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/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
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 enum io_uring_cmd_flags
{
191 IO_URING_F_NONBLOCK
= 1,
192 IO_URING_F_COMPLETE_DEFER
= 2,
195 struct io_mapped_ubuf
{
198 unsigned int nr_bvecs
;
199 unsigned long acct_pages
;
200 struct bio_vec bvec
[];
205 struct io_overflow_cqe
{
206 struct io_uring_cqe cqe
;
207 struct list_head list
;
210 struct io_fixed_file
{
211 /* file * with additional FFS_* flags */
212 unsigned long file_ptr
;
216 struct list_head list
;
221 struct io_mapped_ubuf
*buf
;
225 struct io_file_table
{
226 /* two level table */
227 struct io_fixed_file
**files
;
230 struct io_rsrc_node
{
231 struct percpu_ref refs
;
232 struct list_head node
;
233 struct list_head rsrc_list
;
234 struct io_rsrc_data
*rsrc_data
;
235 struct llist_node llist
;
239 typedef void (rsrc_put_fn
)(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
);
241 struct io_rsrc_data
{
242 struct io_ring_ctx
*ctx
;
247 struct completion done
;
252 struct list_head list
;
258 struct io_restriction
{
259 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
260 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
261 u8 sqe_flags_allowed
;
262 u8 sqe_flags_required
;
267 IO_SQ_THREAD_SHOULD_STOP
= 0,
268 IO_SQ_THREAD_SHOULD_PARK
,
273 atomic_t park_pending
;
276 /* ctx's that are using this sqd */
277 struct list_head ctx_list
;
279 struct task_struct
*thread
;
280 struct wait_queue_head wait
;
282 unsigned sq_thread_idle
;
288 struct completion exited
;
289 struct callback_head
*park_task_work
;
292 #define IO_IOPOLL_BATCH 8
293 #define IO_COMPL_BATCH 32
294 #define IO_REQ_CACHE_SIZE 32
295 #define IO_REQ_ALLOC_BATCH 8
297 struct io_comp_state
{
298 struct io_kiocb
*reqs
[IO_COMPL_BATCH
];
300 unsigned int locked_free_nr
;
301 /* inline/task_work completion list, under ->uring_lock */
302 struct list_head free_list
;
303 /* IRQ completion list, under ->completion_lock */
304 struct list_head locked_free_list
;
307 struct io_submit_link
{
308 struct io_kiocb
*head
;
309 struct io_kiocb
*last
;
312 struct io_submit_state
{
313 struct blk_plug plug
;
314 struct io_submit_link link
;
317 * io_kiocb alloc cache
319 void *reqs
[IO_REQ_CACHE_SIZE
];
320 unsigned int free_reqs
;
325 * Batch completion logic
327 struct io_comp_state comp
;
330 * File reference cache
334 unsigned int file_refs
;
335 unsigned int ios_left
;
340 struct percpu_ref refs
;
341 } ____cacheline_aligned_in_smp
;
345 unsigned int compat
: 1;
346 unsigned int drain_next
: 1;
347 unsigned int eventfd_async
: 1;
348 unsigned int restricted
: 1;
351 * Ring buffer of indices into array of io_uring_sqe, which is
352 * mmapped by the application using the IORING_OFF_SQES offset.
354 * This indirection could e.g. be used to assign fixed
355 * io_uring_sqe entries to operations and only submit them to
356 * the queue when needed.
358 * The kernel modifies neither the indices array nor the entries
362 unsigned cached_sq_head
;
365 unsigned sq_thread_idle
;
366 unsigned cached_sq_dropped
;
367 unsigned cached_cq_overflow
;
368 unsigned long sq_check_overflow
;
370 /* hashed buffered write serialization */
371 struct io_wq_hash
*hash_map
;
373 struct list_head defer_list
;
374 struct list_head timeout_list
;
375 struct list_head cq_overflow_list
;
377 struct io_uring_sqe
*sq_sqes
;
378 } ____cacheline_aligned_in_smp
;
381 struct mutex uring_lock
;
382 wait_queue_head_t wait
;
383 } ____cacheline_aligned_in_smp
;
385 struct io_submit_state submit_state
;
387 struct io_rings
*rings
;
389 /* Only used for accounting purposes */
390 struct mm_struct
*mm_account
;
392 const struct cred
*sq_creds
; /* cred used for __io_sq_thread() */
393 struct io_sq_data
*sq_data
; /* if using sq thread polling */
395 struct wait_queue_head sqo_sq_wait
;
396 struct list_head sqd_list
;
399 * If used, fixed file set. Writers must ensure that ->refs is dead,
400 * readers must ensure that ->refs is alive as long as the file* is
401 * used. Only updated through io_uring_register(2).
403 struct io_rsrc_data
*file_data
;
404 struct io_file_table file_table
;
405 unsigned nr_user_files
;
407 /* if used, fixed mapped user buffers */
408 struct io_rsrc_data
*buf_data
;
409 unsigned nr_user_bufs
;
410 struct io_mapped_ubuf
**user_bufs
;
412 struct user_struct
*user
;
414 struct completion ref_comp
;
416 #if defined(CONFIG_UNIX)
417 struct socket
*ring_sock
;
420 struct xarray io_buffers
;
422 struct xarray personalities
;
426 unsigned cached_cq_tail
;
429 atomic_t cq_timeouts
;
430 unsigned cq_last_tm_flush
;
432 unsigned long cq_check_overflow
;
433 struct wait_queue_head cq_wait
;
434 struct fasync_struct
*cq_fasync
;
435 struct eventfd_ctx
*cq_ev_fd
;
436 } ____cacheline_aligned_in_smp
;
439 spinlock_t completion_lock
;
442 * ->iopoll_list is protected by the ctx->uring_lock for
443 * io_uring instances that don't use IORING_SETUP_SQPOLL.
444 * For SQPOLL, only the single threaded io_sq_thread() will
445 * manipulate the list, hence no extra locking is needed there.
447 struct list_head iopoll_list
;
448 struct hlist_head
*cancel_hash
;
449 unsigned cancel_hash_bits
;
450 bool poll_multi_file
;
451 } ____cacheline_aligned_in_smp
;
453 struct delayed_work rsrc_put_work
;
454 struct llist_head rsrc_put_llist
;
455 struct list_head rsrc_ref_list
;
456 spinlock_t rsrc_ref_lock
;
457 struct io_rsrc_node
*rsrc_node
;
458 struct io_rsrc_node
*rsrc_backup_node
;
460 struct io_restriction restrictions
;
463 struct callback_head
*exit_task_work
;
465 /* Keep this last, we don't need it for the fast path */
466 struct work_struct exit_work
;
467 struct list_head tctx_list
;
470 struct io_uring_task
{
471 /* submission side */
473 struct wait_queue_head wait
;
474 const struct io_ring_ctx
*last
;
476 struct percpu_counter inflight
;
477 atomic_t inflight_tracked
;
480 spinlock_t task_lock
;
481 struct io_wq_work_list task_list
;
482 unsigned long task_state
;
483 struct callback_head task_work
;
487 * First field must be the file pointer in all the
488 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
490 struct io_poll_iocb
{
492 struct wait_queue_head
*head
;
496 struct wait_queue_entry wait
;
499 struct io_poll_update
{
505 bool update_user_data
;
513 struct io_timeout_data
{
514 struct io_kiocb
*req
;
515 struct hrtimer timer
;
516 struct timespec64 ts
;
517 enum hrtimer_mode mode
;
522 struct sockaddr __user
*addr
;
523 int __user
*addr_len
;
525 unsigned long nofile
;
545 struct list_head list
;
546 /* head of the link, used by linked timeouts only */
547 struct io_kiocb
*head
;
550 struct io_timeout_rem
{
555 struct timespec64 ts
;
560 /* NOTE: kiocb has the file as the first member, so don't do it here */
568 struct sockaddr __user
*addr
;
575 struct compat_msghdr __user
*umsg_compat
;
576 struct user_msghdr __user
*umsg
;
582 struct io_buffer
*kbuf
;
588 struct filename
*filename
;
590 unsigned long nofile
;
593 struct io_rsrc_update
{
619 struct epoll_event event
;
623 struct file
*file_out
;
624 struct file
*file_in
;
631 struct io_provide_buf
{
645 const char __user
*filename
;
646 struct statx __user
*buffer
;
658 struct filename
*oldpath
;
659 struct filename
*newpath
;
667 struct filename
*filename
;
670 struct io_completion
{
672 struct list_head list
;
676 struct io_async_connect
{
677 struct sockaddr_storage address
;
680 struct io_async_msghdr
{
681 struct iovec fast_iov
[UIO_FASTIOV
];
682 /* points to an allocated iov, if NULL we use fast_iov instead */
683 struct iovec
*free_iov
;
684 struct sockaddr __user
*uaddr
;
686 struct sockaddr_storage addr
;
690 struct iovec fast_iov
[UIO_FASTIOV
];
691 const struct iovec
*free_iovec
;
692 struct iov_iter iter
;
694 struct wait_page_queue wpq
;
698 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
699 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
700 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
701 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
702 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
703 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
709 REQ_F_LINK_TIMEOUT_BIT
,
710 REQ_F_NEED_CLEANUP_BIT
,
712 REQ_F_BUFFER_SELECTED_BIT
,
713 REQ_F_LTIMEOUT_ACTIVE_BIT
,
714 REQ_F_COMPLETE_INLINE_BIT
,
716 REQ_F_DONT_REISSUE_BIT
,
717 /* keep async read/write and isreg together and in order */
718 REQ_F_ASYNC_READ_BIT
,
719 REQ_F_ASYNC_WRITE_BIT
,
722 /* not a real bit, just to check we're not overflowing the space */
728 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
729 /* drain existing IO first */
730 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
732 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
733 /* doesn't sever on completion < 0 */
734 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
736 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
737 /* IOSQE_BUFFER_SELECT */
738 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
740 /* fail rest of links */
741 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
742 /* on inflight list, should be cancelled and waited on exit reliably */
743 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
744 /* read/write uses file position */
745 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
746 /* must not punt to workers */
747 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
748 /* has or had linked timeout */
749 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
751 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
752 /* already went through poll handler */
753 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
754 /* buffer already selected */
755 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
756 /* linked timeout is active, i.e. prepared by link's head */
757 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
758 /* completion is deferred through io_comp_state */
759 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
760 /* caller should reissue async */
761 REQ_F_REISSUE
= BIT(REQ_F_REISSUE_BIT
),
762 /* don't attempt request reissue, see io_rw_reissue() */
763 REQ_F_DONT_REISSUE
= BIT(REQ_F_DONT_REISSUE_BIT
),
764 /* supports async reads */
765 REQ_F_ASYNC_READ
= BIT(REQ_F_ASYNC_READ_BIT
),
766 /* supports async writes */
767 REQ_F_ASYNC_WRITE
= BIT(REQ_F_ASYNC_WRITE_BIT
),
769 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
773 struct io_poll_iocb poll
;
774 struct io_poll_iocb
*double_poll
;
777 struct io_task_work
{
778 struct io_wq_work_node node
;
779 task_work_func_t func
;
783 * NOTE! Each of the iocb union members has the file pointer
784 * as the first entry in their struct definition. So you can
785 * access the file pointer through any of the sub-structs,
786 * or directly as just 'ki_filp' in this struct.
