1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp
;
109 u32 tail ____cacheline_aligned_in_smp
;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq
, cq
;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask
, cq_ring_mask
;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries
, cq_ring_entries
;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
190 struct io_mapped_ubuf
{
193 struct bio_vec
*bvec
;
194 unsigned int nr_bvecs
;
195 unsigned long acct_pages
;
198 struct fixed_file_table
{
202 struct fixed_file_ref_node
{
203 struct percpu_ref refs
;
204 struct list_head node
;
205 struct list_head file_list
;
206 struct fixed_file_data
*file_data
;
207 struct llist_node llist
;
210 struct fixed_file_data
{
211 struct fixed_file_table
*table
;
212 struct io_ring_ctx
*ctx
;
214 struct fixed_file_ref_node
*node
;
215 struct percpu_ref refs
;
216 struct completion done
;
217 struct list_head ref_list
;
222 struct list_head list
;
228 struct io_restriction
{
229 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
230 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
231 u8 sqe_flags_allowed
;
232 u8 sqe_flags_required
;
240 /* ctx's that are using this sqd */
241 struct list_head ctx_list
;
242 struct list_head ctx_new_list
;
243 struct mutex ctx_lock
;
245 struct task_struct
*thread
;
246 struct wait_queue_head wait
;
251 struct percpu_ref refs
;
252 } ____cacheline_aligned_in_smp
;
256 unsigned int compat
: 1;
257 unsigned int limit_mem
: 1;
258 unsigned int cq_overflow_flushed
: 1;
259 unsigned int drain_next
: 1;
260 unsigned int eventfd_async
: 1;
261 unsigned int restricted
: 1;
264 * Ring buffer of indices into array of io_uring_sqe, which is
265 * mmapped by the application using the IORING_OFF_SQES offset.
267 * This indirection could e.g. be used to assign fixed
268 * io_uring_sqe entries to operations and only submit them to
269 * the queue when needed.
271 * The kernel modifies neither the indices array nor the entries
275 unsigned cached_sq_head
;
278 unsigned sq_thread_idle
;
279 unsigned cached_sq_dropped
;
280 unsigned cached_cq_overflow
;
281 unsigned long sq_check_overflow
;
283 struct list_head defer_list
;
284 struct list_head timeout_list
;
285 struct list_head cq_overflow_list
;
287 wait_queue_head_t inflight_wait
;
288 struct io_uring_sqe
*sq_sqes
;
289 } ____cacheline_aligned_in_smp
;
291 struct io_rings
*rings
;
297 * For SQPOLL usage - we hold a reference to the parent task, so we
298 * have access to the ->files
300 struct task_struct
*sqo_task
;
302 /* Only used for accounting purposes */
303 struct mm_struct
*mm_account
;
305 #ifdef CONFIG_BLK_CGROUP
306 struct cgroup_subsys_state
*sqo_blkcg_css
;
309 struct io_sq_data
*sq_data
; /* if using sq thread polling */
311 struct wait_queue_head sqo_sq_wait
;
312 struct wait_queue_entry sqo_wait_entry
;
313 struct list_head sqd_list
;
316 * If used, fixed file set. Writers must ensure that ->refs is dead,
317 * readers must ensure that ->refs is alive as long as the file* is
318 * used. Only updated through io_uring_register(2).
320 struct fixed_file_data
*file_data
;
321 unsigned nr_user_files
;
323 /* if used, fixed mapped user buffers */
324 unsigned nr_user_bufs
;
325 struct io_mapped_ubuf
*user_bufs
;
327 struct user_struct
*user
;
329 const struct cred
*creds
;
333 unsigned int sessionid
;
336 struct completion ref_comp
;
337 struct completion sq_thread_comp
;
339 /* if all else fails... */
340 struct io_kiocb
*fallback_req
;
342 #if defined(CONFIG_UNIX)
343 struct socket
*ring_sock
;
346 struct idr io_buffer_idr
;
348 struct idr personality_idr
;
351 unsigned cached_cq_tail
;
354 atomic_t cq_timeouts
;
355 unsigned long cq_check_overflow
;
356 struct wait_queue_head cq_wait
;
357 struct fasync_struct
*cq_fasync
;
358 struct eventfd_ctx
*cq_ev_fd
;
359 } ____cacheline_aligned_in_smp
;
362 struct mutex uring_lock
;
363 wait_queue_head_t wait
;
364 } ____cacheline_aligned_in_smp
;
367 spinlock_t completion_lock
;
370 * ->iopoll_list is protected by the ctx->uring_lock for
371 * io_uring instances that don't use IORING_SETUP_SQPOLL.
372 * For SQPOLL, only the single threaded io_sq_thread() will
373 * manipulate the list, hence no extra locking is needed there.
375 struct list_head iopoll_list
;
376 struct hlist_head
*cancel_hash
;
377 unsigned cancel_hash_bits
;
378 bool poll_multi_file
;
380 spinlock_t inflight_lock
;
381 struct list_head inflight_list
;
382 } ____cacheline_aligned_in_smp
;
384 struct delayed_work file_put_work
;
385 struct llist_head file_put_llist
;
387 struct work_struct exit_work
;
388 struct io_restriction restrictions
;
392 * First field must be the file pointer in all the
393 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
395 struct io_poll_iocb
{
398 struct wait_queue_head
*head
;
404 struct wait_queue_entry wait
;
409 struct file
*put_file
;
413 struct io_timeout_data
{
414 struct io_kiocb
*req
;
415 struct hrtimer timer
;
416 struct timespec64 ts
;
417 enum hrtimer_mode mode
;
422 struct sockaddr __user
*addr
;
423 int __user
*addr_len
;
425 unsigned long nofile
;
445 struct list_head list
;
448 struct io_timeout_rem
{
454 /* NOTE: kiocb has the file as the first member, so don't do it here */
462 struct sockaddr __user
*addr
;
469 struct user_msghdr __user
*umsg
;
475 struct io_buffer
*kbuf
;
481 struct filename
*filename
;
483 unsigned long nofile
;
486 struct io_files_update
{
512 struct epoll_event event
;
516 struct file
*file_out
;
517 struct file
*file_in
;
524 struct io_provide_buf
{
538 const char __user
*filename
;
539 struct statx __user
*buffer
;
542 struct io_completion
{
544 struct list_head list
;
548 struct io_async_connect
{
549 struct sockaddr_storage address
;
552 struct io_async_msghdr
{
553 struct iovec fast_iov
[UIO_FASTIOV
];
555 struct sockaddr __user
*uaddr
;
557 struct sockaddr_storage addr
;
561 struct iovec fast_iov
[UIO_FASTIOV
];
562 const struct iovec
*free_iovec
;
563 struct iov_iter iter
;
565 struct wait_page_queue wpq
;
569 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
570 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
571 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
572 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
573 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
574 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
581 REQ_F_LINK_TIMEOUT_BIT
,
583 REQ_F_NEED_CLEANUP_BIT
,
585 REQ_F_BUFFER_SELECTED_BIT
,
586 REQ_F_NO_FILE_TABLE_BIT
,
587 REQ_F_WORK_INITIALIZED_BIT
,
588 REQ_F_LTIMEOUT_ACTIVE_BIT
,
590 /* not a real bit, just to check we're not overflowing the space */
596 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
597 /* drain existing IO first */
598 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
600 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
601 /* doesn't sever on completion < 0 */
602 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
604 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
605 /* IOSQE_BUFFER_SELECT */
606 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
609 REQ_F_LINK_HEAD
= BIT(REQ_F_LINK_HEAD_BIT
),
610 /* fail rest of links */
611 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
612 /* on inflight list */
613 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
614 /* read/write uses file position */
615 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
616 /* must not punt to workers */
617 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
618 /* has or had linked timeout */
619 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
621 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
623 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
624 /* already went through poll handler */
625 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
626 /* buffer already selected */
627 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
628 /* doesn't need file table for this request */
629 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
630 /* io_wq_work is initialized */
631 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
632 /* linked timeout is active, i.e. prepared by link's head */
633 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
637 struct io_poll_iocb poll
;
638 struct io_poll_iocb
*double_poll
;
642 * NOTE! Each of the iocb union members has the file pointer
643 * as the first entry in their struct definition. So you can
644 * access the file pointer through any of the sub-structs,
645 * or directly as just 'ki_filp' in this struct.
651 struct io_poll_iocb poll
;
652 struct io_accept accept
;
654 struct io_cancel cancel
;
655 struct io_timeout timeout
;
656 struct io_timeout_rem timeout_rem
;
657 struct io_connect connect
;
658 struct io_sr_msg sr_msg
;
660 struct io_close close
;
661 struct io_files_update files_update
;
662 struct io_fadvise fadvise
;
663 struct io_madvise madvise
;
664 struct io_epoll epoll
;
665 struct io_splice splice
;
666 struct io_provide_buf pbuf
;
667 struct io_statx statx
;
668 /* use only after cleaning per-op data, see io_clean_op() */
669 struct io_completion
compl;
672 /* opcode allocated if it needs to store data for async defer */
675 /* polled IO has completed */
681 struct io_ring_ctx
*ctx
;
684 struct task_struct
*task
;
687 struct list_head link_list
;
690 * 1. used with ctx->iopoll_list with reads/writes
691 * 2. to track reqs with ->files (see io_op_def::file_table)
693 struct list_head inflight_entry
;
695 struct percpu_ref
*fixed_file_refs
;
696 struct callback_head task_work
;
697 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
698 struct hlist_node hash_node
;
699 struct async_poll
*apoll
;
700 struct io_wq_work work
;
703 struct io_defer_entry
{
704 struct list_head list
;
705 struct io_kiocb
*req
;
709 #define IO_IOPOLL_BATCH 8
711 struct io_comp_state
{
713 struct list_head list
;
714 struct io_ring_ctx
*ctx
;
717 struct io_submit_state
{
718 struct blk_plug plug
;
721 * io_kiocb alloc cache
723 void *reqs
[IO_IOPOLL_BATCH
];
724 unsigned int free_reqs
;
727 * Batch completion logic
729 struct io_comp_state comp
;
732 * File reference cache
736 unsigned int has_refs
;
737 unsigned int ios_left
;
741 /* needs req->file assigned */
742 unsigned needs_file
: 1;
743 /* don't fail if file grab fails */
744 unsigned needs_file_no_error
: 1;
745 /* hash wq insertion if file is a regular file */
746 unsigned hash_reg_file
: 1;
747 /* unbound wq insertion if file is a non-regular file */
748 unsigned unbound_nonreg_file
: 1;
749 /* opcode is not supported by this kernel */
750 unsigned not_supported
: 1;
751 /* set if opcode supports polled "wait" */
753 unsigned pollout
: 1;
754 /* op supports buffer selection */
755 unsigned buffer_select
: 1;
756 /* must always have async data allocated */
757 unsigned needs_async_data
: 1;
758 /* size of async data needed, if any */
759 unsigned short async_size
;
763 static const struct io_op_def io_op_defs
[] = {
764 [IORING_OP_NOP
] = {},
765 [IORING_OP_READV
] = {
767 .unbound_nonreg_file
= 1,
770 .needs_async_data
= 1,
771 .async_size
= sizeof(struct io_async_rw
),
772 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
774 [IORING_OP_WRITEV
] = {
777 .unbound_nonreg_file
= 1,
779 .needs_async_data
= 1,
780 .async_size
= sizeof(struct io_async_rw
),
781 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
784 [IORING_OP_FSYNC
] = {
786 .work_flags
= IO_WQ_WORK_BLKCG
,
788 [IORING_OP_READ_FIXED
] = {
790 .unbound_nonreg_file
= 1,
792 .async_size
= sizeof(struct io_async_rw
),
793 .work_flags
= IO_WQ_WORK_BLKCG
,
795 [IORING_OP_WRITE_FIXED
] = {
798 .unbound_nonreg_file
= 1,
800 .async_size
= sizeof(struct io_async_rw
),
801 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
803 [IORING_OP_POLL_ADD
] = {
805 .unbound_nonreg_file
= 1,
807 [IORING_OP_POLL_REMOVE
] = {},
808 [IORING_OP_SYNC_FILE_RANGE
] = {
810 .work_flags
= IO_WQ_WORK_BLKCG
,
812 [IORING_OP_SENDMSG
] = {
814 .unbound_nonreg_file
= 1,
816 .needs_async_data
= 1,
817 .async_size
= sizeof(struct io_async_msghdr
),
818 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
821 [IORING_OP_RECVMSG
] = {
823 .unbound_nonreg_file
= 1,
826 .needs_async_data
= 1,
827 .async_size
= sizeof(struct io_async_msghdr
),
828 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
831 [IORING_OP_TIMEOUT
] = {
832 .needs_async_data
= 1,
833 .async_size
= sizeof(struct io_timeout_data
),
834 .work_flags
= IO_WQ_WORK_MM
,
836 [IORING_OP_TIMEOUT_REMOVE
] = {},
837 [IORING_OP_ACCEPT
] = {
839 .unbound_nonreg_file
= 1,
841 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_FILES
,
843 [IORING_OP_ASYNC_CANCEL
] = {},
844 [IORING_OP_LINK_TIMEOUT
] = {
845 .needs_async_data
= 1,
846 .async_size
= sizeof(struct io_timeout_data
),
847 .work_flags
= IO_WQ_WORK_MM
,
849 [IORING_OP_CONNECT
] = {
851 .unbound_nonreg_file
= 1,
853 .needs_async_data
= 1,
854 .async_size
= sizeof(struct io_async_connect
),
855 .work_flags
= IO_WQ_WORK_MM
,
857 [IORING_OP_FALLOCATE
] = {
859 .work_flags
= IO_WQ_WORK_BLKCG
| IO_WQ_WORK_FSIZE
,
861 [IORING_OP_OPENAT
] = {
862 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
|
865 [IORING_OP_CLOSE
] = {
867 .needs_file_no_error
= 1,
868 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_BLKCG
,
870 [IORING_OP_FILES_UPDATE
] = {
871 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
,
873 [IORING_OP_STATX
] = {
874 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_MM
|
875 IO_WQ_WORK_FS
| IO_WQ_WORK_BLKCG
,
879 .unbound_nonreg_file
= 1,
882 .async_size
= sizeof(struct io_async_rw
),
883 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
885 [IORING_OP_WRITE
] = {
887 .unbound_nonreg_file
= 1,
889 .async_size
= sizeof(struct io_async_rw
),
890 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
|
893 [IORING_OP_FADVISE
] = {
895 .work_flags
= IO_WQ_WORK_BLKCG
,
897 [IORING_OP_MADVISE
] = {
898 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
902 .unbound_nonreg_file
= 1,
904 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
908 .unbound_nonreg_file
= 1,
911 .work_flags
= IO_WQ_WORK_MM
| IO_WQ_WORK_BLKCG
,
913 [IORING_OP_OPENAT2
] = {
914 .work_flags
= IO_WQ_WORK_FILES
| IO_WQ_WORK_FS
|
917 [IORING_OP_EPOLL_CTL
] = {
918 .unbound_nonreg_file
= 1,
919 .work_flags
= IO_WQ_WORK_FILES
,
921 [IORING_OP_SPLICE
] = {
924 .unbound_nonreg_file
= 1,
925 .work_flags
= IO_WQ_WORK_BLKCG
,
927 [IORING_OP_PROVIDE_BUFFERS
] = {},
928 [IORING_OP_REMOVE_BUFFERS
] = {},
932 .unbound_nonreg_file
= 1,
936 enum io_mem_account
{
941 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
942 struct io_comp_state
*cs
);
943 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
944 static void io_put_req(struct io_kiocb
*req
);
945 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
946 static void io_double_put_req(struct io_kiocb
*req
);
947 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
948 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
949 static void io_queue_linked_timeout(struct io_kiocb
*req
);
950 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
951 struct io_uring_files_update
*ip
,
953 static void __io_clean_op(struct io_kiocb
*req
);
954 static struct file
*io_file_get(struct io_submit_state
*state
,
955 struct io_kiocb
*req
, int fd
, bool fixed
);
956 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
);
957 static void io_file_put_work(struct work_struct
*work
);
959 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
960 struct iovec
**iovec
, struct iov_iter
*iter
,
962 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
963 const struct iovec
*fast_iov
,
964 struct iov_iter
*iter
, bool force
);
966 static struct kmem_cache
*req_cachep
;
968 static const struct file_operations io_uring_fops
;
970 struct sock
*io_uring_get_socket(struct file
*file
)
972 #if defined(CONFIG_UNIX)
973 if (file
->f_op
== &io_uring_fops
) {
974 struct io_ring_ctx
*ctx
= file
->private_data
;
976 return ctx
->ring_sock
->sk
;
981 EXPORT_SYMBOL(io_uring_get_socket
);
983 static inline void io_clean_op(struct io_kiocb
*req
)
985 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
|
990 static void io_sq_thread_drop_mm(void)
992 struct mm_struct
*mm
= current
->mm
;
995 kthread_unuse_mm(mm
);
1000 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1003 if (unlikely(!(ctx
->flags
& IORING_SETUP_SQPOLL
) ||
1004 !ctx
->sqo_task
->mm
||
1005 !mmget_not_zero(ctx
->sqo_task
->mm
)))
1007 kthread_use_mm(ctx
->sqo_task
->mm
);
1013 static int io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
,
1014 struct io_kiocb
*req
)
1016 if (!(io_op_defs
[req
->opcode
].work_flags
& IO_WQ_WORK_MM
))
1018 return __io_sq_thread_acquire_mm(ctx
);
1021 static void io_sq_thread_associate_blkcg(struct io_ring_ctx
*ctx
,
1022 struct cgroup_subsys_state
**cur_css
)
1025 #ifdef CONFIG_BLK_CGROUP
1026 /* puts the old one when swapping */
1027 if (*cur_css
!= ctx
->sqo_blkcg_css
) {
1028 kthread_associate_blkcg(ctx
->sqo_blkcg_css
);
1029 *cur_css
= ctx
->sqo_blkcg_css
;
1034 static void io_sq_thread_unassociate_blkcg(void)
1036 #ifdef CONFIG_BLK_CGROUP
1037 kthread_associate_blkcg(NULL
);
1041 static inline void req_set_fail_links(struct io_kiocb
*req
)
1043 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1044 req
->flags
|= REQ_F_FAIL_LINK
;
1048 * None of these are dereferenced, they are simply used to check if any of
1049 * them have changed. If we're under current and check they are still the
1050 * same, we're fine to grab references to them for actual out-of-line use.
1052 static void io_init_identity(struct io_identity
*id
)
1054 id
->files
= current
->files
;
1055 id
->mm
= current
->mm
;
1056 #ifdef CONFIG_BLK_CGROUP
1058 id
->blkcg_css
= blkcg_css();
1061 id
->creds
= current_cred();
1062 id
->nsproxy
= current
->nsproxy
;
1063 id
->fs
= current
->fs
;
1064 id
->fsize
= rlimit(RLIMIT_FSIZE
);
1066 id
->loginuid
= current
->loginuid
;
1067 id
->sessionid
= current
->sessionid
;
1069 refcount_set(&id
->count
, 1);
1072 static inline void __io_req_init_async(struct io_kiocb
*req
)
1074 memset(&req
->work
, 0, sizeof(req
->work
));
1075 req
->flags
|= REQ_F_WORK_INITIALIZED
;
1079 * Note: must call io_req_init_async() for the first time you
1080 * touch any members of io_wq_work.
1082 static inline void io_req_init_async(struct io_kiocb
*req
)
1084 struct io_uring_task
*tctx
= current
->io_uring
;
1086 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
1089 __io_req_init_async(req
);
1091 /* Grab a ref if this isn't our static identity */
1092 req
->work
.identity
= tctx
->identity
;
1093 if (tctx
->identity
!= &tctx
->__identity
)
1094 refcount_inc(&req
->work
.identity
->count
);
1097 static inline bool io_async_submit(struct io_ring_ctx
*ctx
)
1099 return ctx
->flags
& IORING_SETUP_SQPOLL
;
1102 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1104 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1106 complete(&ctx
->ref_comp
);
1109 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1111 return !req
->timeout
.off
;
1114 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1116 struct io_ring_ctx
*ctx
;
1119 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1123 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
1124 if (!ctx
->fallback_req
)
1128 * Use 5 bits less than the max cq entries, that should give us around
1129 * 32 entries per hash list if totally full and uniformly spread.
1131 hash_bits
= ilog2(p
->cq_entries
);
1135 ctx
->cancel_hash_bits
= hash_bits
;
1136 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1138 if (!ctx
->cancel_hash
)
1140 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1142 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1143 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1146 ctx
->flags
= p
->flags
;
1147 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1148 INIT_LIST_HEAD(&ctx
->sqd_list
);
1149 init_waitqueue_head(&ctx
->cq_wait
);
1150 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1151 init_completion(&ctx
->ref_comp
);
1152 init_completion(&ctx
->sq_thread_comp
);
1153 idr_init(&ctx
->io_buffer_idr
);
1154 idr_init(&ctx
->personality_idr
);
1155 mutex_init(&ctx
->uring_lock
);
1156 init_waitqueue_head(&ctx
->wait
);
1157 spin_lock_init(&ctx
->completion_lock
);
1158 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1159 INIT_LIST_HEAD(&ctx
->defer_list
);
1160 INIT_LIST_HEAD(&ctx
->timeout_list
);
1161 init_waitqueue_head(&ctx
->inflight_wait
);
1162 spin_lock_init(&ctx
->inflight_lock
);
1163 INIT_LIST_HEAD(&ctx
->inflight_list
);
1164 INIT_DELAYED_WORK(&ctx
->file_put_work
, io_file_put_work
);
1165 init_llist_head(&ctx
->file_put_llist
);
1168 if (ctx
->fallback_req
)
1169 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1170 kfree(ctx
->cancel_hash
);
1175 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1177 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1178 struct io_ring_ctx
*ctx
= req
->ctx
;
1180 return seq
!= ctx
->cached_cq_tail
1181 + READ_ONCE(ctx
->cached_cq_overflow
);
1187 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1189 struct io_rings
*rings
= ctx
->rings
;
1191 /* order cqe stores with ring update */
1192 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1194 if (wq_has_sleeper(&ctx
->cq_wait
)) {
1195 wake_up_interruptible(&ctx
->cq_wait
);
1196 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1200 static void io_put_identity(struct io_uring_task
*tctx
, struct io_kiocb
*req
)
1202 if (req
->work
.identity
== &tctx
->__identity
)
1204 if (refcount_dec_and_test(&req
->work
.identity
->count
))
1205 kfree(req
->work
.identity
);
1208 static void io_req_clean_work(struct io_kiocb
*req
)
1210 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1213 req
->flags
&= ~REQ_F_WORK_INITIALIZED
;
1215 if (req
->work
.flags
& IO_WQ_WORK_MM
) {
1216 mmdrop(req
->work
.identity
->mm
);
1217 req
->work
.flags
&= ~IO_WQ_WORK_MM
;
1219 #ifdef CONFIG_BLK_CGROUP
1220 if (req
->work
.flags
& IO_WQ_WORK_BLKCG
) {
1221 css_put(req
->work
.identity
->blkcg_css
);
1222 req
->work
.flags
&= ~IO_WQ_WORK_BLKCG
;
1225 if (req
->work
.flags
& IO_WQ_WORK_CREDS
) {
1226 put_cred(req
->work
.identity
->creds
);
1227 req
->work
.flags
&= ~IO_WQ_WORK_CREDS
;
1229 if (req
->work
.flags
& IO_WQ_WORK_FS
) {
1230 struct fs_struct
*fs
= req
->work
.identity
->fs
;
1232 spin_lock(&req
->work
.identity
->fs
->lock
);
1235 spin_unlock(&req
->work
.identity
->fs
->lock
);
1238 req
->work
.flags
&= ~IO_WQ_WORK_FS
;
1241 io_put_identity(req
->task
->io_uring
, req
);
1245 * Create a private copy of io_identity, since some fields don't match
1246 * the current context.
1248 static bool io_identity_cow(struct io_kiocb
*req
)
1250 struct io_uring_task
*tctx
= current
->io_uring
;
1251 const struct cred
*creds
= NULL
;
1252 struct io_identity
*id
;
1254 if (req
->work
.flags
& IO_WQ_WORK_CREDS
)
1255 creds
= req
->work
.identity
->creds
;
1257 id
= kmemdup(req
->work
.identity
, sizeof(*id
), GFP_KERNEL
);
1258 if (unlikely(!id
)) {
1259 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1264 * We can safely just re-init the creds we copied Either the field
1265 * matches the current one, or we haven't grabbed it yet. The only
1266 * exception is ->creds, through registered personalities, so handle
1267 * that one separately.
1269 io_init_identity(id
);
1271 req
->work
.identity
->creds
= creds
;
1273 /* add one for this request */
1274 refcount_inc(&id
->count
);
1276 /* drop old identity, assign new one. one ref for req, one for tctx */
1277 if (req
->work
.identity
!= tctx
->identity
&&
1278 refcount_sub_and_test(2, &req
->work
.identity
->count
))
1279 kfree(req
->work
.identity
);
1281 req
->work
.identity
= id
;
1282 tctx
->identity
= id
;
1286 static bool io_grab_identity(struct io_kiocb
*req
)
1288 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1289 struct io_identity
*id
= req
->work
.identity
;
1290 struct io_ring_ctx
*ctx
= req
->ctx
;
1292 if (def
->work_flags
& IO_WQ_WORK_FSIZE
) {
1293 if (id
->fsize
!= rlimit(RLIMIT_FSIZE
))
1295 req
->work
.flags
|= IO_WQ_WORK_FSIZE
;
1298 if (!(req
->work
.flags
& IO_WQ_WORK_FILES
) &&
1299 (def
->work_flags
& IO_WQ_WORK_FILES
) &&
1300 !(req
->flags
& REQ_F_NO_FILE_TABLE
)) {
1301 if (id
->files
!= current
->files
||
1302 id
->nsproxy
!= current
->nsproxy
)
1304 atomic_inc(&id
->files
->count
);
1305 get_nsproxy(id
->nsproxy
);
1306 req
->flags
|= REQ_F_INFLIGHT
;
1308 spin_lock_irq(&ctx
->inflight_lock
);
1309 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1310 spin_unlock_irq(&ctx
->inflight_lock
);
1311 req
->work
.flags
|= IO_WQ_WORK_FILES
;
1313 #ifdef CONFIG_BLK_CGROUP
1314 if (!(req
->work
.flags
& IO_WQ_WORK_BLKCG
) &&
1315 (def
->work_flags
& IO_WQ_WORK_BLKCG
)) {
1317 if (id
->blkcg_css
!= blkcg_css()) {
1322 * This should be rare, either the cgroup is dying or the task
1323 * is moving cgroups. Just punt to root for the handful of ios.
