list
;
5362 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5363 hlist_add_head(&req
->hash_node
, list
);
5366 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5367 struct io_poll_iocb
*poll
,
5368 struct io_poll_table
*ipt
, __poll_t mask
,
5369 wait_queue_func_t wake_func
)
5370 __acquires(&ctx
->completion_lock
)
5372 struct io_ring_ctx
*ctx
= req
->ctx
;
5373 bool cancel
= false;
5375 INIT_HLIST_NODE(&req
->hash_node
);
5376 io_init_poll_iocb(poll
, mask
, wake_func
);
5377 poll
->file
= req
->file
;
5378 poll
->wait
.private = req
;
5380 ipt
->pt
._key
= mask
;
5382 ipt
->error
= -EINVAL
;
5384 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5386 spin_lock_irq(&ctx
->completion_lock
);
5387 if (likely(poll
->head
)) {
5388 spin_lock(&poll
->head
->lock
);
5389 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5395 if (mask
|| ipt
->error
)
5396 list_del_init(&poll
->wait
.entry
);
5398 WRITE_ONCE(poll
->canceled
, true);
5399 else if (!poll
->done
) /* actually waiting for an event */
5400 io_poll_req_insert(req
);
5401 spin_unlock(&poll
->head
->lock
);
5407 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5409 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5410 struct io_ring_ctx
*ctx
= req
->ctx
;
5411 struct async_poll
*apoll
;
5412 struct io_poll_table ipt
;
5416 if (!req
->file
|| !file_can_poll(req
->file
))
5418 if (req
->flags
& REQ_F_POLLED
)
5422 else if (def
->pollout
)
5426 /* if we can't nonblock try, then no point in arming a poll handler */
5427 if (!io_file_supports_async(req
->file
, rw
))
5430 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5431 if (unlikely(!apoll
))
5433 apoll
->double_poll
= NULL
;
5435 req
->flags
|= REQ_F_POLLED
;
5440 mask
|= POLLIN
| POLLRDNORM
;
5442 mask
|= POLLOUT
| POLLWRNORM
;
5444 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5445 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5446 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5449 mask
|= POLLERR
| POLLPRI
;
5451 ipt
.pt
._qproc
= io_async_queue_proc
;
5453 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5455 if (ret
|| ipt
.error
) {
5456 io_poll_remove_double(req
);
5457 spin_unlock_irq(&ctx
->completion_lock
);
5458 kfree(apoll
->double_poll
);
5462 spin_unlock_irq(&ctx
->completion_lock
);
5463 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5464 apoll
->poll
.events
);
5468 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5469 struct io_poll_iocb
*poll
)
5471 bool do_complete
= false;
5473 spin_lock(&poll
->head
->lock
);
5474 WRITE_ONCE(poll
->canceled
, true);
5475 if (!list_empty(&poll
->wait
.entry
)) {
5476 list_del_init(&poll
->wait
.entry
);
5479 spin_unlock(&poll
->head
->lock
);
5480 hash_del(&req
->hash_node
);
5484 static bool io_poll_remove_one(struct io_kiocb
*req
)
5488 io_poll_remove_double(req
);
5490 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5491 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5493 struct async_poll
*apoll
= req
->apoll
;
5495 /* non-poll requests have submit ref still */
5496 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5499 kfree(apoll
->double_poll
);
5505 io_cqring_fill_event(req
, -ECANCELED
);
5506 io_commit_cqring(req
->ctx
);
5507 req_set_fail_links(req
);
5508 io_put_req_deferred(req
, 1);
5515 * Returns true if we found and killed one or more poll requests
5517 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5518 struct files_struct
*files
)
5520 struct hlist_node
*tmp
;
5521 struct io_kiocb
*req
;
5524 spin_lock_irq(&ctx
->completion_lock
);
5525 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5526 struct hlist_head
*list
;
5528 list
= &ctx
->cancel_hash
[i
];
5529 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5530 if (io_match_task(req
, tsk
, files
))
5531 posted
+= io_poll_remove_one(req
);
5534 spin_unlock_irq(&ctx
->completion_lock
);
5537 io_cqring_ev_posted(ctx
);
5542 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5544 struct hlist_head
*list
;
5545 struct io_kiocb
*req
;
5547 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5548 hlist_for_each_entry(req
, list
, hash_node
) {
5549 if (sqe_addr
!= req
->user_data
)
5551 if (io_poll_remove_one(req
))
5559 static int io_poll_remove_prep(struct io_kiocb
*req
,
5560 const struct io_uring_sqe
*sqe
)
5562 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5564 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5568 req
->poll_remove
.addr
= READ_ONCE(sqe
->addr
);
5573 * Find a running poll command that matches one specified in sqe->addr,
5574 * and remove it if found.
5576 static int io_poll_remove(struct io_kiocb
*req
)
5578 struct io_ring_ctx
*ctx
= req
->ctx
;
5581 spin_lock_irq(&ctx
->completion_lock
);
5582 ret
= io_poll_cancel(ctx
, req
->poll_remove
.addr
);
5583 spin_unlock_irq(&ctx
->completion_lock
);
5586 req_set_fail_links(req
);
5587 io_req_complete(req
, ret
);
5591 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5594 struct io_kiocb
*req
= wait
->private;
5595 struct io_poll_iocb
*poll
= &req
->poll
;
5597 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5600 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5601 struct poll_table_struct
*p
)
5603 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5605 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5608 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5610 struct io_poll_iocb
*poll
= &req
->poll
;
5613 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5615 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5618 events
= READ_ONCE(sqe
->poll32_events
);
5620 events
= swahw32(events
);
5622 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5623 (events
& EPOLLEXCLUSIVE
);
5627 static int io_poll_add(struct io_kiocb
*req
)
5629 struct io_poll_iocb
*poll
= &req
->poll
;
5630 struct io_ring_ctx
*ctx
= req
->ctx
;
5631 struct io_poll_table ipt
;
5634 ipt
.pt
._qproc
= io_poll_queue_proc
;
5636 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5639 if (mask
) { /* no async, we'd stolen it */
5641 io_poll_complete(req
, mask
, 0);
5643 spin_unlock_irq(&ctx
->completion_lock
);
5646 io_cqring_ev_posted(ctx
);
5652 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5654 struct io_timeout_data
*data
= container_of(timer
,
5655 struct io_timeout_data
, timer
);
5656 struct io_kiocb
*req
= data
->req
;
5657 struct io_ring_ctx
*ctx
= req
->ctx
;
5658 unsigned long flags
;
5660 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5661 list_del_init(&req
->timeout
.list
);
5662 atomic_set(&req
->ctx
->cq_timeouts
,
5663 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5665 io_cqring_fill_event(req
, -ETIME
);
5666 io_commit_cqring(ctx
);
5667 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5669 io_cqring_ev_posted(ctx
);
5670 req_set_fail_links(req
);
5672 return HRTIMER_NORESTART
;
5675 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5678 struct io_timeout_data
*io
;
5679 struct io_kiocb
*req
;
5682 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5683 if (user_data
== req
->user_data
) {
5690 return ERR_PTR(ret
);
5692 io
= req
->async_data
;
5693 ret
= hrtimer_try_to_cancel(&io
->timer
);
5695 return ERR_PTR(-EALREADY
);
5696 list_del_init(&req
->timeout
.list
);
5700 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5702 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5705 return PTR_ERR(req
);
5707 req_set_fail_links(req
);
5708 io_cqring_fill_event(req
, -ECANCELED
);
5709 io_put_req_deferred(req
, 1);
5713 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5714 struct timespec64
*ts
, enum hrtimer_mode mode
)
5716 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5717 struct io_timeout_data
*data
;
5720 return PTR_ERR(req
);
5722 req
->timeout
.off
= 0; /* noseq */
5723 data
= req
->async_data
;
5724 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5725 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5726 data
->timer
.function
= io_timeout_fn
;
5727 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5731 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5732 const struct io_uring_sqe
*sqe
)
5734 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5736 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5738 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5740 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
5743 tr
->addr
= READ_ONCE(sqe
->addr
);
5744 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5745 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5746 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5748 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5750 } else if (tr
->flags
) {
5751 /* timeout removal doesn't support flags */
5758 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
5760 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
5765 * Remove or update an existing timeout command
5767 static int io_timeout_remove(struct io_kiocb
*req
)
5769 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5770 struct io_ring_ctx
*ctx
= req
->ctx
;
5773 spin_lock_irq(&ctx
->completion_lock
);
5774 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
))
5775 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5777 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
,
5778 io_translate_timeout_mode(tr
->flags
));
5780 io_cqring_fill_event(req
, ret
);
5781 io_commit_cqring(ctx
);
5782 spin_unlock_irq(&ctx
->completion_lock
);
5783 io_cqring_ev_posted(ctx
);
5785 req_set_fail_links(req
);
5790 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5791 bool is_timeout_link
)
5793 struct io_timeout_data
*data
;
5795 u32 off
= READ_ONCE(sqe
->off
);
5797 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5799 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5801 if (off
&& is_timeout_link
)
5803 flags
= READ_ONCE(sqe
->timeout_flags
);
5804 if (flags
& ~IORING_TIMEOUT_ABS
)
5807 req
->timeout
.off
= off
;
5809 if (!req
->async_data
&& io_alloc_async_data(req
))
5812 data
= req
->async_data
;
5815 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5818 data
->mode
= io_translate_timeout_mode(flags
);
5819 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5823 static int io_timeout(struct io_kiocb
*req
)
5825 struct io_ring_ctx
*ctx
= req
->ctx
;
5826 struct io_timeout_data
*data
= req
->async_data
;
5827 struct list_head
*entry
;
5828 u32 tail
, off
= req
->timeout
.off
;
5830 spin_lock_irq(&ctx
->completion_lock
);
5833 * sqe->off holds how many events that need to occur for this
5834 * timeout event to be satisfied. If it isn't set, then this is
5835 * a pure timeout request, sequence isn't used.
5837 if (io_is_timeout_noseq(req
)) {
5838 entry
= ctx
->timeout_list
.prev
;
5842 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5843 req
->timeout
.target_seq
= tail
+ off
;
5845 /* Update the last seq here in case io_flush_timeouts() hasn't.
5846 * This is safe because ->completion_lock is held, and submissions
5847 * and completions are never mixed in the same ->completion_lock section.
5849 ctx
->cq_last_tm_flush
= tail
;
5852 * Insertion sort, ensuring the first entry in the list is always
5853 * the one we need first.