792 struct io_poll_iocb poll
;
793 struct io_poll_update poll_update
;
794 struct io_accept accept
;
796 struct io_cancel cancel
;
797 struct io_timeout timeout
;
798 struct io_timeout_rem timeout_rem
;
799 struct io_connect connect
;
800 struct io_sr_msg sr_msg
;
802 struct io_close close
;
803 struct io_rsrc_update rsrc_update
;
804 struct io_fadvise fadvise
;
805 struct io_madvise madvise
;
806 struct io_epoll epoll
;
807 struct io_splice splice
;
808 struct io_provide_buf pbuf
;
809 struct io_statx statx
;
810 struct io_shutdown shutdown
;
811 struct io_rename rename
;
812 struct io_unlink unlink
;
813 /* use only after cleaning per-op data, see io_clean_op() */
814 struct io_completion
compl;
817 /* opcode allocated if it needs to store data for async defer */
820 /* polled IO has completed */
826 struct io_ring_ctx
*ctx
;
829 struct task_struct
*task
;
832 struct io_kiocb
*link
;
833 struct percpu_ref
*fixed_rsrc_refs
;
835 /* used with ctx->iopoll_list with reads/writes */
836 struct list_head inflight_entry
;
838 struct io_task_work io_task_work
;
839 struct callback_head task_work
;
841 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
842 struct hlist_node hash_node
;
843 struct async_poll
*apoll
;
844 struct io_wq_work work
;
845 /* store used ubuf, so we can prevent reloading */
846 struct io_mapped_ubuf
*imu
;
849 struct io_tctx_node
{
850 struct list_head ctx_node
;
851 struct task_struct
*task
;
852 struct io_ring_ctx
*ctx
;
855 struct io_defer_entry
{
856 struct list_head list
;
857 struct io_kiocb
*req
;
862 /* needs req->file assigned */
863 unsigned needs_file
: 1;
864 /* hash wq insertion if file is a regular file */
865 unsigned hash_reg_file
: 1;
866 /* unbound wq insertion if file is a non-regular file */
867 unsigned unbound_nonreg_file
: 1;
868 /* opcode is not supported by this kernel */
869 unsigned not_supported
: 1;
870 /* set if opcode supports polled "wait" */
872 unsigned pollout
: 1;
873 /* op supports buffer selection */
874 unsigned buffer_select
: 1;
875 /* do prep async if is going to be punted */
876 unsigned needs_async_setup
: 1;
877 /* should block plug */
879 /* size of async data needed, if any */
880 unsigned short async_size
;
883 static const struct io_op_def io_op_defs
[] = {
884 [IORING_OP_NOP
] = {},
885 [IORING_OP_READV
] = {
887 .unbound_nonreg_file
= 1,
890 .needs_async_setup
= 1,
892 .async_size
= sizeof(struct io_async_rw
),
894 [IORING_OP_WRITEV
] = {
897 .unbound_nonreg_file
= 1,
899 .needs_async_setup
= 1,
901 .async_size
= sizeof(struct io_async_rw
),
903 [IORING_OP_FSYNC
] = {
906 [IORING_OP_READ_FIXED
] = {
908 .unbound_nonreg_file
= 1,
911 .async_size
= sizeof(struct io_async_rw
),
913 [IORING_OP_WRITE_FIXED
] = {
916 .unbound_nonreg_file
= 1,
919 .async_size
= sizeof(struct io_async_rw
),
921 [IORING_OP_POLL_ADD
] = {
923 .unbound_nonreg_file
= 1,
925 [IORING_OP_POLL_REMOVE
] = {},
926 [IORING_OP_SYNC_FILE_RANGE
] = {
929 [IORING_OP_SENDMSG
] = {
931 .unbound_nonreg_file
= 1,
933 .needs_async_setup
= 1,
934 .async_size
= sizeof(struct io_async_msghdr
),
936 [IORING_OP_RECVMSG
] = {
938 .unbound_nonreg_file
= 1,
941 .needs_async_setup
= 1,
942 .async_size
= sizeof(struct io_async_msghdr
),
944 [IORING_OP_TIMEOUT
] = {
945 .async_size
= sizeof(struct io_timeout_data
),
947 [IORING_OP_TIMEOUT_REMOVE
] = {
948 /* used by timeout updates' prep() */
950 [IORING_OP_ACCEPT
] = {
952 .unbound_nonreg_file
= 1,
955 [IORING_OP_ASYNC_CANCEL
] = {},
956 [IORING_OP_LINK_TIMEOUT
] = {
957 .async_size
= sizeof(struct io_timeout_data
),
959 [IORING_OP_CONNECT
] = {
961 .unbound_nonreg_file
= 1,
963 .needs_async_setup
= 1,
964 .async_size
= sizeof(struct io_async_connect
),
966 [IORING_OP_FALLOCATE
] = {
969 [IORING_OP_OPENAT
] = {},
970 [IORING_OP_CLOSE
] = {},
971 [IORING_OP_FILES_UPDATE
] = {},
972 [IORING_OP_STATX
] = {},
975 .unbound_nonreg_file
= 1,
979 .async_size
= sizeof(struct io_async_rw
),
981 [IORING_OP_WRITE
] = {
983 .unbound_nonreg_file
= 1,
986 .async_size
= sizeof(struct io_async_rw
),
988 [IORING_OP_FADVISE
] = {
991 [IORING_OP_MADVISE
] = {},
994 .unbound_nonreg_file
= 1,
999 .unbound_nonreg_file
= 1,
1003 [IORING_OP_OPENAT2
] = {
1005 [IORING_OP_EPOLL_CTL
] = {
1006 .unbound_nonreg_file
= 1,
1008 [IORING_OP_SPLICE
] = {
1011 .unbound_nonreg_file
= 1,
1013 [IORING_OP_PROVIDE_BUFFERS
] = {},
1014 [IORING_OP_REMOVE_BUFFERS
] = {},
1018 .unbound_nonreg_file
= 1,
1020 [IORING_OP_SHUTDOWN
] = {
1023 [IORING_OP_RENAMEAT
] = {},
1024 [IORING_OP_UNLINKAT
] = {},
1027 static bool io_disarm_next(struct io_kiocb
*req
);
1028 static void io_uring_del_task_file(unsigned long index
);
1029 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1030 struct task_struct
*task
,
1031 struct files_struct
*files
);
1032 static void io_uring_cancel_sqpoll(struct io_sq_data
*sqd
);
1033 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
);
1035 static bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1036 long res
, unsigned int cflags
);
1037 static void io_put_req(struct io_kiocb
*req
);
1038 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1039 static void io_dismantle_req(struct io_kiocb
*req
);
1040 static void io_put_task(struct task_struct
*task
, int nr
);
1041 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1042 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1043 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
1044 struct io_uring_rsrc_update2
*up
,
1046 static void io_clean_op(struct io_kiocb
*req
);
1047 static struct file
*io_file_get(struct io_submit_state
*state
,
1048 struct io_kiocb
*req
, int fd
, bool fixed
);
1049 static void __io_queue_sqe(struct io_kiocb
*req
);
1050 static void io_rsrc_put_work(struct work_struct
*work
);
1052 static void io_req_task_queue(struct io_kiocb
*req
);
1053 static void io_submit_flush_completions(struct io_comp_state
*cs
,
1054 struct io_ring_ctx
*ctx
);
1055 static bool io_poll_remove_waitqs(struct io_kiocb
*req
);
1056 static int io_req_prep_async(struct io_kiocb
*req
);
1058 static struct kmem_cache
*req_cachep
;
1060 static const struct file_operations io_uring_fops
;
1062 struct sock
*io_uring_get_socket(struct file
*file
)
1064 #if defined(CONFIG_UNIX)
1065 if (file
->f_op
== &io_uring_fops
) {
1066 struct io_ring_ctx
*ctx
= file
->private_data
;
1068 return ctx
->ring_sock
->sk
;
1073 EXPORT_SYMBOL(io_uring_get_socket
);
1075 #define io_for_each_link(pos, head) \
1076 for (pos = (head); pos; pos = pos->link)
1078 static inline void io_req_set_rsrc_node(struct io_kiocb
*req
)
1080 struct io_ring_ctx
*ctx
= req
->ctx
;
1082 if (!req
->fixed_rsrc_refs
) {
1083 req
->fixed_rsrc_refs
= &ctx
->rsrc_node
->refs
;
1084 percpu_ref_get(req
->fixed_rsrc_refs
);
1088 static void io_refs_resurrect(struct percpu_ref
*ref
, struct completion
*compl)
1090 bool got
= percpu_ref_tryget(ref
);
1092 /* already at zero, wait for ->release() */
1094 wait_for_completion(compl);
1095 percpu_ref_resurrect(ref
);
1097 percpu_ref_put(ref
);
1100 static bool io_match_task(struct io_kiocb
*head
,
1101 struct task_struct
*task
,
1102 struct files_struct
*files
)
1104 struct io_kiocb
*req
;
1106 if (task
&& head
->task
!= task
)
1111 io_for_each_link(req
, head
) {
1112 if (req
->flags
& REQ_F_INFLIGHT
)
1118 static inline void req_set_fail_links(struct io_kiocb
*req
)
1120 if (req
->flags
& REQ_F_LINK
)
1121 req
->flags
|= REQ_F_FAIL_LINK
;
1124 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1126 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1128 complete(&ctx
->ref_comp
);
1131 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1133 return !req
->timeout
.off
;
1136 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1138 struct io_ring_ctx
*ctx
;
1141 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1146 * Use 5 bits less than the max cq entries, that should give us around
1147 * 32 entries per hash list if totally full and uniformly spread.