1325 if (css_tryget_online(id
->blkcg_css
))
1326 req
->work
.flags
|= IO_WQ_WORK_BLKCG
;
1330 if (!(req
->work
.flags
& IO_WQ_WORK_CREDS
)) {
1331 if (id
->creds
!= current_cred())
1333 get_cred(id
->creds
);
1334 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
1337 if (!uid_eq(current
->loginuid
, id
->loginuid
) ||
1338 current
->sessionid
!= id
->sessionid
)
1341 if (!(req
->work
.flags
& IO_WQ_WORK_FS
) &&
1342 (def
->work_flags
& IO_WQ_WORK_FS
)) {
1343 if (current
->fs
!= id
->fs
)
1345 spin_lock(&id
->fs
->lock
);
1346 if (!id
->fs
->in_exec
) {
1348 req
->work
.flags
|= IO_WQ_WORK_FS
;
1350 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1352 spin_unlock(¤t
->fs
->lock
);
1358 static void io_prep_async_work(struct io_kiocb
*req
)
1360 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1361 struct io_ring_ctx
*ctx
= req
->ctx
;
1362 struct io_identity
*id
;
1364 io_req_init_async(req
);
1365 id
= req
->work
.identity
;
1367 if (req
->flags
& REQ_F_ISREG
) {
1368 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1369 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1371 if (def
->unbound_nonreg_file
)
1372 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1375 /* ->mm can never change on us */
1376 if (!(req
->work
.flags
& IO_WQ_WORK_MM
) &&
1377 (def
->work_flags
& IO_WQ_WORK_MM
)) {
1379 req
->work
.flags
|= IO_WQ_WORK_MM
;
1382 /* if we fail grabbing identity, we must COW, regrab, and retry */
1383 if (io_grab_identity(req
))
1386 if (!io_identity_cow(req
))
1389 /* can't fail at this point */
1390 if (!io_grab_identity(req
))
1394 static void io_prep_async_link(struct io_kiocb
*req
)
1396 struct io_kiocb
*cur
;
1398 io_prep_async_work(req
);
1399 if (req
->flags
& REQ_F_LINK_HEAD
)
1400 list_for_each_entry(cur
, &req
->link_list
, link_list
)
1401 io_prep_async_work(cur
);
1404 static struct io_kiocb
*__io_queue_async_work(struct io_kiocb
*req
)
1406 struct io_ring_ctx
*ctx
= req
->ctx
;
1407 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1409 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1410 &req
->work
, req
->flags
);
1411 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1415 static void io_queue_async_work(struct io_kiocb
*req
)
1417 struct io_kiocb
*link
;
1419 /* init ->work of the whole link before punting */
1420 io_prep_async_link(req
);
1421 link
= __io_queue_async_work(req
);
1424 io_queue_linked_timeout(link
);
1427 static void io_kill_timeout(struct io_kiocb
*req
)
1429 struct io_timeout_data
*io
= req
->async_data
;
1432 ret
= hrtimer_try_to_cancel(&io
->timer
);
1434 atomic_set(&req
->ctx
->cq_timeouts
,
1435 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1436 list_del_init(&req
->timeout
.list
);
1437 io_cqring_fill_event(req
, 0);
1438 io_put_req_deferred(req
, 1);
1442 static bool io_task_match(struct io_kiocb
*req
, struct task_struct
*tsk
)
1444 struct io_ring_ctx
*ctx
= req
->ctx
;
1446 if (!tsk
|| req
->task
== tsk
)
1448 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1449 if (ctx
->sq_data
&& req
->task
== ctx
->sq_data
->thread
)
1456 * Returns true if we found and killed one or more timeouts
1458 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
)
1460 struct io_kiocb
*req
, *tmp
;
1463 spin_lock_irq(&ctx
->completion_lock
);
1464 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1465 if (io_task_match(req
, tsk
)) {
1466 io_kill_timeout(req
);
1470 spin_unlock_irq(&ctx
->completion_lock
);
1471 return canceled
!= 0;
1474 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1477 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1478 struct io_defer_entry
, list
);
1479 struct io_kiocb
*link
;
1481 if (req_need_defer(de
->req
, de
->seq
))
1483 list_del_init(&de
->list
);
1484 /* punt-init is done before queueing for defer */
1485 link
= __io_queue_async_work(de
->req
);
1487 __io_queue_linked_timeout(link
);
1488 /* drop submission reference */
1489 io_put_req_deferred(link
, 1);
1492 } while (!list_empty(&ctx
->defer_list
));
1495 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1497 while (!list_empty(&ctx
->timeout_list
)) {
1498 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1499 struct io_kiocb
, timeout
.list
);
1501 if (io_is_timeout_noseq(req
))
1503 if (req
->timeout
.target_seq
!= ctx
->cached_cq_tail
1504 - atomic_read(&ctx
->cq_timeouts
))
1507 list_del_init(&req
->timeout
.list
);
1508 io_kill_timeout(req
);
1512 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1514 io_flush_timeouts(ctx
);
1515 __io_commit_cqring(ctx
);
1517 if (unlikely(!list_empty(&ctx
->defer_list
)))
1518 __io_queue_deferred(ctx
);
1521 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1523 struct io_rings
*r
= ctx
->rings
;
1525 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1528 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1530 struct io_rings
*rings
= ctx
->rings
;
1533 tail
= ctx
->cached_cq_tail
;
1535 * writes to the cq entry need to come after reading head; the
1536 * control dependency is enough as we're using WRITE_ONCE to
1539 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1542 ctx
->cached_cq_tail
++;
1543 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1546 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1550 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1552 if (!ctx
->eventfd_async
)
1554 return io_wq_current_is_worker();
1557 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1559 if (waitqueue_active(&ctx
->wait
))
1560 wake_up(&ctx
->wait
);
1561 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1562 wake_up(&ctx
->sq_data
->wait
);
1563 if (io_should_trigger_evfd(ctx
))
1564 eventfd_signal(ctx
->cq_ev_fd
, 1);
1567 static void io_cqring_mark_overflow(struct io_ring_ctx
*ctx
)
1569 if (list_empty(&ctx
->cq_overflow_list
)) {
1570 clear_bit(0, &ctx
->sq_check_overflow
);
1571 clear_bit(0, &ctx
->cq_check_overflow
);
1572 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1576 static inline bool io_match_files(struct io_kiocb
*req
,
1577 struct files_struct
*files
)
1581 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
1582 (req
->work
.flags
& IO_WQ_WORK_FILES
))
1583 return req
->work
.identity
->files
== files
;
1587 /* Returns true if there are no backlogged entries after the flush */
1588 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1589 struct task_struct
*tsk
,
1590 struct files_struct
*files
)
1592 struct io_rings
*rings
= ctx
->rings
;
1593 struct io_kiocb
*req
, *tmp
;
1594 struct io_uring_cqe
*cqe
;
1595 unsigned long flags
;
1599 if (list_empty_careful(&ctx
->cq_overflow_list
))
1601 if ((ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
) ==
1602 rings
->cq_ring_entries
))
1606 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1608 /* if force is set, the ring is going away. always drop after that */
1610 ctx
->cq_overflow_flushed
= 1;
1613 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1614 if (tsk
&& req
->task
!= tsk
)
1616 if (!io_match_files(req
, files
))
1619 cqe
= io_get_cqring(ctx
);
1623 list_move(&req
->compl.list
, &list
);
1625 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1626 WRITE_ONCE(cqe
->res
, req
->result
);
1627 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1629 ctx
->cached_cq_overflow
++;
1630 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1631 ctx
->cached_cq_overflow
);
1635 io_commit_cqring(ctx
);
1636 io_cqring_mark_overflow(ctx
);
1638 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1639 io_cqring_ev_posted(ctx
);
1641 while (!list_empty(&list
)) {
1642 req
= list_first_entry(&list
, struct io_kiocb
, compl.list
);
1643 list_del(&req
->compl.list
);
1650 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1652 struct io_ring_ctx
*ctx
= req
->ctx
;
1653 struct io_uring_cqe
*cqe
;
1655 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1658 * If we can't get a cq entry, userspace overflowed the
1659 * submission (by quite a lot). Increment the overflow count in
1662 cqe
= io_get_cqring(ctx
);
1664 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1665 WRITE_ONCE(cqe
->res
, res
);
1666 WRITE_ONCE(cqe
->flags
, cflags
);
1667 } else if (ctx
->cq_overflow_flushed
|| req
->task
->io_uring
->in_idle
) {
1669 * If we're in ring overflow flush mode, or in task cancel mode,
1670 * then we cannot store the request for later flushing, we need
1671 * to drop it on the floor.
1673 ctx
->cached_cq_overflow
++;
1674 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1676 if (list_empty(&ctx
->cq_overflow_list
)) {
1677 set_bit(0, &ctx
->sq_check_overflow
);
1678 set_bit(0, &ctx
->cq_check_overflow
);
1679 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1683 req
->compl.cflags
= cflags
;
1684 refcount_inc(&req
->refs
);
1685 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1689 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1691 __io_cqring_fill_event(req
, res
, 0);
1694 static void io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1696 struct io_ring_ctx
*ctx
= req
->ctx
;
1697 unsigned long flags
;
1699 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1700 __io_cqring_fill_event(req
, res
, cflags
);
1701 io_commit_cqring(ctx
);
1702 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1704 io_cqring_ev_posted(ctx
);
1707 static void io_submit_flush_completions(struct io_comp_state
*cs
)
1709 struct io_ring_ctx
*ctx
= cs
->ctx
;
1711 spin_lock_irq(&ctx
->completion_lock
);
1712 while (!list_empty(&cs
->list
)) {
1713 struct io_kiocb
*req
;
1715 req
= list_first_entry(&cs
->list
, struct io_kiocb
, compl.list
);
1716 list_del(&req
->compl.list
);
1717 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
1720 * io_free_req() doesn't care about completion_lock unless one
1721 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1722 * because of a potential deadlock with req->work.fs->lock
1724 if (req
->flags
& (REQ_F_FAIL_LINK
|REQ_F_LINK_TIMEOUT
1725 |REQ_F_WORK_INITIALIZED
)) {
1726 spin_unlock_irq(&ctx
->completion_lock
);
1728 spin_lock_irq(&ctx
->completion_lock
);
1733 io_commit_cqring(ctx
);
1734 spin_unlock_irq(&ctx
->completion_lock
);
1736 io_cqring_ev_posted(ctx
);
1740 static void __io_req_complete(struct io_kiocb
*req
, long res
, unsigned cflags
,
1741 struct io_comp_state
*cs
)
1744 io_cqring_add_event(req
, res
, cflags
);
1749 req
->compl.cflags
= cflags
;
1750 list_add_tail(&req
->compl.list
, &cs
->list
);
1752 io_submit_flush_completions(cs
);
1756 static void io_req_complete(struct io_kiocb
*req
, long res
)
1758 __io_req_complete(req
, res
, 0, NULL
);
1761 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1763 return req
== (struct io_kiocb
*)
1764 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1767 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1769 struct io_kiocb
*req
;
1771 req
= ctx
->fallback_req
;
1772 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1778 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1779 struct io_submit_state
*state
)
1781 if (!state
->free_reqs
) {
1782 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1786 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1787 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
1790 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1791 * retry single alloc to be on the safe side.
1793 if (unlikely(ret
<= 0)) {
1794 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1795 if (!state
->reqs
[0])
1799 state
->free_reqs
= ret
;
1803 return state
->reqs
[state
->free_reqs
];
1805 return io_get_fallback_req(ctx
);
1808 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
1812 percpu_ref_put(req
->fixed_file_refs
);
1817 static void io_dismantle_req(struct io_kiocb
*req
)
1821 if (req
->async_data
)
1822 kfree(req
->async_data
);
1824 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
1826 io_req_clean_work(req
);
1829 static void __io_free_req(struct io_kiocb
*req
)
1831 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1832 struct io_ring_ctx
*ctx
= req
->ctx
;
1834 io_dismantle_req(req
);
1836 percpu_counter_dec(&tctx
->inflight
);
1838 wake_up(&tctx
->wait
);
1839 put_task_struct(req
->task
);
1841 if (likely(!io_is_fallback_req(req
)))
1842 kmem_cache_free(req_cachep
, req
);
1844 clear_bit_unlock(0, (unsigned long *) &ctx
->fallback_req
);
1845 percpu_ref_put(&ctx
->refs
);
1848 static bool io_link_cancel_timeout(struct io_kiocb
*req
)
1850 struct io_timeout_data
*io
= req
->async_data
;
1851 struct io_ring_ctx
*ctx
= req
->ctx
;
1854 ret
= hrtimer_try_to_cancel(&io
->timer
);
1856 io_cqring_fill_event(req
, -ECANCELED
);
1857 io_commit_cqring(ctx
);
1858 req
->flags
&= ~REQ_F_LINK_HEAD
;
1859 io_put_req_deferred(req
, 1);
1866 static bool __io_kill_linked_timeout(struct io_kiocb
*req
)
1868 struct io_kiocb
*link
;
1871 if (list_empty(&req
->link_list
))
1873 link
= list_first_entry(&req
->link_list
, struct io_kiocb
, link_list
);
1874 if (link
->opcode
!= IORING_OP_LINK_TIMEOUT
)
1877 * Can happen if a linked timeout fired and link had been like
1878 * req -> link t-out -> link t-out [-> ...]
1880 if (!(link
->flags
& REQ_F_LTIMEOUT_ACTIVE
))
1883 list_del_init(&link
->link_list
);
1884 wake_ev
= io_link_cancel_timeout(link
);
1885 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1889 static void io_kill_linked_timeout(struct io_kiocb
*req
)
1891 struct io_ring_ctx
*ctx
= req
->ctx
;
1892 unsigned long flags
;
1895 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1896 wake_ev
= __io_kill_linked_timeout(req
);
1897 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1900 io_cqring_ev_posted(ctx
);
1903 static struct io_kiocb
*io_req_link_next(struct io_kiocb
*req
)
1905 struct io_kiocb
*nxt
;
1908 * The list should never be empty when we are called here. But could
1909 * potentially happen if the chain is messed up, check to be on the
1912 if (unlikely(list_empty(&req
->link_list
)))
1915 nxt
= list_first_entry(&req
->link_list
, struct io_kiocb
, link_list
);
1916 list_del_init(&req
->link_list
);
1917 if (!list_empty(&nxt
->link_list
))
1918 nxt
->flags
|= REQ_F_LINK_HEAD
;
1923 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1925 static void io_fail_links(struct io_kiocb
*req
)
1927 struct io_ring_ctx
*ctx
= req
->ctx
;
1928 unsigned long flags
;
1930 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1931 while (!list_empty(&req
->link_list
)) {
1932 struct io_kiocb
*link
= list_first_entry(&req
->link_list
,
1933 struct io_kiocb
, link_list
);
1935 list_del_init(&link
->link_list
);
1936 trace_io_uring_fail_link(req
, link
);
1938 io_cqring_fill_event(link
, -ECANCELED
);
1941 * It's ok to free under spinlock as they're not linked anymore,
1942 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1945 if (link
->flags
& REQ_F_WORK_INITIALIZED
)
1946 io_put_req_deferred(link
, 2);
1948 io_double_put_req(link
);
1951 io_commit_cqring(ctx
);
1952 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1954 io_cqring_ev_posted(ctx
);
1957 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
1959 req
->flags
&= ~REQ_F_LINK_HEAD
;
1960 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
1961 io_kill_linked_timeout(req
);
1964 * If LINK is set, we have dependent requests in this chain. If we
1965 * didn't fail this request, queue the first one up, moving any other
1966 * dependencies to the next request. In case of failure, fail the rest
1969 if (likely(!(req
->flags
& REQ_F_FAIL_LINK
)))
1970 return io_req_link_next(req
);
1975 static struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1977 if (likely(!(req
->flags
& REQ_F_LINK_HEAD
)))
1979 return __io_req_find_next(req
);
1982 static int io_req_task_work_add(struct io_kiocb
*req
, bool twa_signal_ok
)
1984 struct task_struct
*tsk
= req
->task
;
1985 struct io_ring_ctx
*ctx
= req
->ctx
;
1988 if (tsk
->flags
& PF_EXITING
)
1992 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1993 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1994 * processing task_work. There's no reliable way to tell if TWA_RESUME
1998 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
) && twa_signal_ok
)
1999 notify
= TWA_SIGNAL
;
2001 ret
= task_work_add(tsk
, &req
->task_work
, notify
);
2003 wake_up_process(tsk
);
2008 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
2010 struct io_ring_ctx
*ctx
= req
->ctx
;
2012 spin_lock_irq(&ctx
->completion_lock
);
2013 io_cqring_fill_event(req
, error
);
2014 io_commit_cqring(ctx
);
2015 spin_unlock_irq(&ctx
->completion_lock
);
2017 io_cqring_ev_posted(ctx
);
2018 req_set_fail_links(req
);
2019 io_double_put_req(req
);
2022 static void io_req_task_cancel(struct callback_head
*cb
)
2024 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2025 struct io_ring_ctx
*ctx
= req
->ctx
;
2027 __io_req_task_cancel(req
, -ECANCELED
);
2028 percpu_ref_put(&ctx
->refs
);
2031 static void __io_req_task_submit(struct io_kiocb
*req
)
2033 struct io_ring_ctx
*ctx
= req
->ctx
;
2035 if (!__io_sq_thread_acquire_mm(ctx
)) {
2036 mutex_lock(&ctx
->uring_lock
);
2037 __io_queue_sqe(req
, NULL
);
2038 mutex_unlock(&ctx
->uring_lock
);
2040 __io_req_task_cancel(req
, -EFAULT
);
2044 static void io_req_task_submit(struct callback_head
*cb
)
2046 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2047 struct io_ring_ctx
*ctx
= req
->ctx
;
2049 __io_req_task_submit(req
);
2050 percpu_ref_put(&ctx
->refs
);
2053 static void io_req_task_queue(struct io_kiocb
*req
)
2057 init_task_work(&req
->task_work
, io_req_task_submit
);
2058 percpu_ref_get(&req
->ctx
->refs
);
2060 ret
= io_req_task_work_add(req
, true);
2061 if (unlikely(ret
)) {
2062 struct task_struct
*tsk
;
2064 init_task_work(&req
->task_work
, io_req_task_cancel
);
2065 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2066 task_work_add(tsk
, &req
->task_work
, 0);
2067 wake_up_process(tsk
);
2071 static void io_queue_next(struct io_kiocb
*req
)
2073 struct io_kiocb
*nxt
= io_req_find_next(req
);
2076 io_req_task_queue(nxt
);
2079 static void io_free_req(struct io_kiocb
*req
)
2086 void *reqs
[IO_IOPOLL_BATCH
];
2089 struct task_struct
*task
;
2093 static inline void io_init_req_batch(struct req_batch
*rb
)
2100 static void __io_req_free_batch_flush(struct io_ring_ctx
*ctx
,
2101 struct req_batch
*rb
)
2103 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
2104 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
2108 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2109 struct req_batch
*rb
)
2112 __io_req_free_batch_flush(ctx
, rb
);
2114 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2116 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2117 put_task_struct_many(rb
->task
, rb
->task_refs
);
2122 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
)
2124 if (unlikely(io_is_fallback_req(req
))) {
2128 if (req
->flags
& REQ_F_LINK_HEAD
)
2131 if (req
->task
!= rb
->task
) {
2133 struct io_uring_task
*tctx
= rb
->task
->io_uring
;
2135 percpu_counter_sub(&tctx
->inflight
, rb
->task_refs
);
2136 put_task_struct_many(rb
->task
, rb
->task_refs
);
2138 rb
->task
= req
->task
;
2143 io_dismantle_req(req
);
2144 rb
->reqs
[rb
->to_free
++] = req
;
2145 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
2146 __io_req_free_batch_flush(req
->ctx
, rb
);
2150 * Drop reference to request, return next in chain (if there is one) if this
2151 * was the last reference to this request.
2153 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2155 struct io_kiocb
*nxt
= NULL
;
2157 if (refcount_dec_and_test(&req
->refs
)) {
2158 nxt
= io_req_find_next(req
);
2164 static void io_put_req(struct io_kiocb
*req
)
2166 if (refcount_dec_and_test(&req
->refs
))
2170 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2172 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2177 static void io_free_req_deferred(struct io_kiocb
*req
)
2181 init_task_work(&req
->task_work
, io_put_req_deferred_cb
);
2182 ret
= io_req_task_work_add(req
, true);
2183 if (unlikely(ret
)) {
2184 struct task_struct
*tsk
;
2186 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2187 task_work_add(tsk
, &req
->task_work
, 0);
2188 wake_up_process(tsk
);
2192 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2194 if (refcount_sub_and_test(refs
, &req
->refs
))
2195 io_free_req_deferred(req
);
2198 static struct io_wq_work
*io_steal_work(struct io_kiocb
*req
)
2200 struct io_kiocb
*nxt
;
2203 * A ref is owned by io-wq in which context we're. So, if that's the
2204 * last one, it's safe to steal next work. False negatives are Ok,
2205 * it just will be re-punted async in io_put_work()
2207 if (refcount_read(&req
->refs
) != 1)
2210 nxt
= io_req_find_next(req
);
2211 return nxt
? &nxt
->work
: NULL
;
2214 static void io_double_put_req(struct io_kiocb
*req
)
2216 /* drop both submit and complete references */
2217 if (refcount_sub_and_test(2, &req
->refs
))
2221 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
, bool noflush
)
2223 struct io_rings
*rings
= ctx
->rings
;
2225 if (test_bit(0, &ctx
->cq_check_overflow
)) {
2227 * noflush == true is from the waitqueue handler, just ensure
2228 * we wake up the task, and the next invocation will flush the
2229 * entries. We cannot safely to it from here.
2231 if (noflush
&& !list_empty(&ctx
->cq_overflow_list
))
2234 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2237 /* See comment at the top of this file */
2239 return ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
);
2242 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2244 struct io_rings
*rings
= ctx
->rings
;
2246 /* make sure SQ entry isn't read before tail */
2247 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2250 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2252 unsigned int cflags
;
2254 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2255 cflags
|= IORING_CQE_F_BUFFER
;
2256 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2261 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2263 struct io_buffer
*kbuf
;
2265 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2266 return io_put_kbuf(req
, kbuf
);
2269 static inline bool io_run_task_work(void)
2272 * Not safe to run on exiting task, and the task_work handling will
2273 * not add work to such a task.
2275 if (unlikely(current
->flags
& PF_EXITING
))
2277 if (current
->task_works
) {
2278 __set_current_state(TASK_RUNNING
);
2286 static void io_iopoll_queue(struct list_head
*again
)
2288 struct io_kiocb
*req
;
2291 req
= list_first_entry(again
, struct io_kiocb
, inflight_entry
);
2292 list_del(&req
->inflight_entry
);
2293 __io_complete_rw(req
, -EAGAIN
, 0, NULL
);
2294 } while (!list_empty(again
));
2298 * Find and free completed poll iocbs
2300 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2301 struct list_head
*done
)
2303 struct req_batch rb
;
2304 struct io_kiocb
*req
;
2307 /* order with ->result store in io_complete_rw_iopoll() */
2310 io_init_req_batch(&rb
);
2311 while (!list_empty(done
)) {
2314 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2315 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2317 req
->iopoll_completed
= 0;
2318 list_move_tail(&req
->inflight_entry
, &again
);
2321 list_del(&req
->inflight_entry
);
2323 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2324 cflags
= io_put_rw_kbuf(req
);
2326 __io_cqring_fill_event(req
, req
->result
, cflags
);
2329 if (refcount_dec_and_test(&req
->refs
))
2330 io_req_free_batch(&rb
, req
);
2333 io_commit_cqring(ctx
);
2334 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
2335 io_cqring_ev_posted(ctx
);
2336 io_req_free_batch_finish(ctx
, &rb
);
2338 if (!list_empty(&again
))
2339 io_iopoll_queue(&again
);
2342 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2345 struct io_kiocb
*req
, *tmp
;
2351 * Only spin for completions if we don't have multiple devices hanging
2352 * off our complete list, and we're under the requested amount.
2354 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2357 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2358 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2361 * Move completed and retryable entries to our local lists.
2362 * If we find a request that requires polling, break out
2363 * and complete those lists first, if we have entries there.
2365 if (READ_ONCE(req
->iopoll_completed
)) {
2366 list_move_tail(&req
->inflight_entry
, &done
);
2369 if (!list_empty(&done
))
2372 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2376 /* iopoll may have completed current req */
2377 if (READ_ONCE(req
->iopoll_completed
))
2378 list_move_tail(&req
->inflight_entry
, &done
);
2385 if (!list_empty(&done
))
2386 io_iopoll_complete(ctx
, nr_events
, &done
);
2392 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2393 * non-spinning poll check - we'll still enter the driver poll loop, but only
2394 * as a non-spinning completion check.
2396 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2399 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2402 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2405 if (*nr_events
>= min
)
2413 * We can't just wait for polled events to come to us, we have to actively
2414 * find and complete them.
2416 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2418 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2421 mutex_lock(&ctx
->uring_lock
);
2422 while (!list_empty(&ctx
->iopoll_list
)) {
2423 unsigned int nr_events
= 0;
2425 io_do_iopoll(ctx
, &nr_events
, 0);
2427 /* let it sleep and repeat later if can't complete a request */
2431 * Ensure we allow local-to-the-cpu processing to take place,
2432 * in this case we need to ensure that we reap all events.
2433 * Also let task_work, etc. to progress by releasing the mutex
2435 if (need_resched()) {
2436 mutex_unlock(&ctx
->uring_lock
);
2438 mutex_lock(&ctx
->uring_lock
);
2441 mutex_unlock(&ctx
->uring_lock
);
2444 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2446 unsigned int nr_events
= 0;
2447 int iters
= 0, ret
= 0;
2450 * We disallow the app entering submit/complete with polling, but we
2451 * still need to lock the ring to prevent racing with polled issue
2452 * that got punted to a workqueue.