5855 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5856 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5859 if (io_is_timeout_noseq(nxt
))
5861 /* nxt.seq is behind @tail, otherwise would've been completed */
5862 if (off
>= nxt
->timeout
.target_seq
- tail
)
5866 list_add(&req
->timeout
.list
, entry
);
5867 data
->timer
.function
= io_timeout_fn
;
5868 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5869 spin_unlock_irq(&ctx
->completion_lock
);
5873 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5875 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5877 return req
->user_data
== (unsigned long) data
;
5880 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5882 enum io_wq_cancel cancel_ret
;
5885 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5886 switch (cancel_ret
) {
5887 case IO_WQ_CANCEL_OK
:
5890 case IO_WQ_CANCEL_RUNNING
:
5893 case IO_WQ_CANCEL_NOTFOUND
:
5901 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5902 struct io_kiocb
*req
, __u64 sqe_addr
,
5905 unsigned long flags
;
5908 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5909 if (ret
!= -ENOENT
) {
5910 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5914 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5915 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5918 ret
= io_poll_cancel(ctx
, sqe_addr
);
5922 io_cqring_fill_event(req
, ret
);
5923 io_commit_cqring(ctx
);
5924 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5925 io_cqring_ev_posted(ctx
);
5928 req_set_fail_links(req
);
5932 static int io_async_cancel_prep(struct io_kiocb
*req
,
5933 const struct io_uring_sqe
*sqe
)
5935 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5937 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5939 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5942 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5946 static int io_async_cancel(struct io_kiocb
*req
)
5948 struct io_ring_ctx
*ctx
= req
->ctx
;
5950 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
5954 static int io_rsrc_update_prep(struct io_kiocb
*req
,
5955 const struct io_uring_sqe
*sqe
)
5957 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
5959 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5961 if (sqe
->ioprio
|| sqe
->rw_flags
)
5964 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
5965 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
5966 if (!req
->rsrc_update
.nr_args
)
5968 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
5972 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
,
5973 struct io_comp_state
*cs
)
5975 struct io_ring_ctx
*ctx
= req
->ctx
;
5976 struct io_uring_rsrc_update up
;
5982 up
.offset
= req
->rsrc_update
.offset
;
5983 up
.data
= req
->rsrc_update
.arg
;
5985 mutex_lock(&ctx
->uring_lock
);
5986 ret
= __io_sqe_files_update(ctx
, &up
, req
->rsrc_update
.nr_args
);
5987 mutex_unlock(&ctx
->uring_lock
);
5990 req_set_fail_links(req
);
5991 __io_req_complete(req
, ret
, 0, cs
);
5995 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5997 switch (req
->opcode
) {
6000 case IORING_OP_READV
:
6001 case IORING_OP_READ_FIXED
:
6002 case IORING_OP_READ
:
6003 return io_read_prep(req
, sqe
);
6004 case IORING_OP_WRITEV
:
6005 case IORING_OP_WRITE_FIXED
:
6006 case IORING_OP_WRITE
:
6007 return io_write_prep(req
, sqe
);
6008 case IORING_OP_POLL_ADD
:
6009 return io_poll_add_prep(req
, sqe
);
6010 case IORING_OP_POLL_REMOVE
:
6011 return io_poll_remove_prep(req
, sqe
);
6012 case IORING_OP_FSYNC
:
6013 return io_prep_fsync(req
, sqe
);
6014 case IORING_OP_SYNC_FILE_RANGE
:
6015 return io_prep_sfr(req
, sqe
);
6016 case IORING_OP_SENDMSG
:
6017 case IORING_OP_SEND
:
6018 return io_sendmsg_prep(req
, sqe
);
6019 case IORING_OP_RECVMSG
:
6020 case IORING_OP_RECV
:
6021 return io_recvmsg_prep(req
, sqe
);
6022 case IORING_OP_CONNECT
:
6023 return io_connect_prep(req
, sqe
);
6024 case IORING_OP_TIMEOUT
:
6025 return io_timeout_prep(req
, sqe
, false);
6026 case IORING_OP_TIMEOUT_REMOVE
:
6027 return io_timeout_remove_prep(req
, sqe
);
6028 case IORING_OP_ASYNC_CANCEL
:
6029 return io_async_cancel_prep(req
, sqe
);
6030 case IORING_OP_LINK_TIMEOUT
:
6031 return io_timeout_prep(req
, sqe
, true);
6032 case IORING_OP_ACCEPT
:
6033 return io_accept_prep(req
, sqe
);
6034 case IORING_OP_FALLOCATE
:
6035 return io_fallocate_prep(req
, sqe
);
6036 case IORING_OP_OPENAT
:
6037 return io_openat_prep(req
, sqe
);
6038 case IORING_OP_CLOSE
:
6039 return io_close_prep(req
, sqe
);
6040 case IORING_OP_FILES_UPDATE
:
6041 return io_rsrc_update_prep(req
, sqe
);
6042 case IORING_OP_STATX
:
6043 return io_statx_prep(req
, sqe
);
6044 case IORING_OP_FADVISE
:
6045 return io_fadvise_prep(req
, sqe
);
6046 case IORING_OP_MADVISE
:
6047 return io_madvise_prep(req
, sqe
);
6048 case IORING_OP_OPENAT2
:
6049 return io_openat2_prep(req
, sqe
);
6050 case IORING_OP_EPOLL_CTL
:
6051 return io_epoll_ctl_prep(req
, sqe
);
6052 case IORING_OP_SPLICE
:
6053 return io_splice_prep(req
, sqe
);
6054 case IORING_OP_PROVIDE_BUFFERS
:
6055 return io_provide_buffers_prep(req
, sqe
);
6056 case IORING_OP_REMOVE_BUFFERS
:
6057 return io_remove_buffers_prep(req
, sqe
);
6059 return io_tee_prep(req
, sqe
);
6060 case IORING_OP_SHUTDOWN
:
6061 return io_shutdown_prep(req
, sqe
);
6062 case IORING_OP_RENAMEAT
:
6063 return io_renameat_prep(req
, sqe
);
6064 case IORING_OP_UNLINKAT
:
6065 return io_unlinkat_prep(req
, sqe
);
6068 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6073 static int io_req_defer_prep(struct io_kiocb
*req
,
6074 const struct io_uring_sqe
*sqe
)
6078 if (io_alloc_async_data(req
))
6080 return io_req_prep(req
, sqe
);
6083 static u32
io_get_sequence(struct io_kiocb
*req
)
6085 struct io_kiocb
*pos
;
6086 struct io_ring_ctx
*ctx
= req
->ctx
;
6087 u32 total_submitted
, nr_reqs
= 0;
6089 io_for_each_link(pos
, req
)
6092 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
6093 return total_submitted
- nr_reqs
;
6096 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6098 struct io_ring_ctx
*ctx
= req
->ctx
;
6099 struct io_defer_entry
*de
;
6103 /* Still need defer if there is pending req in defer list. */
6104 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6105 !(req
->flags
& REQ_F_IO_DRAIN
)))
6108 seq
= io_get_sequence(req
);
6109 /* Still a chance to pass the sequence check */
6110 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6113 if (!req
->async_data
) {
6114 ret
= io_req_defer_prep(req
, sqe
);
6118 io_prep_async_link(req
);
6119 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6123 spin_lock_irq(&ctx
->completion_lock
);
6124 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6125 spin_unlock_irq(&ctx
->completion_lock
);
6127 io_queue_async_work(req
);
6128 return -EIOCBQUEUED
;
6131 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6134 list_add_tail(&de
->list
, &ctx
->defer_list
);
6135 spin_unlock_irq(&ctx
->completion_lock
);
6136 return -EIOCBQUEUED
;
6139 static void __io_clean_op(struct io_kiocb
*req
)
6141 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6142 switch (req
->opcode
) {
6143 case IORING_OP_READV
:
6144 case IORING_OP_READ_FIXED
:
6145 case IORING_OP_READ
:
6146 kfree((void *)(unsigned long)req
->rw
.addr
);
6148 case IORING_OP_RECVMSG
:
6149 case IORING_OP_RECV
:
6150 kfree(req
->sr_msg
.kbuf
);
6153 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
6156 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6157 switch (req
->opcode
) {
6158 case IORING_OP_READV
:
6159 case IORING_OP_READ_FIXED
:
6160 case IORING_OP_READ
:
6161 case IORING_OP_WRITEV
:
6162 case IORING_OP_WRITE_FIXED
:
6163 case IORING_OP_WRITE
: {
6164 struct io_async_rw
*io
= req
->async_data
;
6166 kfree(io
->free_iovec
);
6169 case IORING_OP_RECVMSG
:
6170 case IORING_OP_SENDMSG
: {
6171 struct io_async_msghdr
*io
= req
->async_data
;
6172 if (io
->iov
!= io
->fast_iov
)
6176 case IORING_OP_SPLICE
:
6178 io_put_file(req
, req
->splice
.file_in
,
6179 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
6181 case IORING_OP_OPENAT
:
6182 case IORING_OP_OPENAT2
:
6183 if (req
->open
.filename
)
6184 putname(req
->open
.filename
);
6186 case IORING_OP_RENAMEAT
:
6187 putname(req
->rename
.oldpath
);
6188 putname(req
->rename
.newpath
);
6190 case IORING_OP_UNLINKAT
:
6191 putname(req
->unlink
.filename
);
6194 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6198 static int io_issue_sqe(struct io_kiocb
*req
, bool force_nonblock
,
6199 struct io_comp_state
*cs
)
6201 struct io_ring_ctx
*ctx
= req
->ctx
;
6204 switch (req
->opcode
) {
6206 ret
= io_nop(req
, cs
);
6208 case IORING_OP_READV
:
6209 case IORING_OP_READ_FIXED
:
6210 case IORING_OP_READ
:
6211 ret
= io_read(req
, force_nonblock
, cs
);
6213 case IORING_OP_WRITEV
:
6214 case IORING_OP_WRITE_FIXED
:
6215 case IORING_OP_WRITE
:
6216 ret
= io_write(req
, force_nonblock
, cs
);
6218 case IORING_OP_FSYNC
:
6219 ret
= io_fsync(req
, force_nonblock
);
6221 case IORING_OP_POLL_ADD
:
6222 ret
= io_poll_add(req
);
6224 case IORING_OP_POLL_REMOVE
:
6225 ret
= io_poll_remove(req
);
6227 case IORING_OP_SYNC_FILE_RANGE
:
6228 ret
= io_sync_file_range(req
, force_nonblock
);
6230 case IORING_OP_SENDMSG
:
6231 ret
= io_sendmsg(req
, force_nonblock
, cs
);
6233 case IORING_OP_SEND
:
6234 ret
= io_send(req
, force_nonblock
, cs
);
6236 case IORING_OP_RECVMSG
:
6237 ret
= io_recvmsg(req
, force_nonblock
, cs
);
6239 case IORING_OP_RECV
:
6240 ret
= io_recv(req
, force_nonblock
, cs
);
6242 case IORING_OP_TIMEOUT
:
6243 ret
= io_timeout(req
);
6245 case IORING_OP_TIMEOUT_REMOVE
:
6246 ret
= io_timeout_remove(req
);
6248 case IORING_OP_ACCEPT
:
6249 ret
= io_accept(req
, force_nonblock
, cs
);
6251 case IORING_OP_CONNECT
:
6252 ret
= io_connect(req
, force_nonblock
, cs
);
6254 case IORING_OP_ASYNC_CANCEL
:
6255 ret
= io_async_cancel(req
);
6257 case IORING_OP_FALLOCATE
:
6258 ret
= io_fallocate(req
, force_nonblock
);
6260 case IORING_OP_OPENAT
:
6261 ret
= io_openat(req
, force_nonblock
);
6263 case IORING_OP_CLOSE
:
6264 ret
= io_close(req
, force_nonblock
, cs
);
6266 case IORING_OP_FILES_UPDATE
:
6267 ret
= io_files_update(req
, force_nonblock
, cs
);
6269 case IORING_OP_STATX
:
6270 ret
= io_statx(req
, force_nonblock
);
6272 case IORING_OP_FADVISE
:
6273 ret
= io_fadvise(req
, force_nonblock
);
6275 case IORING_OP_MADVISE
:
6276 ret
= io_madvise(req
, force_nonblock
);
6278 case IORING_OP_OPENAT2
:
6279 ret
= io_openat2(req
, force_nonblock
);
6281 case IORING_OP_EPOLL_CTL
:
6282 ret
= io_epoll_ctl(req
, force_nonblock
, cs
);
6284 case IORING_OP_SPLICE
:
6285 ret
= io_splice(req
, force_nonblock
);
6287 case IORING_OP_PROVIDE_BUFFERS
:
6288 ret
= io_provide_buffers(req
, force_nonblock
, cs
);
6290 case IORING_OP_REMOVE_BUFFERS
:
6291 ret
= io_remove_buffers(req
, force_nonblock
, cs
);
6294 ret
= io_tee(req
, force_nonblock
);
6296 case IORING_OP_SHUTDOWN
:
6297 ret
= io_shutdown(req
, force_nonblock
);
6299 case IORING_OP_RENAMEAT
:
6300 ret
= io_renameat(req
, force_nonblock
);
6302 case IORING_OP_UNLINKAT
:
6303 ret
= io_unlinkat(req
, force_nonblock
);
6313 /* If the op doesn't have a file, we're not polling for it */
6314 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6315 const bool in_async
= io_wq_current_is_worker();
6317 /* workqueue context doesn't hold uring_lock, grab it now */
6319 mutex_lock(&ctx
->uring_lock
);
6321 io_iopoll_req_issued(req
, in_async
);
6324 mutex_unlock(&ctx
->uring_lock
);
6330 static void io_wq_submit_work(struct io_wq_work
*work
)
6332 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6333 struct io_kiocb
*timeout
;
6336 timeout
= io_prep_linked_timeout(req
);
6338 io_queue_linked_timeout(timeout
);
6340 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6345 ret
= io_issue_sqe(req
, false, NULL
);
6347 * We can get EAGAIN for polled IO even though we're
6348 * forcing a sync submission from here, since we can't
6349 * wait for request slots on the block side.
6358 struct io_ring_ctx
*lock_ctx
= NULL
;
6360 if (req
->ctx
->flags
& IORING_SETUP_IOPOLL
)
6361 lock_ctx
= req
->ctx
;
6364 * io_iopoll_complete() does not hold completion_lock to
6365 * complete polled io, so here for polled io, we can not call
6366 * io_req_complete() directly, otherwise there maybe concurrent
6367 * access to cqring, defer_list, etc, which is not safe. Given
6368 * that io_iopoll_complete() is always called under uring_lock,
6369 * so here for polled io, we also get uring_lock to complete
6373 mutex_lock(&lock_ctx
->uring_lock
);
6375 req_set_fail_links(req
);
6376 io_req_complete(req
, ret
);
6379 mutex_unlock(&lock_ctx
->uring_lock
);
6383 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6386 struct fixed_rsrc_table
*table
;
6388 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6389 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6392 static struct file
*io_file_get(struct io_submit_state
*state
,
6393 struct io_kiocb
*req
, int fd
, bool fixed
)
6395 struct io_ring_ctx
*ctx
= req
->ctx
;
6399 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6401 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6402 file
= io_file_from_index(ctx
, fd
);
6403 io_set_resource_node(req
);
6405 trace_io_uring_file_get(ctx
, fd
);
6406 file
= __io_file_get(state
, fd
);
6409 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6410 io_req_track_inflight(req
);
6414 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6416 struct io_timeout_data
*data
= container_of(timer
,
6417 struct io_timeout_data
, timer
);
6418 struct io_kiocb
*prev
, *req
= data
->req
;
6419 struct io_ring_ctx
*ctx
= req
->ctx
;
6420 unsigned long flags
;
6422 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6423 prev
= req
->timeout
.head
;
6424 req
->timeout
.head
= NULL
;
6427 * We don't expect the list to be empty, that will only happen if we
6428 * race with the completion of the linked work.