1149 hash_bits
= ilog2(p
->cq_entries
);
1153 ctx
->cancel_hash_bits
= hash_bits
;
1154 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1156 if (!ctx
->cancel_hash
)
1158 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1160 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1161 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1164 ctx
->flags
= p
->flags
;
1165 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1166 INIT_LIST_HEAD(&ctx
->sqd_list
);
1167 init_waitqueue_head(&ctx
->cq_wait
);
1168 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1169 init_completion(&ctx
->ref_comp
);
1170 xa_init_flags(&ctx
->io_buffers
, XA_FLAGS_ALLOC1
);
1171 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1172 mutex_init(&ctx
->uring_lock
);
1173 init_waitqueue_head(&ctx
->wait
);
1174 spin_lock_init(&ctx
->completion_lock
);
1175 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1176 INIT_LIST_HEAD(&ctx
->defer_list
);
1177 INIT_LIST_HEAD(&ctx
->timeout_list
);
1178 spin_lock_init(&ctx
->rsrc_ref_lock
);
1179 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1180 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1181 init_llist_head(&ctx
->rsrc_put_llist
);
1182 INIT_LIST_HEAD(&ctx
->tctx_list
);
1183 INIT_LIST_HEAD(&ctx
->submit_state
.comp
.free_list
);
1184 INIT_LIST_HEAD(&ctx
->submit_state
.comp
.locked_free_list
);
1187 kfree(ctx
->cancel_hash
);
1192 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1194 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1195 struct io_ring_ctx
*ctx
= req
->ctx
;
1197 return seq
+ ctx
->cq_extra
!= ctx
->cached_cq_tail
1198 + READ_ONCE(ctx
->cached_cq_overflow
);
1204 static void io_req_track_inflight(struct io_kiocb
*req
)
1206 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1207 req
->flags
|= REQ_F_INFLIGHT
;
1208 atomic_inc(¤t
->io_uring
->inflight_tracked
);
1212 static void io_prep_async_work(struct io_kiocb
*req
)
1214 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1215 struct io_ring_ctx
*ctx
= req
->ctx
;
1217 if (!req
->work
.creds
)
1218 req
->work
.creds
= get_current_cred();
1220 req
->work
.list
.next
= NULL
;
1221 req
->work
.flags
= 0;
1222 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1223 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1225 if (req
->flags
& REQ_F_ISREG
) {
1226 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1227 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1228 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
1229 if (def
->unbound_nonreg_file
)
1230 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1233 switch (req
->opcode
) {
1234 case IORING_OP_SPLICE
:
1236 if (!S_ISREG(file_inode(req
->splice
.file_in
)->i_mode
))
1237 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1242 static void io_prep_async_link(struct io_kiocb
*req
)
1244 struct io_kiocb
*cur
;
1246 io_for_each_link(cur
, req
)
1247 io_prep_async_work(cur
);
1250 static void io_queue_async_work(struct io_kiocb
*req
)
1252 struct io_ring_ctx
*ctx
= req
->ctx
;
1253 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1254 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1257 BUG_ON(!tctx
->io_wq
);
1259 /* init ->work of the whole link before punting */
1260 io_prep_async_link(req
);
1261 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1262 &req
->work
, req
->flags
);
1263 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
1265 io_queue_linked_timeout(link
);
1268 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1269 __must_hold(&req
->ctx
->completion_lock
)
1271 struct io_timeout_data
*io
= req
->async_data
;
1273 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1274 atomic_set(&req
->ctx
->cq_timeouts
,
1275 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1276 list_del_init(&req
->timeout
.list
);
1277 io_cqring_fill_event(req
->ctx
, req
->user_data
, status
, 0);
1278 io_put_req_deferred(req
, 1);
1282 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1285 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1286 struct io_defer_entry
, list
);
1288 if (req_need_defer(de
->req
, de
->seq
))
1290 list_del_init(&de
->list
);
1291 io_req_task_queue(de
->req
);
1293 } while (!list_empty(&ctx
->defer_list
));
1296 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1300 if (list_empty(&ctx
->timeout_list
))
1303 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1306 u32 events_needed
, events_got
;
1307 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1308 struct io_kiocb
, timeout
.list
);
1310 if (io_is_timeout_noseq(req
))
1314 * Since seq can easily wrap around over time, subtract
1315 * the last seq at which timeouts were flushed before comparing.
1316 * Assuming not more than 2^31-1 events have happened since,
1317 * these subtractions won't have wrapped, so we can check if
1318 * target is in [last_seq, current_seq] by comparing the two.
1320 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1321 events_got
= seq
- ctx
->cq_last_tm_flush
;
1322 if (events_got
< events_needed
)
1325 list_del_init(&req
->timeout
.list
);
1326 io_kill_timeout(req
, 0);
1327 } while (!list_empty(&ctx
->timeout_list
));
1329 ctx
->cq_last_tm_flush
= seq
;
1332 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1334 io_flush_timeouts(ctx
);
1336 /* order cqe stores with ring update */
1337 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1339 if (unlikely(!list_empty(&ctx
->defer_list
)))
1340 __io_queue_deferred(ctx
);
1343 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1345 struct io_rings
*r
= ctx
->rings
;
1347 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1350 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1352 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1355 static inline struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1357 struct io_rings
*rings
= ctx
->rings
;
1361 * writes to the cq entry need to come after reading head; the
1362 * control dependency is enough as we're using WRITE_ONCE to
1365 if (__io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1368 tail
= ctx
->cached_cq_tail
++;
1369 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1372 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1374 if (likely(!ctx
->cq_ev_fd
))
1376 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1378 return !ctx
->eventfd_async
|| io_wq_current_is_worker();
1381 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1383 /* see waitqueue_active() comment */
1386 if (waitqueue_active(&ctx
->wait
))
1387 wake_up(&ctx
->wait
);
1388 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1389 wake_up(&ctx
->sq_data
->wait
);
1390 if (io_should_trigger_evfd(ctx
))
1391 eventfd_signal(ctx
->cq_ev_fd
, 1);
1392 if (waitqueue_active(&ctx
->cq_wait
)) {
1393 wake_up_interruptible(&ctx
->cq_wait
);
1394 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1398 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1400 /* see waitqueue_active() comment */
1403 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1404 if (waitqueue_active(&ctx
->wait
))
1405 wake_up(&ctx
->wait
);
1407 if (io_should_trigger_evfd(ctx
))
1408 eventfd_signal(ctx
->cq_ev_fd
, 1);
1409 if (waitqueue_active(&ctx
->cq_wait
)) {
1410 wake_up_interruptible(&ctx
->cq_wait
);
1411 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1415 /* Returns true if there are no backlogged entries after the flush */
1416 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1418 struct io_rings
*rings
= ctx
->rings
;
1419 unsigned long flags
;
1420 bool all_flushed
, posted
;
1422 if (!force
&& __io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1426 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1427 while (!list_empty(&ctx
->cq_overflow_list
)) {
1428 struct io_uring_cqe
*cqe
= io_get_cqring(ctx
);
1429 struct io_overflow_cqe
*ocqe
;
1433 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
1434 struct io_overflow_cqe
, list
);
1436 memcpy(cqe
, &ocqe
->cqe
, sizeof(*cqe
));
1438 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1439 ++ctx
->cached_cq_overflow
);
1441 list_del(&ocqe
->list
);
1445 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1447 clear_bit(0, &ctx
->sq_check_overflow
);
1448 clear_bit(0, &ctx
->cq_check_overflow
);
1449 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1453 io_commit_cqring(ctx
);
1454 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1456 io_cqring_ev_posted(ctx
);
1460 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1464 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1465 /* iopoll syncs against uring_lock, not completion_lock */
1466 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1467 mutex_lock(&ctx
->uring_lock
);
1468 ret
= __io_cqring_overflow_flush(ctx
, force
);
1469 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1470 mutex_unlock(&ctx
->uring_lock
);
1477 * Shamelessly stolen from the mm implementation of page reference checking,
1478 * see commit f958d7b528b1 for details.