2454 mutex_lock(&ctx
->uring_lock
);
2457 * Don't enter poll loop if we already have events pending.
2458 * If we do, we can potentially be spinning for commands that
2459 * already triggered a CQE (eg in error).
2461 if (io_cqring_events(ctx
, false))
2465 * If a submit got punted to a workqueue, we can have the
2466 * application entering polling for a command before it gets
2467 * issued. That app will hold the uring_lock for the duration
2468 * of the poll right here, so we need to take a breather every
2469 * now and then to ensure that the issue has a chance to add
2470 * the poll to the issued list. Otherwise we can spin here
2471 * forever, while the workqueue is stuck trying to acquire the
2474 if (!(++iters
& 7)) {
2475 mutex_unlock(&ctx
->uring_lock
);
2477 mutex_lock(&ctx
->uring_lock
);
2480 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2484 } while (min
&& !nr_events
&& !need_resched());
2486 mutex_unlock(&ctx
->uring_lock
);
2490 static void kiocb_end_write(struct io_kiocb
*req
)
2493 * Tell lockdep we inherited freeze protection from submission
2496 if (req
->flags
& REQ_F_ISREG
) {
2497 struct inode
*inode
= file_inode(req
->file
);
2499 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2501 file_end_write(req
->file
);
2504 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
,
2505 struct io_comp_state
*cs
)
2507 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2510 if (kiocb
->ki_flags
& IOCB_WRITE
)
2511 kiocb_end_write(req
);
2513 if (res
!= req
->result
)
2514 req_set_fail_links(req
);
2515 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2516 cflags
= io_put_rw_kbuf(req
);
2517 __io_req_complete(req
, res
, cflags
, cs
);
2521 static bool io_resubmit_prep(struct io_kiocb
*req
, int error
)
2523 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2524 ssize_t ret
= -ECANCELED
;
2525 struct iov_iter iter
;
2533 switch (req
->opcode
) {
2534 case IORING_OP_READV
:
2535 case IORING_OP_READ_FIXED
:
2536 case IORING_OP_READ
:
2539 case IORING_OP_WRITEV
:
2540 case IORING_OP_WRITE_FIXED
:
2541 case IORING_OP_WRITE
:
2545 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2550 if (!req
->async_data
) {
2551 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2554 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2562 req_set_fail_links(req
);
2563 io_req_complete(req
, ret
);
2568 static bool io_rw_reissue(struct io_kiocb
*req
, long res
)
2571 umode_t mode
= file_inode(req
->file
)->i_mode
;
2574 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2576 if ((res
!= -EAGAIN
&& res
!= -EOPNOTSUPP
) || io_wq_current_is_worker())
2579 ret
= io_sq_thread_acquire_mm(req
->ctx
, req
);
2581 if (io_resubmit_prep(req
, ret
)) {
2582 refcount_inc(&req
->refs
);
2583 io_queue_async_work(req
);
2591 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2592 struct io_comp_state
*cs
)
2594 if (!io_rw_reissue(req
, res
))
2595 io_complete_rw_common(&req
->rw
.kiocb
, res
, cs
);
2598 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2600 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2602 __io_complete_rw(req
, res
, res2
, NULL
);
2605 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2607 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2609 if (kiocb
->ki_flags
& IOCB_WRITE
)
2610 kiocb_end_write(req
);
2612 if (res
!= -EAGAIN
&& res
!= req
->result
)
2613 req_set_fail_links(req
);
2615 WRITE_ONCE(req
->result
, res
);
2616 /* order with io_poll_complete() checking ->result */
2618 WRITE_ONCE(req
->iopoll_completed
, 1);
2622 * After the iocb has been issued, it's safe to be found on the poll list.
2623 * Adding the kiocb to the list AFTER submission ensures that we don't
2624 * find it from a io_iopoll_getevents() thread before the issuer is done
2625 * accessing the kiocb cookie.
2627 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2629 struct io_ring_ctx
*ctx
= req
->ctx
;
2632 * Track whether we have multiple files in our lists. This will impact
2633 * how we do polling eventually, not spinning if we're on potentially
2634 * different devices.
2636 if (list_empty(&ctx
->iopoll_list
)) {
2637 ctx
->poll_multi_file
= false;
2638 } else if (!ctx
->poll_multi_file
) {
2639 struct io_kiocb
*list_req
;
2641 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2643 if (list_req
->file
!= req
->file
)
2644 ctx
->poll_multi_file
= true;
2648 * For fast devices, IO may have already completed. If it has, add
2649 * it to the front so we find it first.
2651 if (READ_ONCE(req
->iopoll_completed
))
2652 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2654 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2656 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2657 wq_has_sleeper(&ctx
->sq_data
->wait
))
2658 wake_up(&ctx
->sq_data
->wait
);
2661 static void __io_state_file_put(struct io_submit_state
*state
)
2663 if (state
->has_refs
)
2664 fput_many(state
->file
, state
->has_refs
);
2668 static inline void io_state_file_put(struct io_submit_state
*state
)
2671 __io_state_file_put(state
);
2675 * Get as many references to a file as we have IOs left in this submission,
2676 * assuming most submissions are for one file, or at least that each file
2677 * has more than one submission.
2679 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2685 if (state
->fd
== fd
) {
2689 __io_state_file_put(state
);
2691 state
->file
= fget_many(fd
, state
->ios_left
);
2696 state
->has_refs
= state
->ios_left
- 1;
2700 static bool io_bdev_nowait(struct block_device
*bdev
)
2703 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2710 * If we tracked the file through the SCM inflight mechanism, we could support
2711 * any file. For now, just ensure that anything potentially problematic is done
2714 static bool io_file_supports_async(struct file
*file
, int rw
)
2716 umode_t mode
= file_inode(file
)->i_mode
;
2718 if (S_ISBLK(mode
)) {
2719 if (io_bdev_nowait(file
->f_inode
->i_bdev
))
2723 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2725 if (S_ISREG(mode
)) {
2726 if (io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2727 file
->f_op
!= &io_uring_fops
)
2732 /* any ->read/write should understand O_NONBLOCK */
2733 if (file
->f_flags
& O_NONBLOCK
)
2736 if (!(file
->f_mode
& FMODE_NOWAIT
))
2740 return file
->f_op
->read_iter
!= NULL
;
2742 return file
->f_op
->write_iter
!= NULL
;
2745 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2747 struct io_ring_ctx
*ctx
= req
->ctx
;
2748 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2752 if (S_ISREG(file_inode(req
->file
)->i_mode
))
2753 req
->flags
|= REQ_F_ISREG
;
2755 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2756 if (kiocb
->ki_pos
== -1 && !(req
->file
->f_mode
& FMODE_STREAM
)) {
2757 req
->flags
|= REQ_F_CUR_POS
;
2758 kiocb
->ki_pos
= req
->file
->f_pos
;
2760 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2761 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2762 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2766 ioprio
= READ_ONCE(sqe
->ioprio
);
2768 ret
= ioprio_check_cap(ioprio
);
2772 kiocb
->ki_ioprio
= ioprio
;
2774 kiocb
->ki_ioprio
= get_current_ioprio();
2776 /* don't allow async punt if RWF_NOWAIT was requested */
2777 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2778 req
->flags
|= REQ_F_NOWAIT
;
2780 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2781 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2782 !kiocb
->ki_filp
->f_op
->iopoll
)
2785 kiocb
->ki_flags
|= IOCB_HIPRI
;
2786 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2787 req
->iopoll_completed
= 0;
2789 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2791 kiocb
->ki_complete
= io_complete_rw
;
2794 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2795 req
->rw
.len
= READ_ONCE(sqe
->len
);
2796 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2800 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2806 case -ERESTARTNOINTR
:
2807 case -ERESTARTNOHAND
:
2808 case -ERESTART_RESTARTBLOCK
:
2810 * We can't just restart the syscall, since previously
2811 * submitted sqes may already be in progress. Just fail this
2817 kiocb
->ki_complete(kiocb
, ret
, 0);
2821 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2822 struct io_comp_state
*cs
)
2824 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2825 struct io_async_rw
*io
= req
->async_data
;
2827 /* add previously done IO, if any */
2828 if (io
&& io
->bytes_done
> 0) {
2830 ret
= io
->bytes_done
;
2832 ret
+= io
->bytes_done
;
2835 if (req
->flags
& REQ_F_CUR_POS
)
2836 req
->file
->f_pos
= kiocb
->ki_pos
;
2837 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2838 __io_complete_rw(req
, ret
, 0, cs
);
2840 io_rw_done(kiocb
, ret
);
2843 static ssize_t
io_import_fixed(struct io_kiocb
*req
, int rw
,
2844 struct iov_iter
*iter
)
2846 struct io_ring_ctx
*ctx
= req
->ctx
;
2847 size_t len
= req
->rw
.len
;
2848 struct io_mapped_ubuf
*imu
;
2849 u16 index
, buf_index
= req
->buf_index
;
2853 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
2855 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
2856 imu
= &ctx
->user_bufs
[index
];
2857 buf_addr
= req
->rw
.addr
;
2860 if (buf_addr
+ len
< buf_addr
)
2862 /* not inside the mapped region */
2863 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
2867 * May not be a start of buffer, set size appropriately
2868 * and advance us to the beginning.
2870 offset
= buf_addr
- imu
->ubuf
;
2871 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
2875 * Don't use iov_iter_advance() here, as it's really slow for
2876 * using the latter parts of a big fixed buffer - it iterates
2877 * over each segment manually. We can cheat a bit here, because
2880 * 1) it's a BVEC iter, we set it up
2881 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2882 * first and last bvec
2884 * So just find our index, and adjust the iterator afterwards.
2885 * If the offset is within the first bvec (or the whole first
2886 * bvec, just use iov_iter_advance(). This makes it easier
2887 * since we can just skip the first segment, which may not
2888 * be PAGE_SIZE aligned.
2890 const struct bio_vec
*bvec
= imu
->bvec
;
2892 if (offset
<= bvec
->bv_len
) {
2893 iov_iter_advance(iter
, offset
);
2895 unsigned long seg_skip
;
2897 /* skip first vec */
2898 offset
-= bvec
->bv_len
;
2899 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
2901 iter
->bvec
= bvec
+ seg_skip
;
2902 iter
->nr_segs
-= seg_skip
;
2903 iter
->count
-= bvec
->bv_len
+ offset
;
2904 iter
->iov_offset
= offset
& ~PAGE_MASK
;
2911 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2914 mutex_unlock(&ctx
->uring_lock
);
2917 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2920 * "Normal" inline submissions always hold the uring_lock, since we
2921 * grab it from the system call. Same is true for the SQPOLL offload.
2922 * The only exception is when we've detached the request and issue it
2923 * from an async worker thread, grab the lock for that case.
2926 mutex_lock(&ctx
->uring_lock
);
2929 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
2930 int bgid
, struct io_buffer
*kbuf
,
2933 struct io_buffer
*head
;
2935 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2938 io_ring_submit_lock(req
->ctx
, needs_lock
);
2940 lockdep_assert_held(&req
->ctx
->uring_lock
);
2942 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
2944 if (!list_empty(&head
->list
)) {
2945 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
2947 list_del(&kbuf
->list
);
2950 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
2952 if (*len
> kbuf
->len
)
2955 kbuf
= ERR_PTR(-ENOBUFS
);
2958 io_ring_submit_unlock(req
->ctx
, needs_lock
);
2963 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
2966 struct io_buffer
*kbuf
;
2969 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2970 bgid
= req
->buf_index
;
2971 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
2974 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
2975 req
->flags
|= REQ_F_BUFFER_SELECTED
;
2976 return u64_to_user_ptr(kbuf
->addr
);
2979 #ifdef CONFIG_COMPAT
2980 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
2983 struct compat_iovec __user
*uiov
;
2984 compat_ssize_t clen
;
2988 uiov
= u64_to_user_ptr(req
->rw
.addr
);
2989 if (!access_ok(uiov
, sizeof(*uiov
)))
2991 if (__get_user(clen
, &uiov
->iov_len
))
2997 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2999 return PTR_ERR(buf
);
3000 iov
[0].iov_base
= buf
;
3001 iov
[0].iov_len
= (compat_size_t
) len
;
3006 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3009 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3013 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3016 len
= iov
[0].iov_len
;
3019 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3021 return PTR_ERR(buf
);
3022 iov
[0].iov_base
= buf
;
3023 iov
[0].iov_len
= len
;
3027 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3030 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3031 struct io_buffer
*kbuf
;
3033 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3034 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3035 iov
[0].iov_len
= kbuf
->len
;
3040 else if (req
->rw
.len
> 1)
3043 #ifdef CONFIG_COMPAT
3044 if (req
->ctx
->compat
)
3045 return io_compat_import(req
, iov
, needs_lock
);
3048 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3051 static ssize_t
__io_import_iovec(int rw
, struct io_kiocb
*req
,
3052 struct iovec
**iovec
, struct iov_iter
*iter
,
3055 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3056 size_t sqe_len
= req
->rw
.len
;
3060 opcode
= req
->opcode
;
3061 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3063 return io_import_fixed(req
, rw
, iter
);
3066 /* buffer index only valid with fixed read/write, or buffer select */
3067 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3070 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3071 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3072 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3074 return PTR_ERR(buf
);
3075 req
->rw
.len
= sqe_len
;
3078 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3080 return ret
< 0 ? ret
: sqe_len
;
3083 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3084 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3086 ret
= (*iovec
)->iov_len
;
3087 iov_iter_init(iter
, rw
, *iovec
, 1, ret
);
3093 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3097 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
3098 struct iovec
**iovec
, struct iov_iter
*iter
,
3101 struct io_async_rw
*iorw
= req
->async_data
;
3104 return __io_import_iovec(rw
, req
, iovec
, iter
, needs_lock
);
3106 return iov_iter_count(&iorw
->iter
);
3109 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3111 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3115 * For files that don't have ->read_iter() and ->write_iter(), handle them
3116 * by looping over ->read() or ->write() manually.
3118 static ssize_t
loop_rw_iter(int rw
, struct file
*file
, struct kiocb
*kiocb
,
3119 struct iov_iter
*iter
)
3124 * Don't support polled IO through this interface, and we can't
3125 * support non-blocking either. For the latter, this just causes
3126 * the kiocb to be handled from an async context.
3128 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3130 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3133 while (iov_iter_count(iter
)) {
3137 if (!iov_iter_is_bvec(iter
)) {
3138 iovec
= iov_iter_iovec(iter
);
3140 /* fixed buffers import bvec */
3141 iovec
.iov_base
= kmap(iter
->bvec
->bv_page
)
3143 iovec
.iov_len
= min(iter
->count
,
3144 iter
->bvec
->bv_len
- iter
->iov_offset
);
3148 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3149 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3151 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3152 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3155 if (iov_iter_is_bvec(iter
))
3156 kunmap(iter
->bvec
->bv_page
);
3164 if (nr
!= iovec
.iov_len
)
3166 iov_iter_advance(iter
, nr
);
3172 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3173 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3175 struct io_async_rw
*rw
= req
->async_data
;
3177 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3178 rw
->free_iovec
= iovec
;
3180 /* can only be fixed buffers, no need to do anything */
3181 if (iter
->type
== ITER_BVEC
)
3184 unsigned iov_off
= 0;
3186 rw
->iter
.iov
= rw
->fast_iov
;
3187 if (iter
->iov
!= fast_iov
) {
3188 iov_off
= iter
->iov
- fast_iov
;
3189 rw
->iter
.iov
+= iov_off
;
3191 if (rw
->fast_iov
!= fast_iov
)
3192 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3193 sizeof(struct iovec
) * iter
->nr_segs
);
3195 req
->flags
|= REQ_F_NEED_CLEANUP
;
3199 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3201 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3202 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3203 return req
->async_data
== NULL
;
3206 static int io_alloc_async_data(struct io_kiocb
*req
)
3208 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3211 return __io_alloc_async_data(req
);
3214 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3215 const struct iovec
*fast_iov
,
3216 struct iov_iter
*iter
, bool force
)
3218 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3220 if (!req
->async_data
) {
3221 if (__io_alloc_async_data(req
))
3224 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3229 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3231 struct io_async_rw
*iorw
= req
->async_data
;
3232 struct iovec
*iov
= iorw
->fast_iov
;
3235 ret
= __io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3236 if (unlikely(ret
< 0))
3239 iorw
->bytes_done
= 0;
3240 iorw
->free_iovec
= iov
;
3242 req
->flags
|= REQ_F_NEED_CLEANUP
;
3246 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3250 ret
= io_prep_rw(req
, sqe
);
3254 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3257 /* either don't need iovec imported or already have it */
3258 if (!req
->async_data
)
3260 return io_rw_prep_async(req
, READ
);
3264 * This is our waitqueue callback handler, registered through lock_page_async()
3265 * when we initially tried to do the IO with the iocb armed our waitqueue.
3266 * This gets called when the page is unlocked, and we generally expect that to
3267 * happen when the page IO is completed and the page is now uptodate. This will
3268 * queue a task_work based retry of the operation, attempting to copy the data
3269 * again. If the latter fails because the page was NOT uptodate, then we will
3270 * do a thread based blocking retry of the operation. That's the unexpected
3273 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3274 int sync
, void *arg
)
3276 struct wait_page_queue
*wpq
;
3277 struct io_kiocb
*req
= wait
->private;
3278 struct wait_page_key
*key
= arg
;
3281 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3283 if (!wake_page_match(wpq
, key
))
3286 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3287 list_del_init(&wait
->entry
);
3289 init_task_work(&req
->task_work
, io_req_task_submit
);
3290 percpu_ref_get(&req
->ctx
->refs
);
3292 /* submit ref gets dropped, acquire a new one */
3293 refcount_inc(&req
->refs
);
3294 ret
= io_req_task_work_add(req
, true);
3295 if (unlikely(ret
)) {
3296 struct task_struct
*tsk
;
3298 /* queue just for cancelation */
3299 init_task_work(&req
->task_work
, io_req_task_cancel
);
3300 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
3301 task_work_add(tsk
, &req
->task_work
, 0);
3302 wake_up_process(tsk
);
3308 * This controls whether a given IO request should be armed for async page
3309 * based retry. If we return false here, the request is handed to the async
3310 * worker threads for retry. If we're doing buffered reads on a regular file,
3311 * we prepare a private wait_page_queue entry and retry the operation. This
3312 * will either succeed because the page is now uptodate and unlocked, or it
3313 * will register a callback when the page is unlocked at IO completion. Through
3314 * that callback, io_uring uses task_work to setup a retry of the operation.
3315 * That retry will attempt the buffered read again. The retry will generally
3316 * succeed, or in rare cases where it fails, we then fall back to using the
3317 * async worker threads for a blocking retry.
3319 static bool io_rw_should_retry(struct io_kiocb
*req
)
3321 struct io_async_rw
*rw
= req
->async_data
;
3322 struct wait_page_queue
*wait
= &rw
->wpq
;
3323 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3325 /* never retry for NOWAIT, we just complete with -EAGAIN */
3326 if (req
->flags
& REQ_F_NOWAIT
)
3329 /* Only for buffered IO */
3330 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3334 * just use poll if we can, and don't attempt if the fs doesn't
3335 * support callback based unlocks
3337 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3340 wait
->wait
.func
= io_async_buf_func
;
3341 wait
->wait
.private = req
;
3342 wait
->wait
.flags
= 0;
3343 INIT_LIST_HEAD(&wait
->wait
.entry
);
3344 kiocb
->ki_flags
|= IOCB_WAITQ
;
3345 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3346 kiocb
->ki_waitq
= wait
;
3350 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3352 if (req
->file
->f_op
->read_iter
)
3353 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3354 else if (req
->file
->f_op
->read
)
3355 return loop_rw_iter(READ
, req
->file
, &req
->rw
.kiocb
, iter
);
3360 static int io_read(struct io_kiocb
*req
, bool force_nonblock
,
3361 struct io_comp_state
*cs
)
3363 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3364 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3365 struct iov_iter __iter
, *iter
= &__iter
;
3366 struct io_async_rw
*rw
= req
->async_data
;
3367 ssize_t io_size
, ret
, ret2
;
3374 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3377 iov_count
= iov_iter_count(iter
);
3379 req
->result
= io_size
;
3382 /* Ensure we clear previously set non-block flag */
3383 if (!force_nonblock
)
3384 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3386 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3389 /* If the file doesn't support async, just async punt */
3390 no_async
= force_nonblock
&& !io_file_supports_async(req
->file
, READ
);
3394 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), iov_count
);
3398 ret
= io_iter_do_read(req
, iter
);
3402 } else if (ret
== -EIOCBQUEUED
) {
3405 } else if (ret
== -EAGAIN
) {
3406 /* IOPOLL retry should happen for io-wq threads */
3407 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3409 /* no retry on NONBLOCK marked file */
3410 if (req
->file
->f_flags
& O_NONBLOCK
)
3412 /* some cases will consume bytes even on error returns */
3413 iov_iter_revert(iter
, iov_count
- iov_iter_count(iter
));
3416 } else if (ret
< 0) {
3417 /* make sure -ERESTARTSYS -> -EINTR is done */
3421 /* read it all, or we did blocking attempt. no retry. */
3422 if (!iov_iter_count(iter
) || !force_nonblock
||
3423 (req
->file
->f_flags
& O_NONBLOCK
))
3428 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3435 rw
= req
->async_data
;
3436 /* it's copied and will be cleaned with ->io */
3438 /* now use our persistent iterator, if we aren't already */
3441 rw
->bytes_done
+= ret
;
3442 /* if we can retry, do so with the callbacks armed */
3443 if (!io_rw_should_retry(req
)) {
3444 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3449 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3450 * get -EIOCBQUEUED, then we'll get a notification when the desired
3451 * page gets unlocked. We can also get a partial read here, and if we
3452 * do, then just retry at the new offset.
3454 ret
= io_iter_do_read(req
, iter
);
3455 if (ret
== -EIOCBQUEUED
) {
3458 } else if (ret
> 0 && ret
< io_size
) {
3459 /* we got some bytes, but not all. retry. */
3463 kiocb_done(kiocb
, ret
, cs
);
3466 /* it's reportedly faster than delegating the null check to kfree() */
3472 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3476 ret
= io_prep_rw(req
, sqe
);
3480 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3483 /* either don't need iovec imported or already have it */
3484 if (!req
->async_data
)
3486 return io_rw_prep_async(req
, WRITE
);
3489 static int io_write(struct io_kiocb
*req
, bool force_nonblock
,
3490 struct io_comp_state
*cs
)
3492 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3493 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3494 struct iov_iter __iter
, *iter
= &__iter
;
3495 struct io_async_rw
*rw
= req
->async_data
;
3497 ssize_t ret
, ret2
, io_size
;
3502 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3505 iov_count
= iov_iter_count(iter
);
3507 req
->result
= io_size
;
3509 /* Ensure we clear previously set non-block flag */
3510 if (!force_nonblock
)
3511 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3513 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3515 /* If the file doesn't support async, just async punt */
3516 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3519 /* file path doesn't support NOWAIT for non-direct_IO */
3520 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3521 (req
->flags
& REQ_F_ISREG
))
3524 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), iov_count
);
3529 * Open-code file_start_write here to grab freeze protection,
3530 * which will be released by another thread in
3531 * io_complete_rw(). Fool lockdep by telling it the lock got
3532 * released so that it doesn't complain about the held lock when
3533 * we return to userspace.
3535 if (req
->flags
& REQ_F_ISREG
) {
3536 __sb_start_write(file_inode(req
->file
)->i_sb
,
3537 SB_FREEZE_WRITE
, true);
3538 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3541 kiocb
->ki_flags
|= IOCB_WRITE
;
3543 if (req
->file
->f_op
->write_iter
)
3544 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3545 else if (req
->file
->f_op
->write
)
3546 ret2
= loop_rw_iter(WRITE
, req
->file
, kiocb
, iter
);
3551 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3552 * retry them without IOCB_NOWAIT.
3554 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3556 /* no retry on NONBLOCK marked file */
3557 if (ret2
== -EAGAIN
&& (req
->file
->f_flags
& O_NONBLOCK
))
3559 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3560 /* IOPOLL retry should happen for io-wq threads */
3561 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3564 kiocb_done(kiocb
, ret2
, cs
);
3567 /* some cases will consume bytes even on error returns */
3568 iov_iter_revert(iter
, iov_count
- iov_iter_count(iter
));
3569 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3574 /* it's reportedly faster than delegating the null check to kfree() */
3580 static int __io_splice_prep(struct io_kiocb
*req
,
3581 const struct io_uring_sqe
*sqe
)
3583 struct io_splice
* sp
= &req
->splice
;
3584 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3586 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3590 sp
->len
= READ_ONCE(sqe
->len
);
3591 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3593 if (unlikely(sp
->flags
& ~valid_flags
))
3596 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3597 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3600 req
->flags
|= REQ_F_NEED_CLEANUP
;
3602 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3604 * Splice operation will be punted aync, and here need to
3605 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3607 io_req_init_async(req
);
3608 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3614 static int io_tee_prep(struct io_kiocb
*req
,
3615 const struct io_uring_sqe
*sqe
)
3617 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3619 return __io_splice_prep(req
, sqe
);
3622 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3624 struct io_splice
*sp
= &req
->splice
;
3625 struct file
*in
= sp
->file_in
;
3626 struct file
*out
= sp
->file_out
;
3627 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3633 ret
= do_tee(in
, out
, sp
->len
, flags
);
3635 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3636 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3639 req_set_fail_links(req
);
3640 io_req_complete(req
, ret
);
3644 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3646 struct io_splice
* sp
= &req
->splice
;
3648 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3649 sp
->off_out
= READ_ONCE(sqe
->off
);
3650 return __io_splice_prep(req
, sqe
);
3653 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3655 struct io_splice
*sp
= &req
->splice
;
3656 struct file
*in
= sp
->file_in
;
3657 struct file
*out
= sp
->file_out
;
3658 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3659 loff_t
*poff_in
, *poff_out
;
3665 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3666 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3669 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3671 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3672 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3675 req_set_fail_links(req
);
3676 io_req_complete(req
, ret
);
3681 * IORING_OP_NOP just posts a completion event, nothing else.