6430 if (prev
&& refcount_inc_not_zero(&prev
->refs
))
6431 io_remove_next_linked(prev
);
6434 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6437 req_set_fail_links(prev
);
6438 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6439 io_put_req_deferred(prev
, 1);
6441 io_req_complete_post(req
, -ETIME
, 0);
6442 io_put_req_deferred(req
, 1);
6444 return HRTIMER_NORESTART
;
6447 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6450 * If the back reference is NULL, then our linked request finished
6451 * before we got a chance to setup the timer
6453 if (req
->timeout
.head
) {
6454 struct io_timeout_data
*data
= req
->async_data
;
6456 data
->timer
.function
= io_link_timeout_fn
;
6457 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6462 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6464 struct io_ring_ctx
*ctx
= req
->ctx
;
6466 spin_lock_irq(&ctx
->completion_lock
);
6467 __io_queue_linked_timeout(req
);
6468 spin_unlock_irq(&ctx
->completion_lock
);
6470 /* drop submission reference */
6474 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6476 struct io_kiocb
*nxt
= req
->link
;
6478 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6479 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6482 nxt
->timeout
.head
= req
;
6483 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6484 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6488 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
)
6490 struct io_kiocb
*linked_timeout
;
6491 const struct cred
*old_creds
= NULL
;
6495 linked_timeout
= io_prep_linked_timeout(req
);
6497 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
6498 (req
->work
.flags
& IO_WQ_WORK_CREDS
) &&
6499 req
->work
.identity
->creds
!= current_cred()) {
6501 revert_creds(old_creds
);
6502 if (old_creds
== req
->work
.identity
->creds
)
6503 old_creds
= NULL
; /* restored original creds */
6505 old_creds
= override_creds(req
->work
.identity
->creds
);
6508 ret
= io_issue_sqe(req
, true, cs
);
6511 * We async punt it if the file wasn't marked NOWAIT, or if the file
6512 * doesn't support non-blocking read/write attempts
6514 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6515 if (!io_arm_poll_handler(req
)) {
6517 * Queued up for async execution, worker will release
6518 * submit reference when the iocb is actually submitted.
6520 io_queue_async_work(req
);
6524 io_queue_linked_timeout(linked_timeout
);
6525 } else if (likely(!ret
)) {
6526 /* drop submission reference */
6527 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
6528 list_add_tail(&req
->compl.list
, &cs
->list
);
6530 io_submit_flush_completions(cs
);
6533 req
= io_put_req_find_next(req
);
6537 io_queue_linked_timeout(linked_timeout
);
6540 if (!(req
->flags
& REQ_F_FORCE_ASYNC
))
6542 io_queue_async_work(req
);
6545 /* un-prep timeout, so it'll be killed as any other linked */
6546 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
6547 req_set_fail_links(req
);
6549 io_req_complete(req
, ret
);
6553 revert_creds(old_creds
);
6556 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6557 struct io_comp_state
*cs
)
6561 ret
= io_req_defer(req
, sqe
);
6563 if (ret
!= -EIOCBQUEUED
) {
6565 req_set_fail_links(req
);
6567 io_req_complete(req
, ret
);
6569 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6570 if (!req
->async_data
) {
6571 ret
= io_req_defer_prep(req
, sqe
);
6575 io_queue_async_work(req
);
6578 ret
= io_req_prep(req
, sqe
);
6582 __io_queue_sqe(req
, cs
);
6586 static inline void io_queue_link_head(struct io_kiocb
*req
,
6587 struct io_comp_state
*cs
)
6589 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
6591 io_req_complete(req
, -ECANCELED
);
6593 io_queue_sqe(req
, NULL
, cs
);
6596 struct io_submit_link
{
6597 struct io_kiocb
*head
;
6598 struct io_kiocb
*last
;
6601 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6602 struct io_submit_link
*link
, struct io_comp_state
*cs
)
6604 struct io_ring_ctx
*ctx
= req
->ctx
;
6608 * If we already have a head request, queue this one for async
6609 * submittal once the head completes. If we don't have a head but
6610 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6611 * submitted sync once the chain is complete. If none of those
6612 * conditions are true (normal request), then just queue it.
6615 struct io_kiocb
*head
= link
->head
;
6618 * Taking sequential execution of a link, draining both sides
6619 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6620 * requests in the link. So, it drains the head and the
6621 * next after the link request. The last one is done via
6622 * drain_next flag to persist the effect across calls.
6624 if (req
->flags
& REQ_F_IO_DRAIN
) {
6625 head
->flags
|= REQ_F_IO_DRAIN
;
6626 ctx
->drain_next
= 1;
6628 ret
= io_req_defer_prep(req
, sqe
);
6629 if (unlikely(ret
)) {
6630 /* fail even hard links since we don't submit */
6631 head
->flags
|= REQ_F_FAIL_LINK
;
6634 trace_io_uring_link(ctx
, req
, head
);
6635 link
->last
->link
= req
;
6638 /* last request of a link, enqueue the link */
6639 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6640 io_queue_link_head(head
, cs
);
6644 if (unlikely(ctx
->drain_next
)) {
6645 req
->flags
|= REQ_F_IO_DRAIN
;
6646 ctx
->drain_next
= 0;
6648 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6649 ret
= io_req_defer_prep(req
, sqe
);
6651 req
->flags
|= REQ_F_FAIL_LINK
;
6655 io_queue_sqe(req
, sqe
, cs
);
6663 * Batched submission is done, ensure local IO is flushed out.
6665 static void io_submit_state_end(struct io_submit_state
*state
)
6667 if (!list_empty(&state
->comp
.list
))
6668 io_submit_flush_completions(&state
->comp
);
6669 if (state
->plug_started
)
6670 blk_finish_plug(&state
->plug
);
6671 io_state_file_put(state
);
6672 if (state
->free_reqs
)
6673 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6677 * Start submission side cache.
6679 static void io_submit_state_start(struct io_submit_state
*state
,
6680 struct io_ring_ctx
*ctx
, unsigned int max_ios
)
6682 state
->plug_started
= false;
6684 INIT_LIST_HEAD(&state
->comp
.list
);
6685 state
->comp
.ctx
= ctx
;
6686 state
->free_reqs
= 0;
6687 state
->file_refs
= 0;
6688 state
->ios_left
= max_ios
;
6691 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6693 struct io_rings
*rings
= ctx
->rings
;
6696 * Ensure any loads from the SQEs are done at this point,
6697 * since once we write the new head, the application could
6698 * write new data to them.
6700 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6704 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6705 * that is mapped by userspace. This means that care needs to be taken to
6706 * ensure that reads are stable, as we cannot rely on userspace always
6707 * being a good citizen. If members of the sqe are validated and then later
6708 * used, it's important that those reads are done through READ_ONCE() to
6709 * prevent a re-load down the line.
6711 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6713 u32
*sq_array
= ctx
->sq_array
;
6717 * The cached sq head (or cq tail) serves two purposes:
6719 * 1) allows us to batch the cost of updating the user visible
6721 * 2) allows the kernel side to track the head on its own, even
6722 * though the application is the one updating it.
6724 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6725 if (likely(head
< ctx
->sq_entries
))
6726 return &ctx
->sq_sqes
[head
];
6728 /* drop invalid entries */
6729 ctx
->cached_sq_dropped
++;
6730 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6734 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6736 ctx
->cached_sq_head
++;
6740 * Check SQE restrictions (opcode and flags).
6742 * Returns 'true' if SQE is allowed, 'false' otherwise.
6744 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6745 struct io_kiocb
*req
,
6746 unsigned int sqe_flags
)
6748 if (!ctx
->restricted
)
6751 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6754 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6755 ctx
->restrictions
.sqe_flags_required
)
6758 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6759 ctx
->restrictions
.sqe_flags_required
))
6765 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6766 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6767 IOSQE_BUFFER_SELECT)
6769 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6770 const struct io_uring_sqe
*sqe
,
6771 struct io_submit_state
*state
)
6773 unsigned int sqe_flags
;
6776 req
->opcode
= READ_ONCE(sqe
->opcode
);
6777 req
->user_data
= READ_ONCE(sqe
->user_data
);
6778 req
->async_data
= NULL
;
6783 req
->fixed_rsrc_refs
= NULL
;
6784 /* one is dropped after submission, the other at completion */
6785 refcount_set(&req
->refs
, 2);
6786 req
->task
= current
;
6789 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6792 if (unlikely(io_sq_thread_acquire_mm_files(ctx
, req
)))
6795 sqe_flags
= READ_ONCE(sqe
->flags
);
6796 /* enforce forwards compatibility on users */
6797 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6800 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6803 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6804 !io_op_defs
[req
->opcode
].buffer_select
)
6807 id
= READ_ONCE(sqe
->personality
);
6809 struct io_identity
*iod
;
6811 iod
= idr_find(&ctx
->personality_idr
, id
);
6814 refcount_inc(&iod
->count
);
6816 __io_req_init_async(req
);
6817 get_cred(iod
->creds
);
6818 req
->work
.identity
= iod
;
6819 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
6822 /* same numerical values with corresponding REQ_F_*, safe to copy */
6823 req
->flags
|= sqe_flags
;
6826 * Plug now if we have more than 1 IO left after this, and the target
6827 * is potentially a read/write to block based storage.
6829 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6830 io_op_defs
[req
->opcode
].plug
) {
6831 blk_start_plug(&state
->plug
);
6832 state
->plug_started
= true;
6836 if (io_op_defs
[req
->opcode
].needs_file
) {
6837 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6839 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6840 if (unlikely(!req
->file
))
6848 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6850 struct io_submit_state state
;
6851 struct io_submit_link link
;
6852 int i
, submitted
= 0;
6854 /* if we have a backlog and couldn't flush it all, return BUSY */
6855 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6856 if (!__io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6860 /* make sure SQ entry isn't read before tail */
6861 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6863 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6866 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6867 refcount_add(nr
, ¤t
->usage
);
6869 io_submit_state_start(&state
, ctx
, nr
);
6872 for (i
= 0; i
< nr
; i
++) {
6873 const struct io_uring_sqe
*sqe
;
6874 struct io_kiocb
*req
;
6877 sqe
= io_get_sqe(ctx
);
6878 if (unlikely(!sqe
)) {
6879 io_consume_sqe(ctx
);
6882 req
= io_alloc_req(ctx
, &state
);
6883 if (unlikely(!req
)) {
6885 submitted
= -EAGAIN
;
6888 io_consume_sqe(ctx
);
6889 /* will complete beyond this point, count as submitted */
6892 err
= io_init_req(ctx
, req
, sqe
, &state
);
6893 if (unlikely(err
)) {
6896 io_req_complete(req
, err
);
6900 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6901 true, ctx
->flags
& IORING_SETUP_SQPOLL
);
6902 err
= io_submit_sqe(req
, sqe
, &link
, &state
.comp
);
6907 if (unlikely(submitted
!= nr
)) {
6908 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6909 struct io_uring_task
*tctx
= current
->io_uring
;
6910 int unused
= nr
- ref_used
;
6912 percpu_ref_put_many(&ctx
->refs
, unused
);
6913 percpu_counter_sub(&tctx
->inflight
, unused
);
6914 put_task_struct_many(current
, unused
);
6917 io_queue_link_head(link
.head
, &state
.comp
);
6918 io_submit_state_end(&state
);
6920 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6921 io_commit_sqring(ctx
);
6926 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6928 /* Tell userspace we may need a wakeup call */
6929 spin_lock_irq(&ctx
->completion_lock
);
6930 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6931 spin_unlock_irq(&ctx
->completion_lock
);
6934 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6936 spin_lock_irq(&ctx
->completion_lock
);
6937 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6938 spin_unlock_irq(&ctx
->completion_lock
);
6941 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
6943 unsigned int to_submit
;
6946 to_submit
= io_sqring_entries(ctx
);
6947 /* if we're handling multiple rings, cap submit size for fairness */
6948 if (cap_entries
&& to_submit
> 8)
6951 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
6952 unsigned nr_events
= 0;
6954 mutex_lock(&ctx
->uring_lock
);
6955 if (!list_empty(&ctx
->iopoll_list
))
6956 io_do_iopoll(ctx
, &nr_events
, 0);
6958 if (to_submit
&& !ctx
->sqo_dead
&&
6959 likely(!percpu_ref_is_dying(&ctx
->refs
)))
6960 ret
= io_submit_sqes(ctx
, to_submit
);
6961 mutex_unlock(&ctx
->uring_lock
);
6964 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
6965 wake_up(&ctx
->sqo_sq_wait
);
6970 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
6972 struct io_ring_ctx
*ctx
;
6973 unsigned sq_thread_idle
= 0;
6975 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6976 if (sq_thread_idle
< ctx
->sq_thread_idle
)
6977 sq_thread_idle
= ctx
->sq_thread_idle
;
6980 sqd
->sq_thread_idle
= sq_thread_idle
;
6983 static void io_sqd_init_new(struct io_sq_data
*sqd
)
6985 struct io_ring_ctx
*ctx
;
6987 while (!list_empty(&sqd
->ctx_new_list
)) {
6988 ctx
= list_first_entry(&sqd
->ctx_new_list
, struct io_ring_ctx
, sqd_list
);
6989 list_move_tail(&ctx
->sqd_list
, &sqd
->ctx_list
);
6990 complete(&ctx
->sq_thread_comp
);
6993 io_sqd_update_thread_idle(sqd
);
6996 static int io_sq_thread(void *data
)
6998 struct cgroup_subsys_state
*cur_css
= NULL
;
6999 struct files_struct
*old_files
= current
->files
;
7000 struct nsproxy
*old_nsproxy
= current
->nsproxy
;
7001 const struct cred
*old_cred
= NULL
;
7002 struct io_sq_data
*sqd
= data
;
7003 struct io_ring_ctx
*ctx
;
7004 unsigned long timeout
= 0;
7008 current
->files
= NULL
;
7009 current
->nsproxy
= NULL
;
7010 task_unlock(current
);
7012 while (!kthread_should_stop()) {
7014 bool cap_entries
, sqt_spin
, needs_sched
;
7017 * Any changes to the sqd lists are synchronized through the
7018 * kthread parking. This synchronizes the thread vs users,
7019 * the users are synchronized on the sqd->ctx_lock.