1480 #define req_ref_zero_or_close_to_overflow(req) \
1481 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1483 static inline bool req_ref_inc_not_zero(struct io_kiocb
*req
)
1485 return atomic_inc_not_zero(&req
->refs
);
1488 static inline bool req_ref_sub_and_test(struct io_kiocb
*req
, int refs
)
1490 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1491 return atomic_sub_and_test(refs
, &req
->refs
);
1494 static inline bool req_ref_put_and_test(struct io_kiocb
*req
)
1496 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1497 return atomic_dec_and_test(&req
->refs
);
1500 static inline void req_ref_put(struct io_kiocb
*req
)
1502 WARN_ON_ONCE(req_ref_put_and_test(req
));
1505 static inline void req_ref_get(struct io_kiocb
*req
)
1507 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1508 atomic_inc(&req
->refs
);
1511 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
1512 long res
, unsigned int cflags
)
1514 struct io_overflow_cqe
*ocqe
;
1516 ocqe
= kmalloc(sizeof(*ocqe
), GFP_ATOMIC
| __GFP_ACCOUNT
);
1519 * If we're in ring overflow flush mode, or in task cancel mode,
1520 * or cannot allocate an overflow entry, then we need to drop it
1523 WRITE_ONCE(ctx
->rings
->cq_overflow
, ++ctx
->cached_cq_overflow
);
1526 if (list_empty(&ctx
->cq_overflow_list
)) {
1527 set_bit(0, &ctx
->sq_check_overflow
);
1528 set_bit(0, &ctx
->cq_check_overflow
);
1529 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1531 ocqe
->cqe
.user_data
= user_data
;
1532 ocqe
->cqe
.res
= res
;
1533 ocqe
->cqe
.flags
= cflags
;
1534 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
1538 static inline bool __io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1539 long res
, unsigned int cflags
)
1541 struct io_uring_cqe
*cqe
;
1543 trace_io_uring_complete(ctx
, user_data
, res
, cflags
);
1546 * If we can't get a cq entry, userspace overflowed the
1547 * submission (by quite a lot). Increment the overflow count in
1550 cqe
= io_get_cqring(ctx
);
1552 WRITE_ONCE(cqe
->user_data
, user_data
);
1553 WRITE_ONCE(cqe
->res
, res
);
1554 WRITE_ONCE(cqe
->flags
, cflags
);
1557 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
);
1560 /* not as hot to bloat with inlining */
1561 static noinline
bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1562 long res
, unsigned int cflags
)
1564 return __io_cqring_fill_event(ctx
, user_data
, res
, cflags
);
1567 static void io_req_complete_post(struct io_kiocb
*req
, long res
,
1568 unsigned int cflags
)
1570 struct io_ring_ctx
*ctx
= req
->ctx
;
1571 unsigned long flags
;
1573 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1574 __io_cqring_fill_event(ctx
, req
->user_data
, res
, cflags
);
1576 * If we're the last reference to this request, add to our locked
1579 if (req_ref_put_and_test(req
)) {
1580 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
1582 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1583 if (req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_FAIL_LINK
))
1584 io_disarm_next(req
);
1586 io_req_task_queue(req
->link
);
1590 io_dismantle_req(req
);
1591 io_put_task(req
->task
, 1);
1592 list_add(&req
->compl.list
, &cs
->locked_free_list
);
1593 cs
->locked_free_nr
++;
1595 if (!percpu_ref_tryget(&ctx
->refs
))
1598 io_commit_cqring(ctx
);
1599 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1602 io_cqring_ev_posted(ctx
);
1603 percpu_ref_put(&ctx
->refs
);
1607 static inline bool io_req_needs_clean(struct io_kiocb
*req
)
1609 return req
->flags
& (REQ_F_BUFFER_SELECTED
| REQ_F_NEED_CLEANUP
|
1610 REQ_F_POLLED
| REQ_F_INFLIGHT
);
1613 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1614 unsigned int cflags
)
1616 if (io_req_needs_clean(req
))
1619 req
->compl.cflags
= cflags
;
1620 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1623 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1624 long res
, unsigned cflags
)
1626 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1627 io_req_complete_state(req
, res
, cflags
);
1629 io_req_complete_post(req
, res
, cflags
);
1632 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1634 __io_req_complete(req
, 0, res
, 0);
1637 static void io_req_complete_failed(struct io_kiocb
*req
, long res
)
1639 req_set_fail_links(req
);
1641 io_req_complete_post(req
, res
, 0);
1644 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1645 struct io_comp_state
*cs
)
1647 spin_lock_irq(&ctx
->completion_lock
);
1648 list_splice_init(&cs
->locked_free_list
, &cs
->free_list
);
1649 cs
->locked_free_nr
= 0;
1650 spin_unlock_irq(&ctx
->completion_lock
);
1653 /* Returns true IFF there are requests in the cache */
1654 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1656 struct io_submit_state
*state
= &ctx
->submit_state
;
1657 struct io_comp_state
*cs
= &state
->comp
;
1661 * If we have more than a batch's worth of requests in our IRQ side
1662 * locked cache, grab the lock and move them over to our submission
1665 if (READ_ONCE(cs
->locked_free_nr
) > IO_COMPL_BATCH
)
1666 io_flush_cached_locked_reqs(ctx
, cs
);
1668 nr
= state
->free_reqs
;
1669 while (!list_empty(&cs
->free_list
)) {
1670 struct io_kiocb
*req
= list_first_entry(&cs
->free_list
,
1671 struct io_kiocb
, compl.list
);
1673 list_del(&req
->compl.list
);
1674 state
->reqs
[nr
++] = req
;
1675 if (nr
== ARRAY_SIZE(state
->reqs
))
1679 state
->free_reqs
= nr
;
1683 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1685 struct io_submit_state
*state
= &ctx
->submit_state
;
1687 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH
> ARRAY_SIZE(state
->reqs
));
1689 if (!state
->free_reqs
) {
1690 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1693 if (io_flush_cached_reqs(ctx
))
1696 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1700 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1701 * retry single alloc to be on the safe side.
1703 if (unlikely(ret
<= 0)) {
1704 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1705 if (!state
->reqs
[0])
1709 state
->free_reqs
= ret
;
1713 return state
->reqs
[state
->free_reqs
];
1716 static inline void io_put_file(struct file
*file
)
1722 static void io_dismantle_req(struct io_kiocb
*req
)
1724 unsigned int flags
= req
->flags
;
1726 if (io_req_needs_clean(req
))
1728 if (!(flags
& REQ_F_FIXED_FILE
))
1729 io_put_file(req
->file
);
1730 if (req
->fixed_rsrc_refs
)
1731 percpu_ref_put(req
->fixed_rsrc_refs
);
1732 if (req
->async_data
)
1733 kfree(req
->async_data
);
1734 if (req
->work
.creds
) {
1735 put_cred(req
->work
.creds
);
1736 req
->work
.creds
= NULL
;
1740 /* must to be called somewhat shortly after putting a request */
1741 static inline void io_put_task(struct task_struct
*task
, int nr
)
1743 struct io_uring_task
*tctx
= task
->io_uring
;
1745 percpu_counter_sub(&tctx
->inflight
, nr
);
1746 if (unlikely(atomic_read(&tctx
->in_idle
)))
1747 wake_up(&tctx
->wait
);
1748 put_task_struct_many(task
, nr
);
1751 static void __io_free_req(struct io_kiocb
*req
)
1753 struct io_ring_ctx
*ctx
= req
->ctx
;
1755 io_dismantle_req(req
);
1756 io_put_task(req
->task
, 1);
1758 kmem_cache_free(req_cachep
, req
);
1759 percpu_ref_put(&ctx
->refs
);
1762 static inline void io_remove_next_linked(struct io_kiocb
*req
)
1764 struct io_kiocb
*nxt
= req
->link
;
1766 req
->link
= nxt
->link
;
1770 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
1771 __must_hold(&req
->ctx
->completion_lock
)
1773 struct io_kiocb
*link
= req
->link
;
1776 * Can happen if a linked timeout fired and link had been like
1777 * req -> link t-out -> link t-out [-> ...]
1779 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
1780 struct io_timeout_data
*io
= link
->async_data
;
1782 io_remove_next_linked(req
);
1783 link
->timeout
.head
= NULL
;
1784 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1785 io_cqring_fill_event(link
->ctx
, link
->user_data
,
1787 io_put_req_deferred(link
, 1);
1794 static void io_fail_links(struct io_kiocb
*req
)
1795 __must_hold(&req
->ctx
->completion_lock
)
1797 struct io_kiocb
*nxt
, *link
= req
->link
;
1804 trace_io_uring_fail_link(req
, link
);
1805 io_cqring_fill_event(link
->ctx
, link
->user_data
, -ECANCELED
, 0);
1806 io_put_req_deferred(link
, 2);
1811 static bool io_disarm_next(struct io_kiocb
*req
)
1812 __must_hold(&req
->ctx
->completion_lock
)
1814 bool posted
= false;
1816 if (likely(req
->flags
& REQ_F_LINK_TIMEOUT
))
1817 posted
= io_kill_linked_timeout(req
);
1818 if (unlikely((req
->flags
& REQ_F_FAIL_LINK
) &&
1819 !(req
->flags
& REQ_F_HARDLINK
))) {
1820 posted
|= (req
->link
!= NULL
);
1826 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
1828 struct io_kiocb
*nxt
;
1831 * If LINK is set, we have dependent requests in this chain. If we
1832 * didn't fail this request, queue the first one up, moving any other
1833 * dependencies to the next request. In case of failure, fail the rest
1836 if (req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_FAIL_LINK
)) {
1837 struct io_ring_ctx
*ctx
= req
->ctx
;
1838 unsigned long flags
;
1841 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1842 posted
= io_disarm_next(req
);
1844 io_commit_cqring(req
->ctx
);
1845 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1847 io_cqring_ev_posted(ctx
);
1854 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1856 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
1858 return __io_req_find_next(req
);
1861 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
)
1865 if (ctx
->submit_state
.comp
.nr
) {
1866 mutex_lock(&ctx
->uring_lock
);
1867 io_submit_flush_completions(&ctx
->submit_state
.comp
, ctx
);
1868 mutex_unlock(&ctx
->uring_lock
);
1870 percpu_ref_put(&ctx
->refs
);
1873 static bool __tctx_task_work(struct io_uring_task
*tctx
)
1875 struct io_ring_ctx
*ctx
= NULL
;
1876 struct io_wq_work_list list
;
1877 struct io_wq_work_node
*node
;
1879 if (wq_list_empty(&tctx
->task_list
))
1882 spin_lock_irq(&tctx
->task_lock
);
1883 list
= tctx
->task_list
;
1884 INIT_WQ_LIST(&tctx
->task_list
);
1885 spin_unlock_irq(&tctx
->task_lock
);
1889 struct io_wq_work_node
*next
= node
->next
;
1890 struct io_kiocb
*req
;
1892 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1893 if (req
->ctx
!= ctx
) {
1894 ctx_flush_and_put(ctx
);
1896 percpu_ref_get(&ctx
->refs
);
1899 req
->task_work
.func(&req
->task_work
);
1903 ctx_flush_and_put(ctx
);
1904 return list
.first
!= NULL
;
1907 static void tctx_task_work(struct callback_head
*cb
)
1909 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
, task_work
);
1911 clear_bit(0, &tctx
->task_state
);
1913 while (__tctx_task_work(tctx
))
1917 static int io_req_task_work_add(struct io_kiocb
*req
)
1919 struct task_struct
*tsk
= req
->task
;
1920 struct io_uring_task
*tctx
= tsk
->io_uring
;
1921 enum task_work_notify_mode notify
;
1922 struct io_wq_work_node
*node
, *prev
;
1923 unsigned long flags
;
1926 if (unlikely(tsk
->flags
& PF_EXITING
))
1929 WARN_ON_ONCE(!tctx
);
1931 spin_lock_irqsave(&tctx
->task_lock
, flags
);
1932 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
1933 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
1935 /* task_work already pending, we're done */
1936 if (test_bit(0, &tctx
->task_state
) ||
1937 test_and_set_bit(0, &tctx
->task_state
))
1941 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1942 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1943 * processing task_work. There's no reliable way to tell if TWA_RESUME
1946 notify
= (req
->ctx
->flags
& IORING_SETUP_SQPOLL
) ? TWA_NONE
: TWA_SIGNAL
;
1948 if (!task_work_add(tsk
, &tctx
->task_work
, notify
)) {
1949 wake_up_process(tsk
);
1954 * Slow path - we failed, find and delete work. if the work is not
1955 * in the list, it got run and we're fine.