3683 static int io_nop(struct io_kiocb
*req
, struct io_comp_state
*cs
)
3685 struct io_ring_ctx
*ctx
= req
->ctx
;
3687 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3690 __io_req_complete(req
, 0, 0, cs
);
3694 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3696 struct io_ring_ctx
*ctx
= req
->ctx
;
3701 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3703 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3706 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3707 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3710 req
->sync
.off
= READ_ONCE(sqe
->off
);
3711 req
->sync
.len
= READ_ONCE(sqe
->len
);
3715 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3717 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3720 /* fsync always requires a blocking context */
3724 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3725 end
> 0 ? end
: LLONG_MAX
,
3726 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3728 req_set_fail_links(req
);
3729 io_req_complete(req
, ret
);
3733 static int io_fallocate_prep(struct io_kiocb
*req
,
3734 const struct io_uring_sqe
*sqe
)
3736 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3738 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3741 req
->sync
.off
= READ_ONCE(sqe
->off
);
3742 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3743 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3747 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
3751 /* fallocate always requiring blocking context */
3754 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3757 req_set_fail_links(req
);
3758 io_req_complete(req
, ret
);
3762 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3764 const char __user
*fname
;
3767 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3769 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3772 /* open.how should be already initialised */
3773 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
3774 req
->open
.how
.flags
|= O_LARGEFILE
;
3776 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3777 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3778 req
->open
.filename
= getname(fname
);
3779 if (IS_ERR(req
->open
.filename
)) {
3780 ret
= PTR_ERR(req
->open
.filename
);
3781 req
->open
.filename
= NULL
;
3784 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
3785 req
->flags
|= REQ_F_NEED_CLEANUP
;
3789 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3793 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3795 mode
= READ_ONCE(sqe
->len
);
3796 flags
= READ_ONCE(sqe
->open_flags
);
3797 req
->open
.how
= build_open_how(flags
, mode
);
3798 return __io_openat_prep(req
, sqe
);
3801 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3803 struct open_how __user
*how
;
3807 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3809 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3810 len
= READ_ONCE(sqe
->len
);
3811 if (len
< OPEN_HOW_SIZE_VER0
)
3814 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
3819 return __io_openat_prep(req
, sqe
);
3822 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
3824 struct open_flags op
;
3831 ret
= build_open_flags(&req
->open
.how
, &op
);
3835 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
3839 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
3842 ret
= PTR_ERR(file
);
3844 fsnotify_open(file
);
3845 fd_install(ret
, file
);
3848 putname(req
->open
.filename
);
3849 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3851 req_set_fail_links(req
);
3852 io_req_complete(req
, ret
);
3856 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
3858 return io_openat2(req
, force_nonblock
);
3861 static int io_remove_buffers_prep(struct io_kiocb
*req
,
3862 const struct io_uring_sqe
*sqe
)
3864 struct io_provide_buf
*p
= &req
->pbuf
;
3867 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
3870 tmp
= READ_ONCE(sqe
->fd
);
3871 if (!tmp
|| tmp
> USHRT_MAX
)
3874 memset(p
, 0, sizeof(*p
));
3876 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3880 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
3881 int bgid
, unsigned nbufs
)
3885 /* shouldn't happen */
3889 /* the head kbuf is the list itself */
3890 while (!list_empty(&buf
->list
)) {
3891 struct io_buffer
*nxt
;
3893 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
3894 list_del(&nxt
->list
);
3901 idr_remove(&ctx
->io_buffer_idr
, bgid
);
3906 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
,
3907 struct io_comp_state
*cs
)
3909 struct io_provide_buf
*p
= &req
->pbuf
;
3910 struct io_ring_ctx
*ctx
= req
->ctx
;
3911 struct io_buffer
*head
;
3914 io_ring_submit_lock(ctx
, !force_nonblock
);
3916 lockdep_assert_held(&ctx
->uring_lock
);
3919 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3921 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
3923 io_ring_submit_lock(ctx
, !force_nonblock
);
3925 req_set_fail_links(req
);
3926 __io_req_complete(req
, ret
, 0, cs
);
3930 static int io_provide_buffers_prep(struct io_kiocb
*req
,
3931 const struct io_uring_sqe
*sqe
)
3933 struct io_provide_buf
*p
= &req
->pbuf
;
3936 if (sqe
->ioprio
|| sqe
->rw_flags
)
3939 tmp
= READ_ONCE(sqe
->fd
);
3940 if (!tmp
|| tmp
> USHRT_MAX
)
3943 p
->addr
= READ_ONCE(sqe
->addr
);
3944 p
->len
= READ_ONCE(sqe
->len
);
3946 if (!access_ok(u64_to_user_ptr(p
->addr
), (p
->len
* p
->nbufs
)))
3949 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3950 tmp
= READ_ONCE(sqe
->off
);
3951 if (tmp
> USHRT_MAX
)
3957 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
3959 struct io_buffer
*buf
;
3960 u64 addr
= pbuf
->addr
;
3961 int i
, bid
= pbuf
->bid
;
3963 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
3964 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
3969 buf
->len
= pbuf
->len
;
3974 INIT_LIST_HEAD(&buf
->list
);
3977 list_add_tail(&buf
->list
, &(*head
)->list
);
3981 return i
? i
: -ENOMEM
;
3984 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
,
3985 struct io_comp_state
*cs
)
3987 struct io_provide_buf
*p
= &req
->pbuf
;
3988 struct io_ring_ctx
*ctx
= req
->ctx
;
3989 struct io_buffer
*head
, *list
;
3992 io_ring_submit_lock(ctx
, !force_nonblock
);
3994 lockdep_assert_held(&ctx
->uring_lock
);
3996 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3998 ret
= io_add_buffers(p
, &head
);
4003 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
4006 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4011 io_ring_submit_unlock(ctx
, !force_nonblock
);
4013 req_set_fail_links(req
);
4014 __io_req_complete(req
, ret
, 0, cs
);
4018 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4019 const struct io_uring_sqe
*sqe
)
4021 #if defined(CONFIG_EPOLL)
4022 if (sqe
->ioprio
|| sqe
->buf_index
)
4024 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4027 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4028 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4029 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4031 if (ep_op_has_event(req
->epoll
.op
)) {
4032 struct epoll_event __user
*ev
;
4034 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4035 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4045 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
,
4046 struct io_comp_state
*cs
)
4048 #if defined(CONFIG_EPOLL)
4049 struct io_epoll
*ie
= &req
->epoll
;
4052 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4053 if (force_nonblock
&& ret
== -EAGAIN
)
4057 req_set_fail_links(req
);
4058 __io_req_complete(req
, ret
, 0, cs
);
4065 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4067 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4068 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4070 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4073 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4074 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4075 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4082 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
4084 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4085 struct io_madvise
*ma
= &req
->madvise
;
4091 ret
= do_madvise(ma
->addr
, ma
->len
, ma
->advice
);
4093 req_set_fail_links(req
);
4094 io_req_complete(req
, ret
);
4101 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4103 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4105 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4108 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4109 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4110 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4114 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
4116 struct io_fadvise
*fa
= &req
->fadvise
;
4119 if (force_nonblock
) {
4120 switch (fa
->advice
) {
4121 case POSIX_FADV_NORMAL
:
4122 case POSIX_FADV_RANDOM
:
4123 case POSIX_FADV_SEQUENTIAL
:
4130 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4132 req_set_fail_links(req
);
4133 io_req_complete(req
, ret
);
4137 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4139 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4141 if (sqe
->ioprio
|| sqe
->buf_index
)
4143 if (req
->flags
& REQ_F_FIXED_FILE
)
4146 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4147 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4148 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4149 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4150 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4155 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
4157 struct io_statx
*ctx
= &req
->statx
;
4160 if (force_nonblock
) {
4161 /* only need file table for an actual valid fd */
4162 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4163 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4167 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4171 req_set_fail_links(req
);
4172 io_req_complete(req
, ret
);
4176 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4179 * If we queue this for async, it must not be cancellable. That would
4180 * leave the 'file' in an undeterminate state, and here need to modify
4181 * io_wq_work.flags, so initialize io_wq_work firstly.
4183 io_req_init_async(req
);
4184 req
->work
.flags
|= IO_WQ_WORK_NO_CANCEL
;
4186 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
4188 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4189 sqe
->rw_flags
|| sqe
->buf_index
)
4191 if (req
->flags
& REQ_F_FIXED_FILE
)
4194 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4195 if ((req
->file
&& req
->file
->f_op
== &io_uring_fops
))
4198 req
->close
.put_file
= NULL
;
4202 static int io_close(struct io_kiocb
*req
, bool force_nonblock
,
4203 struct io_comp_state
*cs
)
4205 struct io_close
*close
= &req
->close
;
4208 /* might be already done during nonblock submission */
4209 if (!close
->put_file
) {
4210 ret
= __close_fd_get_file(close
->fd
, &close
->put_file
);
4212 return (ret
== -ENOENT
) ? -EBADF
: ret
;
4215 /* if the file has a flush method, be safe and punt to async */
4216 if (close
->put_file
->f_op
->flush
&& force_nonblock
) {
4217 /* was never set, but play safe */
4218 req
->flags
&= ~REQ_F_NOWAIT
;
4219 /* avoid grabbing files - we don't need the files */
4220 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4224 /* No ->flush() or already async, safely close from here */
4225 ret
= filp_close(close
->put_file
, req
->work
.identity
->files
);
4227 req_set_fail_links(req
);
4228 fput(close
->put_file
);
4229 close
->put_file
= NULL
;
4230 __io_req_complete(req
, ret
, 0, cs
);
4234 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4236 struct io_ring_ctx
*ctx
= req
->ctx
;
4241 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4243 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4246 req
->sync
.off
= READ_ONCE(sqe
->off
);
4247 req
->sync
.len
= READ_ONCE(sqe
->len
);
4248 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4252 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
4256 /* sync_file_range always requires a blocking context */
4260 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4263 req_set_fail_links(req
);
4264 io_req_complete(req
, ret
);
4268 #if defined(CONFIG_NET)
4269 static int io_setup_async_msg(struct io_kiocb
*req
,
4270 struct io_async_msghdr
*kmsg
)
4272 struct io_async_msghdr
*async_msg
= req
->async_data
;
4276 if (io_alloc_async_data(req
)) {
4277 if (kmsg
->iov
!= kmsg
->fast_iov
)
4281 async_msg
= req
->async_data
;
4282 req
->flags
|= REQ_F_NEED_CLEANUP
;
4283 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4287 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4288 struct io_async_msghdr
*iomsg
)
4290 iomsg
->iov
= iomsg
->fast_iov
;
4291 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4292 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4293 req
->sr_msg
.msg_flags
, &iomsg
->iov
);
4296 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4298 struct io_async_msghdr
*async_msg
= req
->async_data
;
4299 struct io_sr_msg
*sr
= &req
->sr_msg
;
4302 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4305 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4306 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4307 sr
->len
= READ_ONCE(sqe
->len
);
4309 #ifdef CONFIG_COMPAT
4310 if (req
->ctx
->compat
)
4311 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4314 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4316 ret
= io_sendmsg_copy_hdr(req
, async_msg
);
4318 req
->flags
|= REQ_F_NEED_CLEANUP
;
4322 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4323 struct io_comp_state
*cs
)
4325 struct io_async_msghdr iomsg
, *kmsg
;
4326 struct socket
*sock
;
4330 sock
= sock_from_file(req
->file
, &ret
);
4331 if (unlikely(!sock
))
4334 if (req
->async_data
) {
4335 kmsg
= req
->async_data
;
4336 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4337 /* if iov is set, it's allocated already */
4339 kmsg
->iov
= kmsg
->fast_iov
;
4340 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4342 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4348 flags
= req
->sr_msg
.msg_flags
;
4349 if (flags
& MSG_DONTWAIT
)
4350 req
->flags
|= REQ_F_NOWAIT
;
4351 else if (force_nonblock
)
4352 flags
|= MSG_DONTWAIT
;
4354 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4355 if (force_nonblock
&& ret
== -EAGAIN
)
4356 return io_setup_async_msg(req
, kmsg
);
4357 if (ret
== -ERESTARTSYS
)
4360 if (kmsg
->iov
!= kmsg
->fast_iov
)
4362 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4364 req_set_fail_links(req
);
4365 __io_req_complete(req
, ret
, 0, cs
);
4369 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4370 struct io_comp_state
*cs
)
4372 struct io_sr_msg
*sr
= &req
->sr_msg
;
4375 struct socket
*sock
;
4379 sock
= sock_from_file(req
->file
, &ret
);
4380 if (unlikely(!sock
))
4383 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4387 msg
.msg_name
= NULL
;
4388 msg
.msg_control
= NULL
;
4389 msg
.msg_controllen
= 0;
4390 msg
.msg_namelen
= 0;
4392 flags
= req
->sr_msg
.msg_flags
;
4393 if (flags
& MSG_DONTWAIT
)
4394 req
->flags
|= REQ_F_NOWAIT
;
4395 else if (force_nonblock
)
4396 flags
|= MSG_DONTWAIT
;
4398 msg
.msg_flags
= flags
;
4399 ret
= sock_sendmsg(sock
, &msg
);
4400 if (force_nonblock
&& ret
== -EAGAIN
)
4402 if (ret
== -ERESTARTSYS
)
4406 req_set_fail_links(req
);
4407 __io_req_complete(req
, ret
, 0, cs
);
4411 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4412 struct io_async_msghdr
*iomsg
)
4414 struct io_sr_msg
*sr
= &req
->sr_msg
;
4415 struct iovec __user
*uiov
;
4419 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4420 &iomsg
->uaddr
, &uiov
, &iov_len
);
4424 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4427 if (copy_from_user(iomsg
->iov
, uiov
, sizeof(*uiov
)))
4429 sr
->len
= iomsg
->iov
[0].iov_len
;
4430 iov_iter_init(&iomsg
->msg
.msg_iter
, READ
, iomsg
->iov
, 1,
4434 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4435 &iomsg
->iov
, &iomsg
->msg
.msg_iter
,
4444 #ifdef CONFIG_COMPAT
4445 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4446 struct io_async_msghdr
*iomsg
)
4448 struct compat_msghdr __user
*msg_compat
;
4449 struct io_sr_msg
*sr
= &req
->sr_msg
;
4450 struct compat_iovec __user
*uiov
;
4455 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4456 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4461 uiov
= compat_ptr(ptr
);
4462 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4463 compat_ssize_t clen
;
4467 if (!access_ok(uiov
, sizeof(*uiov
)))
4469 if (__get_user(clen
, &uiov
->iov_len
))
4473 sr
->len
= iomsg
->iov
[0].iov_len
;
4476 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4477 UIO_FASTIOV
, &iomsg
->iov
,
4478 &iomsg
->msg
.msg_iter
, true);
4487 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4488 struct io_async_msghdr
*iomsg
)
4490 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4491 iomsg
->iov
= iomsg
->fast_iov
;
4493 #ifdef CONFIG_COMPAT
4494 if (req
->ctx
->compat
)
4495 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4498 return __io_recvmsg_copy_hdr(req
, iomsg
);
4501 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4504 struct io_sr_msg
*sr
= &req
->sr_msg
;
4505 struct io_buffer
*kbuf
;
4507 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4512 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4516 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4518 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4521 static int io_recvmsg_prep(struct io_kiocb
*req
,
4522 const struct io_uring_sqe
*sqe
)
4524 struct io_async_msghdr
*async_msg
= req
->async_data
;
4525 struct io_sr_msg
*sr
= &req
->sr_msg
;
4528 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4531 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4532 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4533 sr
->len
= READ_ONCE(sqe
->len
);
4534 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4536 #ifdef CONFIG_COMPAT
4537 if (req
->ctx
->compat
)
4538 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4541 if (!async_msg
|| !io_op_defs
[req
->opcode
].needs_async_data
)
4543 ret
= io_recvmsg_copy_hdr(req
, async_msg
);
4545 req
->flags
|= REQ_F_NEED_CLEANUP
;
4549 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
4550 struct io_comp_state
*cs
)
4552 struct io_async_msghdr iomsg
, *kmsg
;
4553 struct socket
*sock
;
4554 struct io_buffer
*kbuf
;
4556 int ret
, cflags
= 0;
4558 sock
= sock_from_file(req
->file
, &ret
);
4559 if (unlikely(!sock
))
4562 if (req
->async_data
) {
4563 kmsg
= req
->async_data
;
4564 kmsg
->msg
.msg_name
= &kmsg
->addr
;
4565 /* if iov is set, it's allocated already */
4567 kmsg
->iov
= kmsg
->fast_iov
;
4568 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4570 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4576 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4577 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4579 return PTR_ERR(kbuf
);
4580 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4581 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4582 1, req
->sr_msg
.len
);
4585 flags
= req
->sr_msg
.msg_flags
;
4586 if (flags
& MSG_DONTWAIT
)
4587 req
->flags
|= REQ_F_NOWAIT
;
4588 else if (force_nonblock
)
4589 flags
|= MSG_DONTWAIT
;
4591 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4592 kmsg
->uaddr
, flags
);
4593 if (force_nonblock
&& ret
== -EAGAIN
)
4594 return io_setup_async_msg(req
, kmsg
);
4595 if (ret
== -ERESTARTSYS
)
4598 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4599 cflags
= io_put_recv_kbuf(req
);
4600 if (kmsg
->iov
!= kmsg
->fast_iov
)
4602 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4604 req_set_fail_links(req
);
4605 __io_req_complete(req
, ret
, cflags
, cs
);
4609 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
4610 struct io_comp_state
*cs
)
4612 struct io_buffer
*kbuf
;
4613 struct io_sr_msg
*sr
= &req
->sr_msg
;
4615 void __user
*buf
= sr
->buf
;
4616 struct socket
*sock
;
4619 int ret
, cflags
= 0;
4621 sock
= sock_from_file(req
->file
, &ret
);
4622 if (unlikely(!sock
))
4625 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4626 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4628 return PTR_ERR(kbuf
);
4629 buf
= u64_to_user_ptr(kbuf
->addr
);
4632 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4636 msg
.msg_name
= NULL
;
4637 msg
.msg_control
= NULL
;
4638 msg
.msg_controllen
= 0;
4639 msg
.msg_namelen
= 0;
4640 msg
.msg_iocb
= NULL
;
4643 flags
= req
->sr_msg
.msg_flags
;
4644 if (flags
& MSG_DONTWAIT
)
4645 req
->flags
|= REQ_F_NOWAIT
;
4646 else if (force_nonblock
)
4647 flags
|= MSG_DONTWAIT
;
4649 ret
= sock_recvmsg(sock
, &msg
, flags
);
4650 if (force_nonblock
&& ret
== -EAGAIN
)
4652 if (ret
== -ERESTARTSYS
)
4655 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4656 cflags
= io_put_recv_kbuf(req
);
4658 req_set_fail_links(req
);
4659 __io_req_complete(req
, ret
, cflags
, cs
);
4663 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4665 struct io_accept
*accept
= &req
->accept
;
4667 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
4669 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4672 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4673 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4674 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4675 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4679 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
4680 struct io_comp_state
*cs
)
4682 struct io_accept
*accept
= &req
->accept
;
4683 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4686 if (req
->file
->f_flags
& O_NONBLOCK
)
4687 req
->flags
|= REQ_F_NOWAIT
;
4689 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4690 accept
->addr_len
, accept
->flags
,
4692 if (ret
== -EAGAIN
&& force_nonblock
)
4695 if (ret
== -ERESTARTSYS
)
4697 req_set_fail_links(req
);
4699 __io_req_complete(req
, ret
, 0, cs
);
4703 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4705 struct io_connect
*conn
= &req
->connect
;
4706 struct io_async_connect
*io
= req
->async_data
;
4708 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
4710 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4713 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4714 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4719 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
4723 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
4724 struct io_comp_state
*cs
)
4726 struct io_async_connect __io
, *io
;
4727 unsigned file_flags
;
4730 if (req
->async_data
) {
4731 io
= req
->async_data
;
4733 ret
= move_addr_to_kernel(req
->connect
.addr
,
4734 req
->connect
.addr_len
,
4741 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4743 ret
= __sys_connect_file(req
->file
, &io
->address
,
4744 req
->connect
.addr_len
, file_flags
);
4745 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4746 if (req
->async_data
)
4748 if (io_alloc_async_data(req
)) {
4752 io
= req
->async_data
;
4753 memcpy(req
->async_data
, &__io
, sizeof(__io
));
4756 if (ret
== -ERESTARTSYS
)
4760 req_set_fail_links(req
);
4761 __io_req_complete(req
, ret
, 0, cs
);
4764 #else /* !CONFIG_NET */
4765 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4770 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
,
4771 struct io_comp_state
*cs
)
4776 static int io_send(struct io_kiocb
*req
, bool force_nonblock
,
4777 struct io_comp_state
*cs
)
4782 static int io_recvmsg_prep(struct io_kiocb
*req
,
4783 const struct io_uring_sqe
*sqe
)
4788 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
,
4789 struct io_comp_state
*cs
)
4794 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
,
4795 struct io_comp_state
*cs
)
4800 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4805 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
,
4806 struct io_comp_state
*cs
)
4811 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4816 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
,
4817 struct io_comp_state
*cs
)
4821 #endif /* CONFIG_NET */
4823 struct io_poll_table
{
4824 struct poll_table_struct pt
;
4825 struct io_kiocb
*req
;
4829 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
4830 __poll_t mask
, task_work_func_t func
)
4835 /* for instances that support it check for an event match first: */
4836 if (mask
&& !(mask
& poll
->events
))
4839 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
4841 list_del_init(&poll
->wait
.entry
);
4844 init_task_work(&req
->task_work
, func
);
4845 percpu_ref_get(&req
->ctx
->refs
);
4848 * If we using the signalfd wait_queue_head for this wakeup, then
4849 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4850 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4851 * either, as the normal wakeup will suffice.
4853 twa_signal_ok
= (poll
->head
!= &req
->task
->sighand
->signalfd_wqh
);
4856 * If this fails, then the task is exiting. When a task exits, the
4857 * work gets canceled, so just cancel this request as well instead
4858 * of executing it. We can't safely execute it anyway, as we may not
4859 * have the needed state needed for it anyway.
4861 ret
= io_req_task_work_add(req
, twa_signal_ok
);
4862 if (unlikely(ret
)) {
4863 struct task_struct
*tsk
;
4865 WRITE_ONCE(poll
->canceled
, true);
4866 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
4867 task_work_add(tsk
, &req
->task_work
, 0);
4868 wake_up_process(tsk
);
4873 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
4874 __acquires(&req
->ctx
->completion_lock
)
4876 struct io_ring_ctx
*ctx
= req
->ctx
;
4878 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4879 struct poll_table_struct pt
= { ._key
= poll
->events
};
4881 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
4884 spin_lock_irq(&ctx
->completion_lock
);
4885 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4886 add_wait_queue(poll
->head
, &poll
->wait
);
4893 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
4895 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4896 if (req
->opcode
== IORING_OP_POLL_ADD
)
4897 return req
->async_data
;
4898 return req
->apoll
->double_poll
;
4901 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
4903 if (req
->opcode
== IORING_OP_POLL_ADD
)
4905 return &req
->apoll
->poll
;
4908 static void io_poll_remove_double(struct io_kiocb
*req
)
4910 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
4912 lockdep_assert_held(&req
->ctx
->completion_lock
);
4914 if (poll
&& poll
->head
) {
4915 struct wait_queue_head
*head
= poll
->head
;
4917 spin_lock(&head
->lock
);
4918 list_del_init(&poll
->wait
.entry
);
4919 if (poll
->wait
.private)
4920 refcount_dec(&req
->refs
);
4922 spin_unlock(&head
->lock
);
4926 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
4928 struct io_ring_ctx
*ctx
= req
->ctx
;
4930 io_poll_remove_double(req
);
4931 req
->poll
.done
= true;
4932 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
4933 io_commit_cqring(ctx
);
4936 static void io_poll_task_func(struct callback_head
*cb
)
4938 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4939 struct io_ring_ctx
*ctx
= req
->ctx
;
4940 struct io_kiocb
*nxt
;
4942 if (io_poll_rewait(req
, &req
->poll
)) {
4943 spin_unlock_irq(&ctx
->completion_lock
);
4945 hash_del(&req
->hash_node
);
4946 io_poll_complete(req
, req
->result
, 0);
4947 spin_unlock_irq(&ctx
->completion_lock
);
4949 nxt
= io_put_req_find_next(req
);
4950 io_cqring_ev_posted(ctx
);
4952 __io_req_task_submit(nxt
);
4955 percpu_ref_put(&ctx
->refs
);
4958 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
4959 int sync
, void *key
)
4961 struct io_kiocb
*req
= wait
->private;
4962 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
4963 __poll_t mask
= key_to_poll(key
);
4965 /* for instances that support it check for an event match first: */
4966 if (mask
&& !(mask
& poll
->events
))
4969 list_del_init(&wait
->entry
);
4971 if (poll
&& poll
->head
) {
4974 spin_lock(&poll
->head
->lock
);
4975 done
= list_empty(&poll
->wait
.entry
);
4977 list_del_init(&poll
->wait
.entry
);
4978 /* make sure double remove sees this as being gone */
4979 wait
->private = NULL
;
4980 spin_unlock(&poll
->head
->lock
);
4982 __io_async_wake(req
, poll
, mask
, io_poll_task_func
);
4984 refcount_dec(&req
->refs
);
4988 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
4989 wait_queue_func_t wake_func
)
4993 poll
->canceled
= false;
4994 poll
->events
= events
;
4995 INIT_LIST_HEAD(&poll
->wait
.entry
);
4996 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
4999 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5000 struct wait_queue_head
*head
,
5001 struct io_poll_iocb
**poll_ptr
)
5003 struct io_kiocb
*req
= pt
->req
;
5006 * If poll->head is already set, it's because the file being polled
5007 * uses multiple waitqueues for poll handling (eg one for read, one
5008 * for write). Setup a separate io_poll_iocb if this happens.