7021 if (kthread_should_park()) {
7024 * When sq thread is unparked, in case the previous park operation
7025 * comes from io_put_sq_data(), which means that sq thread is going
7026 * to be stopped, so here needs to have a check.
7028 if (kthread_should_stop())
7032 if (unlikely(!list_empty(&sqd
->ctx_new_list
))) {
7033 io_sqd_init_new(sqd
);
7034 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7038 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7039 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7040 if (current
->cred
!= ctx
->creds
) {
7042 revert_creds(old_cred
);
7043 old_cred
= override_creds(ctx
->creds
);
7045 io_sq_thread_associate_blkcg(ctx
, &cur_css
);
7047 current
->loginuid
= ctx
->loginuid
;
7048 current
->sessionid
= ctx
->sessionid
;
7051 ret
= __io_sq_thread(ctx
, cap_entries
);
7052 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7055 io_sq_thread_drop_mm_files();
7058 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7060 io_sq_thread_drop_mm_files();
7063 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7068 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7069 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7070 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7071 !list_empty_careful(&ctx
->iopoll_list
)) {
7072 needs_sched
= false;
7075 if (io_sqring_entries(ctx
)) {
7076 needs_sched
= false;
7081 if (needs_sched
&& !kthread_should_park()) {
7082 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7083 io_ring_set_wakeup_flag(ctx
);
7086 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7087 io_ring_clear_wakeup_flag(ctx
);
7090 finish_wait(&sqd
->wait
, &wait
);
7091 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7095 io_sq_thread_drop_mm_files();
7098 io_sq_thread_unassociate_blkcg();
7100 revert_creds(old_cred
);
7103 current
->files
= old_files
;
7104 current
->nsproxy
= old_nsproxy
;
7105 task_unlock(current
);
7112 struct io_wait_queue
{
7113 struct wait_queue_entry wq
;
7114 struct io_ring_ctx
*ctx
;
7116 unsigned nr_timeouts
;
7119 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7121 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7124 * Wake up if we have enough events, or if a timeout occurred since we
7125 * started waiting. For timeouts, we always want to return to userspace,
7126 * regardless of event count.
7128 return io_cqring_events(ctx
) >= iowq
->to_wait
||
7129 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7132 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7133 int wake_flags
, void *key
)
7135 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7139 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7140 * the task, and the next invocation will do it.
7142 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->cq_check_overflow
))
7143 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7147 static int io_run_task_work_sig(void)
7149 if (io_run_task_work())
7151 if (!signal_pending(current
))
7153 if (test_tsk_thread_flag(current
, TIF_NOTIFY_SIGNAL
))
7154 return -ERESTARTSYS
;
7158 /* when returns >0, the caller should retry */
7159 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7160 struct io_wait_queue
*iowq
,
7161 signed long *timeout
)
7165 /* make sure we run task_work before checking for signals */
7166 ret
= io_run_task_work_sig();
7167 if (ret
|| io_should_wake(iowq
))
7169 /* let the caller flush overflows, retry */
7170 if (test_bit(0, &ctx
->cq_check_overflow
))
7173 *timeout
= schedule_timeout(*timeout
);
7174 return !*timeout
? -ETIME
: 1;
7178 * Wait until events become available, if we don't already have some. The
7179 * application must reap them itself, as they reside on the shared cq ring.
7181 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7182 const sigset_t __user
*sig
, size_t sigsz
,
7183 struct __kernel_timespec __user
*uts
)
7185 struct io_wait_queue iowq
= {
7188 .func
= io_wake_function
,
7189 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
7192 .to_wait
= min_events
,
7194 struct io_rings
*rings
= ctx
->rings
;
7195 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
7199 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7200 if (io_cqring_events(ctx
) >= min_events
)
7202 if (!io_run_task_work())
7207 #ifdef CONFIG_COMPAT
7208 if (in_compat_syscall())
7209 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7213 ret
= set_user_sigmask(sig
, sigsz
);
7220 struct timespec64 ts
;
7222 if (get_timespec64(&ts
, uts
))
7224 timeout
= timespec64_to_jiffies(&ts
);
7227 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7228 trace_io_uring_cqring_wait(ctx
, min_events
);
7230 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
7231 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
7232 TASK_INTERRUPTIBLE
);
7233 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
7234 finish_wait(&ctx
->wait
, &iowq
.wq
);
7237 restore_saved_sigmask_unless(ret
== -EINTR
);
7239 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7242 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7244 #if defined(CONFIG_UNIX)
7245 if (ctx
->ring_sock
) {
7246 struct sock
*sock
= ctx
->ring_sock
->sk
;
7247 struct sk_buff
*skb
;
7249 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7255 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7258 file
= io_file_from_index(ctx
, i
);
7265 static void io_rsrc_data_ref_zero(struct percpu_ref
*ref
)
7267 struct fixed_rsrc_data
*data
;
7269 data
= container_of(ref
, struct fixed_rsrc_data
, refs
);
7270 complete(&data
->done
);
7273 static inline void io_rsrc_ref_lock(struct io_ring_ctx
*ctx
)
7275 spin_lock_bh(&ctx
->rsrc_ref_lock
);
7278 static inline void io_rsrc_ref_unlock(struct io_ring_ctx
*ctx
)
7280 spin_unlock_bh(&ctx
->rsrc_ref_lock
);
7283 static void io_sqe_rsrc_set_node(struct io_ring_ctx
*ctx
,
7284 struct fixed_rsrc_data
*rsrc_data
,
7285 struct fixed_rsrc_ref_node
*ref_node
)
7287 io_rsrc_ref_lock(ctx
);
7288 rsrc_data
->node
= ref_node
;
7289 list_add_tail(&ref_node
->node
, &ctx
->rsrc_ref_list
);
7290 io_rsrc_ref_unlock(ctx
);
7291 percpu_ref_get(&rsrc_data
->refs
);
7294 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data
*data
,
7295 struct io_ring_ctx
*ctx
,
7296 struct fixed_rsrc_ref_node
*backup_node
)
7298 struct fixed_rsrc_ref_node
*ref_node
;
7301 io_rsrc_ref_lock(ctx
);
7302 ref_node
= data
->node
;
7303 io_rsrc_ref_unlock(ctx
);
7305 percpu_ref_kill(&ref_node
->refs
);
7307 percpu_ref_kill(&data
->refs
);
7309 /* wait for all refs nodes to complete */
7310 flush_delayed_work(&ctx
->rsrc_put_work
);
7312 ret
= wait_for_completion_interruptible(&data
->done
);
7315 ret
= io_run_task_work_sig();
7317 percpu_ref_resurrect(&data
->refs
);
7318 reinit_completion(&data
->done
);
7319 io_sqe_rsrc_set_node(ctx
, data
, backup_node
);
7324 destroy_fixed_rsrc_ref_node(backup_node
);
7328 static struct fixed_rsrc_data
*alloc_fixed_rsrc_data(struct io_ring_ctx
*ctx
)
7330 struct fixed_rsrc_data
*data
;
7332 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7336 if (percpu_ref_init(&data
->refs
, io_rsrc_data_ref_zero
,
7337 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
)) {
7342 init_completion(&data
->done
);
7346 static void free_fixed_rsrc_data(struct fixed_rsrc_data
*data
)
7348 percpu_ref_exit(&data
->refs
);
7353 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7355 struct fixed_rsrc_data
*data
= ctx
->file_data
;
7356 struct fixed_rsrc_ref_node
*backup_node
;
7357 unsigned nr_tables
, i
;
7362 backup_node
= alloc_fixed_rsrc_ref_node(ctx
);
7365 init_fixed_file_ref_node(ctx
, backup_node
);
7367 ret
= io_rsrc_ref_quiesce(data
, ctx
, backup_node
);
7371 __io_sqe_files_unregister(ctx
);
7372 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
7373 for (i
= 0; i
< nr_tables
; i
++)
7374 kfree(data
->table
[i
].files
);
7375 free_fixed_rsrc_data(data
);
7376 ctx
->file_data
= NULL
;
7377 ctx
->nr_user_files
= 0;
7381 static void io_put_sq_data(struct io_sq_data
*sqd
)
7383 if (refcount_dec_and_test(&sqd
->refs
)) {
7385 * The park is a bit of a work-around, without it we get
7386 * warning spews on shutdown with SQPOLL set and affinity
7387 * set to a single CPU.
7390 kthread_park(sqd
->thread
);
7391 kthread_stop(sqd
->thread
);
7398 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7400 struct io_ring_ctx
*ctx_attach
;
7401 struct io_sq_data
*sqd
;
7404 f
= fdget(p
->wq_fd
);
7406 return ERR_PTR(-ENXIO
);
7407 if (f
.file
->f_op
!= &io_uring_fops
) {
7409 return ERR_PTR(-EINVAL
);
7412 ctx_attach
= f
.file
->private_data
;
7413 sqd
= ctx_attach
->sq_data
;
7416 return ERR_PTR(-EINVAL
);
7419 refcount_inc(&sqd
->refs
);
7424 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
)
7426 struct io_sq_data
*sqd
;
7428 if (p
->flags
& IORING_SETUP_ATTACH_WQ
)
7429 return io_attach_sq_data(p
);
7431 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7433 return ERR_PTR(-ENOMEM
);
7435 refcount_set(&sqd
->refs
, 1);
7436 INIT_LIST_HEAD(&sqd
->ctx_list
);
7437 INIT_LIST_HEAD(&sqd
->ctx_new_list
);
7438 mutex_init(&sqd
->ctx_lock
);
7439 mutex_init(&sqd
->lock
);
7440 init_waitqueue_head(&sqd
->wait
);
7444 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7445 __releases(&sqd
->lock
)
7449 kthread_unpark(sqd
->thread
);
7450 mutex_unlock(&sqd
->lock
);
7453 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7454 __acquires(&sqd
->lock
)
7458 mutex_lock(&sqd
->lock
);
7459 kthread_park(sqd
->thread
);
7462 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
7464 struct io_sq_data
*sqd
= ctx
->sq_data
;
7469 * We may arrive here from the error branch in
7470 * io_sq_offload_create() where the kthread is created
7471 * without being waked up, thus wake it up now to make
7472 * sure the wait will complete.
7474 wake_up_process(sqd
->thread
);
7475 wait_for_completion(&ctx
->sq_thread_comp
);
7477 io_sq_thread_park(sqd
);
7480 mutex_lock(&sqd
->ctx_lock
);
7481 list_del(&ctx
->sqd_list
);
7482 io_sqd_update_thread_idle(sqd
);
7483 mutex_unlock(&sqd
->ctx_lock
);
7486 io_sq_thread_unpark(sqd
);
7488 io_put_sq_data(sqd
);
7489 ctx
->sq_data
= NULL
;
7493 static void io_finish_async(struct io_ring_ctx
*ctx
)
7495 io_sq_thread_stop(ctx
);
7498 io_wq_destroy(ctx
->io_wq
);
7503 #if defined(CONFIG_UNIX)
7505 * Ensure the UNIX gc is aware of our file set, so we are certain that
7506 * the io_uring can be safely unregistered on process exit, even if we have
7507 * loops in the file referencing.
7509 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7511 struct sock
*sk
= ctx
->ring_sock
->sk
;
7512 struct scm_fp_list
*fpl
;
7513 struct sk_buff
*skb
;
7516 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7520 skb
= alloc_skb(0, GFP_KERNEL
);
7529 fpl
->user
= get_uid(ctx
->user
);
7530 for (i
= 0; i
< nr
; i
++) {
7531 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7535 fpl
->fp
[nr_files
] = get_file(file
);
7536 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7541 fpl
->max
= SCM_MAX_FD
;
7542 fpl
->count
= nr_files
;
7543 UNIXCB(skb
).fp
= fpl
;
7544 skb
->destructor
= unix_destruct_scm
;
7545 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7546 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7548 for (i
= 0; i
< nr_files
; i
++)
7559 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7560 * causes regular reference counting to break down. We rely on the UNIX
7561 * garbage collection to take care of this problem for us.