1957 spin_lock_irqsave(&tctx
->task_lock
, flags
);
1958 wq_list_for_each(node
, prev
, &tctx
->task_list
) {
1959 if (&req
->io_task_work
.node
== node
) {
1960 wq_list_del(&tctx
->task_list
, node
, prev
);
1965 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
1966 clear_bit(0, &tctx
->task_state
);
1970 static bool io_run_task_work_head(struct callback_head
**work_head
)
1972 struct callback_head
*work
, *next
;
1973 bool executed
= false;
1976 work
= xchg(work_head
, NULL
);
1992 static void io_task_work_add_head(struct callback_head
**work_head
,
1993 struct callback_head
*task_work
)
1995 struct callback_head
*head
;
1998 head
= READ_ONCE(*work_head
);
1999 task_work
->next
= head
;
2000 } while (cmpxchg(work_head
, head
, task_work
) != head
);
2003 static void io_req_task_work_add_fallback(struct io_kiocb
*req
,
2004 task_work_func_t cb
)
2006 init_task_work(&req
->task_work
, cb
);
2007 io_task_work_add_head(&req
->ctx
->exit_task_work
, &req
->task_work
);
2010 static void io_req_task_cancel(struct callback_head
*cb
)
2012 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2013 struct io_ring_ctx
*ctx
= req
->ctx
;
2015 /* ctx is guaranteed to stay alive while we hold uring_lock */
2016 mutex_lock(&ctx
->uring_lock
);
2017 io_req_complete_failed(req
, req
->result
);
2018 mutex_unlock(&ctx
->uring_lock
);
2021 static void __io_req_task_submit(struct io_kiocb
*req
)
2023 struct io_ring_ctx
*ctx
= req
->ctx
;
2025 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2026 mutex_lock(&ctx
->uring_lock
);
2027 if (!(current
->flags
& PF_EXITING
) && !current
->in_execve
)
2028 __io_queue_sqe(req
);
2030 io_req_complete_failed(req
, -EFAULT
);
2031 mutex_unlock(&ctx
->uring_lock
);
2034 static void io_req_task_submit(struct callback_head
*cb
)
2036 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2038 __io_req_task_submit(req
);
2041 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2044 req
->task_work
.func
= io_req_task_cancel
;
2046 if (unlikely(io_req_task_work_add(req
)))
2047 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2050 static void io_req_task_queue(struct io_kiocb
*req
)
2052 req
->task_work
.func
= io_req_task_submit
;
2054 if (unlikely(io_req_task_work_add(req
)))
2055 io_req_task_queue_fail(req
, -ECANCELED
);
2058 static inline void io_queue_next(struct io_kiocb
*req
)
2060 struct io_kiocb
*nxt
= io_req_find_next(req
);
2063 io_req_task_queue(nxt
);
2066 static void io_free_req(struct io_kiocb
*req
)
2073 struct task_struct
*task
;
2078 static inline void io_init_req_batch(struct req_batch
*rb
)
2085 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2086 struct req_batch
*rb
)
2089 io_put_task(rb
->task
, rb
->task_refs
);
2091 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2094 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2095 struct io_submit_state
*state
)
2098 io_dismantle_req(req
);
2100 if (req
->task
!= rb
->task
) {
2102 io_put_task(rb
->task
, rb
->task_refs
);
2103 rb
->task
= req
->task
;
2109 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2110 state
->reqs
[state
->free_reqs
++] = req
;
2112 list_add(&req
->compl.list
, &state
->comp
.free_list
);
2115 static void io_submit_flush_completions(struct io_comp_state
*cs
,
2116 struct io_ring_ctx
*ctx
)
2119 struct io_kiocb
*req
;
2120 struct req_batch rb
;
2122 io_init_req_batch(&rb
);
2123 spin_lock_irq(&ctx
->completion_lock
);
2124 for (i
= 0; i
< nr
; i
++) {
2126 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2129 io_commit_cqring(ctx
);
2130 spin_unlock_irq(&ctx
->completion_lock
);
2132 io_cqring_ev_posted(ctx
);
2133 for (i
= 0; i
< nr
; i
++) {
2136 /* submission and completion refs */
2137 if (req_ref_sub_and_test(req
, 2))
2138 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2141 io_req_free_batch_finish(ctx
, &rb
);
2146 * Drop reference to request, return next in chain (if there is one) if this
2147 * was the last reference to this request.
2149 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2151 struct io_kiocb
*nxt
= NULL
;
2153 if (req_ref_put_and_test(req
)) {
2154 nxt
= io_req_find_next(req
);
2160 static inline void io_put_req(struct io_kiocb
*req
)
2162 if (req_ref_put_and_test(req
))
2166 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2168 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2173 static void io_free_req_deferred(struct io_kiocb
*req
)
2175 req
->task_work
.func
= io_put_req_deferred_cb
;
2176 if (unlikely(io_req_task_work_add(req
)))
2177 io_req_task_work_add_fallback(req
, io_put_req_deferred_cb
);
2180 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2182 if (req_ref_sub_and_test(req
, refs
))
2183 io_free_req_deferred(req
);
2186 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2188 /* See comment at the top of this file */
2190 return __io_cqring_events(ctx
);
2193 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2195 struct io_rings
*rings
= ctx
->rings
;
2197 /* make sure SQ entry isn't read before tail */
2198 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2201 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2203 unsigned int cflags
;
2205 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2206 cflags
|= IORING_CQE_F_BUFFER
;
2207 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2212 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2214 struct io_buffer
*kbuf
;
2216 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2217 return io_put_kbuf(req
, kbuf
);
2220 static inline bool io_run_task_work(void)
2223 * Not safe to run on exiting task, and the task_work handling will
2224 * not add work to such a task.
2226 if (unlikely(current
->flags
& PF_EXITING
))
2228 if (current
->task_works
) {
2229 __set_current_state(TASK_RUNNING
);
2238 * Find and free completed poll iocbs
2240 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2241 struct list_head
*done
)
2243 struct req_batch rb
;
2244 struct io_kiocb
*req
;
2246 /* order with ->result store in io_complete_rw_iopoll() */
2249 io_init_req_batch(&rb
);
2250 while (!list_empty(done
)) {
2253 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2254 list_del(&req
->inflight_entry
);
2256 if (READ_ONCE(req
->result
) == -EAGAIN
&&
2257 !(req
->flags
& REQ_F_DONT_REISSUE
)) {
2258 req
->iopoll_completed
= 0;
2260 io_queue_async_work(req
);
2264 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2265 cflags
= io_put_rw_kbuf(req
);
2267 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
, cflags
);
2270 if (req_ref_put_and_test(req
))
2271 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2274 io_commit_cqring(ctx
);
2275 io_cqring_ev_posted_iopoll(ctx
);
2276 io_req_free_batch_finish(ctx
, &rb
);
2279 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2282 struct io_kiocb
*req
, *tmp
;
2288 * Only spin for completions if we don't have multiple devices hanging
2289 * off our complete list, and we're under the requested amount.
2291 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2294 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2295 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2298 * Move completed and retryable entries to our local lists.
2299 * If we find a request that requires polling, break out
2300 * and complete those lists first, if we have entries there.
2302 if (READ_ONCE(req
->iopoll_completed
)) {
2303 list_move_tail(&req
->inflight_entry
, &done
);
2306 if (!list_empty(&done
))
2309 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2313 /* iopoll may have completed current req */
2314 if (READ_ONCE(req
->iopoll_completed
))
2315 list_move_tail(&req
->inflight_entry
, &done
);
2322 if (!list_empty(&done
))
2323 io_iopoll_complete(ctx
, nr_events
, &done
);
2329 * We can't just wait for polled events to come to us, we have to actively
2330 * find and complete them.
2332 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2334 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2337 mutex_lock(&ctx
->uring_lock
);
2338 while (!list_empty(&ctx
->iopoll_list
)) {
2339 unsigned int nr_events
= 0;
2341 io_do_iopoll(ctx
, &nr_events
, 0);
2343 /* let it sleep and repeat later if can't complete a request */
2347 * Ensure we allow local-to-the-cpu processing to take place,
2348 * in this case we need to ensure that we reap all events.
2349 * Also let task_work, etc. to progress by releasing the mutex
2351 if (need_resched()) {
2352 mutex_unlock(&ctx
->uring_lock
);
2354 mutex_lock(&ctx
->uring_lock
);
2357 mutex_unlock(&ctx
->uring_lock
);
2360 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2362 unsigned int nr_events
= 0;
2366 * We disallow the app entering submit/complete with polling, but we
2367 * still need to lock the ring to prevent racing with polled issue
2368 * that got punted to a workqueue.
2370 mutex_lock(&ctx
->uring_lock
);
2372 * Don't enter poll loop if we already have events pending.
2373 * If we do, we can potentially be spinning for commands that
2374 * already triggered a CQE (eg in error).
2376 if (test_bit(0, &ctx
->cq_check_overflow
))
2377 __io_cqring_overflow_flush(ctx
, false);
2378 if (io_cqring_events(ctx
))
2382 * If a submit got punted to a workqueue, we can have the
2383 * application entering polling for a command before it gets
2384 * issued. That app will hold the uring_lock for the duration
2385 * of the poll right here, so we need to take a breather every
2386 * now and then to ensure that the issue has a chance to add
2387 * the poll to the issued list. Otherwise we can spin here
2388 * forever, while the workqueue is stuck trying to acquire the
2391 if (list_empty(&ctx
->iopoll_list
)) {
2392 mutex_unlock(&ctx
->uring_lock
);
2394 mutex_lock(&ctx
->uring_lock
);
2396 if (list_empty(&ctx
->iopoll_list
))
2399 ret
= io_do_iopoll(ctx
, &nr_events
, min
);
2400 } while (!ret
&& nr_events
< min
&& !need_resched());
2402 mutex_unlock(&ctx
->uring_lock
);
2406 static void kiocb_end_write(struct io_kiocb
*req
)
2409 * Tell lockdep we inherited freeze protection from submission
2412 if (req
->flags
& REQ_F_ISREG
) {
2413 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
2415 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
2421 static bool io_resubmit_prep(struct io_kiocb
*req
)
2423 struct io_async_rw
*rw
= req
->async_data
;
2426 return !io_req_prep_async(req
);
2427 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2428 iov_iter_revert(&rw
->iter
, req
->result
- iov_iter_count(&rw
->iter
));
2432 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2434 umode_t mode
= file_inode(req
->file
)->i_mode
;
2435 struct io_ring_ctx
*ctx
= req
->ctx
;
2437 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2439 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2440 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2443 * If ref is dying, we might be running poll reap from the exit work.
2444 * Don't attempt to reissue from that path, just let it fail with
2447 if (percpu_ref_is_dying(&ctx
->refs
))
2452 static bool io_resubmit_prep(struct io_kiocb
*req
)
2456 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2462 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2463 unsigned int issue_flags
)
2467 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
)
2468 kiocb_end_write(req
);
2469 if (res
!= req
->result
) {
2470 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
2471 io_rw_should_reissue(req
)) {
2472 req
->flags
|= REQ_F_REISSUE
;
2475 req_set_fail_links(req
);
2477 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2478 cflags
= io_put_rw_kbuf(req
);
2479 __io_req_complete(req
, issue_flags
, res
, cflags
);
2482 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2484 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2486 __io_complete_rw(req
, res
, res2
, 0);
2489 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2491 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2493 if (kiocb
->ki_flags
& IOCB_WRITE
)
2494 kiocb_end_write(req
);
2495 if (unlikely(res
!= req
->result
)) {
2496 if (!(res
== -EAGAIN
&& io_rw_should_reissue(req
) &&
2497 io_resubmit_prep(req
))) {
2498 req_set_fail_links(req
);
2499 req
->flags
|= REQ_F_DONT_REISSUE
;
2503 WRITE_ONCE(req
->result
, res
);
2504 /* order with io_iopoll_complete() checking ->result */
2506 WRITE_ONCE(req
->iopoll_completed
, 1);
2510 * After the iocb has been issued, it's safe to be found on the poll list.