5010 if (unlikely(poll
->head
)) {
5011 struct io_poll_iocb
*poll_one
= poll
;
5013 /* already have a 2nd entry, fail a third attempt */
5015 pt
->error
= -EINVAL
;
5018 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5020 pt
->error
= -ENOMEM
;
5023 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5024 refcount_inc(&req
->refs
);
5025 poll
->wait
.private = req
;
5032 if (poll
->events
& EPOLLEXCLUSIVE
)
5033 add_wait_queue_exclusive(head
, &poll
->wait
);
5035 add_wait_queue(head
, &poll
->wait
);
5038 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5039 struct poll_table_struct
*p
)
5041 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5042 struct async_poll
*apoll
= pt
->req
->apoll
;
5044 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5047 static void io_async_task_func(struct callback_head
*cb
)
5049 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5050 struct async_poll
*apoll
= req
->apoll
;
5051 struct io_ring_ctx
*ctx
= req
->ctx
;
5053 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5055 if (io_poll_rewait(req
, &apoll
->poll
)) {
5056 spin_unlock_irq(&ctx
->completion_lock
);
5057 percpu_ref_put(&ctx
->refs
);
5061 /* If req is still hashed, it cannot have been canceled. Don't check. */
5062 if (hash_hashed(&req
->hash_node
))
5063 hash_del(&req
->hash_node
);
5065 io_poll_remove_double(req
);
5066 spin_unlock_irq(&ctx
->completion_lock
);
5068 if (!READ_ONCE(apoll
->poll
.canceled
))
5069 __io_req_task_submit(req
);
5071 __io_req_task_cancel(req
, -ECANCELED
);
5073 percpu_ref_put(&ctx
->refs
);
5074 kfree(apoll
->double_poll
);
5078 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5081 struct io_kiocb
*req
= wait
->private;
5082 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5084 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5087 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5090 static void io_poll_req_insert(struct io_kiocb
*req
)
5092 struct io_ring_ctx
*ctx
= req
->ctx
;
5093 struct hlist_head
*list
;
5095 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5096 hlist_add_head(&req
->hash_node
, list
);
5099 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5100 struct io_poll_iocb
*poll
,
5101 struct io_poll_table
*ipt
, __poll_t mask
,
5102 wait_queue_func_t wake_func
)
5103 __acquires(&ctx
->completion_lock
)
5105 struct io_ring_ctx
*ctx
= req
->ctx
;
5106 bool cancel
= false;
5108 INIT_HLIST_NODE(&req
->hash_node
);
5109 io_init_poll_iocb(poll
, mask
, wake_func
);
5110 poll
->file
= req
->file
;
5111 poll
->wait
.private = req
;
5113 ipt
->pt
._key
= mask
;
5115 ipt
->error
= -EINVAL
;
5117 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5119 spin_lock_irq(&ctx
->completion_lock
);
5120 if (likely(poll
->head
)) {
5121 spin_lock(&poll
->head
->lock
);
5122 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5128 if (mask
|| ipt
->error
)
5129 list_del_init(&poll
->wait
.entry
);
5131 WRITE_ONCE(poll
->canceled
, true);
5132 else if (!poll
->done
) /* actually waiting for an event */
5133 io_poll_req_insert(req
);
5134 spin_unlock(&poll
->head
->lock
);
5140 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5142 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5143 struct io_ring_ctx
*ctx
= req
->ctx
;
5144 struct async_poll
*apoll
;
5145 struct io_poll_table ipt
;
5149 if (!req
->file
|| !file_can_poll(req
->file
))
5151 if (req
->flags
& REQ_F_POLLED
)
5155 else if (def
->pollout
)
5159 /* if we can't nonblock try, then no point in arming a poll handler */
5160 if (!io_file_supports_async(req
->file
, rw
))
5163 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5164 if (unlikely(!apoll
))
5166 apoll
->double_poll
= NULL
;
5168 req
->flags
|= REQ_F_POLLED
;
5173 mask
|= POLLIN
| POLLRDNORM
;
5175 mask
|= POLLOUT
| POLLWRNORM
;
5177 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5178 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5179 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5182 mask
|= POLLERR
| POLLPRI
;
5184 ipt
.pt
._qproc
= io_async_queue_proc
;
5186 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5188 if (ret
|| ipt
.error
) {
5189 io_poll_remove_double(req
);
5190 spin_unlock_irq(&ctx
->completion_lock
);
5191 kfree(apoll
->double_poll
);
5195 spin_unlock_irq(&ctx
->completion_lock
);
5196 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5197 apoll
->poll
.events
);
5201 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5202 struct io_poll_iocb
*poll
)
5204 bool do_complete
= false;
5206 spin_lock(&poll
->head
->lock
);
5207 WRITE_ONCE(poll
->canceled
, true);
5208 if (!list_empty(&poll
->wait
.entry
)) {
5209 list_del_init(&poll
->wait
.entry
);
5212 spin_unlock(&poll
->head
->lock
);
5213 hash_del(&req
->hash_node
);
5217 static bool io_poll_remove_one(struct io_kiocb
*req
)
5221 io_poll_remove_double(req
);
5223 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5224 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5226 struct async_poll
*apoll
= req
->apoll
;
5228 /* non-poll requests have submit ref still */
5229 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5232 kfree(apoll
->double_poll
);
5238 io_cqring_fill_event(req
, -ECANCELED
);
5239 io_commit_cqring(req
->ctx
);
5240 req_set_fail_links(req
);
5241 io_put_req_deferred(req
, 1);
5248 * Returns true if we found and killed one or more poll requests
5250 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
)
5252 struct hlist_node
*tmp
;
5253 struct io_kiocb
*req
;
5256 spin_lock_irq(&ctx
->completion_lock
);
5257 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5258 struct hlist_head
*list
;
5260 list
= &ctx
->cancel_hash
[i
];
5261 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5262 if (io_task_match(req
, tsk
))
5263 posted
+= io_poll_remove_one(req
);
5266 spin_unlock_irq(&ctx
->completion_lock
);
5269 io_cqring_ev_posted(ctx
);
5274 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5276 struct hlist_head
*list
;
5277 struct io_kiocb
*req
;
5279 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5280 hlist_for_each_entry(req
, list
, hash_node
) {
5281 if (sqe_addr
!= req
->user_data
)
5283 if (io_poll_remove_one(req
))
5291 static int io_poll_remove_prep(struct io_kiocb
*req
,
5292 const struct io_uring_sqe
*sqe
)
5294 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5296 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5300 req
->poll
.addr
= READ_ONCE(sqe
->addr
);
5305 * Find a running poll command that matches one specified in sqe->addr,
5306 * and remove it if found.
5308 static int io_poll_remove(struct io_kiocb
*req
)
5310 struct io_ring_ctx
*ctx
= req
->ctx
;
5314 addr
= req
->poll
.addr
;
5315 spin_lock_irq(&ctx
->completion_lock
);
5316 ret
= io_poll_cancel(ctx
, addr
);
5317 spin_unlock_irq(&ctx
->completion_lock
);
5320 req_set_fail_links(req
);
5321 io_req_complete(req
, ret
);
5325 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5328 struct io_kiocb
*req
= wait
->private;
5329 struct io_poll_iocb
*poll
= &req
->poll
;
5331 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5334 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5335 struct poll_table_struct
*p
)
5337 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5339 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5342 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5344 struct io_poll_iocb
*poll
= &req
->poll
;
5347 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5349 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5352 events
= READ_ONCE(sqe
->poll32_events
);
5354 events
= swahw32(events
);
5356 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5357 (events
& EPOLLEXCLUSIVE
);
5361 static int io_poll_add(struct io_kiocb
*req
)
5363 struct io_poll_iocb
*poll
= &req
->poll
;
5364 struct io_ring_ctx
*ctx
= req
->ctx
;
5365 struct io_poll_table ipt
;
5368 ipt
.pt
._qproc
= io_poll_queue_proc
;
5370 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5373 if (mask
) { /* no async, we'd stolen it */
5375 io_poll_complete(req
, mask
, 0);
5377 spin_unlock_irq(&ctx
->completion_lock
);
5380 io_cqring_ev_posted(ctx
);
5386 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5388 struct io_timeout_data
*data
= container_of(timer
,
5389 struct io_timeout_data
, timer
);
5390 struct io_kiocb
*req
= data
->req
;
5391 struct io_ring_ctx
*ctx
= req
->ctx
;
5392 unsigned long flags
;
5394 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5395 list_del_init(&req
->timeout
.list
);
5396 atomic_set(&req
->ctx
->cq_timeouts
,
5397 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5399 io_cqring_fill_event(req
, -ETIME
);
5400 io_commit_cqring(ctx
);
5401 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5403 io_cqring_ev_posted(ctx
);
5404 req_set_fail_links(req
);
5406 return HRTIMER_NORESTART
;
5409 static int __io_timeout_cancel(struct io_kiocb
*req
)
5411 struct io_timeout_data
*io
= req
->async_data
;
5414 ret
= hrtimer_try_to_cancel(&io
->timer
);
5417 list_del_init(&req
->timeout
.list
);
5419 req_set_fail_links(req
);
5420 io_cqring_fill_event(req
, -ECANCELED
);
5421 io_put_req_deferred(req
, 1);
5425 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5427 struct io_kiocb
*req
;
5430 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5431 if (user_data
== req
->user_data
) {
5440 return __io_timeout_cancel(req
);
5443 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5444 const struct io_uring_sqe
*sqe
)
5446 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5448 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5450 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->timeout_flags
)
5453 req
->timeout_rem
.addr
= READ_ONCE(sqe
->addr
);
5458 * Remove or update an existing timeout command
5460 static int io_timeout_remove(struct io_kiocb
*req
)
5462 struct io_ring_ctx
*ctx
= req
->ctx
;
5465 spin_lock_irq(&ctx
->completion_lock
);
5466 ret
= io_timeout_cancel(ctx
, req
->timeout_rem
.addr
);
5468 io_cqring_fill_event(req
, ret
);
5469 io_commit_cqring(ctx
);
5470 spin_unlock_irq(&ctx
->completion_lock
);
5471 io_cqring_ev_posted(ctx
);
5473 req_set_fail_links(req
);
5478 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5479 bool is_timeout_link
)
5481 struct io_timeout_data
*data
;
5483 u32 off
= READ_ONCE(sqe
->off
);
5485 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5487 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5489 if (off
&& is_timeout_link
)
5491 flags
= READ_ONCE(sqe
->timeout_flags
);
5492 if (flags
& ~IORING_TIMEOUT_ABS
)
5495 req
->timeout
.off
= off
;
5497 if (!req
->async_data
&& io_alloc_async_data(req
))
5500 data
= req
->async_data
;
5503 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5506 if (flags
& IORING_TIMEOUT_ABS
)
5507 data
->mode
= HRTIMER_MODE_ABS
;
5509 data
->mode
= HRTIMER_MODE_REL
;
5511 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5515 static int io_timeout(struct io_kiocb
*req
)
5517 struct io_ring_ctx
*ctx
= req
->ctx
;
5518 struct io_timeout_data
*data
= req
->async_data
;
5519 struct list_head
*entry
;
5520 u32 tail
, off
= req
->timeout
.off
;
5522 spin_lock_irq(&ctx
->completion_lock
);
5525 * sqe->off holds how many events that need to occur for this
5526 * timeout event to be satisfied. If it isn't set, then this is
5527 * a pure timeout request, sequence isn't used.
5529 if (io_is_timeout_noseq(req
)) {
5530 entry
= ctx
->timeout_list
.prev
;
5534 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5535 req
->timeout
.target_seq
= tail
+ off
;
5538 * Insertion sort, ensuring the first entry in the list is always
5539 * the one we need first.
5541 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5542 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5545 if (io_is_timeout_noseq(nxt
))
5547 /* nxt.seq is behind @tail, otherwise would've been completed */
5548 if (off
>= nxt
->timeout
.target_seq
- tail
)
5552 list_add(&req
->timeout
.list
, entry
);
5553 data
->timer
.function
= io_timeout_fn
;
5554 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5555 spin_unlock_irq(&ctx
->completion_lock
);
5559 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5561 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5563 return req
->user_data
== (unsigned long) data
;
5566 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5568 enum io_wq_cancel cancel_ret
;
5571 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5572 switch (cancel_ret
) {
5573 case IO_WQ_CANCEL_OK
:
5576 case IO_WQ_CANCEL_RUNNING
:
5579 case IO_WQ_CANCEL_NOTFOUND
:
5587 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5588 struct io_kiocb
*req
, __u64 sqe_addr
,
5591 unsigned long flags
;
5594 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5595 if (ret
!= -ENOENT
) {
5596 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5600 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5601 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5604 ret
= io_poll_cancel(ctx
, sqe_addr
);
5608 io_cqring_fill_event(req
, ret
);
5609 io_commit_cqring(ctx
);
5610 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5611 io_cqring_ev_posted(ctx
);
5614 req_set_fail_links(req
);
5618 static int io_async_cancel_prep(struct io_kiocb
*req
,
5619 const struct io_uring_sqe
*sqe
)
5621 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5623 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5625 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5628 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5632 static int io_async_cancel(struct io_kiocb
*req
)
5634 struct io_ring_ctx
*ctx
= req
->ctx
;
5636 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
5640 static int io_files_update_prep(struct io_kiocb
*req
,
5641 const struct io_uring_sqe
*sqe
)
5643 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
5645 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5647 if (sqe
->ioprio
|| sqe
->rw_flags
)
5650 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
5651 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
5652 if (!req
->files_update
.nr_args
)
5654 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
5658 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
,
5659 struct io_comp_state
*cs
)
5661 struct io_ring_ctx
*ctx
= req
->ctx
;
5662 struct io_uring_files_update up
;
5668 up
.offset
= req
->files_update
.offset
;
5669 up
.fds
= req
->files_update
.arg
;
5671 mutex_lock(&ctx
->uring_lock
);
5672 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
5673 mutex_unlock(&ctx
->uring_lock
);
5676 req_set_fail_links(req
);
5677 __io_req_complete(req
, ret
, 0, cs
);
5681 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5683 switch (req
->opcode
) {
5686 case IORING_OP_READV
:
5687 case IORING_OP_READ_FIXED
:
5688 case IORING_OP_READ
:
5689 return io_read_prep(req
, sqe
);
5690 case IORING_OP_WRITEV
:
5691 case IORING_OP_WRITE_FIXED
:
5692 case IORING_OP_WRITE
:
5693 return io_write_prep(req
, sqe
);
5694 case IORING_OP_POLL_ADD
:
5695 return io_poll_add_prep(req
, sqe
);
5696 case IORING_OP_POLL_REMOVE
:
5697 return io_poll_remove_prep(req
, sqe
);
5698 case IORING_OP_FSYNC
:
5699 return io_prep_fsync(req
, sqe
);
5700 case IORING_OP_SYNC_FILE_RANGE
:
5701 return io_prep_sfr(req
, sqe
);
5702 case IORING_OP_SENDMSG
:
5703 case IORING_OP_SEND
:
5704 return io_sendmsg_prep(req
, sqe
);
5705 case IORING_OP_RECVMSG
:
5706 case IORING_OP_RECV
:
5707 return io_recvmsg_prep(req
, sqe
);
5708 case IORING_OP_CONNECT
:
5709 return io_connect_prep(req
, sqe
);
5710 case IORING_OP_TIMEOUT
:
5711 return io_timeout_prep(req
, sqe
, false);
5712 case IORING_OP_TIMEOUT_REMOVE
:
5713 return io_timeout_remove_prep(req
, sqe
);
5714 case IORING_OP_ASYNC_CANCEL
:
5715 return io_async_cancel_prep(req
, sqe
);
5716 case IORING_OP_LINK_TIMEOUT
:
5717 return io_timeout_prep(req
, sqe
, true);
5718 case IORING_OP_ACCEPT
:
5719 return io_accept_prep(req
, sqe
);
5720 case IORING_OP_FALLOCATE
:
5721 return io_fallocate_prep(req
, sqe
);
5722 case IORING_OP_OPENAT
:
5723 return io_openat_prep(req
, sqe
);
5724 case IORING_OP_CLOSE
:
5725 return io_close_prep(req
, sqe
);
5726 case IORING_OP_FILES_UPDATE
:
5727 return io_files_update_prep(req
, sqe
);
5728 case IORING_OP_STATX
:
5729 return io_statx_prep(req
, sqe
);
5730 case IORING_OP_FADVISE
:
5731 return io_fadvise_prep(req
, sqe
);
5732 case IORING_OP_MADVISE
:
5733 return io_madvise_prep(req
, sqe
);
5734 case IORING_OP_OPENAT2
:
5735 return io_openat2_prep(req
, sqe
);
5736 case IORING_OP_EPOLL_CTL
:
5737 return io_epoll_ctl_prep(req
, sqe
);
5738 case IORING_OP_SPLICE
:
5739 return io_splice_prep(req
, sqe
);
5740 case IORING_OP_PROVIDE_BUFFERS
:
5741 return io_provide_buffers_prep(req
, sqe
);
5742 case IORING_OP_REMOVE_BUFFERS
:
5743 return io_remove_buffers_prep(req
, sqe
);
5745 return io_tee_prep(req
, sqe
);
5748 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5753 static int io_req_defer_prep(struct io_kiocb
*req
,
5754 const struct io_uring_sqe
*sqe
)
5758 if (io_alloc_async_data(req
))
5760 return io_req_prep(req
, sqe
);
5763 static u32
io_get_sequence(struct io_kiocb
*req
)
5765 struct io_kiocb
*pos
;
5766 struct io_ring_ctx
*ctx
= req
->ctx
;
5767 u32 total_submitted
, nr_reqs
= 1;
5769 if (req
->flags
& REQ_F_LINK_HEAD
)
5770 list_for_each_entry(pos
, &req
->link_list
, link_list
)
5773 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
5774 return total_submitted
- nr_reqs
;
5777 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5779 struct io_ring_ctx
*ctx
= req
->ctx
;
5780 struct io_defer_entry
*de
;
5784 /* Still need defer if there is pending req in defer list. */
5785 if (likely(list_empty_careful(&ctx
->defer_list
) &&
5786 !(req
->flags
& REQ_F_IO_DRAIN
)))
5789 seq
= io_get_sequence(req
);
5790 /* Still a chance to pass the sequence check */
5791 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
5794 if (!req
->async_data
) {
5795 ret
= io_req_defer_prep(req
, sqe
);
5799 io_prep_async_link(req
);
5800 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
5804 spin_lock_irq(&ctx
->completion_lock
);
5805 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
5806 spin_unlock_irq(&ctx
->completion_lock
);
5808 io_queue_async_work(req
);
5809 return -EIOCBQUEUED
;
5812 trace_io_uring_defer(ctx
, req
, req
->user_data
);
5815 list_add_tail(&de
->list
, &ctx
->defer_list
);
5816 spin_unlock_irq(&ctx
->completion_lock
);
5817 return -EIOCBQUEUED
;
5820 static void io_req_drop_files(struct io_kiocb
*req
)
5822 struct io_ring_ctx
*ctx
= req
->ctx
;
5823 unsigned long flags
;
5825 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
5826 list_del(&req
->inflight_entry
);
5827 if (waitqueue_active(&ctx
->inflight_wait
))
5828 wake_up(&ctx
->inflight_wait
);
5829 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
5830 req
->flags
&= ~REQ_F_INFLIGHT
;
5831 put_files_struct(req
->work
.identity
->files
);
5832 put_nsproxy(req
->work
.identity
->nsproxy
);
5833 req
->work
.flags
&= ~IO_WQ_WORK_FILES
;
5836 static void __io_clean_op(struct io_kiocb
*req
)
5838 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
5839 switch (req
->opcode
) {
5840 case IORING_OP_READV
:
5841 case IORING_OP_READ_FIXED
:
5842 case IORING_OP_READ
:
5843 kfree((void *)(unsigned long)req
->rw
.addr
);
5845 case IORING_OP_RECVMSG
:
5846 case IORING_OP_RECV
:
5847 kfree(req
->sr_msg
.kbuf
);
5850 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
5853 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
5854 switch (req
->opcode
) {
5855 case IORING_OP_READV
:
5856 case IORING_OP_READ_FIXED
:
5857 case IORING_OP_READ
:
5858 case IORING_OP_WRITEV
:
5859 case IORING_OP_WRITE_FIXED
:
5860 case IORING_OP_WRITE
: {
5861 struct io_async_rw
*io
= req
->async_data
;
5863 kfree(io
->free_iovec
);
5866 case IORING_OP_RECVMSG
:
5867 case IORING_OP_SENDMSG
: {
5868 struct io_async_msghdr
*io
= req
->async_data
;
5869 if (io
->iov
!= io
->fast_iov
)
5873 case IORING_OP_SPLICE
:
5875 io_put_file(req
, req
->splice
.file_in
,
5876 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
5878 case IORING_OP_OPENAT
:
5879 case IORING_OP_OPENAT2
:
5880 if (req
->open
.filename
)
5881 putname(req
->open
.filename
);
5884 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5887 if (req
->flags
& REQ_F_INFLIGHT
)
5888 io_req_drop_files(req
);
5891 static int io_issue_sqe(struct io_kiocb
*req
, bool force_nonblock
,
5892 struct io_comp_state
*cs
)
5894 struct io_ring_ctx
*ctx
= req
->ctx
;
5897 switch (req
->opcode
) {
5899 ret
= io_nop(req
, cs
);
5901 case IORING_OP_READV
:
5902 case IORING_OP_READ_FIXED
:
5903 case IORING_OP_READ
:
5904 ret
= io_read(req
, force_nonblock
, cs
);
5906 case IORING_OP_WRITEV
:
5907 case IORING_OP_WRITE_FIXED
:
5908 case IORING_OP_WRITE
:
5909 ret
= io_write(req
, force_nonblock
, cs
);
5911 case IORING_OP_FSYNC
:
5912 ret
= io_fsync(req
, force_nonblock
);
5914 case IORING_OP_POLL_ADD
:
5915 ret
= io_poll_add(req
);
5917 case IORING_OP_POLL_REMOVE
:
5918 ret
= io_poll_remove(req
);
5920 case IORING_OP_SYNC_FILE_RANGE
:
5921 ret
= io_sync_file_range(req
, force_nonblock
);
5923 case IORING_OP_SENDMSG
:
5924 ret
= io_sendmsg(req
, force_nonblock
, cs
);
5926 case IORING_OP_SEND
:
5927 ret
= io_send(req
, force_nonblock
, cs
);
5929 case IORING_OP_RECVMSG
:
5930 ret
= io_recvmsg(req
, force_nonblock
, cs
);
5932 case IORING_OP_RECV
:
5933 ret
= io_recv(req
, force_nonblock
, cs
);
5935 case IORING_OP_TIMEOUT
:
5936 ret
= io_timeout(req
);
5938 case IORING_OP_TIMEOUT_REMOVE
:
5939 ret
= io_timeout_remove(req
);
5941 case IORING_OP_ACCEPT
:
5942 ret
= io_accept(req
, force_nonblock
, cs
);
5944 case IORING_OP_CONNECT
:
5945 ret
= io_connect(req
, force_nonblock
, cs
);
5947 case IORING_OP_ASYNC_CANCEL
:
5948 ret
= io_async_cancel(req
);
5950 case IORING_OP_FALLOCATE
:
5951 ret
= io_fallocate(req
, force_nonblock
);
5953 case IORING_OP_OPENAT
:
5954 ret
= io_openat(req
, force_nonblock
);
5956 case IORING_OP_CLOSE
:
5957 ret
= io_close(req
, force_nonblock
, cs
);
5959 case IORING_OP_FILES_UPDATE
:
5960 ret
= io_files_update(req
, force_nonblock
, cs
);
5962 case IORING_OP_STATX
:
5963 ret
= io_statx(req
, force_nonblock
);
5965 case IORING_OP_FADVISE
:
5966 ret
= io_fadvise(req
, force_nonblock
);
5968 case IORING_OP_MADVISE
:
5969 ret
= io_madvise(req
, force_nonblock
);
5971 case IORING_OP_OPENAT2
:
5972 ret
= io_openat2(req
, force_nonblock
);
5974 case IORING_OP_EPOLL_CTL
:
5975 ret
= io_epoll_ctl(req
, force_nonblock
, cs
);
5977 case IORING_OP_SPLICE
:
5978 ret
= io_splice(req
, force_nonblock
);
5980 case IORING_OP_PROVIDE_BUFFERS
:
5981 ret
= io_provide_buffers(req
, force_nonblock
, cs
);
5983 case IORING_OP_REMOVE_BUFFERS
:
5984 ret
= io_remove_buffers(req
, force_nonblock
, cs
);
5987 ret
= io_tee(req
, force_nonblock
);
5997 /* If the op doesn't have a file, we're not polling for it */
5998 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
5999 const bool in_async
= io_wq_current_is_worker();
6001 /* workqueue context doesn't hold uring_lock, grab it now */
6003 mutex_lock(&ctx
->uring_lock
);
6005 io_iopoll_req_issued(req
);
6008 mutex_unlock(&ctx
->uring_lock
);
6014 static struct io_wq_work
*io_wq_submit_work(struct io_wq_work
*work
)
6016 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6017 struct io_kiocb
*timeout
;
6020 timeout
= io_prep_linked_timeout(req
);
6022 io_queue_linked_timeout(timeout
);
6024 /* if NO_CANCEL is set, we must still run the work */
6025 if ((work
->flags
& (IO_WQ_WORK_CANCEL
|IO_WQ_WORK_NO_CANCEL
)) ==
6026 IO_WQ_WORK_CANCEL
) {
6032 ret
= io_issue_sqe(req
, false, NULL
);
6034 * We can get EAGAIN for polled IO even though we're
6035 * forcing a sync submission from here, since we can't
6036 * wait for request slots on the block side.