7563 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7565 unsigned left
, total
;
7569 left
= ctx
->nr_user_files
;
7571 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7573 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7577 total
+= this_files
;
7583 while (total
< ctx
->nr_user_files
) {
7584 struct file
*file
= io_file_from_index(ctx
, total
);
7594 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7600 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data
*file_data
,
7601 unsigned nr_tables
, unsigned nr_files
)
7605 for (i
= 0; i
< nr_tables
; i
++) {
7606 struct fixed_rsrc_table
*table
= &file_data
->table
[i
];
7607 unsigned this_files
;
7609 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7610 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7614 nr_files
-= this_files
;
7620 for (i
= 0; i
< nr_tables
; i
++) {
7621 struct fixed_rsrc_table
*table
= &file_data
->table
[i
];
7622 kfree(table
->files
);
7627 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
7629 struct file
*file
= prsrc
->file
;
7630 #if defined(CONFIG_UNIX)
7631 struct sock
*sock
= ctx
->ring_sock
->sk
;
7632 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7633 struct sk_buff
*skb
;
7636 __skb_queue_head_init(&list
);
7639 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7640 * remove this entry and rearrange the file array.
7642 skb
= skb_dequeue(head
);
7644 struct scm_fp_list
*fp
;
7646 fp
= UNIXCB(skb
).fp
;
7647 for (i
= 0; i
< fp
->count
; i
++) {
7650 if (fp
->fp
[i
] != file
)
7653 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7654 left
= fp
->count
- 1 - i
;
7656 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7657 left
* sizeof(struct file
*));
7664 __skb_queue_tail(&list
, skb
);
7674 __skb_queue_tail(&list
, skb
);
7676 skb
= skb_dequeue(head
);
7679 if (skb_peek(&list
)) {
7680 spin_lock_irq(&head
->lock
);
7681 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7682 __skb_queue_tail(head
, skb
);
7683 spin_unlock_irq(&head
->lock
);
7690 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node
*ref_node
)
7692 struct fixed_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
7693 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
7694 struct io_rsrc_put
*prsrc
, *tmp
;
7696 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
7697 list_del(&prsrc
->list
);
7698 ref_node
->rsrc_put(ctx
, prsrc
);
7702 percpu_ref_exit(&ref_node
->refs
);
7704 percpu_ref_put(&rsrc_data
->refs
);
7707 static void io_rsrc_put_work(struct work_struct
*work
)
7709 struct io_ring_ctx
*ctx
;
7710 struct llist_node
*node
;
7712 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
7713 node
= llist_del_all(&ctx
->rsrc_put_llist
);
7716 struct fixed_rsrc_ref_node
*ref_node
;
7717 struct llist_node
*next
= node
->next
;
7719 ref_node
= llist_entry(node
, struct fixed_rsrc_ref_node
, llist
);
7720 __io_rsrc_put_work(ref_node
);
7725 static struct file
**io_fixed_file_slot(struct fixed_rsrc_data
*file_data
,
7728 struct fixed_rsrc_table
*table
;
7730 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7731 return &table
->files
[i
& IORING_FILE_TABLE_MASK
];
7734 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7736 struct fixed_rsrc_ref_node
*ref_node
;
7737 struct fixed_rsrc_data
*data
;
7738 struct io_ring_ctx
*ctx
;
7739 bool first_add
= false;
7742 ref_node
= container_of(ref
, struct fixed_rsrc_ref_node
, refs
);
7743 data
= ref_node
->rsrc_data
;
7746 io_rsrc_ref_lock(ctx
);
7747 ref_node
->done
= true;
7749 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7750 ref_node
= list_first_entry(&ctx
->rsrc_ref_list
,
7751 struct fixed_rsrc_ref_node
, node
);
7752 /* recycle ref nodes in order */
7753 if (!ref_node
->done
)
7755 list_del(&ref_node
->node
);
7756 first_add
|= llist_add(&ref_node
->llist
, &ctx
->rsrc_put_llist
);
7758 io_rsrc_ref_unlock(ctx
);
7760 if (percpu_ref_is_dying(&data
->refs
))
7764 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, 0);
7766 queue_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
7769 static struct fixed_rsrc_ref_node
*alloc_fixed_rsrc_ref_node(
7770 struct io_ring_ctx
*ctx
)
7772 struct fixed_rsrc_ref_node
*ref_node
;
7774 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7778 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7783 INIT_LIST_HEAD(&ref_node
->node
);
7784 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7785 ref_node
->done
= false;
7789 static void init_fixed_file_ref_node(struct io_ring_ctx
*ctx
,
7790 struct fixed_rsrc_ref_node
*ref_node
)
7792 ref_node
->rsrc_data
= ctx
->file_data
;
7793 ref_node
->rsrc_put
= io_ring_file_put
;
7796 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node
*ref_node
)
7798 percpu_ref_exit(&ref_node
->refs
);
7803 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7806 __s32 __user
*fds
= (__s32 __user
*) arg
;
7807 unsigned nr_tables
, i
;
7809 int fd
, ret
= -ENOMEM
;
7810 struct fixed_rsrc_ref_node
*ref_node
;
7811 struct fixed_rsrc_data
*file_data
;
7817 if (nr_args
> IORING_MAX_FIXED_FILES
)
7820 file_data
= alloc_fixed_rsrc_data(ctx
);
7823 ctx
->file_data
= file_data
;
7825 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7826 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7828 if (!file_data
->table
)
7831 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7834 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7835 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7839 /* allow sparse sets */
7849 * Don't allow io_uring instances to be registered. If UNIX
7850 * isn't enabled, then this causes a reference cycle and this
7851 * instance can never get freed. If UNIX is enabled we'll
7852 * handle it just fine, but there's still no point in allowing
7853 * a ring fd as it doesn't support regular read/write anyway.
7855 if (file
->f_op
== &io_uring_fops
) {
7859 *io_fixed_file_slot(file_data
, i
) = file
;
7862 ret
= io_sqe_files_scm(ctx
);
7864 io_sqe_files_unregister(ctx
);
7868 ref_node
= alloc_fixed_rsrc_ref_node(ctx
);
7870 io_sqe_files_unregister(ctx
);
7873 init_fixed_file_ref_node(ctx
, ref_node
);
7875 io_sqe_rsrc_set_node(ctx
, file_data
, ref_node
);
7878 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7879 file
= io_file_from_index(ctx
, i
);
7883 for (i
= 0; i
< nr_tables
; i
++)
7884 kfree(file_data
->table
[i
].files
);
7885 ctx
->nr_user_files
= 0;
7887 free_fixed_rsrc_data(ctx
->file_data
);
7888 ctx
->file_data
= NULL
;
7892 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7895 #if defined(CONFIG_UNIX)
7896 struct sock
*sock
= ctx
->ring_sock
->sk
;
7897 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7898 struct sk_buff
*skb
;
7901 * See if we can merge this file into an existing skb SCM_RIGHTS
7902 * file set. If there's no room, fall back to allocating a new skb
7903 * and filling it in.
7905 spin_lock_irq(&head
->lock
);
7906 skb
= skb_peek(head
);
7908 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7910 if (fpl
->count
< SCM_MAX_FD
) {
7911 __skb_unlink(skb
, head
);
7912 spin_unlock_irq(&head
->lock
);
7913 fpl
->fp
[fpl
->count
] = get_file(file
);
7914 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7916 spin_lock_irq(&head
->lock
);
7917 __skb_queue_head(head
, skb
);
7922 spin_unlock_irq(&head
->lock
);
7929 return __io_sqe_files_scm(ctx
, 1, index
);
7935 static int io_queue_rsrc_removal(struct fixed_rsrc_data
*data
, void *rsrc
)
7937 struct io_rsrc_put
*prsrc
;
7938 struct fixed_rsrc_ref_node
*ref_node
= data
->node
;
7940 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
7945 list_add(&prsrc
->list
, &ref_node
->rsrc_list
);
7950 static inline int io_queue_file_removal(struct fixed_rsrc_data
*data
,
7953 return io_queue_rsrc_removal(data
, (void *)file
);
7956 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7957 struct io_uring_rsrc_update
*up
,
7960 struct fixed_rsrc_data
*data
= ctx
->file_data
;
7961 struct fixed_rsrc_ref_node
*ref_node
;
7962 struct file
*file
, **file_slot
;
7966 bool needs_switch
= false;
7968 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7970 if (done
> ctx
->nr_user_files
)
7973 ref_node
= alloc_fixed_rsrc_ref_node(ctx
);
7976 init_fixed_file_ref_node(ctx
, ref_node
);
7978 fds
= u64_to_user_ptr(up
->data
);
7979 for (done
= 0; done
< nr_args
; done
++) {
7981 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7985 if (fd
== IORING_REGISTER_FILES_SKIP
)
7988 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
7989 file_slot
= io_fixed_file_slot(ctx
->file_data
, i
);
7992 err
= io_queue_file_removal(data
, *file_slot
);
7996 needs_switch
= true;
8005 * Don't allow io_uring instances to be registered. If
8006 * UNIX isn't enabled, then this causes a reference
8007 * cycle and this instance can never get freed. If UNIX
8008 * is enabled we'll handle it just fine, but there's
8009 * still no point in allowing a ring fd as it doesn't
8010 * support regular read/write anyway.
8012 if (file
->f_op
== &io_uring_fops
) {
8017 err
= io_sqe_file_register(ctx
, file
, i
);
8027 percpu_ref_kill(&data
->node
->refs
);
8028 io_sqe_rsrc_set_node(ctx
, data
, ref_node
);
8030 destroy_fixed_rsrc_ref_node(ref_node
);
8032 return done
? done
: err
;
8035 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
8038 struct io_uring_rsrc_update up
;
8040 if (!ctx
->file_data
)
8044 if (copy_from_user(&up
, arg
, sizeof(up
)))
8049 return __io_sqe_files_update(ctx
, &up
, nr_args
);
8052 static struct io_wq_work
*io_free_work(struct io_wq_work
*work
)
8054 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8056 req
= io_put_req_find_next(req
);
8057 return req
? &req
->work
: NULL
;
8060 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
8061 struct io_uring_params
*p
)
8063 struct io_wq_data data
;
8065 struct io_ring_ctx
*ctx_attach
;
8066 unsigned int concurrency
;
8069 data
.user
= ctx
->user
;
8070 data
.free_work
= io_free_work
;
8071 data
.do_work
= io_wq_submit_work
;
8073 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
8074 /* Do QD, or 4 * CPUS, whatever is smallest */
8075 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8077 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
8078 if (IS_ERR(ctx
->io_wq
)) {
8079 ret
= PTR_ERR(ctx
->io_wq
);
8085 f
= fdget(p
->wq_fd
);
8089 if (f
.file
->f_op
!= &io_uring_fops
) {
8094 ctx_attach
= f
.file
->private_data
;
8095 /* @io_wq is protected by holding the fd */
8096 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
8101 ctx
->io_wq
= ctx_attach
->io_wq
;
8107 static int io_uring_alloc_task_context(struct task_struct
*task
)
8109 struct io_uring_task
*tctx
;
8112 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
8113 if (unlikely(!tctx
))
8116 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8117 if (unlikely(ret
)) {
8123 init_waitqueue_head(&tctx
->wait
);
8125 atomic_set(&tctx
->in_idle
, 0);
8126 tctx
->sqpoll
= false;
8127 io_init_identity(&tctx
->__identity
);
8128 tctx
->identity
= &tctx
->__identity
;
8129 task
->io_uring
= tctx
;
8133 void __io_uring_free(struct task_struct
*tsk
)
8135 struct io_uring_task
*tctx
= tsk
->io_uring
;
8137 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8138 WARN_ON_ONCE(refcount_read(&tctx
->identity
->count
) != 1);
8139 if (tctx
->identity
!= &tctx
->__identity
)
8140 kfree(tctx
->identity
);
8141 percpu_counter_destroy(&tctx
->inflight
);
8143 tsk
->io_uring
= NULL
;
8146 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8147 struct io_uring_params
*p
)
8151 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8152 struct io_sq_data
*sqd
;
8155 if (!capable(CAP_SYS_ADMIN
) && !capable(CAP_SYS_NICE
))
8158 sqd
= io_get_sq_data(p
);
8165 io_sq_thread_park(sqd
);
8166 mutex_lock(&sqd
->ctx_lock
);
8167 list_add(&ctx
->sqd_list
, &sqd
->ctx_new_list
);
8168 mutex_unlock(&sqd
->ctx_lock
);
8169 io_sq_thread_unpark(sqd
);
8171 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8172 if (!ctx
->sq_thread_idle
)
8173 ctx
->sq_thread_idle
= HZ
;
8178 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8179 int cpu
= p
->sq_thread_cpu
;
8182 if (cpu
>= nr_cpu_ids
)
8184 if (!cpu_online(cpu
))
8187 sqd
->thread
= kthread_create_on_cpu(io_sq_thread
, sqd
,
8188 cpu
, "io_uring-sq");
8190 sqd
->thread
= kthread_create(io_sq_thread
, sqd
,
8193 if (IS_ERR(sqd
->thread
)) {
8194 ret
= PTR_ERR(sqd
->thread
);
8198 ret
= io_uring_alloc_task_context(sqd
->thread
);
8201 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8202 /* Can't have SQ_AFF without SQPOLL */
8208 ret
= io_init_wq_offload(ctx
, p
);
8214 io_finish_async(ctx
);
8218 static void io_sq_offload_start(struct io_ring_ctx
*ctx
)
8220 struct io_sq_data
*sqd
= ctx
->sq_data
;
8222 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && sqd
->thread
)
8223 wake_up_process(sqd
->thread
);
8226 static inline void __io_unaccount_mem(struct user_struct
*user
,
8227 unsigned long nr_pages
)
8229 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8232 static inline int __io_account_mem(struct user_struct
*user
,
8233 unsigned long nr_pages
)
8235 unsigned long page_limit
, cur_pages
, new_pages
;
8237 /* Don't allow more pages than we can safely lock */
8238 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8241 cur_pages
= atomic_long_read(&user
->locked_vm
);
8242 new_pages
= cur_pages
+ nr_pages
;
8243 if (new_pages
> page_limit
)
8245 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8246 new_pages
) != cur_pages
);
8251 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8252 enum io_mem_account acct
)
8255 __io_unaccount_mem(ctx
->user
, nr_pages
);
8257 if (ctx
->mm_account
) {
8258 if (acct
== ACCT_LOCKED
) {
8259 mmap_write_lock(ctx
->mm_account
);
8260 ctx
->mm_account
->locked_vm
-= nr_pages
;
8261 mmap_write_unlock(ctx
->mm_account
);
8262 }else if (acct
== ACCT_PINNED
) {
8263 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8268 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
8269 enum io_mem_account acct
)
8273 if (ctx
->limit_mem
) {
8274 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8279 if (ctx
->mm_account
) {
8280 if (acct
== ACCT_LOCKED
) {
8281 mmap_write_lock(ctx
->mm_account
);
8282 ctx
->mm_account
->locked_vm
+= nr_pages
;
8283 mmap_write_unlock(ctx
->mm_account
);
8284 } else if (acct
== ACCT_PINNED
) {
8285 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8292 static void io_mem_free(void *ptr
)
8299 page
= virt_to_head_page(ptr
);
8300 if (put_page_testzero(page
))
8301 free_compound_page(page
);
8304 static void *io_mem_alloc(size_t size
)
8306 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8309 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8312 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8315 struct io_rings
*rings
;
8316 size_t off
, sq_array_size
;
8318 off
= struct_size(rings
, cqes
, cq_entries
);
8319 if (off
== SIZE_MAX
)
8323 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8331 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8332 if (sq_array_size
== SIZE_MAX
)
8335 if (check_add_overflow(off
, sq_array_size
, &off
))
8341 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
8345 pages
= (size_t)1 << get_order(
8346 rings_size(sq_entries
, cq_entries
, NULL
));
8347 pages
+= (size_t)1 << get_order(
8348 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
8353 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8357 if (!ctx
->user_bufs
)
8360 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8361 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8363 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
8364 unpin_user_page(imu
->bvec
[j
].bv_page
);
8366 if (imu
->acct_pages
)
8367 io_unaccount_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8372 kfree(ctx
->user_bufs
);
8373 ctx
->user_bufs
= NULL
;
8374 ctx
->nr_user_bufs
= 0;
8378 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8379 void __user
*arg
, unsigned index
)
8381 struct iovec __user
*src
;
8383 #ifdef CONFIG_COMPAT
8385 struct compat_iovec __user
*ciovs
;
8386 struct compat_iovec ciov
;
8388 ciovs
= (struct compat_iovec __user
*) arg
;
8389 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8392 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8393 dst
->iov_len
= ciov
.iov_len
;
8397 src
= (struct iovec __user
*) arg
;
8398 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8404 * Not super efficient, but this is just a registration time. And we do cache
8405 * the last compound head, so generally we'll only do a full search if we don't
8408 * We check if the given compound head page has already been accounted, to
8409 * avoid double accounting it. This allows us to account the full size of the
8410 * page, not just the constituent pages of a huge page.