2511 * Adding the kiocb to the list AFTER submission ensures that we don't
2512 * find it from a io_do_iopoll() thread before the issuer is done
2513 * accessing the kiocb cookie.
2515 static void io_iopoll_req_issued(struct io_kiocb
*req
, bool in_async
)
2517 struct io_ring_ctx
*ctx
= req
->ctx
;
2520 * Track whether we have multiple files in our lists. This will impact
2521 * how we do polling eventually, not spinning if we're on potentially
2522 * different devices.
2524 if (list_empty(&ctx
->iopoll_list
)) {
2525 ctx
->poll_multi_file
= false;
2526 } else if (!ctx
->poll_multi_file
) {
2527 struct io_kiocb
*list_req
;
2529 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2531 if (list_req
->file
!= req
->file
)
2532 ctx
->poll_multi_file
= true;
2536 * For fast devices, IO may have already completed. If it has, add
2537 * it to the front so we find it first.
2539 if (READ_ONCE(req
->iopoll_completed
))
2540 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2542 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2545 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2546 * task context or in io worker task context. If current task context is
2547 * sq thread, we don't need to check whether should wake up sq thread.
2549 if (in_async
&& (ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2550 wq_has_sleeper(&ctx
->sq_data
->wait
))
2551 wake_up(&ctx
->sq_data
->wait
);
2554 static inline void io_state_file_put(struct io_submit_state
*state
)
2556 if (state
->file_refs
) {
2557 fput_many(state
->file
, state
->file_refs
);
2558 state
->file_refs
= 0;
2563 * Get as many references to a file as we have IOs left in this submission,
2564 * assuming most submissions are for one file, or at least that each file
2565 * has more than one submission.
2567 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2572 if (state
->file_refs
) {
2573 if (state
->fd
== fd
) {
2577 io_state_file_put(state
);
2579 state
->file
= fget_many(fd
, state
->ios_left
);
2580 if (unlikely(!state
->file
))
2584 state
->file_refs
= state
->ios_left
- 1;
2588 static bool io_bdev_nowait(struct block_device
*bdev
)
2590 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2594 * If we tracked the file through the SCM inflight mechanism, we could support
2595 * any file. For now, just ensure that anything potentially problematic is done
2598 static bool __io_file_supports_async(struct file
*file
, int rw
)
2600 umode_t mode
= file_inode(file
)->i_mode
;
2602 if (S_ISBLK(mode
)) {
2603 if (IS_ENABLED(CONFIG_BLOCK
) &&
2604 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2608 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2610 if (S_ISREG(mode
)) {
2611 if (IS_ENABLED(CONFIG_BLOCK
) &&
2612 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2613 file
->f_op
!= &io_uring_fops
)
2618 /* any ->read/write should understand O_NONBLOCK */
2619 if (file
->f_flags
& O_NONBLOCK
)
2622 if (!(file
->f_mode
& FMODE_NOWAIT
))
2626 return file
->f_op
->read_iter
!= NULL
;
2628 return file
->f_op
->write_iter
!= NULL
;
2631 static bool io_file_supports_async(struct io_kiocb
*req
, int rw
)
2633 if (rw
== READ
&& (req
->flags
& REQ_F_ASYNC_READ
))
2635 else if (rw
== WRITE
&& (req
->flags
& REQ_F_ASYNC_WRITE
))
2638 return __io_file_supports_async(req
->file
, rw
);
2641 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2643 struct io_ring_ctx
*ctx
= req
->ctx
;
2644 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2645 struct file
*file
= req
->file
;
2649 if (!(req
->flags
& REQ_F_ISREG
) && S_ISREG(file_inode(file
)->i_mode
))
2650 req
->flags
|= REQ_F_ISREG
;
2652 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2653 if (kiocb
->ki_pos
== -1 && !(file
->f_mode
& FMODE_STREAM
)) {
2654 req
->flags
|= REQ_F_CUR_POS
;
2655 kiocb
->ki_pos
= file
->f_pos
;
2657 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2658 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2659 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2663 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2664 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) || (file
->f_flags
& O_NONBLOCK
))
2665 req
->flags
|= REQ_F_NOWAIT
;
2667 ioprio
= READ_ONCE(sqe
->ioprio
);
2669 ret
= ioprio_check_cap(ioprio
);
2673 kiocb
->ki_ioprio
= ioprio
;
2675 kiocb
->ki_ioprio
= get_current_ioprio();
2677 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2678 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2679 !kiocb
->ki_filp
->f_op
->iopoll
)
2682 kiocb
->ki_flags
|= IOCB_HIPRI
;
2683 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2684 req
->iopoll_completed
= 0;
2686 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2688 kiocb
->ki_complete
= io_complete_rw
;
2691 if (req
->opcode
== IORING_OP_READ_FIXED
||
2692 req
->opcode
== IORING_OP_WRITE_FIXED
) {
2694 io_req_set_rsrc_node(req
);
2697 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2698 req
->rw
.len
= READ_ONCE(sqe
->len
);
2699 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2703 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2709 case -ERESTARTNOINTR
:
2710 case -ERESTARTNOHAND
:
2711 case -ERESTART_RESTARTBLOCK
:
2713 * We can't just restart the syscall, since previously
2714 * submitted sqes may already be in progress. Just fail this
2720 kiocb
->ki_complete(kiocb
, ret
, 0);
2724 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2725 unsigned int issue_flags
)
2727 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2728 struct io_async_rw
*io
= req
->async_data
;
2729 bool check_reissue
= kiocb
->ki_complete
== io_complete_rw
;
2731 /* add previously done IO, if any */
2732 if (io
&& io
->bytes_done
> 0) {
2734 ret
= io
->bytes_done
;
2736 ret
+= io
->bytes_done
;
2739 if (req
->flags
& REQ_F_CUR_POS
)
2740 req
->file
->f_pos
= kiocb
->ki_pos
;
2741 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2742 __io_complete_rw(req
, ret
, 0, issue_flags
);
2744 io_rw_done(kiocb
, ret
);
2746 if (check_reissue
&& req
->flags
& REQ_F_REISSUE
) {
2747 req
->flags
&= ~REQ_F_REISSUE
;
2748 if (io_resubmit_prep(req
)) {
2750 io_queue_async_work(req
);
2754 req_set_fail_links(req
);
2755 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2756 cflags
= io_put_rw_kbuf(req
);
2757 __io_req_complete(req
, issue_flags
, ret
, cflags
);
2762 static int __io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
2763 struct io_mapped_ubuf
*imu
)
2765 size_t len
= req
->rw
.len
;
2766 u64 buf_end
, buf_addr
= req
->rw
.addr
;
2769 if (unlikely(check_add_overflow(buf_addr
, (u64
)len
, &buf_end
)))
2771 /* not inside the mapped region */
2772 if (unlikely(buf_addr
< imu
->ubuf
|| buf_end
> imu
->ubuf_end
))
2776 * May not be a start of buffer, set size appropriately
2777 * and advance us to the beginning.
2779 offset
= buf_addr
- imu
->ubuf
;
2780 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
2784 * Don't use iov_iter_advance() here, as it's really slow for
2785 * using the latter parts of a big fixed buffer - it iterates
2786 * over each segment manually. We can cheat a bit here, because
2789 * 1) it's a BVEC iter, we set it up
2790 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2791 * first and last bvec
2793 * So just find our index, and adjust the iterator afterwards.
2794 * If the offset is within the first bvec (or the whole first
2795 * bvec, just use iov_iter_advance(). This makes it easier
2796 * since we can just skip the first segment, which may not
2797 * be PAGE_SIZE aligned.
2799 const struct bio_vec
*bvec
= imu
->bvec
;
2801 if (offset
<= bvec
->bv_len
) {
2802 iov_iter_advance(iter
, offset
);
2804 unsigned long seg_skip
;
2806 /* skip first vec */
2807 offset
-= bvec
->bv_len
;
2808 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
2810 iter
->bvec
= bvec
+ seg_skip
;
2811 iter
->nr_segs
-= seg_skip
;
2812 iter
->count
-= bvec
->bv_len
+ offset
;
2813 iter
->iov_offset
= offset
& ~PAGE_MASK
;
2820 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
2822 struct io_ring_ctx
*ctx
= req
->ctx
;
2823 struct io_mapped_ubuf
*imu
= req
->imu
;
2824 u16 index
, buf_index
= req
->buf_index
;
2827 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
2829 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
2830 imu
= READ_ONCE(ctx
->user_bufs
[index
]);
2833 return __io_import_fixed(req
, rw
, iter
, imu
);
2836 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2839 mutex_unlock(&ctx
->uring_lock
);
2842 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2845 * "Normal" inline submissions always hold the uring_lock, since we
2846 * grab it from the system call. Same is true for the SQPOLL offload.
2847 * The only exception is when we've detached the request and issue it
2848 * from an async worker thread, grab the lock for that case.