6045 req_set_fail_links(req
);
6046 io_req_complete(req
, ret
);
6049 return io_steal_work(req
);
6052 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6055 struct fixed_file_table
*table
;
6057 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6058 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6061 static struct file
*io_file_get(struct io_submit_state
*state
,
6062 struct io_kiocb
*req
, int fd
, bool fixed
)
6064 struct io_ring_ctx
*ctx
= req
->ctx
;
6068 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6070 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6071 file
= io_file_from_index(ctx
, fd
);
6073 req
->fixed_file_refs
= &ctx
->file_data
->node
->refs
;
6074 percpu_ref_get(req
->fixed_file_refs
);
6077 trace_io_uring_file_get(ctx
, fd
);
6078 file
= __io_file_get(state
, fd
);
6084 static int io_req_set_file(struct io_submit_state
*state
, struct io_kiocb
*req
,
6089 fixed
= (req
->flags
& REQ_F_FIXED_FILE
) != 0;
6090 if (unlikely(!fixed
&& io_async_submit(req
->ctx
)))
6093 req
->file
= io_file_get(state
, req
, fd
, fixed
);
6094 if (req
->file
|| io_op_defs
[req
->opcode
].needs_file_no_error
)
6099 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6101 struct io_timeout_data
*data
= container_of(timer
,
6102 struct io_timeout_data
, timer
);
6103 struct io_kiocb
*req
= data
->req
;
6104 struct io_ring_ctx
*ctx
= req
->ctx
;
6105 struct io_kiocb
*prev
= NULL
;
6106 unsigned long flags
;
6108 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6111 * We don't expect the list to be empty, that will only happen if we
6112 * race with the completion of the linked work.
6114 if (!list_empty(&req
->link_list
)) {
6115 prev
= list_entry(req
->link_list
.prev
, struct io_kiocb
,
6117 if (refcount_inc_not_zero(&prev
->refs
))
6118 list_del_init(&req
->link_list
);
6123 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6126 req_set_fail_links(prev
);
6127 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6130 io_req_complete(req
, -ETIME
);
6132 return HRTIMER_NORESTART
;
6135 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6138 * If the list is now empty, then our linked request finished before
6139 * we got a chance to setup the timer
6141 if (!list_empty(&req
->link_list
)) {
6142 struct io_timeout_data
*data
= req
->async_data
;
6144 data
->timer
.function
= io_link_timeout_fn
;
6145 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6150 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6152 struct io_ring_ctx
*ctx
= req
->ctx
;
6154 spin_lock_irq(&ctx
->completion_lock
);
6155 __io_queue_linked_timeout(req
);
6156 spin_unlock_irq(&ctx
->completion_lock
);
6158 /* drop submission reference */
6162 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6164 struct io_kiocb
*nxt
;
6166 if (!(req
->flags
& REQ_F_LINK_HEAD
))
6168 if (req
->flags
& REQ_F_LINK_TIMEOUT
)
6171 nxt
= list_first_entry_or_null(&req
->link_list
, struct io_kiocb
,
6173 if (!nxt
|| nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6176 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6177 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6181 static void __io_queue_sqe(struct io_kiocb
*req
, struct io_comp_state
*cs
)
6183 struct io_kiocb
*linked_timeout
;
6184 struct io_kiocb
*nxt
;
6185 const struct cred
*old_creds
= NULL
;
6189 linked_timeout
= io_prep_linked_timeout(req
);
6191 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) &&
6192 (req
->work
.flags
& IO_WQ_WORK_CREDS
) &&
6193 req
->work
.identity
->creds
!= current_cred()) {
6195 revert_creds(old_creds
);
6196 if (old_creds
== req
->work
.identity
->creds
)
6197 old_creds
= NULL
; /* restored original creds */
6199 old_creds
= override_creds(req
->work
.identity
->creds
);
6202 ret
= io_issue_sqe(req
, true, cs
);
6205 * We async punt it if the file wasn't marked NOWAIT, or if the file
6206 * doesn't support non-blocking read/write attempts
6208 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6209 if (!io_arm_poll_handler(req
)) {
6212 * Queued up for async execution, worker will release
6213 * submit reference when the iocb is actually submitted.
6215 io_queue_async_work(req
);
6219 io_queue_linked_timeout(linked_timeout
);
6223 if (unlikely(ret
)) {
6224 /* un-prep timeout, so it'll be killed as any other linked */
6225 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
6226 req_set_fail_links(req
);
6228 io_req_complete(req
, ret
);
6232 /* drop submission reference */
6233 nxt
= io_put_req_find_next(req
);
6235 io_queue_linked_timeout(linked_timeout
);
6240 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6241 linked_timeout
= NULL
;
6248 revert_creds(old_creds
);
6251 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6252 struct io_comp_state
*cs
)
6256 ret
= io_req_defer(req
, sqe
);
6258 if (ret
!= -EIOCBQUEUED
) {
6260 req_set_fail_links(req
);
6262 io_req_complete(req
, ret
);
6264 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6265 if (!req
->async_data
) {
6266 ret
= io_req_defer_prep(req
, sqe
);
6272 * Never try inline submit of IOSQE_ASYNC is set, go straight
6273 * to async execution.
6275 io_req_init_async(req
);
6276 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
6277 io_queue_async_work(req
);
6280 ret
= io_req_prep(req
, sqe
);
6284 __io_queue_sqe(req
, cs
);
6288 static inline void io_queue_link_head(struct io_kiocb
*req
,
6289 struct io_comp_state
*cs
)
6291 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
6293 io_req_complete(req
, -ECANCELED
);
6295 io_queue_sqe(req
, NULL
, cs
);
6298 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6299 struct io_kiocb
**link
, struct io_comp_state
*cs
)
6301 struct io_ring_ctx
*ctx
= req
->ctx
;
6305 * If we already have a head request, queue this one for async
6306 * submittal once the head completes. If we don't have a head but
6307 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6308 * submitted sync once the chain is complete. If none of those
6309 * conditions are true (normal request), then just queue it.
6312 struct io_kiocb
*head
= *link
;
6315 * Taking sequential execution of a link, draining both sides
6316 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6317 * requests in the link. So, it drains the head and the
6318 * next after the link request. The last one is done via
6319 * drain_next flag to persist the effect across calls.
6321 if (req
->flags
& REQ_F_IO_DRAIN
) {
6322 head
->flags
|= REQ_F_IO_DRAIN
;
6323 ctx
->drain_next
= 1;
6325 ret
= io_req_defer_prep(req
, sqe
);
6326 if (unlikely(ret
)) {
6327 /* fail even hard links since we don't submit */
6328 head
->flags
|= REQ_F_FAIL_LINK
;
6331 trace_io_uring_link(ctx
, req
, head
);
6332 list_add_tail(&req
->link_list
, &head
->link_list
);
6334 /* last request of a link, enqueue the link */
6335 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6336 io_queue_link_head(head
, cs
);
6340 if (unlikely(ctx
->drain_next
)) {
6341 req
->flags
|= REQ_F_IO_DRAIN
;
6342 ctx
->drain_next
= 0;
6344 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6345 req
->flags
|= REQ_F_LINK_HEAD
;
6346 INIT_LIST_HEAD(&req
->link_list
);
6348 ret
= io_req_defer_prep(req
, sqe
);
6350 req
->flags
|= REQ_F_FAIL_LINK
;
6353 io_queue_sqe(req
, sqe
, cs
);
6361 * Batched submission is done, ensure local IO is flushed out.
6363 static void io_submit_state_end(struct io_submit_state
*state
)
6365 if (!list_empty(&state
->comp
.list
))
6366 io_submit_flush_completions(&state
->comp
);
6367 blk_finish_plug(&state
->plug
);
6368 io_state_file_put(state
);
6369 if (state
->free_reqs
)
6370 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6374 * Start submission side cache.
6376 static void io_submit_state_start(struct io_submit_state
*state
,
6377 struct io_ring_ctx
*ctx
, unsigned int max_ios
)
6379 blk_start_plug(&state
->plug
);
6381 INIT_LIST_HEAD(&state
->comp
.list
);
6382 state
->comp
.ctx
= ctx
;
6383 state
->free_reqs
= 0;
6385 state
->ios_left
= max_ios
;
6388 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6390 struct io_rings
*rings
= ctx
->rings
;
6393 * Ensure any loads from the SQEs are done at this point,
6394 * since once we write the new head, the application could
6395 * write new data to them.
6397 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6401 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6402 * that is mapped by userspace. This means that care needs to be taken to
6403 * ensure that reads are stable, as we cannot rely on userspace always
6404 * being a good citizen. If members of the sqe are validated and then later
6405 * used, it's important that those reads are done through READ_ONCE() to
6406 * prevent a re-load down the line.
6408 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6410 u32
*sq_array
= ctx
->sq_array
;
6414 * The cached sq head (or cq tail) serves two purposes:
6416 * 1) allows us to batch the cost of updating the user visible
6418 * 2) allows the kernel side to track the head on its own, even
6419 * though the application is the one updating it.
6421 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6422 if (likely(head
< ctx
->sq_entries
))
6423 return &ctx
->sq_sqes
[head
];
6425 /* drop invalid entries */
6426 ctx
->cached_sq_dropped
++;
6427 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6431 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6433 ctx
->cached_sq_head
++;
6437 * Check SQE restrictions (opcode and flags).
6439 * Returns 'true' if SQE is allowed, 'false' otherwise.
6441 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6442 struct io_kiocb
*req
,
6443 unsigned int sqe_flags
)
6445 if (!ctx
->restricted
)
6448 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6451 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6452 ctx
->restrictions
.sqe_flags_required
)
6455 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6456 ctx
->restrictions
.sqe_flags_required
))
6462 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6463 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6464 IOSQE_BUFFER_SELECT)
6466 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6467 const struct io_uring_sqe
*sqe
,
6468 struct io_submit_state
*state
)
6470 unsigned int sqe_flags
;
6473 req
->opcode
= READ_ONCE(sqe
->opcode
);
6474 req
->user_data
= READ_ONCE(sqe
->user_data
);
6475 req
->async_data
= NULL
;
6479 /* one is dropped after submission, the other at completion */
6480 refcount_set(&req
->refs
, 2);
6481 req
->task
= current
;
6484 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6487 if (unlikely(io_sq_thread_acquire_mm(ctx
, req
)))
6490 sqe_flags
= READ_ONCE(sqe
->flags
);
6491 /* enforce forwards compatibility on users */
6492 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6495 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6498 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6499 !io_op_defs
[req
->opcode
].buffer_select
)
6502 id
= READ_ONCE(sqe
->personality
);
6504 struct io_identity
*iod
;
6506 iod
= idr_find(&ctx
->personality_idr
, id
);
6509 refcount_inc(&iod
->count
);
6511 __io_req_init_async(req
);
6512 get_cred(iod
->creds
);
6513 req
->work
.identity
= iod
;
6514 req
->work
.flags
|= IO_WQ_WORK_CREDS
;
6517 /* same numerical values with corresponding REQ_F_*, safe to copy */
6518 req
->flags
|= sqe_flags
;
6520 if (!io_op_defs
[req
->opcode
].needs_file
)
6523 ret
= io_req_set_file(state
, req
, READ_ONCE(sqe
->fd
));
6528 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6530 struct io_submit_state state
;
6531 struct io_kiocb
*link
= NULL
;
6532 int i
, submitted
= 0;
6534 /* if we have a backlog and couldn't flush it all, return BUSY */
6535 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6536 if (!list_empty(&ctx
->cq_overflow_list
) &&
6537 !io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6541 /* make sure SQ entry isn't read before tail */
6542 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6544 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6547 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6548 refcount_add(nr
, ¤t
->usage
);
6550 io_submit_state_start(&state
, ctx
, nr
);
6552 for (i
= 0; i
< nr
; i
++) {
6553 const struct io_uring_sqe
*sqe
;
6554 struct io_kiocb
*req
;
6557 sqe
= io_get_sqe(ctx
);
6558 if (unlikely(!sqe
)) {
6559 io_consume_sqe(ctx
);
6562 req
= io_alloc_req(ctx
, &state
);
6563 if (unlikely(!req
)) {
6565 submitted
= -EAGAIN
;
6568 io_consume_sqe(ctx
);
6569 /* will complete beyond this point, count as submitted */
6572 err
= io_init_req(ctx
, req
, sqe
, &state
);
6573 if (unlikely(err
)) {
6576 io_req_complete(req
, err
);
6580 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6581 true, io_async_submit(ctx
));
6582 err
= io_submit_sqe(req
, sqe
, &link
, &state
.comp
);
6587 if (unlikely(submitted
!= nr
)) {
6588 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6589 struct io_uring_task
*tctx
= current
->io_uring
;
6590 int unused
= nr
- ref_used
;
6592 percpu_ref_put_many(&ctx
->refs
, unused
);
6593 percpu_counter_sub(&tctx
->inflight
, unused
);
6594 put_task_struct_many(current
, unused
);
6597 io_queue_link_head(link
, &state
.comp
);
6598 io_submit_state_end(&state
);
6600 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6601 io_commit_sqring(ctx
);
6606 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6608 /* Tell userspace we may need a wakeup call */
6609 spin_lock_irq(&ctx
->completion_lock
);
6610 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6611 spin_unlock_irq(&ctx
->completion_lock
);
6614 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6616 spin_lock_irq(&ctx
->completion_lock
);
6617 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6618 spin_unlock_irq(&ctx
->completion_lock
);
6621 static int io_sq_wake_function(struct wait_queue_entry
*wqe
, unsigned mode
,
6622 int sync
, void *key
)
6624 struct io_ring_ctx
*ctx
= container_of(wqe
, struct io_ring_ctx
, sqo_wait_entry
);
6627 ret
= autoremove_wake_function(wqe
, mode
, sync
, key
);
6629 unsigned long flags
;
6631 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6632 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6633 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6644 static enum sq_ret
__io_sq_thread(struct io_ring_ctx
*ctx
,
6645 unsigned long start_jiffies
, bool cap_entries
)
6647 unsigned long timeout
= start_jiffies
+ ctx
->sq_thread_idle
;
6648 struct io_sq_data
*sqd
= ctx
->sq_data
;
6649 unsigned int to_submit
;
6653 if (!list_empty(&ctx
->iopoll_list
)) {
6654 unsigned nr_events
= 0;
6656 mutex_lock(&ctx
->uring_lock
);
6657 if (!list_empty(&ctx
->iopoll_list
) && !need_resched())
6658 io_do_iopoll(ctx
, &nr_events
, 0);
6659 mutex_unlock(&ctx
->uring_lock
);
6662 to_submit
= io_sqring_entries(ctx
);
6665 * If submit got -EBUSY, flag us as needing the application
6666 * to enter the kernel to reap and flush events.
6668 if (!to_submit
|| ret
== -EBUSY
|| need_resched()) {
6670 * Drop cur_mm before scheduling, we can't hold it for
6671 * long periods (or over schedule()). Do this before
6672 * adding ourselves to the waitqueue, as the unuse/drop
6675 io_sq_thread_drop_mm();
6678 * We're polling. If we're within the defined idle
6679 * period, then let us spin without work before going
6680 * to sleep. The exception is if we got EBUSY doing
6681 * more IO, we should wait for the application to
6682 * reap events and wake us up.
6684 if (!list_empty(&ctx
->iopoll_list
) || need_resched() ||
6685 (!time_after(jiffies
, timeout
) && ret
!= -EBUSY
&&
6686 !percpu_ref_is_dying(&ctx
->refs
)))
6689 prepare_to_wait(&sqd
->wait
, &ctx
->sqo_wait_entry
,
6690 TASK_INTERRUPTIBLE
);
6693 * While doing polled IO, before going to sleep, we need
6694 * to check if there are new reqs added to iopoll_list,
6695 * it is because reqs may have been punted to io worker
6696 * and will be added to iopoll_list later, hence check
6697 * the iopoll_list again.
6699 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6700 !list_empty_careful(&ctx
->iopoll_list
)) {
6701 finish_wait(&sqd
->wait
, &ctx
->sqo_wait_entry
);
6705 to_submit
= io_sqring_entries(ctx
);
6706 if (!to_submit
|| ret
== -EBUSY
)
6710 finish_wait(&sqd
->wait
, &ctx
->sqo_wait_entry
);
6711 io_ring_clear_wakeup_flag(ctx
);
6713 /* if we're handling multiple rings, cap submit size for fairness */
6714 if (cap_entries
&& to_submit
> 8)
6717 mutex_lock(&ctx
->uring_lock
);
6718 if (likely(!percpu_ref_is_dying(&ctx
->refs
)))
6719 ret
= io_submit_sqes(ctx
, to_submit
);
6720 mutex_unlock(&ctx
->uring_lock
);
6722 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
6723 wake_up(&ctx
->sqo_sq_wait
);
6725 return SQT_DID_WORK
;
6728 static void io_sqd_init_new(struct io_sq_data
*sqd
)
6730 struct io_ring_ctx
*ctx
;
6732 while (!list_empty(&sqd
->ctx_new_list
)) {
6733 ctx
= list_first_entry(&sqd
->ctx_new_list
, struct io_ring_ctx
, sqd_list
);
6734 init_wait(&ctx
->sqo_wait_entry
);
6735 ctx
->sqo_wait_entry
.func
= io_sq_wake_function
;
6736 list_move_tail(&ctx
->sqd_list
, &sqd
->ctx_list
);
6737 complete(&ctx
->sq_thread_comp
);
6741 static int io_sq_thread(void *data
)
6743 struct cgroup_subsys_state
*cur_css
= NULL
;
6744 const struct cred
*old_cred
= NULL
;
6745 struct io_sq_data
*sqd
= data
;
6746 struct io_ring_ctx
*ctx
;
6747 unsigned long start_jiffies
;
6749 start_jiffies
= jiffies
;
6750 while (!kthread_should_stop()) {
6751 enum sq_ret ret
= 0;
6755 * Any changes to the sqd lists are synchronized through the
6756 * kthread parking. This synchronizes the thread vs users,
6757 * the users are synchronized on the sqd->ctx_lock.
6759 if (kthread_should_park())
6762 if (unlikely(!list_empty(&sqd
->ctx_new_list
)))
6763 io_sqd_init_new(sqd
);
6765 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
6767 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6768 if (current
->cred
!= ctx
->creds
) {
6770 revert_creds(old_cred
);
6771 old_cred
= override_creds(ctx
->creds
);
6773 io_sq_thread_associate_blkcg(ctx
, &cur_css
);
6775 current
->loginuid
= ctx
->loginuid
;
6776 current
->sessionid
= ctx
->sessionid
;
6779 ret
|= __io_sq_thread(ctx
, start_jiffies
, cap_entries
);
6781 io_sq_thread_drop_mm();
6784 if (ret
& SQT_SPIN
) {
6787 } else if (ret
== SQT_IDLE
) {
6788 if (kthread_should_park())
6790 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6791 io_ring_set_wakeup_flag(ctx
);
6793 start_jiffies
= jiffies
;
6794 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6795 io_ring_clear_wakeup_flag(ctx
);
6802 io_sq_thread_unassociate_blkcg();
6804 revert_creds(old_cred
);
6811 struct io_wait_queue
{
6812 struct wait_queue_entry wq
;
6813 struct io_ring_ctx
*ctx
;
6815 unsigned nr_timeouts
;
6818 static inline bool io_should_wake(struct io_wait_queue
*iowq
, bool noflush
)
6820 struct io_ring_ctx
*ctx
= iowq
->ctx
;
6823 * Wake up if we have enough events, or if a timeout occurred since we
6824 * started waiting. For timeouts, we always want to return to userspace,
6825 * regardless of event count.
6827 return io_cqring_events(ctx
, noflush
) >= iowq
->to_wait
||
6828 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
6831 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
6832 int wake_flags
, void *key
)
6834 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
6837 /* use noflush == true, as we can't safely rely on locking context */
6838 if (!io_should_wake(iowq
, true))
6841 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
6844 static int io_run_task_work_sig(void)
6846 if (io_run_task_work())
6848 if (!signal_pending(current
))
6850 if (current
->jobctl
& JOBCTL_TASK_WORK
) {
6851 spin_lock_irq(¤t
->sighand
->siglock
);
6852 current
->jobctl
&= ~JOBCTL_TASK_WORK
;
6853 recalc_sigpending();
6854 spin_unlock_irq(¤t
->sighand
->siglock
);
6861 * Wait until events become available, if we don't already have some. The
6862 * application must reap them itself, as they reside on the shared cq ring.
6864 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
6865 const sigset_t __user
*sig
, size_t sigsz
)
6867 struct io_wait_queue iowq
= {
6870 .func
= io_wake_function
,
6871 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
6874 .to_wait
= min_events
,
6876 struct io_rings
*rings
= ctx
->rings
;
6880 if (io_cqring_events(ctx
, false) >= min_events
)
6882 if (!io_run_task_work())
6887 #ifdef CONFIG_COMPAT
6888 if (in_compat_syscall())
6889 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
6893 ret
= set_user_sigmask(sig
, sigsz
);
6899 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
6900 trace_io_uring_cqring_wait(ctx
, min_events
);
6902 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
6903 TASK_INTERRUPTIBLE
);
6904 /* make sure we run task_work before checking for signals */
6905 ret
= io_run_task_work_sig();
6910 if (io_should_wake(&iowq
, false))
6914 finish_wait(&ctx
->wait
, &iowq
.wq
);
6916 restore_saved_sigmask_unless(ret
== -EINTR
);
6918 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
6921 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6923 #if defined(CONFIG_UNIX)
6924 if (ctx
->ring_sock
) {
6925 struct sock
*sock
= ctx
->ring_sock
->sk
;
6926 struct sk_buff
*skb
;
6928 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
6934 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6937 file
= io_file_from_index(ctx
, i
);
6944 static void io_file_ref_kill(struct percpu_ref
*ref
)
6946 struct fixed_file_data
*data
;
6948 data
= container_of(ref
, struct fixed_file_data
, refs
);
6949 complete(&data
->done
);
6952 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6954 struct fixed_file_data
*data
= ctx
->file_data
;
6955 struct fixed_file_ref_node
*ref_node
= NULL
;
6956 unsigned nr_tables
, i
;
6961 spin_lock(&data
->lock
);
6962 if (!list_empty(&data
->ref_list
))
6963 ref_node
= list_first_entry(&data
->ref_list
,
6964 struct fixed_file_ref_node
, node
);
6965 spin_unlock(&data
->lock
);
6967 percpu_ref_kill(&ref_node
->refs
);
6969 percpu_ref_kill(&data
->refs
);
6971 /* wait for all refs nodes to complete */
6972 flush_delayed_work(&ctx
->file_put_work
);
6973 wait_for_completion(&data
->done
);
6975 __io_sqe_files_unregister(ctx
);
6976 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
6977 for (i
= 0; i
< nr_tables
; i
++)
6978 kfree(data
->table
[i
].files
);
6980 percpu_ref_exit(&data
->refs
);
6982 ctx
->file_data
= NULL
;
6983 ctx
->nr_user_files
= 0;
6987 static void io_put_sq_data(struct io_sq_data
*sqd
)
6989 if (refcount_dec_and_test(&sqd
->refs
)) {
6991 * The park is a bit of a work-around, without it we get
6992 * warning spews on shutdown with SQPOLL set and affinity
6993 * set to a single CPU.
6996 kthread_park(sqd
->thread
);
6997 kthread_stop(sqd
->thread
);
7004 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7006 struct io_ring_ctx
*ctx_attach
;
7007 struct io_sq_data
*sqd
;
7010 f
= fdget(p
->wq_fd
);
7012 return ERR_PTR(-ENXIO
);
7013 if (f
.file
->f_op
!= &io_uring_fops
) {
7015 return ERR_PTR(-EINVAL
);
7018 ctx_attach
= f
.file
->private_data
;
7019 sqd
= ctx_attach
->sq_data
;
7022 return ERR_PTR(-EINVAL
);
7025 refcount_inc(&sqd
->refs
);
7030 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
)
7032 struct io_sq_data
*sqd
;
7034 if (p
->flags
& IORING_SETUP_ATTACH_WQ
)
7035 return io_attach_sq_data(p
);
7037 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7039 return ERR_PTR(-ENOMEM
);
7041 refcount_set(&sqd
->refs
, 1);
7042 INIT_LIST_HEAD(&sqd
->ctx_list
);
7043 INIT_LIST_HEAD(&sqd
->ctx_new_list
);
7044 mutex_init(&sqd
->ctx_lock
);
7045 mutex_init(&sqd
->lock
);
7046 init_waitqueue_head(&sqd
->wait
);
7050 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7051 __releases(&sqd
->lock
)
7055 kthread_unpark(sqd
->thread
);
7056 mutex_unlock(&sqd
->lock
);
7059 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7060 __acquires(&sqd
->lock
)
7064 mutex_lock(&sqd
->lock
);
7065 kthread_park(sqd
->thread
);
7068 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
7070 struct io_sq_data
*sqd
= ctx
->sq_data
;
7075 * We may arrive here from the error branch in
7076 * io_sq_offload_create() where the kthread is created
7077 * without being waked up, thus wake it up now to make
7078 * sure the wait will complete.
7080 wake_up_process(sqd
->thread
);
7081 wait_for_completion(&ctx
->sq_thread_comp
);
7083 io_sq_thread_park(sqd
);
7086 mutex_lock(&sqd
->ctx_lock
);
7087 list_del(&ctx
->sqd_list
);
7088 mutex_unlock(&sqd
->ctx_lock
);
7091 finish_wait(&sqd
->wait
, &ctx
->sqo_wait_entry
);
7092 io_sq_thread_unpark(sqd
);
7095 io_put_sq_data(sqd
);
7096 ctx
->sq_data
= NULL
;
7100 static void io_finish_async(struct io_ring_ctx
*ctx
)
7102 io_sq_thread_stop(ctx
);
7105 io_wq_destroy(ctx
->io_wq
);
7110 #if defined(CONFIG_UNIX)
7112 * Ensure the UNIX gc is aware of our file set, so we are certain that
7113 * the io_uring can be safely unregistered on process exit, even if we have
7114 * loops in the file referencing.
7116 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7118 struct sock
*sk
= ctx
->ring_sock
->sk
;
7119 struct scm_fp_list
*fpl
;
7120 struct sk_buff
*skb
;
7123 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7127 skb
= alloc_skb(0, GFP_KERNEL
);
7136 fpl
->user
= get_uid(ctx
->user
);
7137 for (i
= 0; i
< nr
; i
++) {
7138 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7142 fpl
->fp
[nr_files
] = get_file(file
);
7143 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7148 fpl
->max
= SCM_MAX_FD
;
7149 fpl
->count
= nr_files
;
7150 UNIXCB(skb
).fp
= fpl
;
7151 skb
->destructor
= unix_destruct_scm
;
7152 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7153 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7155 for (i
= 0; i
< nr_files
; i
++)
7166 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7167 * causes regular reference counting to break down. We rely on the UNIX
7168 * garbage collection to take care of this problem for us.