8412 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8413 int nr_pages
, struct page
*hpage
)
8417 /* check current page array */
8418 for (i
= 0; i
< nr_pages
; i
++) {
8419 if (!PageCompound(pages
[i
]))
8421 if (compound_head(pages
[i
]) == hpage
)
8425 /* check previously registered pages */
8426 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8427 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8429 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8430 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8432 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8440 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8441 int nr_pages
, struct io_mapped_ubuf
*imu
,
8442 struct page
**last_hpage
)
8446 for (i
= 0; i
< nr_pages
; i
++) {
8447 if (!PageCompound(pages
[i
])) {
8452 hpage
= compound_head(pages
[i
]);
8453 if (hpage
== *last_hpage
)
8455 *last_hpage
= hpage
;
8456 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8458 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8462 if (!imu
->acct_pages
)
8465 ret
= io_account_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8467 imu
->acct_pages
= 0;
8471 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
8472 struct io_mapped_ubuf
*imu
,
8473 struct page
**last_hpage
)
8475 struct vm_area_struct
**vmas
= NULL
;
8476 struct page
**pages
= NULL
;
8477 unsigned long off
, start
, end
, ubuf
;
8479 int ret
, pret
, nr_pages
, i
;
8481 ubuf
= (unsigned long) iov
->iov_base
;
8482 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8483 start
= ubuf
>> PAGE_SHIFT
;
8484 nr_pages
= end
- start
;
8488 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
8492 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
8497 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8503 mmap_read_lock(current
->mm
);
8504 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
8506 if (pret
== nr_pages
) {
8507 /* don't support file backed memory */
8508 for (i
= 0; i
< nr_pages
; i
++) {
8509 struct vm_area_struct
*vma
= vmas
[i
];
8512 !is_file_hugepages(vma
->vm_file
)) {
8518 ret
= pret
< 0 ? pret
: -EFAULT
;
8520 mmap_read_unlock(current
->mm
);
8523 * if we did partial map, or found file backed vmas,
8524 * release any pages we did get
8527 unpin_user_pages(pages
, pret
);
8532 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
8534 unpin_user_pages(pages
, pret
);
8539 off
= ubuf
& ~PAGE_MASK
;
8540 size
= iov
->iov_len
;
8541 for (i
= 0; i
< nr_pages
; i
++) {
8544 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8545 imu
->bvec
[i
].bv_page
= pages
[i
];
8546 imu
->bvec
[i
].bv_len
= vec_len
;
8547 imu
->bvec
[i
].bv_offset
= off
;
8551 /* store original address for later verification */
8553 imu
->len
= iov
->iov_len
;
8554 imu
->nr_bvecs
= nr_pages
;
8562 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
8566 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8569 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8571 if (!ctx
->user_bufs
)
8577 static int io_buffer_validate(struct iovec
*iov
)
8580 * Don't impose further limits on the size and buffer
8581 * constraints here, we'll -EINVAL later when IO is
8582 * submitted if they are wrong.
8584 if (!iov
->iov_base
|| !iov
->iov_len
)
8587 /* arbitrary limit, but we need something */
8588 if (iov
->iov_len
> SZ_1G
)
8594 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8595 unsigned int nr_args
)
8599 struct page
*last_hpage
= NULL
;
8601 ret
= io_buffers_map_alloc(ctx
, nr_args
);
8605 for (i
= 0; i
< nr_args
; i
++) {
8606 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8608 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8612 ret
= io_buffer_validate(&iov
);
8616 ret
= io_sqe_buffer_register(ctx
, &iov
, imu
, &last_hpage
);
8620 ctx
->nr_user_bufs
++;
8624 io_sqe_buffers_unregister(ctx
);
8629 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8631 __s32 __user
*fds
= arg
;
8637 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8640 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8641 if (IS_ERR(ctx
->cq_ev_fd
)) {
8642 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8643 ctx
->cq_ev_fd
= NULL
;
8650 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8652 if (ctx
->cq_ev_fd
) {
8653 eventfd_ctx_put(ctx
->cq_ev_fd
);
8654 ctx
->cq_ev_fd
= NULL
;
8661 static int __io_destroy_buffers(int id
, void *p
, void *data
)
8663 struct io_ring_ctx
*ctx
= data
;
8664 struct io_buffer
*buf
= p
;
8666 __io_remove_buffers(ctx
, buf
, id
, -1U);
8670 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8672 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
8673 idr_destroy(&ctx
->io_buffer_idr
);
8676 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8678 io_finish_async(ctx
);
8679 io_sqe_buffers_unregister(ctx
);
8681 if (ctx
->sqo_task
) {
8682 put_task_struct(ctx
->sqo_task
);
8683 ctx
->sqo_task
= NULL
;
8684 mmdrop(ctx
->mm_account
);
8685 ctx
->mm_account
= NULL
;
8688 #ifdef CONFIG_BLK_CGROUP
8689 if (ctx
->sqo_blkcg_css
)
8690 css_put(ctx
->sqo_blkcg_css
);
8693 io_sqe_files_unregister(ctx
);
8694 io_eventfd_unregister(ctx
);
8695 io_destroy_buffers(ctx
);
8696 idr_destroy(&ctx
->personality_idr
);
8698 #if defined(CONFIG_UNIX)
8699 if (ctx
->ring_sock
) {
8700 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8701 sock_release(ctx
->ring_sock
);
8705 io_mem_free(ctx
->rings
);
8706 io_mem_free(ctx
->sq_sqes
);
8708 percpu_ref_exit(&ctx
->refs
);
8709 free_uid(ctx
->user
);
8710 put_cred(ctx
->creds
);
8711 kfree(ctx
->cancel_hash
);
8712 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
8716 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8718 struct io_ring_ctx
*ctx
= file
->private_data
;
8721 poll_wait(file
, &ctx
->cq_wait
, wait
);
8723 * synchronizes with barrier from wq_has_sleeper call in
8727 if (!io_sqring_full(ctx
))
8728 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8729 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
8730 if (io_cqring_events(ctx
))
8731 mask
|= EPOLLIN
| EPOLLRDNORM
;
8736 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8738 struct io_ring_ctx
*ctx
= file
->private_data
;
8740 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8743 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8745 struct io_identity
*iod
;
8747 iod
= idr_remove(&ctx
->personality_idr
, id
);
8749 put_cred(iod
->creds
);
8750 if (refcount_dec_and_test(&iod
->count
))
8758 static int io_remove_personalities(int id
, void *p
, void *data
)
8760 struct io_ring_ctx
*ctx
= data
;
8762 io_unregister_personality(ctx
, id
);
8766 static void io_ring_exit_work(struct work_struct
*work
)
8768 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
8772 * If we're doing polled IO and end up having requests being
8773 * submitted async (out-of-line), then completions can come in while
8774 * we're waiting for refs to drop. We need to reap these manually,
8775 * as nobody else will be looking for them.
8778 io_uring_try_cancel_requests(ctx
, NULL
, NULL
);
8779 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8780 io_ring_ctx_free(ctx
);
8783 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8785 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8787 return req
->ctx
== data
;
8790 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8792 mutex_lock(&ctx
->uring_lock
);
8793 percpu_ref_kill(&ctx
->refs
);
8795 if (WARN_ON_ONCE((ctx
->flags
& IORING_SETUP_SQPOLL
) && !ctx
->sqo_dead
))
8798 /* if force is set, the ring is going away. always drop after that */
8799 ctx
->cq_overflow_flushed
= 1;
8801 __io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8802 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
8803 mutex_unlock(&ctx
->uring_lock
);
8805 io_kill_timeouts(ctx
, NULL
, NULL
);
8806 io_poll_remove_all(ctx
, NULL
, NULL
);
8809 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8811 /* if we failed setting up the ctx, we might not have any rings */
8812 io_iopoll_try_reap_events(ctx
);
8815 * Do this upfront, so we won't have a grace period where the ring
8816 * is closed but resources aren't reaped yet. This can cause
8817 * spurious failure in setting up a new ring.
8819 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
8822 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8824 * Use system_unbound_wq to avoid spawning tons of event kworkers
8825 * if we're exiting a ton of rings at the same time. It just adds
8826 * noise and overhead, there's no discernable change in runtime
8827 * over using system_wq.
8829 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8832 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8834 struct io_ring_ctx
*ctx
= file
->private_data
;
8836 file
->private_data
= NULL
;
8837 io_ring_ctx_wait_and_kill(ctx
);
8841 struct io_task_cancel
{
8842 struct task_struct
*task
;
8843 struct files_struct
*files
;
8846 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8848 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8849 struct io_task_cancel
*cancel
= data
;
8852 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8853 unsigned long flags
;
8854 struct io_ring_ctx
*ctx
= req
->ctx
;
8856 /* protect against races with linked timeouts */
8857 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8858 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8859 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8861 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8866 static void io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8867 struct task_struct
*task
,
8868 struct files_struct
*files
)
8870 struct io_defer_entry
*de
= NULL
;
8873 spin_lock_irq(&ctx
->completion_lock
);
8874 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8875 if (io_match_task(de
->req
, task
, files
)) {
8876 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8880 spin_unlock_irq(&ctx
->completion_lock
);
8882 while (!list_empty(&list
)) {
8883 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8884 list_del_init(&de
->list
);
8885 req_set_fail_links(de
->req
);
8886 io_put_req(de
->req
);
8887 io_req_complete(de
->req
, -ECANCELED
);
8892 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
8893 struct task_struct
*task
,
8894 struct files_struct
*files
)
8896 struct io_task_cancel cancel
= { .task
= task
, .files
= files
, };
8899 enum io_wq_cancel cret
;
8903 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
,
8905 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8908 /* SQPOLL thread does its own polling */
8909 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) && !files
) {
8910 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8911 io_iopoll_try_reap_events(ctx
);
8916 ret
|= io_poll_remove_all(ctx
, task
, files
);
8917 ret
|= io_kill_timeouts(ctx
, task
, files
);
8918 ret
|= io_run_task_work();
8919 io_cqring_overflow_flush(ctx
, true, task
, files
);
8926 static int io_uring_count_inflight(struct io_ring_ctx
*ctx
,
8927 struct task_struct
*task
,
8928 struct files_struct
*files
)
8930 struct io_kiocb
*req
;
8933 spin_lock_irq(&ctx
->inflight_lock
);
8934 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
)
8935 cnt
+= io_match_task(req
, task
, files
);
8936 spin_unlock_irq(&ctx
->inflight_lock
);
8940 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
8941 struct task_struct
*task
,
8942 struct files_struct
*files
)
8944 while (!list_empty_careful(&ctx
->inflight_list
)) {
8948 inflight
= io_uring_count_inflight(ctx
, task
, files
);
8952 io_uring_try_cancel_requests(ctx
, task
, files
);
8953 prepare_to_wait(&task
->io_uring
->wait
, &wait
,
8954 TASK_UNINTERRUPTIBLE
);
8955 if (inflight
== io_uring_count_inflight(ctx
, task
, files
))
8957 finish_wait(&task
->io_uring
->wait
, &wait
);
8961 static void io_disable_sqo_submit(struct io_ring_ctx
*ctx
)
8963 mutex_lock(&ctx
->uring_lock
);
8965 mutex_unlock(&ctx
->uring_lock
);
8967 /* make sure callers enter the ring to get error */
8969 io_ring_set_wakeup_flag(ctx
);
8973 * We need to iteratively cancel requests, in case a request has dependent
8974 * hard links. These persist even for failure of cancelations, hence keep
8975 * looping until none are found.