2851 mutex_lock(&ctx
->uring_lock
);
2854 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
2855 int bgid
, struct io_buffer
*kbuf
,
2858 struct io_buffer
*head
;
2860 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2863 io_ring_submit_lock(req
->ctx
, needs_lock
);
2865 lockdep_assert_held(&req
->ctx
->uring_lock
);
2867 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
2869 if (!list_empty(&head
->list
)) {
2870 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
2872 list_del(&kbuf
->list
);
2875 xa_erase(&req
->ctx
->io_buffers
, bgid
);
2877 if (*len
> kbuf
->len
)
2880 kbuf
= ERR_PTR(-ENOBUFS
);
2883 io_ring_submit_unlock(req
->ctx
, needs_lock
);
2888 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
2891 struct io_buffer
*kbuf
;
2894 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2895 bgid
= req
->buf_index
;
2896 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
2899 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
2900 req
->flags
|= REQ_F_BUFFER_SELECTED
;
2901 return u64_to_user_ptr(kbuf
->addr
);
2904 #ifdef CONFIG_COMPAT
2905 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
2908 struct compat_iovec __user
*uiov
;
2909 compat_ssize_t clen
;
2913 uiov
= u64_to_user_ptr(req
->rw
.addr
);
2914 if (!access_ok(uiov
, sizeof(*uiov
)))
2916 if (__get_user(clen
, &uiov
->iov_len
))
2922 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2924 return PTR_ERR(buf
);
2925 iov
[0].iov_base
= buf
;
2926 iov
[0].iov_len
= (compat_size_t
) len
;
2931 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2934 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
2938 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
2941 len
= iov
[0].iov_len
;
2944 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2946 return PTR_ERR(buf
);
2947 iov
[0].iov_base
= buf
;
2948 iov
[0].iov_len
= len
;
2952 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2955 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
2956 struct io_buffer
*kbuf
;
2958 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2959 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
2960 iov
[0].iov_len
= kbuf
->len
;
2963 if (req
->rw
.len
!= 1)
2966 #ifdef CONFIG_COMPAT
2967 if (req
->ctx
->compat
)
2968 return io_compat_import(req
, iov
, needs_lock
);
2971 return __io_iov_buffer_select(req
, iov
, needs_lock
);
2974 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
2975 struct iov_iter
*iter
, bool needs_lock
)
2977 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
2978 size_t sqe_len
= req
->rw
.len
;
2979 u8 opcode
= req
->opcode
;
2982 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
2984 return io_import_fixed(req
, rw
, iter
);
2987 /* buffer index only valid with fixed read/write, or buffer select */
2988 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
2991 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
2992 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
2993 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
2995 return PTR_ERR(buf
);
2996 req
->rw
.len
= sqe_len
;
2999 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3004 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3005 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3007 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3012 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3016 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3018 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3022 * For files that don't have ->read_iter() and ->write_iter(), handle them
3023 * by looping over ->read() or ->write() manually.
3025 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3027 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3028 struct file
*file
= req
->file
;
3032 * Don't support polled IO through this interface, and we can't
3033 * support non-blocking either. For the latter, this just causes
3034 * the kiocb to be handled from an async context.
3036 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3038 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3041 while (iov_iter_count(iter
)) {
3045 if (!iov_iter_is_bvec(iter
)) {
3046 iovec
= iov_iter_iovec(iter
);
3048 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3049 iovec
.iov_len
= req
->rw
.len
;
3053 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3054 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3056 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3057 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3066 if (nr
!= iovec
.iov_len
)
3070 iov_iter_advance(iter
, nr
);
3076 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3077 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3079 struct io_async_rw
*rw
= req
->async_data
;
3081 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3082 rw
->free_iovec
= iovec
;
3084 /* can only be fixed buffers, no need to do anything */
3085 if (iov_iter_is_bvec(iter
))
3088 unsigned iov_off
= 0;
3090 rw
->iter
.iov
= rw
->fast_iov
;
3091 if (iter
->iov
!= fast_iov
) {
3092 iov_off
= iter
->iov
- fast_iov
;
3093 rw
->iter
.iov
+= iov_off
;
3095 if (rw
->fast_iov
!= fast_iov
)
3096 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3097 sizeof(struct iovec
) * iter
->nr_segs
);
3099 req
->flags
|= REQ_F_NEED_CLEANUP
;
3103 static inline int io_alloc_async_data(struct io_kiocb
*req
)
3105 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3106 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3107 return req
->async_data
== NULL
;
3110 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3111 const struct iovec
*fast_iov
,
3112 struct iov_iter
*iter
, bool force
)
3114 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_setup
)
3116 if (!req
->async_data
) {
3117 if (io_alloc_async_data(req
)) {
3122 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3127 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3129 struct io_async_rw
*iorw
= req
->async_data
;
3130 struct iovec
*iov
= iorw
->fast_iov
;
3133 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3134 if (unlikely(ret
< 0))
3137 iorw
->bytes_done
= 0;
3138 iorw
->free_iovec
= iov
;
3140 req
->flags
|= REQ_F_NEED_CLEANUP
;
3144 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3146 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3148 return io_prep_rw(req
, sqe
);
3152 * This is our waitqueue callback handler, registered through lock_page_async()
3153 * when we initially tried to do the IO with the iocb armed our waitqueue.
3154 * This gets called when the page is unlocked, and we generally expect that to
3155 * happen when the page IO is completed and the page is now uptodate. This will
3156 * queue a task_work based retry of the operation, attempting to copy the data
3157 * again. If the latter fails because the page was NOT uptodate, then we will
3158 * do a thread based blocking retry of the operation. That's the unexpected
3161 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3162 int sync
, void *arg
)
3164 struct wait_page_queue
*wpq
;
3165 struct io_kiocb
*req
= wait
->private;
3166 struct wait_page_key
*key
= arg
;
3168 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3170 if (!wake_page_match(wpq
, key
))
3173 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3174 list_del_init(&wait
->entry
);
3176 /* submit ref gets dropped, acquire a new one */
3178 io_req_task_queue(req
);
3183 * This controls whether a given IO request should be armed for async page
3184 * based retry. If we return false here, the request is handed to the async
3185 * worker threads for retry. If we're doing buffered reads on a regular file,
3186 * we prepare a private wait_page_queue entry and retry the operation. This
3187 * will either succeed because the page is now uptodate and unlocked, or it
3188 * will register a callback when the page is unlocked at IO completion. Through
3189 * that callback, io_uring uses task_work to setup a retry of the operation.
3190 * That retry will attempt the buffered read again. The retry will generally
3191 * succeed, or in rare cases where it fails, we then fall back to using the
3192 * async worker threads for a blocking retry.
3194 static bool io_rw_should_retry(struct io_kiocb
*req
)
3196 struct io_async_rw
*rw
= req
->async_data
;
3197 struct wait_page_queue
*wait
= &rw
->wpq
;
3198 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3200 /* never retry for NOWAIT, we just complete with -EAGAIN */
3201 if (req
->flags
& REQ_F_NOWAIT
)
3204 /* Only for buffered IO */
3205 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3209 * just use poll if we can, and don't attempt if the fs doesn't
3210 * support callback based unlocks
3212 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3215 wait
->wait
.func
= io_async_buf_func
;
3216 wait
->wait
.private = req
;
3217 wait
->wait
.flags
= 0;
3218 INIT_LIST_HEAD(&wait
->wait
.entry
);
3219 kiocb
->ki_flags
|= IOCB_WAITQ
;
3220 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3221 kiocb
->ki_waitq
= wait
;
3225 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3227 if (req
->file
->f_op
->read_iter
)
3228 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3229 else if (req
->file
->f_op
->read
)
3230 return loop_rw_iter(READ
, req
, iter
);
3235 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3237 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3238 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3239 struct iov_iter __iter
, *iter
= &__iter
;
3240 struct io_async_rw
*rw
= req
->async_data
;
3241 ssize_t io_size
, ret
, ret2
;
3242 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3248 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3252 io_size
= iov_iter_count(iter
);
3253 req
->result
= io_size
;
3255 /* Ensure we clear previously set non-block flag */
3256 if (!force_nonblock
)
3257 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3259 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3261 /* If the file doesn't support async, just async punt */
3262 if (force_nonblock
&& !io_file_supports_async(req
, READ
)) {
3263 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3264 return ret
?: -EAGAIN
;
3267 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3268 if (unlikely(ret
)) {
3273 ret
= io_iter_do_read(req
, iter
);
3275 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
3276 req
->flags
&= ~REQ_F_REISSUE
;
3277 /* IOPOLL retry should happen for io-wq threads */
3278 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3280 /* no retry on NONBLOCK nor RWF_NOWAIT */
3281 if (req
->flags
& REQ_F_NOWAIT
)
3283 /* some cases will consume bytes even on error returns */
3284 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3286 } else if (ret
== -EIOCBQUEUED
) {
3288 } else if (ret
<= 0 || ret
== io_size
|| !force_nonblock
||
3289 (req
->flags
& REQ_F_NOWAIT
) || !(req
->flags
& REQ_F_ISREG
)) {
3290 /* read all, failed, already did sync or don't want to retry */
3294 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3299 rw
= req
->async_data
;
3300 /* now use our persistent iterator, if we aren't already */
3305 rw
->bytes_done
+= ret
;
3306 /* if we can retry, do so with the callbacks armed */
3307 if (!io_rw_should_retry(req
)) {
3308 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3313 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3314 * we get -EIOCBQUEUED, then we'll get a notification when the
3315 * desired page gets unlocked. We can also get a partial read
3316 * here, and if we do, then just retry at the new offset.
3318 ret
= io_iter_do_read(req
, iter
);
3319 if (ret
== -EIOCBQUEUED
)
3321 /* we got some bytes, but not all. retry. */
3322 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3323 } while (ret
> 0 && ret
< io_size
);
3325 kiocb_done(kiocb
, ret
, issue_flags
);
3327 /* it's faster to check here then delegate to kfree */
3333 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3335 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3337 return io_prep_rw(req
, sqe
);
3340 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3342 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3343 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3344 struct iov_iter __iter
, *iter
= &__iter
;
3345 struct io_async_rw
*rw
= req
->async_data
;
3346 ssize_t ret
, ret2
, io_size
;
3347 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3353 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3357 io_size
= iov_iter_count(iter
);
3358 req
->result
= io_size
;
3360 /* Ensure we clear previously set non-block flag */
3361 if (!force_nonblock
)
3362 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3364 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3366 /* If the file doesn't support async, just async punt */
3367 if (force_nonblock
&& !io_file_supports_async(req
, WRITE
))
3370 /* file path doesn't support NOWAIT for non-direct_IO */
3371 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3372 (req
->flags
& REQ_F_ISREG
))
3375 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3380 * Open-code file_start_write here to grab freeze protection,
3381 * which will be released by another thread in
3382 * io_complete_rw(). Fool lockdep by telling it the lock got
3383 * released so that it doesn't complain about the held lock when
3384 * we return to userspace.
3386 if (req
->flags
& REQ_F_ISREG
) {
3387 sb_start_write(file_inode(req
->file
)->i_sb
);
3388 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3391 kiocb
->ki_flags
|= IOCB_WRITE
;
3393 if (req
->file
->f_op
->write_iter
)
3394 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3395 else if (req
->file
->f_op
->write
)
3396 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3400 if (req
->flags
& REQ_F_REISSUE
) {
3401 req
->flags
&= ~REQ_F_REISSUE
;
3406 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3407 * retry them without IOCB_NOWAIT.