7170 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7172 unsigned left
, total
;
7176 left
= ctx
->nr_user_files
;
7178 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7180 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7184 total
+= this_files
;
7190 while (total
< ctx
->nr_user_files
) {
7191 struct file
*file
= io_file_from_index(ctx
, total
);
7201 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7207 static int io_sqe_alloc_file_tables(struct fixed_file_data
*file_data
,
7208 unsigned nr_tables
, unsigned nr_files
)
7212 for (i
= 0; i
< nr_tables
; i
++) {
7213 struct fixed_file_table
*table
= &file_data
->table
[i
];
7214 unsigned this_files
;
7216 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7217 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7221 nr_files
-= this_files
;
7227 for (i
= 0; i
< nr_tables
; i
++) {
7228 struct fixed_file_table
*table
= &file_data
->table
[i
];
7229 kfree(table
->files
);
7234 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
7236 #if defined(CONFIG_UNIX)
7237 struct sock
*sock
= ctx
->ring_sock
->sk
;
7238 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7239 struct sk_buff
*skb
;
7242 __skb_queue_head_init(&list
);
7245 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7246 * remove this entry and rearrange the file array.
7248 skb
= skb_dequeue(head
);
7250 struct scm_fp_list
*fp
;
7252 fp
= UNIXCB(skb
).fp
;
7253 for (i
= 0; i
< fp
->count
; i
++) {
7256 if (fp
->fp
[i
] != file
)
7259 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7260 left
= fp
->count
- 1 - i
;
7262 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7263 left
* sizeof(struct file
*));
7270 __skb_queue_tail(&list
, skb
);
7280 __skb_queue_tail(&list
, skb
);
7282 skb
= skb_dequeue(head
);
7285 if (skb_peek(&list
)) {
7286 spin_lock_irq(&head
->lock
);
7287 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7288 __skb_queue_tail(head
, skb
);
7289 spin_unlock_irq(&head
->lock
);
7296 struct io_file_put
{
7297 struct list_head list
;
7301 static void __io_file_put_work(struct fixed_file_ref_node
*ref_node
)
7303 struct fixed_file_data
*file_data
= ref_node
->file_data
;
7304 struct io_ring_ctx
*ctx
= file_data
->ctx
;
7305 struct io_file_put
*pfile
, *tmp
;
7307 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
7308 list_del(&pfile
->list
);
7309 io_ring_file_put(ctx
, pfile
->file
);
7313 spin_lock(&file_data
->lock
);
7314 list_del(&ref_node
->node
);
7315 spin_unlock(&file_data
->lock
);
7317 percpu_ref_exit(&ref_node
->refs
);
7319 percpu_ref_put(&file_data
->refs
);
7322 static void io_file_put_work(struct work_struct
*work
)
7324 struct io_ring_ctx
*ctx
;
7325 struct llist_node
*node
;
7327 ctx
= container_of(work
, struct io_ring_ctx
, file_put_work
.work
);
7328 node
= llist_del_all(&ctx
->file_put_llist
);
7331 struct fixed_file_ref_node
*ref_node
;
7332 struct llist_node
*next
= node
->next
;
7334 ref_node
= llist_entry(node
, struct fixed_file_ref_node
, llist
);
7335 __io_file_put_work(ref_node
);
7340 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
7342 struct fixed_file_ref_node
*ref_node
;
7343 struct io_ring_ctx
*ctx
;
7347 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
7348 ctx
= ref_node
->file_data
->ctx
;
7350 if (percpu_ref_is_dying(&ctx
->file_data
->refs
))
7353 first_add
= llist_add(&ref_node
->llist
, &ctx
->file_put_llist
);
7355 mod_delayed_work(system_wq
, &ctx
->file_put_work
, 0);
7357 queue_delayed_work(system_wq
, &ctx
->file_put_work
, delay
);
7360 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
7361 struct io_ring_ctx
*ctx
)
7363 struct fixed_file_ref_node
*ref_node
;
7365 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7367 return ERR_PTR(-ENOMEM
);
7369 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
7372 return ERR_PTR(-ENOMEM
);
7374 INIT_LIST_HEAD(&ref_node
->node
);
7375 INIT_LIST_HEAD(&ref_node
->file_list
);
7376 ref_node
->file_data
= ctx
->file_data
;
7380 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
7382 percpu_ref_exit(&ref_node
->refs
);
7386 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7389 __s32 __user
*fds
= (__s32 __user
*) arg
;
7390 unsigned nr_tables
, i
;
7392 int fd
, ret
= -ENOMEM
;
7393 struct fixed_file_ref_node
*ref_node
;
7394 struct fixed_file_data
*file_data
;
7400 if (nr_args
> IORING_MAX_FIXED_FILES
)
7403 file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL
);
7406 file_data
->ctx
= ctx
;
7407 init_completion(&file_data
->done
);
7408 INIT_LIST_HEAD(&file_data
->ref_list
);
7409 spin_lock_init(&file_data
->lock
);
7411 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7412 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7414 if (!file_data
->table
)
7417 if (percpu_ref_init(&file_data
->refs
, io_file_ref_kill
,
7418 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
7421 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7423 ctx
->file_data
= file_data
;
7425 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7426 struct fixed_file_table
*table
;
7429 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7433 /* allow sparse sets */
7443 * Don't allow io_uring instances to be registered. If UNIX
7444 * isn't enabled, then this causes a reference cycle and this
7445 * instance can never get freed. If UNIX is enabled we'll
7446 * handle it just fine, but there's still no point in allowing
7447 * a ring fd as it doesn't support regular read/write anyway.
7449 if (file
->f_op
== &io_uring_fops
) {
7453 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7454 index
= i
& IORING_FILE_TABLE_MASK
;
7455 table
->files
[index
] = file
;
7458 ret
= io_sqe_files_scm(ctx
);
7460 io_sqe_files_unregister(ctx
);
7464 ref_node
= alloc_fixed_file_ref_node(ctx
);
7465 if (IS_ERR(ref_node
)) {
7466 io_sqe_files_unregister(ctx
);
7467 return PTR_ERR(ref_node
);
7470 file_data
->node
= ref_node
;
7471 spin_lock(&file_data
->lock
);
7472 list_add(&ref_node
->node
, &file_data
->ref_list
);
7473 spin_unlock(&file_data
->lock
);
7474 percpu_ref_get(&file_data
->refs
);
7477 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7478 file
= io_file_from_index(ctx
, i
);
7482 for (i
= 0; i
< nr_tables
; i
++)
7483 kfree(file_data
->table
[i
].files
);
7484 ctx
->nr_user_files
= 0;
7486 percpu_ref_exit(&file_data
->refs
);
7488 kfree(file_data
->table
);
7490 ctx
->file_data
= NULL
;
7494 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7497 #if defined(CONFIG_UNIX)
7498 struct sock
*sock
= ctx
->ring_sock
->sk
;
7499 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7500 struct sk_buff
*skb
;
7503 * See if we can merge this file into an existing skb SCM_RIGHTS
7504 * file set. If there's no room, fall back to allocating a new skb
7505 * and filling it in.
7507 spin_lock_irq(&head
->lock
);
7508 skb
= skb_peek(head
);
7510 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7512 if (fpl
->count
< SCM_MAX_FD
) {
7513 __skb_unlink(skb
, head
);
7514 spin_unlock_irq(&head
->lock
);
7515 fpl
->fp
[fpl
->count
] = get_file(file
);
7516 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7518 spin_lock_irq(&head
->lock
);
7519 __skb_queue_head(head
, skb
);
7524 spin_unlock_irq(&head
->lock
);
7531 return __io_sqe_files_scm(ctx
, 1, index
);
7537 static int io_queue_file_removal(struct fixed_file_data
*data
,
7540 struct io_file_put
*pfile
;
7541 struct fixed_file_ref_node
*ref_node
= data
->node
;
7543 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
7548 list_add(&pfile
->list
, &ref_node
->file_list
);
7553 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7554 struct io_uring_files_update
*up
,
7557 struct fixed_file_data
*data
= ctx
->file_data
;
7558 struct fixed_file_ref_node
*ref_node
;
7563 bool needs_switch
= false;
7565 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7567 if (done
> ctx
->nr_user_files
)
7570 ref_node
= alloc_fixed_file_ref_node(ctx
);
7571 if (IS_ERR(ref_node
))
7572 return PTR_ERR(ref_node
);
7575 fds
= u64_to_user_ptr(up
->fds
);
7577 struct fixed_file_table
*table
;
7581 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7585 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
7586 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7587 index
= i
& IORING_FILE_TABLE_MASK
;
7588 if (table
->files
[index
]) {
7589 file
= table
->files
[index
];
7590 err
= io_queue_file_removal(data
, file
);
7593 table
->files
[index
] = NULL
;
7594 needs_switch
= true;
7603 * Don't allow io_uring instances to be registered. If
7604 * UNIX isn't enabled, then this causes a reference
7605 * cycle and this instance can never get freed. If UNIX
7606 * is enabled we'll handle it just fine, but there's
7607 * still no point in allowing a ring fd as it doesn't
7608 * support regular read/write anyway.
7610 if (file
->f_op
== &io_uring_fops
) {
7615 table
->files
[index
] = file
;
7616 err
= io_sqe_file_register(ctx
, file
, i
);
7618 table
->files
[index
] = NULL
;
7629 percpu_ref_kill(&data
->node
->refs
);
7630 spin_lock(&data
->lock
);
7631 list_add(&ref_node
->node
, &data
->ref_list
);
7632 data
->node
= ref_node
;
7633 spin_unlock(&data
->lock
);
7634 percpu_ref_get(&ctx
->file_data
->refs
);
7636 destroy_fixed_file_ref_node(ref_node
);
7638 return done
? done
: err
;
7641 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
7644 struct io_uring_files_update up
;
7646 if (!ctx
->file_data
)
7650 if (copy_from_user(&up
, arg
, sizeof(up
)))
7655 return __io_sqe_files_update(ctx
, &up
, nr_args
);
7658 static void io_free_work(struct io_wq_work
*work
)
7660 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7662 /* Consider that io_steal_work() relies on this ref */
7666 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
7667 struct io_uring_params
*p
)
7669 struct io_wq_data data
;
7671 struct io_ring_ctx
*ctx_attach
;
7672 unsigned int concurrency
;
7675 data
.user
= ctx
->user
;
7676 data
.free_work
= io_free_work
;
7677 data
.do_work
= io_wq_submit_work
;
7679 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
7680 /* Do QD, or 4 * CPUS, whatever is smallest */
7681 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
7683 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
7684 if (IS_ERR(ctx
->io_wq
)) {
7685 ret
= PTR_ERR(ctx
->io_wq
);
7691 f
= fdget(p
->wq_fd
);
7695 if (f
.file
->f_op
!= &io_uring_fops
) {
7700 ctx_attach
= f
.file
->private_data
;
7701 /* @io_wq is protected by holding the fd */
7702 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
7707 ctx
->io_wq
= ctx_attach
->io_wq
;
7713 static int io_uring_alloc_task_context(struct task_struct
*task
)
7715 struct io_uring_task
*tctx
;
7718 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
7719 if (unlikely(!tctx
))
7722 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
7723 if (unlikely(ret
)) {
7729 init_waitqueue_head(&tctx
->wait
);
7732 io_init_identity(&tctx
->__identity
);
7733 tctx
->identity
= &tctx
->__identity
;
7734 task
->io_uring
= tctx
;
7738 void __io_uring_free(struct task_struct
*tsk
)
7740 struct io_uring_task
*tctx
= tsk
->io_uring
;
7742 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
7743 WARN_ON_ONCE(refcount_read(&tctx
->identity
->count
) != 1);
7744 if (tctx
->identity
!= &tctx
->__identity
)
7745 kfree(tctx
->identity
);
7746 percpu_counter_destroy(&tctx
->inflight
);
7748 tsk
->io_uring
= NULL
;
7751 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
7752 struct io_uring_params
*p
)
7756 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7757 struct io_sq_data
*sqd
;
7760 if (!capable(CAP_SYS_ADMIN
))
7763 sqd
= io_get_sq_data(p
);
7770 io_sq_thread_park(sqd
);
7771 mutex_lock(&sqd
->ctx_lock
);
7772 list_add(&ctx
->sqd_list
, &sqd
->ctx_new_list
);
7773 mutex_unlock(&sqd
->ctx_lock
);
7774 io_sq_thread_unpark(sqd
);
7776 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
7777 if (!ctx
->sq_thread_idle
)
7778 ctx
->sq_thread_idle
= HZ
;
7783 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7784 int cpu
= p
->sq_thread_cpu
;
7787 if (cpu
>= nr_cpu_ids
)
7789 if (!cpu_online(cpu
))
7792 sqd
->thread
= kthread_create_on_cpu(io_sq_thread
, sqd
,
7793 cpu
, "io_uring-sq");
7795 sqd
->thread
= kthread_create(io_sq_thread
, sqd
,
7798 if (IS_ERR(sqd
->thread
)) {
7799 ret
= PTR_ERR(sqd
->thread
);
7803 ret
= io_uring_alloc_task_context(sqd
->thread
);
7806 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7807 /* Can't have SQ_AFF without SQPOLL */
7813 ret
= io_init_wq_offload(ctx
, p
);
7819 io_finish_async(ctx
);
7823 static void io_sq_offload_start(struct io_ring_ctx
*ctx
)
7825 struct io_sq_data
*sqd
= ctx
->sq_data
;
7827 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && sqd
->thread
)
7828 wake_up_process(sqd
->thread
);
7831 static inline void __io_unaccount_mem(struct user_struct
*user
,
7832 unsigned long nr_pages
)
7834 atomic_long_sub(nr_pages
, &user
->locked_vm
);
7837 static inline int __io_account_mem(struct user_struct
*user
,
7838 unsigned long nr_pages
)
7840 unsigned long page_limit
, cur_pages
, new_pages
;
7842 /* Don't allow more pages than we can safely lock */
7843 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
7846 cur_pages
= atomic_long_read(&user
->locked_vm
);
7847 new_pages
= cur_pages
+ nr_pages
;
7848 if (new_pages
> page_limit
)
7850 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
7851 new_pages
) != cur_pages
);
7856 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
7857 enum io_mem_account acct
)
7860 __io_unaccount_mem(ctx
->user
, nr_pages
);
7862 if (ctx
->mm_account
) {
7863 if (acct
== ACCT_LOCKED
)
7864 ctx
->mm_account
->locked_vm
-= nr_pages
;
7865 else if (acct
== ACCT_PINNED
)
7866 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
7870 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
7871 enum io_mem_account acct
)
7875 if (ctx
->limit_mem
) {
7876 ret
= __io_account_mem(ctx
->user
, nr_pages
);
7881 if (ctx
->mm_account
) {
7882 if (acct
== ACCT_LOCKED
)
7883 ctx
->mm_account
->locked_vm
+= nr_pages
;
7884 else if (acct
== ACCT_PINNED
)
7885 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
7891 static void io_mem_free(void *ptr
)
7898 page
= virt_to_head_page(ptr
);
7899 if (put_page_testzero(page
))
7900 free_compound_page(page
);
7903 static void *io_mem_alloc(size_t size
)
7905 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
7908 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
7911 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
7914 struct io_rings
*rings
;
7915 size_t off
, sq_array_size
;
7917 off
= struct_size(rings
, cqes
, cq_entries
);
7918 if (off
== SIZE_MAX
)
7922 off
= ALIGN(off
, SMP_CACHE_BYTES
);
7930 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
7931 if (sq_array_size
== SIZE_MAX
)
7934 if (check_add_overflow(off
, sq_array_size
, &off
))
7940 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
7944 pages
= (size_t)1 << get_order(
7945 rings_size(sq_entries
, cq_entries
, NULL
));
7946 pages
+= (size_t)1 << get_order(
7947 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
7952 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
7956 if (!ctx
->user_bufs
)
7959 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
7960 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
7962 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
7963 unpin_user_page(imu
->bvec
[j
].bv_page
);
7965 if (imu
->acct_pages
)
7966 io_unaccount_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
7971 kfree(ctx
->user_bufs
);
7972 ctx
->user_bufs
= NULL
;
7973 ctx
->nr_user_bufs
= 0;
7977 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
7978 void __user
*arg
, unsigned index
)
7980 struct iovec __user
*src
;
7982 #ifdef CONFIG_COMPAT
7984 struct compat_iovec __user
*ciovs
;
7985 struct compat_iovec ciov
;
7987 ciovs
= (struct compat_iovec __user
*) arg
;
7988 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
7991 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
7992 dst
->iov_len
= ciov
.iov_len
;
7996 src
= (struct iovec __user
*) arg
;
7997 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8003 * Not super efficient, but this is just a registration time. And we do cache
8004 * the last compound head, so generally we'll only do a full search if we don't
8007 * We check if the given compound head page has already been accounted, to
8008 * avoid double accounting it. This allows us to account the full size of the
8009 * page, not just the constituent pages of a huge page.
8011 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8012 int nr_pages
, struct page
*hpage
)
8016 /* check current page array */
8017 for (i
= 0; i
< nr_pages
; i
++) {
8018 if (!PageCompound(pages
[i
]))
8020 if (compound_head(pages
[i
]) == hpage
)
8024 /* check previously registered pages */
8025 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8026 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8028 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8029 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8031 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8039 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8040 int nr_pages
, struct io_mapped_ubuf
*imu
,
8041 struct page
**last_hpage
)
8045 for (i
= 0; i
< nr_pages
; i
++) {
8046 if (!PageCompound(pages
[i
])) {
8051 hpage
= compound_head(pages
[i
]);
8052 if (hpage
== *last_hpage
)
8054 *last_hpage
= hpage
;
8055 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8057 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8061 if (!imu
->acct_pages
)
8064 ret
= io_account_mem(ctx
, imu
->acct_pages
, ACCT_PINNED
);
8066 imu
->acct_pages
= 0;
8070 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8073 struct vm_area_struct
**vmas
= NULL
;
8074 struct page
**pages
= NULL
;
8075 struct page
*last_hpage
= NULL
;
8076 int i
, j
, got_pages
= 0;
8081 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8084 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8086 if (!ctx
->user_bufs
)
8089 for (i
= 0; i
< nr_args
; i
++) {
8090 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8091 unsigned long off
, start
, end
, ubuf
;
8096 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8101 * Don't impose further limits on the size and buffer
8102 * constraints here, we'll -EINVAL later when IO is
8103 * submitted if they are wrong.
8106 if (!iov
.iov_base
|| !iov
.iov_len
)
8109 /* arbitrary limit, but we need something */
8110 if (iov
.iov_len
> SZ_1G
)
8113 ubuf
= (unsigned long) iov
.iov_base
;
8114 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8115 start
= ubuf
>> PAGE_SHIFT
;
8116 nr_pages
= end
- start
;
8119 if (!pages
|| nr_pages
> got_pages
) {
8122 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
8124 vmas
= kvmalloc_array(nr_pages
,
8125 sizeof(struct vm_area_struct
*),
8127 if (!pages
|| !vmas
) {
8131 got_pages
= nr_pages
;
8134 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8141 mmap_read_lock(current
->mm
);
8142 pret
= pin_user_pages(ubuf
, nr_pages
,
8143 FOLL_WRITE
| FOLL_LONGTERM
,
8145 if (pret
== nr_pages
) {
8146 /* don't support file backed memory */
8147 for (j
= 0; j
< nr_pages
; j
++) {
8148 struct vm_area_struct
*vma
= vmas
[j
];
8151 !is_file_hugepages(vma
->vm_file
)) {
8157 ret
= pret
< 0 ? pret
: -EFAULT
;
8159 mmap_read_unlock(current
->mm
);
8162 * if we did partial map, or found file backed vmas,
8163 * release any pages we did get
8166 unpin_user_pages(pages
, pret
);
8171 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, &last_hpage
);
8173 unpin_user_pages(pages
, pret
);
8178 off
= ubuf
& ~PAGE_MASK
;
8180 for (j
= 0; j
< nr_pages
; j
++) {
8183 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8184 imu
->bvec
[j
].bv_page
= pages
[j
];
8185 imu
->bvec
[j
].bv_len
= vec_len
;
8186 imu
->bvec
[j
].bv_offset
= off
;
8190 /* store original address for later verification */
8192 imu
->len
= iov
.iov_len
;
8193 imu
->nr_bvecs
= nr_pages
;
8195 ctx
->nr_user_bufs
++;
8203 io_sqe_buffer_unregister(ctx
);
8207 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8209 __s32 __user
*fds
= arg
;
8215 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8218 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8219 if (IS_ERR(ctx
->cq_ev_fd
)) {
8220 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8221 ctx
->cq_ev_fd
= NULL
;
8228 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8230 if (ctx
->cq_ev_fd
) {
8231 eventfd_ctx_put(ctx
->cq_ev_fd
);
8232 ctx
->cq_ev_fd
= NULL
;
8239 static int __io_destroy_buffers(int id
, void *p
, void *data
)
8241 struct io_ring_ctx
*ctx
= data
;
8242 struct io_buffer
*buf
= p
;
8244 __io_remove_buffers(ctx
, buf
, id
, -1U);
8248 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8250 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
8251 idr_destroy(&ctx
->io_buffer_idr
);
8254 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8256 io_finish_async(ctx
);
8257 io_sqe_buffer_unregister(ctx
);
8259 if (ctx
->sqo_task
) {
8260 put_task_struct(ctx
->sqo_task
);
8261 ctx
->sqo_task
= NULL
;
8262 mmdrop(ctx
->mm_account
);
8263 ctx
->mm_account
= NULL
;
8266 #ifdef CONFIG_BLK_CGROUP
8267 if (ctx
->sqo_blkcg_css
)
8268 css_put(ctx
->sqo_blkcg_css
);
8271 io_sqe_files_unregister(ctx
);
8272 io_eventfd_unregister(ctx
);
8273 io_destroy_buffers(ctx
);
8274 idr_destroy(&ctx
->personality_idr
);
8276 #if defined(CONFIG_UNIX)
8277 if (ctx
->ring_sock
) {
8278 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8279 sock_release(ctx
->ring_sock
);
8283 io_mem_free(ctx
->rings
);
8284 io_mem_free(ctx
->sq_sqes
);
8286 percpu_ref_exit(&ctx
->refs
);
8287 free_uid(ctx
->user
);
8288 put_cred(ctx
->creds
);
8289 kfree(ctx
->cancel_hash
);
8290 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
8294 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8296 struct io_ring_ctx
*ctx
= file
->private_data
;
8299 poll_wait(file
, &ctx
->cq_wait
, wait
);
8301 * synchronizes with barrier from wq_has_sleeper call in
8305 if (!io_sqring_full(ctx
))
8306 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8307 if (io_cqring_events(ctx
, false))
8308 mask
|= EPOLLIN
| EPOLLRDNORM
;
8313 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8315 struct io_ring_ctx
*ctx
= file
->private_data
;
8317 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8320 static int io_remove_personalities(int id
, void *p
, void *data
)
8322 struct io_ring_ctx
*ctx
= data
;
8323 struct io_identity
*iod
;
8325 iod
= idr_remove(&ctx
->personality_idr
, id
);
8327 put_cred(iod
->creds
);
8328 if (refcount_dec_and_test(&iod
->count
))
8334 static void io_ring_exit_work(struct work_struct
*work
)
8336 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
8340 * If we're doing polled IO and end up having requests being
8341 * submitted async (out-of-line), then completions can come in while
8342 * we're waiting for refs to drop. We need to reap these manually,
8343 * as nobody else will be looking for them.
8347 io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8348 io_iopoll_try_reap_events(ctx
);
8349 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8350 io_ring_ctx_free(ctx
);
8353 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8355 mutex_lock(&ctx
->uring_lock
);
8356 percpu_ref_kill(&ctx
->refs
);
8357 mutex_unlock(&ctx
->uring_lock
);
8359 io_kill_timeouts(ctx
, NULL
);
8360 io_poll_remove_all(ctx
, NULL
);
8363 io_wq_cancel_all(ctx
->io_wq
);
8365 /* if we failed setting up the ctx, we might not have any rings */
8367 io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8368 io_iopoll_try_reap_events(ctx
);
8369 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
8372 * Do this upfront, so we won't have a grace period where the ring
8373 * is closed but resources aren't reaped yet. This can cause
8374 * spurious failure in setting up a new ring.
8376 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
8379 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8381 * Use system_unbound_wq to avoid spawning tons of event kworkers
8382 * if we're exiting a ton of rings at the same time. It just adds
8383 * noise and overhead, there's no discernable change in runtime
8384 * over using system_wq.
8386 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8389 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8391 struct io_ring_ctx
*ctx
= file
->private_data
;
8393 file
->private_data
= NULL
;
8394 io_ring_ctx_wait_and_kill(ctx
);
8398 static bool io_wq_files_match(struct io_wq_work
*work
, void *data
)
8400 struct files_struct
*files
= data
;
8402 return !files
|| ((work
->flags
& IO_WQ_WORK_FILES
) &&
8403 work
->identity
->files
== files
);
8407 * Returns true if 'preq' is the link parent of 'req'
8409 static bool io_match_link(struct io_kiocb
*preq
, struct io_kiocb
*req
)
8411 struct io_kiocb
*link
;
8413 if (!(preq
->flags
& REQ_F_LINK_HEAD
))
8416 list_for_each_entry(link
, &preq
->link_list
, link_list
) {
8424 static bool io_match_link_files(struct io_kiocb
*req
,
8425 struct files_struct
*files
)
8427 struct io_kiocb
*link
;
8429 if (io_match_files(req
, files
))
8431 if (req
->flags
& REQ_F_LINK_HEAD
) {
8432 list_for_each_entry(link
, &req
->link_list
, link_list
) {
8433 if (io_match_files(link
, files
))
8441 * We're looking to cancel 'req' because it's holding on to our files, but
8442 * 'req' could be a link to another request. See if it is, and cancel that
8443 * parent request if so.