8977 static void io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
8978 struct files_struct
*files
)
8980 struct task_struct
*task
= current
;
8982 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
8983 io_disable_sqo_submit(ctx
);
8984 task
= ctx
->sq_data
->thread
;
8985 atomic_inc(&task
->io_uring
->in_idle
);
8986 io_sq_thread_park(ctx
->sq_data
);
8989 io_cancel_defer_files(ctx
, task
, files
);
8991 io_uring_cancel_files(ctx
, task
, files
);
8993 io_uring_try_cancel_requests(ctx
, task
, NULL
);
8995 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
) {
8996 atomic_dec(&task
->io_uring
->in_idle
);
8998 * If the files that are going away are the ones in the thread
8999 * identity, clear them out.
9001 if (task
->io_uring
->identity
->files
== files
)
9002 task
->io_uring
->identity
->files
= NULL
;
9003 io_sq_thread_unpark(ctx
->sq_data
);
9008 * Note that this task has used io_uring. We use it for cancelation purposes.
9010 static int io_uring_add_task_file(struct io_ring_ctx
*ctx
, struct file
*file
)
9012 struct io_uring_task
*tctx
= current
->io_uring
;
9015 if (unlikely(!tctx
)) {
9016 ret
= io_uring_alloc_task_context(current
);
9019 tctx
= current
->io_uring
;
9021 if (tctx
->last
!= file
) {
9022 void *old
= xa_load(&tctx
->xa
, (unsigned long)file
);
9026 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)file
,
9033 /* one and only SQPOLL file note, held by sqo_task */
9034 WARN_ON_ONCE((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
9035 current
!= ctx
->sqo_task
);
9041 * This is race safe in that the task itself is doing this, hence it
9042 * cannot be going through the exit/cancel paths at the same time.
9043 * This cannot be modified while exit/cancel is running.
9045 if (!tctx
->sqpoll
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9046 tctx
->sqpoll
= true;
9052 * Remove this io_uring_file -> task mapping.
9054 static void io_uring_del_task_file(struct file
*file
)
9056 struct io_uring_task
*tctx
= current
->io_uring
;
9058 if (tctx
->last
== file
)
9060 file
= xa_erase(&tctx
->xa
, (unsigned long)file
);
9065 static void io_uring_remove_task_files(struct io_uring_task
*tctx
)
9068 unsigned long index
;
9070 xa_for_each(&tctx
->xa
, index
, file
)
9071 io_uring_del_task_file(file
);
9074 void __io_uring_files_cancel(struct files_struct
*files
)
9076 struct io_uring_task
*tctx
= current
->io_uring
;
9078 unsigned long index
;
9080 /* make sure overflow events are dropped */
9081 atomic_inc(&tctx
->in_idle
);
9082 xa_for_each(&tctx
->xa
, index
, file
)
9083 io_uring_cancel_task_requests(file
->private_data
, files
);
9084 atomic_dec(&tctx
->in_idle
);
9087 io_uring_remove_task_files(tctx
);
9090 static s64
tctx_inflight(struct io_uring_task
*tctx
)
9092 unsigned long index
;
9096 inflight
= percpu_counter_sum(&tctx
->inflight
);
9101 * If we have SQPOLL rings, then we need to iterate and find them, and
9102 * add the pending count for those.
9104 xa_for_each(&tctx
->xa
, index
, file
) {
9105 struct io_ring_ctx
*ctx
= file
->private_data
;
9107 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9108 struct io_uring_task
*__tctx
= ctx
->sqo_task
->io_uring
;
9110 inflight
+= percpu_counter_sum(&__tctx
->inflight
);
9118 * Find any io_uring fd that this task has registered or done IO on, and cancel
9121 void __io_uring_task_cancel(void)
9123 struct io_uring_task
*tctx
= current
->io_uring
;
9127 /* make sure overflow events are dropped */
9128 atomic_inc(&tctx
->in_idle
);
9130 /* trigger io_disable_sqo_submit() */
9132 __io_uring_files_cancel(NULL
);
9135 /* read completions before cancelations */
9136 inflight
= tctx_inflight(tctx
);
9139 __io_uring_files_cancel(NULL
);
9141 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9144 * If we've seen completions, retry without waiting. This
9145 * avoids a race where a completion comes in before we did
9146 * prepare_to_wait().
9148 if (inflight
== tctx_inflight(tctx
))
9150 finish_wait(&tctx
->wait
, &wait
);
9153 atomic_dec(&tctx
->in_idle
);
9155 io_uring_remove_task_files(tctx
);
9158 static int io_uring_flush(struct file
*file
, void *data
)
9160 struct io_uring_task
*tctx
= current
->io_uring
;
9161 struct io_ring_ctx
*ctx
= file
->private_data
;
9163 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
))
9164 io_uring_cancel_task_requests(ctx
, NULL
);
9169 /* we should have cancelled and erased it before PF_EXITING */
9170 WARN_ON_ONCE((current
->flags
& PF_EXITING
) &&
9171 xa_load(&tctx
->xa
, (unsigned long)file
));
9174 * fput() is pending, will be 2 if the only other ref is our potential
9175 * task file note. If the task is exiting, drop regardless of count.
9177 if (atomic_long_read(&file
->f_count
) != 2)
9180 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9181 /* there is only one file note, which is owned by sqo_task */
9182 WARN_ON_ONCE(ctx
->sqo_task
!= current
&&
9183 xa_load(&tctx
->xa
, (unsigned long)file
));
9184 /* sqo_dead check is for when this happens after cancellation */
9185 WARN_ON_ONCE(ctx
->sqo_task
== current
&& !ctx
->sqo_dead
&&
9186 !xa_load(&tctx
->xa
, (unsigned long)file
));
9188 io_disable_sqo_submit(ctx
);
9191 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) || ctx
->sqo_task
== current
)
9192 io_uring_del_task_file(file
);
9196 static void *io_uring_validate_mmap_request(struct file
*file
,
9197 loff_t pgoff
, size_t sz
)
9199 struct io_ring_ctx
*ctx
= file
->private_data
;
9200 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9205 case IORING_OFF_SQ_RING
:
9206 case IORING_OFF_CQ_RING
:
9209 case IORING_OFF_SQES
:
9213 return ERR_PTR(-EINVAL
);
9216 page
= virt_to_head_page(ptr
);
9217 if (sz
> page_size(page
))
9218 return ERR_PTR(-EINVAL
);
9225 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9227 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9231 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9233 return PTR_ERR(ptr
);
9235 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9236 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9239 #else /* !CONFIG_MMU */
9241 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9243 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9246 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9248 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9251 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9252 unsigned long addr
, unsigned long len
,
9253 unsigned long pgoff
, unsigned long flags
)
9257 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9259 return PTR_ERR(ptr
);
9261 return (unsigned long) ptr
;
9264 #endif /* !CONFIG_MMU */
9266 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9272 if (!io_sqring_full(ctx
))
9275 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9277 if (unlikely(ctx
->sqo_dead
)) {
9282 if (!io_sqring_full(ctx
))
9286 } while (!signal_pending(current
));
9288 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9293 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9294 struct __kernel_timespec __user
**ts
,
9295 const sigset_t __user
**sig
)
9297 struct io_uring_getevents_arg arg
;
9300 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9301 * is just a pointer to the sigset_t.
9303 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9304 *sig
= (const sigset_t __user
*) argp
;
9310 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9311 * timespec and sigset_t pointers if good.
9313 if (*argsz
!= sizeof(arg
))
9315 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9317 *sig
= u64_to_user_ptr(arg
.sigmask
);
9318 *argsz
= arg
.sigmask_sz
;
9319 *ts
= u64_to_user_ptr(arg
.ts
);
9323 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9324 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9327 struct io_ring_ctx
*ctx
;
9334 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9335 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
))
9343 if (f
.file
->f_op
!= &io_uring_fops
)
9347 ctx
= f
.file
->private_data
;
9348 if (!percpu_ref_tryget(&ctx
->refs
))
9352 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9356 * For SQ polling, the thread will do all submissions and completions.
9357 * Just return the requested submit count, and wake the thread if
9361 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9362 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
9365 if (unlikely(ctx
->sqo_dead
))
9367 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9368 wake_up(&ctx
->sq_data
->wait
);
9369 if (flags
& IORING_ENTER_SQ_WAIT
) {
9370 ret
= io_sqpoll_wait_sq(ctx
);
9374 submitted
= to_submit
;
9375 } else if (to_submit
) {
9376 ret
= io_uring_add_task_file(ctx
, f
.file
);
9379 mutex_lock(&ctx
->uring_lock
);
9380 submitted
= io_submit_sqes(ctx
, to_submit
);
9381 mutex_unlock(&ctx
->uring_lock
);
9383 if (submitted
!= to_submit
)
9386 if (flags
& IORING_ENTER_GETEVENTS
) {
9387 const sigset_t __user
*sig
;
9388 struct __kernel_timespec __user
*ts
;
9390 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9394 min_complete
= min(min_complete
, ctx
->cq_entries
);
9397 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9398 * space applications don't need to do io completion events
9399 * polling again, they can rely on io_sq_thread to do polling
9400 * work, which can reduce cpu usage and uring_lock contention.
9402 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9403 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9404 ret
= io_iopoll_check(ctx
, min_complete
);
9406 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9411 percpu_ref_put(&ctx
->refs
);
9414 return submitted
? submitted
: ret
;
9417 #ifdef CONFIG_PROC_FS
9418 static int io_uring_show_cred(int id
, void *p
, void *data
)
9420 struct io_identity
*iod
= p
;
9421 const struct cred
*cred
= iod
->creds
;
9422 struct seq_file
*m
= data
;
9423 struct user_namespace
*uns
= seq_user_ns(m
);
9424 struct group_info
*gi
;
9429 seq_printf(m
, "%5d\n", id
);
9430 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9431 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9432 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9433 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9434 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9435 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9436 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9437 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9438 seq_puts(m
, "\n\tGroups:\t");
9439 gi
= cred
->group_info
;
9440 for (g
= 0; g
< gi
->ngroups
; g
++) {
9441 seq_put_decimal_ull(m
, g
? " " : "",
9442 from_kgid_munged(uns
, gi
->gid
[g
]));
9444 seq_puts(m
, "\n\tCapEff:\t");
9445 cap
= cred
->cap_effective
;
9446 CAP_FOR_EACH_U32(__capi
)
9447 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9452 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9454 struct io_sq_data
*sq
= NULL
;
9459 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9460 * since fdinfo case grabs it in the opposite direction of normal use
9461 * cases. If we fail to get the lock, we just don't iterate any
9462 * structures that could be going away outside the io_uring mutex.