8445 static bool io_poll_remove_link(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
)
8447 struct hlist_node
*tmp
;
8448 struct io_kiocb
*preq
;
8452 spin_lock_irq(&ctx
->completion_lock
);
8453 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
8454 struct hlist_head
*list
;
8456 list
= &ctx
->cancel_hash
[i
];
8457 hlist_for_each_entry_safe(preq
, tmp
, list
, hash_node
) {
8458 found
= io_match_link(preq
, req
);
8460 io_poll_remove_one(preq
);
8465 spin_unlock_irq(&ctx
->completion_lock
);
8469 static bool io_timeout_remove_link(struct io_ring_ctx
*ctx
,
8470 struct io_kiocb
*req
)
8472 struct io_kiocb
*preq
;
8475 spin_lock_irq(&ctx
->completion_lock
);
8476 list_for_each_entry(preq
, &ctx
->timeout_list
, timeout
.list
) {
8477 found
= io_match_link(preq
, req
);
8479 __io_timeout_cancel(preq
);
8483 spin_unlock_irq(&ctx
->completion_lock
);
8487 static bool io_cancel_link_cb(struct io_wq_work
*work
, void *data
)
8489 return io_match_link(container_of(work
, struct io_kiocb
, work
), data
);
8492 static void io_attempt_cancel(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
)
8494 enum io_wq_cancel cret
;
8496 /* cancel this particular work, if it's running */
8497 cret
= io_wq_cancel_work(ctx
->io_wq
, &req
->work
);
8498 if (cret
!= IO_WQ_CANCEL_NOTFOUND
)
8501 /* find links that hold this pending, cancel those */
8502 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_link_cb
, req
, true);
8503 if (cret
!= IO_WQ_CANCEL_NOTFOUND
)
8506 /* if we have a poll link holding this pending, cancel that */
8507 if (io_poll_remove_link(ctx
, req
))
8510 /* final option, timeout link is holding this req pending */
8511 io_timeout_remove_link(ctx
, req
);
8514 static void io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8515 struct files_struct
*files
)
8517 struct io_defer_entry
*de
= NULL
;
8520 spin_lock_irq(&ctx
->completion_lock
);
8521 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8522 if (io_match_link_files(de
->req
, files
)) {
8523 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8527 spin_unlock_irq(&ctx
->completion_lock
);
8529 while (!list_empty(&list
)) {
8530 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8531 list_del_init(&de
->list
);
8532 req_set_fail_links(de
->req
);
8533 io_put_req(de
->req
);
8534 io_req_complete(de
->req
, -ECANCELED
);
8540 * Returns true if we found and killed one or more files pinning requests
8542 static bool io_uring_cancel_files(struct io_ring_ctx
*ctx
,
8543 struct files_struct
*files
)
8545 if (list_empty_careful(&ctx
->inflight_list
))
8548 io_cancel_defer_files(ctx
, files
);
8549 /* cancel all at once, should be faster than doing it one by one*/
8550 io_wq_cancel_cb(ctx
->io_wq
, io_wq_files_match
, files
, true);
8552 while (!list_empty_careful(&ctx
->inflight_list
)) {
8553 struct io_kiocb
*cancel_req
= NULL
, *req
;
8556 spin_lock_irq(&ctx
->inflight_lock
);
8557 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
) {
8558 if (files
&& (req
->work
.flags
& IO_WQ_WORK_FILES
) &&
8559 req
->work
.identity
->files
!= files
)
8561 /* req is being completed, ignore */
8562 if (!refcount_inc_not_zero(&req
->refs
))
8568 prepare_to_wait(&ctx
->inflight_wait
, &wait
,
8569 TASK_UNINTERRUPTIBLE
);
8570 spin_unlock_irq(&ctx
->inflight_lock
);
8572 /* We need to keep going until we don't find a matching req */
8575 /* cancel this request, or head link requests */
8576 io_attempt_cancel(ctx
, cancel_req
);
8577 io_put_req(cancel_req
);
8578 /* cancellations _may_ trigger task work */
8581 finish_wait(&ctx
->inflight_wait
, &wait
);
8587 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8589 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8590 struct task_struct
*task
= data
;
8592 return io_task_match(req
, task
);
8595 static bool __io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
8596 struct task_struct
*task
,
8597 struct files_struct
*files
)
8601 ret
= io_uring_cancel_files(ctx
, files
);
8603 enum io_wq_cancel cret
;
8605 cret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
, task
, true);
8606 if (cret
!= IO_WQ_CANCEL_NOTFOUND
)
8609 /* SQPOLL thread does its own polling */
8610 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
8611 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8612 io_iopoll_try_reap_events(ctx
);
8617 ret
|= io_poll_remove_all(ctx
, task
);
8618 ret
|= io_kill_timeouts(ctx
, task
);
8625 * We need to iteratively cancel requests, in case a request has dependent
8626 * hard links. These persist even for failure of cancelations, hence keep
8627 * looping until none are found.
8629 static void io_uring_cancel_task_requests(struct io_ring_ctx
*ctx
,
8630 struct files_struct
*files
)
8632 struct task_struct
*task
= current
;
8634 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) && ctx
->sq_data
)
8635 task
= ctx
->sq_data
->thread
;
8637 io_cqring_overflow_flush(ctx
, true, task
, files
);
8639 while (__io_uring_cancel_task_requests(ctx
, task
, files
)) {
8646 * Note that this task has used io_uring. We use it for cancelation purposes.
8648 static int io_uring_add_task_file(struct file
*file
)
8650 struct io_uring_task
*tctx
= current
->io_uring
;
8652 if (unlikely(!tctx
)) {
8655 ret
= io_uring_alloc_task_context(current
);
8658 tctx
= current
->io_uring
;
8660 if (tctx
->last
!= file
) {
8661 void *old
= xa_load(&tctx
->xa
, (unsigned long)file
);
8665 xa_store(&tctx
->xa
, (unsigned long)file
, file
, GFP_KERNEL
);
8674 * Remove this io_uring_file -> task mapping.
8676 static void io_uring_del_task_file(struct file
*file
)
8678 struct io_uring_task
*tctx
= current
->io_uring
;
8680 if (tctx
->last
== file
)
8682 file
= xa_erase(&tctx
->xa
, (unsigned long)file
);
8687 static void __io_uring_attempt_task_drop(struct file
*file
)
8689 struct file
*old
= xa_load(¤t
->io_uring
->xa
, (unsigned long)file
);
8692 io_uring_del_task_file(file
);
8696 * Drop task note for this file if we're the only ones that hold it after
8699 static void io_uring_attempt_task_drop(struct file
*file
, bool exiting
)
8701 if (!current
->io_uring
)
8704 * fput() is pending, will be 2 if the only other ref is our potential
8705 * task file note. If the task is exiting, drop regardless of count.
8707 if (!exiting
&& atomic_long_read(&file
->f_count
) != 2)
8710 __io_uring_attempt_task_drop(file
);
8713 void __io_uring_files_cancel(struct files_struct
*files
)
8715 struct io_uring_task
*tctx
= current
->io_uring
;
8717 unsigned long index
;
8719 /* make sure overflow events are dropped */
8720 tctx
->in_idle
= true;
8722 xa_for_each(&tctx
->xa
, index
, file
) {
8723 struct io_ring_ctx
*ctx
= file
->private_data
;
8725 io_uring_cancel_task_requests(ctx
, files
);
8727 io_uring_del_task_file(file
);
8732 * Find any io_uring fd that this task has registered or done IO on, and cancel
8735 void __io_uring_task_cancel(void)
8737 struct io_uring_task
*tctx
= current
->io_uring
;
8741 /* make sure overflow events are dropped */
8742 tctx
->in_idle
= true;
8745 /* read completions before cancelations */
8746 inflight
= percpu_counter_sum(&tctx
->inflight
);
8749 __io_uring_files_cancel(NULL
);
8751 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
8754 * If we've seen completions, retry. This avoids a race where
8755 * a completion comes in before we did prepare_to_wait().
8757 if (inflight
!= percpu_counter_sum(&tctx
->inflight
))
8762 finish_wait(&tctx
->wait
, &wait
);
8763 tctx
->in_idle
= false;
8766 static int io_uring_flush(struct file
*file
, void *data
)
8768 struct io_ring_ctx
*ctx
= file
->private_data
;
8771 * If the task is going away, cancel work it may have pending
8773 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
))
8776 io_uring_cancel_task_requests(ctx
, data
);
8777 io_uring_attempt_task_drop(file
, !data
);
8781 static void *io_uring_validate_mmap_request(struct file
*file
,
8782 loff_t pgoff
, size_t sz
)
8784 struct io_ring_ctx
*ctx
= file
->private_data
;
8785 loff_t offset
= pgoff
<< PAGE_SHIFT
;
8790 case IORING_OFF_SQ_RING
:
8791 case IORING_OFF_CQ_RING
:
8794 case IORING_OFF_SQES
:
8798 return ERR_PTR(-EINVAL
);
8801 page
= virt_to_head_page(ptr
);
8802 if (sz
> page_size(page
))
8803 return ERR_PTR(-EINVAL
);
8810 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
8812 size_t sz
= vma
->vm_end
- vma
->vm_start
;
8816 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
8818 return PTR_ERR(ptr
);
8820 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
8821 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
8824 #else /* !CONFIG_MMU */
8826 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
8828 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
8831 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
8833 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
8836 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
8837 unsigned long addr
, unsigned long len
,
8838 unsigned long pgoff
, unsigned long flags
)
8842 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
8844 return PTR_ERR(ptr
);
8846 return (unsigned long) ptr
;
8849 #endif /* !CONFIG_MMU */
8851 static void io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
8856 if (!io_sqring_full(ctx
))
8859 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
8861 if (!io_sqring_full(ctx
))
8865 } while (!signal_pending(current
));
8867 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
8870 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
8871 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
8874 struct io_ring_ctx
*ctx
;
8881 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
8882 IORING_ENTER_SQ_WAIT
))
8890 if (f
.file
->f_op
!= &io_uring_fops
)
8894 ctx
= f
.file
->private_data
;
8895 if (!percpu_ref_tryget(&ctx
->refs
))
8899 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
8903 * For SQ polling, the thread will do all submissions and completions.
8904 * Just return the requested submit count, and wake the thread if
8908 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8909 if (!list_empty_careful(&ctx
->cq_overflow_list
))
8910 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
8911 if (flags
& IORING_ENTER_SQ_WAKEUP
)
8912 wake_up(&ctx
->sq_data
->wait
);
8913 if (flags
& IORING_ENTER_SQ_WAIT
)
8914 io_sqpoll_wait_sq(ctx
);
8915 submitted
= to_submit
;
8916 } else if (to_submit
) {
8917 ret
= io_uring_add_task_file(f
.file
);
8920 mutex_lock(&ctx
->uring_lock
);
8921 submitted
= io_submit_sqes(ctx
, to_submit
);
8922 mutex_unlock(&ctx
->uring_lock
);
8924 if (submitted
!= to_submit
)
8927 if (flags
& IORING_ENTER_GETEVENTS
) {
8928 min_complete
= min(min_complete
, ctx
->cq_entries
);
8931 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8932 * space applications don't need to do io completion events
8933 * polling again, they can rely on io_sq_thread to do polling
8934 * work, which can reduce cpu usage and uring_lock contention.
8936 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
8937 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
8938 ret
= io_iopoll_check(ctx
, min_complete
);
8940 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
8945 percpu_ref_put(&ctx
->refs
);
8948 return submitted
? submitted
: ret
;
8951 #ifdef CONFIG_PROC_FS
8952 static int io_uring_show_cred(int id
, void *p
, void *data
)
8954 const struct cred
*cred
= p
;
8955 struct seq_file
*m
= data
;
8956 struct user_namespace
*uns
= seq_user_ns(m
);
8957 struct group_info
*gi
;
8962 seq_printf(m
, "%5d\n", id
);
8963 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
8964 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
8965 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
8966 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
8967 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
8968 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
8969 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
8970 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
8971 seq_puts(m
, "\n\tGroups:\t");
8972 gi
= cred
->group_info
;
8973 for (g
= 0; g
< gi
->ngroups
; g
++) {
8974 seq_put_decimal_ull(m
, g
? " " : "",
8975 from_kgid_munged(uns
, gi
->gid
[g
]));
8977 seq_puts(m
, "\n\tCapEff:\t");
8978 cap
= cred
->cap_effective
;
8979 CAP_FOR_EACH_U32(__capi
)
8980 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
8985 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
8987 struct io_sq_data
*sq
= NULL
;
8992 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8993 * since fdinfo case grabs it in the opposite direction of normal use
8994 * cases. If we fail to get the lock, we just don't iterate any
8995 * structures that could be going away outside the io_uring mutex.
8997 has_lock
= mutex_trylock(&ctx
->uring_lock
);
8999 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
))
9002 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9003 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9004 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9005 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9006 struct fixed_file_table
*table
;
9009 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
9010 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
9012 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9014 seq_printf(m
, "%5u: <none>\n", i
);
9016 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9017 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9018 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9020 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9021 (unsigned int) buf
->len
);
9023 if (has_lock
&& !idr_is_empty(&ctx
->personality_idr
)) {
9024 seq_printf(m
, "Personalities:\n");
9025 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
9027 seq_printf(m
, "PollList:\n");
9028 spin_lock_irq(&ctx
->completion_lock
);
9029 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9030 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9031 struct io_kiocb
*req
;
9033 hlist_for_each_entry(req
, list
, hash_node
)
9034 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9035 req
->task
->task_works
!= NULL
);
9037 spin_unlock_irq(&ctx
->completion_lock
);
9039 mutex_unlock(&ctx
->uring_lock
);
9042 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9044 struct io_ring_ctx
*ctx
= f
->private_data
;
9046 if (percpu_ref_tryget(&ctx
->refs
)) {
9047 __io_uring_show_fdinfo(ctx
, m
);
9048 percpu_ref_put(&ctx
->refs
);
9053 static const struct file_operations io_uring_fops
= {
9054 .release
= io_uring_release
,
9055 .flush
= io_uring_flush
,
9056 .mmap
= io_uring_mmap
,
9058 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9059 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9061 .poll
= io_uring_poll
,
9062 .fasync
= io_uring_fasync
,
9063 #ifdef CONFIG_PROC_FS
9064 .show_fdinfo
= io_uring_show_fdinfo
,
9068 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9069 struct io_uring_params
*p
)
9071 struct io_rings
*rings
;
9072 size_t size
, sq_array_offset
;
9074 /* make sure these are sane, as we already accounted them */
9075 ctx
->sq_entries
= p
->sq_entries
;
9076 ctx
->cq_entries
= p
->cq_entries
;
9078 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9079 if (size
== SIZE_MAX
)
9082 rings
= io_mem_alloc(size
);
9087 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9088 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9089 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9090 rings
->sq_ring_entries
= p
->sq_entries
;
9091 rings
->cq_ring_entries
= p
->cq_entries
;
9092 ctx
->sq_mask
= rings
->sq_ring_mask
;
9093 ctx
->cq_mask
= rings
->cq_ring_mask
;
9095 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9096 if (size
== SIZE_MAX
) {
9097 io_mem_free(ctx
->rings
);
9102 ctx
->sq_sqes
= io_mem_alloc(size
);
9103 if (!ctx
->sq_sqes
) {
9104 io_mem_free(ctx
->rings
);
9113 * Allocate an anonymous fd, this is what constitutes the application
9114 * visible backing of an io_uring instance. The application mmaps this
9115 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9116 * we have to tie this fd to a socket for file garbage collection purposes.
9118 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
9123 #if defined(CONFIG_UNIX)
9124 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9130 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9134 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9135 O_RDWR
| O_CLOEXEC
);
9139 ret
= PTR_ERR(file
);
9143 #if defined(CONFIG_UNIX)
9144 ctx
->ring_sock
->file
= file
;
9146 if (unlikely(io_uring_add_task_file(file
))) {
9147 file
= ERR_PTR(-ENOMEM
);
9150 fd_install(ret
, file
);
9153 #if defined(CONFIG_UNIX)
9154 sock_release(ctx
->ring_sock
);
9155 ctx
->ring_sock
= NULL
;
9160 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9161 struct io_uring_params __user
*params
)
9163 struct user_struct
*user
= NULL
;
9164 struct io_ring_ctx
*ctx
;
9170 if (entries
> IORING_MAX_ENTRIES
) {
9171 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9173 entries
= IORING_MAX_ENTRIES
;
9177 * Use twice as many entries for the CQ ring. It's possible for the
9178 * application to drive a higher depth than the size of the SQ ring,
9179 * since the sqes are only used at submission time. This allows for
9180 * some flexibility in overcommitting a bit. If the application has
9181 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9182 * of CQ ring entries manually.
9184 p
->sq_entries
= roundup_pow_of_two(entries
);
9185 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9187 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9188 * to a power-of-two, if it isn't already. We do NOT impose
9189 * any cq vs sq ring sizing.
9191 if (p
->cq_entries
< p
->sq_entries
)
9193 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9194 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9196 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9198 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9200 p
->cq_entries
= 2 * p
->sq_entries
;
9203 user
= get_uid(current_user());
9204 limit_mem
= !capable(CAP_IPC_LOCK
);
9207 ret
= __io_account_mem(user
,
9208 ring_pages(p
->sq_entries
, p
->cq_entries
));
9215 ctx
= io_ring_ctx_alloc(p
);
9218 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
9223 ctx
->compat
= in_compat_syscall();
9225 ctx
->creds
= get_current_cred();
9227 ctx
->loginuid
= current
->loginuid
;
9228 ctx
->sessionid
= current
->sessionid
;
9230 ctx
->sqo_task
= get_task_struct(current
);
9233 * This is just grabbed for accounting purposes. When a process exits,
9234 * the mm is exited and dropped before the files, hence we need to hang
9235 * on to this mm purely for the purposes of being able to unaccount
9236 * memory (locked/pinned vm). It's not used for anything else.
9238 mmgrab(current
->mm
);
9239 ctx
->mm_account
= current
->mm
;
9241 #ifdef CONFIG_BLK_CGROUP
9243 * The sq thread will belong to the original cgroup it was inited in.
9244 * If the cgroup goes offline (e.g. disabling the io controller), then
9245 * issued bios will be associated with the closest cgroup later in the
9249 ctx
->sqo_blkcg_css
= blkcg_css();
9250 ret
= css_tryget_online(ctx
->sqo_blkcg_css
);
9253 /* don't init against a dying cgroup, have the user try again */
9254 ctx
->sqo_blkcg_css
= NULL
;
9261 * Account memory _before_ installing the file descriptor. Once
9262 * the descriptor is installed, it can get closed at any time. Also
9263 * do this before hitting the general error path, as ring freeing
9264 * will un-account as well.
9266 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
9268 ctx
->limit_mem
= limit_mem
;
9270 ret
= io_allocate_scq_urings(ctx
, p
);
9274 ret
= io_sq_offload_create(ctx
, p
);
9278 if (!(p
->flags
& IORING_SETUP_R_DISABLED
))
9279 io_sq_offload_start(ctx
);
9281 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9282 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9283 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9284 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9285 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9286 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9287 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9288 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9290 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9291 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9292 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9293 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9294 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9295 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9296 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9297 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9299 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9300 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9301 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9302 IORING_FEAT_POLL_32BITS
;
9304 if (copy_to_user(params
, p
, sizeof(*p
))) {
9310 * Install ring fd as the very last thing, so we don't risk someone
9311 * having closed it before we finish setup
9313 ret
= io_uring_get_fd(ctx
);
9317 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9320 io_ring_ctx_wait_and_kill(ctx
);
9325 * Sets up an aio uring context, and returns the fd. Applications asks for a
9326 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9327 * params structure passed in.
9329 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9331 struct io_uring_params p
;
9334 if (copy_from_user(&p
, params
, sizeof(p
)))
9336 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9341 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9342 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9343 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9344 IORING_SETUP_R_DISABLED
))
9347 return io_uring_create(entries
, &p
, params
);
9350 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9351 struct io_uring_params __user
*, params
)
9353 return io_uring_setup(entries
, params
);
9356 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9358 struct io_uring_probe
*p
;
9362 size
= struct_size(p
, ops
, nr_args
);
9363 if (size
== SIZE_MAX
)
9365 p
= kzalloc(size
, GFP_KERNEL
);
9370 if (copy_from_user(p
, arg
, size
))
9373 if (memchr_inv(p
, 0, size
))
9376 p
->last_op
= IORING_OP_LAST
- 1;
9377 if (nr_args
> IORING_OP_LAST
)
9378 nr_args
= IORING_OP_LAST
;
9380 for (i
= 0; i
< nr_args
; i
++) {
9382 if (!io_op_defs
[i
].not_supported
)
9383 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9388 if (copy_to_user(arg
, p
, size
))
9395 static int io_register_personality(struct io_ring_ctx
*ctx
)
9397 struct io_identity
*id
;
9400 id
= kmalloc(sizeof(*id
), GFP_KERNEL
);
9404 io_init_identity(id
);
9405 id
->creds
= get_current_cred();
9407 ret
= idr_alloc_cyclic(&ctx
->personality_idr
, id
, 1, USHRT_MAX
, GFP_KERNEL
);
9409 put_cred(id
->creds
);
9415 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9417 struct io_identity
*iod
;
9419 iod
= idr_remove(&ctx
->personality_idr
, id
);
9421 put_cred(iod
->creds
);
9422 if (refcount_dec_and_test(&iod
->count
))
9430 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9431 unsigned int nr_args
)
9433 struct io_uring_restriction
*res
;
9437 /* Restrictions allowed only if rings started disabled */
9438 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9441 /* We allow only a single restrictions registration */
9442 if (ctx
->restrictions
.registered
)
9445 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9448 size
= array_size(nr_args
, sizeof(*res
));
9449 if (size
== SIZE_MAX
)
9452 res
= memdup_user(arg
, size
);
9454 return PTR_ERR(res
);
9458 for (i
= 0; i
< nr_args
; i
++) {
9459 switch (res
[i
].opcode
) {
9460 case IORING_RESTRICTION_REGISTER_OP
:
9461 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9466 __set_bit(res
[i
].register_op
,
9467 ctx
->restrictions
.register_op
);
9469 case IORING_RESTRICTION_SQE_OP
:
9470 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9475 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9477 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9478 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9480 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9481 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9490 /* Reset all restrictions if an error happened */
9492 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9494 ctx
->restrictions
.registered
= true;
9500 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
9502 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9505 if (ctx
->restrictions
.registered
)
9506 ctx
->restricted
= 1;
9508 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
9510 io_sq_offload_start(ctx
);
9515 static bool io_register_op_must_quiesce(int op
)
9518 case IORING_UNREGISTER_FILES
:
9519 case IORING_REGISTER_FILES_UPDATE
:
9520 case IORING_REGISTER_PROBE
:
9521 case IORING_REGISTER_PERSONALITY
:
9522 case IORING_UNREGISTER_PERSONALITY
:
9529 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
9530 void __user
*arg
, unsigned nr_args
)
9531 __releases(ctx
->uring_lock
)
9532 __acquires(ctx
->uring_lock
)
9537 * We're inside the ring mutex, if the ref is already dying, then
9538 * someone else killed the ctx or is already going through
9539 * io_uring_register().
9541 if (percpu_ref_is_dying(&ctx
->refs
))
9544 if (io_register_op_must_quiesce(opcode
)) {
9545 percpu_ref_kill(&ctx
->refs
);
9548 * Drop uring mutex before waiting for references to exit. If
9549 * another thread is currently inside io_uring_enter() it might
9550 * need to grab the uring_lock to make progress. If we hold it
9551 * here across the drain wait, then we can deadlock. It's safe
9552 * to drop the mutex here, since no new references will come in
9553 * after we've killed the percpu ref.
9555 mutex_unlock(&ctx
->uring_lock
);
9557 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
9560 ret
= io_run_task_work_sig();
9565 mutex_lock(&ctx
->uring_lock
);
9568 percpu_ref_resurrect(&ctx
->refs
);
9573 if (ctx
->restricted
) {
9574 if (opcode
>= IORING_REGISTER_LAST
) {
9579 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
9586 case IORING_REGISTER_BUFFERS
:
9587 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
9589 case IORING_UNREGISTER_BUFFERS
:
9593 ret
= io_sqe_buffer_unregister(ctx
);
9595 case IORING_REGISTER_FILES
:
9596 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
9598 case IORING_UNREGISTER_FILES
:
9602 ret
= io_sqe_files_unregister(ctx
);
9604 case IORING_REGISTER_FILES_UPDATE
:
9605 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
9607 case IORING_REGISTER_EVENTFD
:
9608 case IORING_REGISTER_EVENTFD_ASYNC
:
9612 ret
= io_eventfd_register(ctx
, arg
);
9615 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
9616 ctx
->eventfd_async
= 1;
9618 ctx
->eventfd_async
= 0;
9620 case IORING_UNREGISTER_EVENTFD
:
9624 ret
= io_eventfd_unregister(ctx
);
9626 case IORING_REGISTER_PROBE
:
9628 if (!arg
|| nr_args
> 256)
9630 ret
= io_probe(ctx
, arg
, nr_args
);
9632 case IORING_REGISTER_PERSONALITY
:
9636 ret
= io_register_personality(ctx
);
9638 case IORING_UNREGISTER_PERSONALITY
:
9642 ret
= io_unregister_personality(ctx
, nr_args
);
9644 case IORING_REGISTER_ENABLE_RINGS
:
9648 ret
= io_register_enable_rings(ctx
);
9650 case IORING_REGISTER_RESTRICTIONS
:
9651 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
9659 if (io_register_op_must_quiesce(opcode
)) {
9660 /* bring the ctx back to life */
9661 percpu_ref_reinit(&ctx
->refs
);
9663 reinit_completion(&ctx
->ref_comp
);
9668 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
9669 void __user
*, arg
, unsigned int, nr_args
)
9671 struct io_ring_ctx
*ctx
;
9680 if (f
.file
->f_op
!= &io_uring_fops
)
9683 ctx
= f
.file
->private_data
;
9685 mutex_lock(&ctx
->uring_lock
);
9686 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
9687 mutex_unlock(&ctx
->uring_lock
);
9688 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
9689 ctx
->cq_ev_fd
!= NULL
, ret
);
9695 static int __init
io_uring_init(void)
9697 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9698 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9699 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9702 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9703 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9704 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
9705 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
9706 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
9707 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
9708 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
9709 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
9710 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
9711 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
9712 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
9713 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
9714 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
9715 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
9716 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
9717 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
9718 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
9719 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
9720 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
9721 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
9722 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
9723 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
9724 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
9725 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
9726 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
9727 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
9728 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
9729 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
9730 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
9731 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
9732 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
9734 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
9735 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
9736 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
9739 __initcall(io_uring_init
);