9464 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9466 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9469 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9470 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9471 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9472 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9473 struct file
*f
= *io_fixed_file_slot(ctx
->file_data
, i
);
9476 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9478 seq_printf(m
, "%5u: <none>\n", i
);
9480 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9481 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9482 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9484 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9485 (unsigned int) buf
->len
);
9487 if (has_lock
&& !idr_is_empty(&ctx
->personality_idr
)) {
9488 seq_printf(m
, "Personalities:\n");
9489 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
9491 seq_printf(m
, "PollList:\n");
9492 spin_lock_irq(&ctx
->completion_lock
);
9493 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9494 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9495 struct io_kiocb
*req
;
9497 hlist_for_each_entry(req
, list
, hash_node
)
9498 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9499 req
->task
->task_works
!= NULL
);
9501 spin_unlock_irq(&ctx
->completion_lock
);
9503 mutex_unlock(&ctx
->uring_lock
);
9506 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9508 struct io_ring_ctx
*ctx
= f
->private_data
;
9510 if (percpu_ref_tryget(&ctx
->refs
)) {
9511 __io_uring_show_fdinfo(ctx
, m
);
9512 percpu_ref_put(&ctx
->refs
);
9517 static const struct file_operations io_uring_fops
= {
9518 .release
= io_uring_release
,
9519 .flush
= io_uring_flush
,
9520 .mmap
= io_uring_mmap
,
9522 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9523 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9525 .poll
= io_uring_poll
,
9526 .fasync
= io_uring_fasync
,
9527 #ifdef CONFIG_PROC_FS
9528 .show_fdinfo
= io_uring_show_fdinfo
,
9532 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9533 struct io_uring_params
*p
)
9535 struct io_rings
*rings
;
9536 size_t size
, sq_array_offset
;
9538 /* make sure these are sane, as we already accounted them */
9539 ctx
->sq_entries
= p
->sq_entries
;
9540 ctx
->cq_entries
= p
->cq_entries
;
9542 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9543 if (size
== SIZE_MAX
)
9546 rings
= io_mem_alloc(size
);
9551 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9552 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9553 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9554 rings
->sq_ring_entries
= p
->sq_entries
;
9555 rings
->cq_ring_entries
= p
->cq_entries
;
9556 ctx
->sq_mask
= rings
->sq_ring_mask
;
9557 ctx
->cq_mask
= rings
->cq_ring_mask
;
9559 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9560 if (size
== SIZE_MAX
) {
9561 io_mem_free(ctx
->rings
);
9566 ctx
->sq_sqes
= io_mem_alloc(size
);
9567 if (!ctx
->sq_sqes
) {
9568 io_mem_free(ctx
->rings
);
9576 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9580 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9584 ret
= io_uring_add_task_file(ctx
, file
);
9589 fd_install(fd
, file
);
9594 * Allocate an anonymous fd, this is what constitutes the application
9595 * visible backing of an io_uring instance. The application mmaps this
9596 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9597 * we have to tie this fd to a socket for file garbage collection purposes.
9599 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9602 #if defined(CONFIG_UNIX)
9605 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9608 return ERR_PTR(ret
);
9611 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9612 O_RDWR
| O_CLOEXEC
);
9613 #if defined(CONFIG_UNIX)
9615 sock_release(ctx
->ring_sock
);
9616 ctx
->ring_sock
= NULL
;
9618 ctx
->ring_sock
->file
= file
;
9624 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9625 struct io_uring_params __user
*params
)
9627 struct user_struct
*user
= NULL
;
9628 struct io_ring_ctx
*ctx
;
9635 if (entries
> IORING_MAX_ENTRIES
) {
9636 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9638 entries
= IORING_MAX_ENTRIES
;
9642 * Use twice as many entries for the CQ ring. It's possible for the
9643 * application to drive a higher depth than the size of the SQ ring,
9644 * since the sqes are only used at submission time. This allows for
9645 * some flexibility in overcommitting a bit. If the application has
9646 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9647 * of CQ ring entries manually.
9649 p
->sq_entries
= roundup_pow_of_two(entries
);
9650 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9652 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9653 * to a power-of-two, if it isn't already. We do NOT impose
9654 * any cq vs sq ring sizing.
9658 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9659 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9661 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9663 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9664 if (p
->cq_entries
< p
->sq_entries
)
9667 p
->cq_entries
= 2 * p
->sq_entries
;
9670 user
= get_uid(current_user());
9671 limit_mem
= !capable(CAP_IPC_LOCK
);
9674 ret
= __io_account_mem(user
,
9675 ring_pages(p
->sq_entries
, p
->cq_entries
));
9682 ctx
= io_ring_ctx_alloc(p
);
9685 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
9690 ctx
->compat
= in_compat_syscall();
9692 ctx
->creds
= get_current_cred();
9694 ctx
->loginuid
= current
->loginuid
;
9695 ctx
->sessionid
= current
->sessionid
;
9697 ctx
->sqo_task
= get_task_struct(current
);
9700 * This is just grabbed for accounting purposes. When a process exits,
9701 * the mm is exited and dropped before the files, hence we need to hang
9702 * on to this mm purely for the purposes of being able to unaccount
9703 * memory (locked/pinned vm). It's not used for anything else.
9705 mmgrab(current
->mm
);
9706 ctx
->mm_account
= current
->mm
;
9708 #ifdef CONFIG_BLK_CGROUP
9710 * The sq thread will belong to the original cgroup it was inited in.
9711 * If the cgroup goes offline (e.g. disabling the io controller), then
9712 * issued bios will be associated with the closest cgroup later in the
9716 ctx
->sqo_blkcg_css
= blkcg_css();
9717 ret
= css_tryget_online(ctx
->sqo_blkcg_css
);
9720 /* don't init against a dying cgroup, have the user try again */
9721 ctx
->sqo_blkcg_css
= NULL
;
9728 * Account memory _before_ installing the file descriptor. Once
9729 * the descriptor is installed, it can get closed at any time. Also
9730 * do this before hitting the general error path, as ring freeing
9731 * will un-account as well.
9733 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
9735 ctx
->limit_mem
= limit_mem
;
9737 ret
= io_allocate_scq_urings(ctx
, p
);
9741 ret
= io_sq_offload_create(ctx
, p
);
9745 if (!(p
->flags
& IORING_SETUP_R_DISABLED
))
9746 io_sq_offload_start(ctx
);
9748 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9749 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9750 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9751 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9752 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9753 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9754 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9755 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9757 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9758 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9759 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9760 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9761 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9762 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9763 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9764 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9766 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9767 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9768 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9769 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9770 IORING_FEAT_EXT_ARG
;
9772 if (copy_to_user(params
, p
, sizeof(*p
))) {
9777 file
= io_uring_get_file(ctx
);
9779 ret
= PTR_ERR(file
);
9784 * Install ring fd as the very last thing, so we don't risk someone
9785 * having closed it before we finish setup
9787 ret
= io_uring_install_fd(ctx
, file
);
9789 io_disable_sqo_submit(ctx
);
9790 /* fput will clean it up */
9795 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9798 io_disable_sqo_submit(ctx
);
9799 io_ring_ctx_wait_and_kill(ctx
);
9804 * Sets up an aio uring context, and returns the fd. Applications asks for a
9805 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9806 * params structure passed in.
9808 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9810 struct io_uring_params p
;
9813 if (copy_from_user(&p
, params
, sizeof(p
)))
9815 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9820 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9821 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9822 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9823 IORING_SETUP_R_DISABLED
))
9826 return io_uring_create(entries
, &p
, params
);
9829 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9830 struct io_uring_params __user
*, params
)
9832 return io_uring_setup(entries
, params
);
9835 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9837 struct io_uring_probe
*p
;
9841 size
= struct_size(p
, ops
, nr_args
);
9842 if (size
== SIZE_MAX
)
9844 p
= kzalloc(size
, GFP_KERNEL
);
9849 if (copy_from_user(p
, arg
, size
))
9852 if (memchr_inv(p
, 0, size
))
9855 p
->last_op
= IORING_OP_LAST
- 1;
9856 if (nr_args
> IORING_OP_LAST
)
9857 nr_args
= IORING_OP_LAST
;
9859 for (i
= 0; i
< nr_args
; i
++) {
9861 if (!io_op_defs
[i
].not_supported
)
9862 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9867 if (copy_to_user(arg
, p
, size
))
9874 static int io_register_personality(struct io_ring_ctx
*ctx
)
9876 struct io_identity
*id
;
9879 id
= kmalloc(sizeof(*id
), GFP_KERNEL
);
9883 io_init_identity(id
);
9884 id
->creds
= get_current_cred();
9886 ret
= idr_alloc_cyclic(&ctx
->personality_idr
, id
, 1, USHRT_MAX
, GFP_KERNEL
);
9888 put_cred(id
->creds
);
9894 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9895 unsigned int nr_args
)
9897 struct io_uring_restriction
*res
;
9901 /* Restrictions allowed only if rings started disabled */
9902 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9905 /* We allow only a single restrictions registration */
9906 if (ctx
->restrictions
.registered
)
9909 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9912 size
= array_size(nr_args
, sizeof(*res
));
9913 if (size
== SIZE_MAX
)
9916 res
= memdup_user(arg
, size
);
9918 return PTR_ERR(res
);
9922 for (i
= 0; i
< nr_args
; i
++) {
9923 switch (res
[i
].opcode
) {
9924 case IORING_RESTRICTION_REGISTER_OP
:
9925 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9930 __set_bit(res
[i
].register_op
,
9931 ctx
->restrictions
.register_op
);
9933 case IORING_RESTRICTION_SQE_OP
:
9934 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9939 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9941 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9942 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9944 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9945 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9954 /* Reset all restrictions if an error happened */
9956 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9958 ctx
->restrictions
.registered
= true;
9964 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
9966 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9969 if (ctx
->restrictions
.registered
)
9970 ctx
->restricted
= 1;
9972 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
9974 io_sq_offload_start(ctx
);
9979 static bool io_register_op_must_quiesce(int op
)
9982 case IORING_UNREGISTER_FILES
:
9983 case IORING_REGISTER_FILES_UPDATE
:
9984 case IORING_REGISTER_PROBE
:
9985 case IORING_REGISTER_PERSONALITY
:
9986 case IORING_UNREGISTER_PERSONALITY
:
9993 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
9994 void __user
*arg
, unsigned nr_args
)
9995 __releases(ctx
->uring_lock
)
9996 __acquires(ctx
->uring_lock
)
10001 * We're inside the ring mutex, if the ref is already dying, then
10002 * someone else killed the ctx or is already going through
10003 * io_uring_register().
10005 if (percpu_ref_is_dying(&ctx
->refs
))
10008 if (io_register_op_must_quiesce(opcode
)) {
10009 percpu_ref_kill(&ctx
->refs
);
10012 * Drop uring mutex before waiting for references to exit. If
10013 * another thread is currently inside io_uring_enter() it might
10014 * need to grab the uring_lock to make progress. If we hold it
10015 * here across the drain wait, then we can deadlock. It's safe
10016 * to drop the mutex here, since no new references will come in
10017 * after we've killed the percpu ref.
10019 mutex_unlock(&ctx
->uring_lock
);
10021 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10024 ret
= io_run_task_work_sig();
10029 mutex_lock(&ctx
->uring_lock
);
10032 percpu_ref_resurrect(&ctx
->refs
);
10037 if (ctx
->restricted
) {
10038 if (opcode
>= IORING_REGISTER_LAST
) {
10043 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
10050 case IORING_REGISTER_BUFFERS
:
10051 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
);
10053 case IORING_UNREGISTER_BUFFERS
:
10055 if (arg
|| nr_args
)
10057 ret
= io_sqe_buffers_unregister(ctx
);
10059 case IORING_REGISTER_FILES
:
10060 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
10062 case IORING_UNREGISTER_FILES
:
10064 if (arg
|| nr_args
)
10066 ret
= io_sqe_files_unregister(ctx
);
10068 case IORING_REGISTER_FILES_UPDATE
:
10069 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
10071 case IORING_REGISTER_EVENTFD
:
10072 case IORING_REGISTER_EVENTFD_ASYNC
:
10076 ret
= io_eventfd_register(ctx
, arg
);
10079 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10080 ctx
->eventfd_async
= 1;
10082 ctx
->eventfd_async
= 0;
10084 case IORING_UNREGISTER_EVENTFD
:
10086 if (arg
|| nr_args
)
10088 ret
= io_eventfd_unregister(ctx
);
10090 case IORING_REGISTER_PROBE
:
10092 if (!arg
|| nr_args
> 256)
10094 ret
= io_probe(ctx
, arg
, nr_args
);
10096 case IORING_REGISTER_PERSONALITY
:
10098 if (arg
|| nr_args
)
10100 ret
= io_register_personality(ctx
);
10102 case IORING_UNREGISTER_PERSONALITY
:
10106 ret
= io_unregister_personality(ctx
, nr_args
);
10108 case IORING_REGISTER_ENABLE_RINGS
:
10110 if (arg
|| nr_args
)
10112 ret
= io_register_enable_rings(ctx
);
10114 case IORING_REGISTER_RESTRICTIONS
:
10115 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10123 if (io_register_op_must_quiesce(opcode
)) {
10124 /* bring the ctx back to life */
10125 percpu_ref_reinit(&ctx
->refs
);
10127 reinit_completion(&ctx
->ref_comp
);
10132 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10133 void __user
*, arg
, unsigned int, nr_args
)
10135 struct io_ring_ctx
*ctx
;
10144 if (f
.file
->f_op
!= &io_uring_fops
)
10147 ctx
= f
.file
->private_data
;
10149 mutex_lock(&ctx
->uring_lock
);
10150 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10151 mutex_unlock(&ctx
->uring_lock
);
10152 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10153 ctx
->cq_ev_fd
!= NULL
, ret
);
10159 static int __init
io_uring_init(void)
10161 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10162 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10163 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10166 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10167 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10168 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10169 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10170 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10171 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10172 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10173 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10174 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10175 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10176 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10177 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10178 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10179 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10180 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10181 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10182 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10183 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10184 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10185 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10186 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10187 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10188 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10189 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10190 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10191 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10192 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10193 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10194 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10195 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10196 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10198 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10199 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
10200 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
10203 __initcall(io_uring_init
);