3409 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3411 /* no retry on NONBLOCK nor RWF_NOWAIT */
3412 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3414 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3415 /* IOPOLL retry should happen for io-wq threads */
3416 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3419 kiocb_done(kiocb
, ret2
, issue_flags
);
3422 /* some cases will consume bytes even on error returns */
3423 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3424 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3425 return ret
?: -EAGAIN
;
3428 /* it's reportedly faster than delegating the null check to kfree() */
3434 static int io_renameat_prep(struct io_kiocb
*req
,
3435 const struct io_uring_sqe
*sqe
)
3437 struct io_rename
*ren
= &req
->rename
;
3438 const char __user
*oldf
, *newf
;
3440 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3443 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3444 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3445 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3446 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3447 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3449 ren
->oldpath
= getname(oldf
);
3450 if (IS_ERR(ren
->oldpath
))
3451 return PTR_ERR(ren
->oldpath
);
3453 ren
->newpath
= getname(newf
);
3454 if (IS_ERR(ren
->newpath
)) {
3455 putname(ren
->oldpath
);
3456 return PTR_ERR(ren
->newpath
);
3459 req
->flags
|= REQ_F_NEED_CLEANUP
;
3463 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3465 struct io_rename
*ren
= &req
->rename
;
3468 if (issue_flags
& IO_URING_F_NONBLOCK
)
3471 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3472 ren
->newpath
, ren
->flags
);
3474 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3476 req_set_fail_links(req
);
3477 io_req_complete(req
, ret
);
3481 static int io_unlinkat_prep(struct io_kiocb
*req
,
3482 const struct io_uring_sqe
*sqe
)
3484 struct io_unlink
*un
= &req
->unlink
;
3485 const char __user
*fname
;
3487 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3490 un
->dfd
= READ_ONCE(sqe
->fd
);
3492 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3493 if (un
->flags
& ~AT_REMOVEDIR
)
3496 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3497 un
->filename
= getname(fname
);
3498 if (IS_ERR(un
->filename
))
3499 return PTR_ERR(un
->filename
);
3501 req
->flags
|= REQ_F_NEED_CLEANUP
;
3505 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3507 struct io_unlink
*un
= &req
->unlink
;
3510 if (issue_flags
& IO_URING_F_NONBLOCK
)
3513 if (un
->flags
& AT_REMOVEDIR
)
3514 ret
= do_rmdir(un
->dfd
, un
->filename
);
3516 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3518 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3520 req_set_fail_links(req
);
3521 io_req_complete(req
, ret
);
3525 static int io_shutdown_prep(struct io_kiocb
*req
,
3526 const struct io_uring_sqe
*sqe
)
3528 #if defined(CONFIG_NET)
3529 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3531 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3535 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3542 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3544 #if defined(CONFIG_NET)
3545 struct socket
*sock
;
3548 if (issue_flags
& IO_URING_F_NONBLOCK
)
3551 sock
= sock_from_file(req
->file
);
3552 if (unlikely(!sock
))
3555 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3557 req_set_fail_links(req
);
3558 io_req_complete(req
, ret
);
3565 static int __io_splice_prep(struct io_kiocb
*req
,
3566 const struct io_uring_sqe
*sqe
)
3568 struct io_splice
* sp
= &req
->splice
;
3569 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3571 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3575 sp
->len
= READ_ONCE(sqe
->len
);
3576 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3578 if (unlikely(sp
->flags
& ~valid_flags
))
3581 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3582 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3585 req
->flags
|= REQ_F_NEED_CLEANUP
;
3589 static int io_tee_prep(struct io_kiocb
*req
,
3590 const struct io_uring_sqe
*sqe
)
3592 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3594 return __io_splice_prep(req
, sqe
);
3597 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
3599 struct io_splice
*sp
= &req
->splice
;
3600 struct file
*in
= sp
->file_in
;
3601 struct file
*out
= sp
->file_out
;
3602 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3605 if (issue_flags
& IO_URING_F_NONBLOCK
)
3608 ret
= do_tee(in
, out
, sp
->len
, flags
);
3610 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
3612 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3615 req_set_fail_links(req
);
3616 io_req_complete(req
, ret
);
3620 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3622 struct io_splice
* sp
= &req
->splice
;
3624 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3625 sp
->off_out
= READ_ONCE(sqe
->off
);
3626 return __io_splice_prep(req
, sqe
);
3629 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
3631 struct io_splice
*sp
= &req
->splice
;
3632 struct file
*in
= sp
->file_in
;
3633 struct file
*out
= sp
->file_out
;
3634 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3635 loff_t
*poff_in
, *poff_out
;
3638 if (issue_flags
& IO_URING_F_NONBLOCK
)
3641 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3642 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3645 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3647 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
3649 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3652 req_set_fail_links(req
);
3653 io_req_complete(req
, ret
);
3658 * IORING_OP_NOP just posts a completion event, nothing else.
3660 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
3662 struct io_ring_ctx
*ctx
= req
->ctx
;
3664 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3667 __io_req_complete(req
, issue_flags
, 0, 0);
3671 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3673 struct io_ring_ctx
*ctx
= req
->ctx
;
3678 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3680 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3683 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3684 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3687 req
->sync
.off
= READ_ONCE(sqe
->off
);
3688 req
->sync
.len
= READ_ONCE(sqe
->len
);
3692 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
3694 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3697 /* fsync always requires a blocking context */
3698 if (issue_flags
& IO_URING_F_NONBLOCK
)
3701 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3702 end
> 0 ? end
: LLONG_MAX
,
3703 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3705 req_set_fail_links(req
);
3706 io_req_complete(req
, ret
);
3710 static int io_fallocate_prep(struct io_kiocb
*req
,
3711 const struct io_uring_sqe
*sqe
)
3713 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3715 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3718 req
->sync
.off
= READ_ONCE(sqe
->off
);
3719 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3720 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3724 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
3728 /* fallocate always requiring blocking context */
3729 if (issue_flags
& IO_URING_F_NONBLOCK
)
3731 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3734 req_set_fail_links(req
);
3735 io_req_complete(req
, ret
);
3739 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3741 const char __user
*fname
;
3744 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3746 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3749 /* open.how should be already initialised */
3750 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
3751 req
->open
.how
.flags
|= O_LARGEFILE
;
3753 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3754 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3755 req
->open
.filename
= getname(fname
);
3756 if (IS_ERR(req
->open
.filename
)) {
3757 ret
= PTR_ERR(req
->open
.filename
);
3758 req
->open
.filename
= NULL
;
3761 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
3762 req
->flags
|= REQ_F_NEED_CLEANUP
;
3766 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3770 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3772 mode
= READ_ONCE(sqe
->len
);
3773 flags
= READ_ONCE(sqe
->open_flags
);
3774 req
->open
.how
= build_open_how(flags
, mode
);
3775 return __io_openat_prep(req
, sqe
);
3778 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3780 struct open_how __user
*how
;
3784 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3786 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3787 len
= READ_ONCE(sqe
->len
);
3788 if (len
< OPEN_HOW_SIZE_VER0
)
3791 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
3796 return __io_openat_prep(req
, sqe
);
3799 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
3801 struct open_flags op
;
3804 bool resolve_nonblock
;
3807 ret
= build_open_flags(&req
->open
.how
, &op
);
3810 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
3811 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
3812 if (issue_flags
& IO_URING_F_NONBLOCK
) {
3814 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3815 * it'll always -EAGAIN
3817 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
3819 op
.lookup_flags
|= LOOKUP_CACHED
;
3820 op
.open_flag
|= O_NONBLOCK
;
3823 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
3827 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
3828 /* only retry if RESOLVE_CACHED wasn't already set by application */
3829 if ((!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)) &&
3830 file
== ERR_PTR(-EAGAIN
)) {
3832 * We could hang on to this 'fd', but seems like marginal
3833 * gain for something that is now known to be a slower path.
3834 * So just put it, and we'll get a new one when we retry.
3842 ret
= PTR_ERR(file
);
3844 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
3845 file
->f_flags
&= ~O_NONBLOCK
;
3846 fsnotify_open(file
);
3847 fd_install(ret
, file
);
3850 putname(req
->open
.filename
);
3851 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3853 req_set_fail_links(req
);
3854 __io_req_complete(req
, issue_flags
, ret
, 0);
3858 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
3860 return io_openat2(req
, issue_flags
);
3863 static int io_remove_buffers_prep(struct io_kiocb
*req
,
3864 const struct io_uring_sqe
*sqe
)
3866 struct io_provide_buf
*p
= &req
->pbuf
;
3869 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
3872 tmp
= READ_ONCE(sqe
->fd
);
3873 if (!tmp
|| tmp
> USHRT_MAX
)
3876 memset(p
, 0, sizeof(*p
));
3878 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3882 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
3883 int bgid
, unsigned nbufs
)
3887 /* shouldn't happen */
3891 /* the head kbuf is the list itself */
3892 while (!list_empty(&buf
->list
)) {
3893 struct io_buffer
*nxt
;
3895 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
3896 list_del(&nxt
->list
);
3903 xa_erase(&ctx
->io_buffers
, bgid
);
3908 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
3910 struct io_provide_buf
*p
= &req
->pbuf
;
3911 struct io_ring_ctx
*ctx
= req
->ctx
;
3912 struct io_buffer
*head
;
3914 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3916 io_ring_submit_lock(ctx
, !force_nonblock
);
3918 lockdep_assert_held(&ctx
->uring_lock
);
3921 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
3923 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
3925 req_set_fail_links(req
);
3927 /* complete before unlock, IOPOLL may need the lock */
3928 __io_req_complete(req
, issue_flags
, ret
, 0);
3929 io_ring_submit_unlock(ctx
, !force_nonblock
);
3933 static int io_provide_buffers_prep(struct io_kiocb
*req
,
3934 const struct io_uring_sqe
*sqe
)
3936 unsigned long size
, tmp_check
;
3937 struct io_provide_buf
*p
= &req
->pbuf
;
3940 if (sqe
->ioprio
|| sqe
->rw_flags
)
3943 tmp
= READ_ONCE(sqe
->fd
);
3944 if (!tmp
|| tmp
> USHRT_MAX
)
3947 p
->addr
= READ_ONCE(sqe
->addr
);
3948 p
->len
= READ_ONCE(sqe
->len
);
3950 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
3953 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
3956 size
= (unsigned long)p
->len
* p
->nbufs
;
3957 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
3960 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3961 tmp
= READ_ONCE(sqe
->off
);
3962 if (tmp
> USHRT_MAX
)
3968 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
3970 struct io_buffer
*buf
;
3971 u64 addr
= pbuf
->addr
;
3972 int i
, bid
= pbuf
->bid
;
3974 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
3975 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
3980 buf
->len
= pbuf
->len
;
3985 INIT_LIST_HEAD(&buf
->list
);
3988 list_add_tail(&buf
->list
, &(*head
)->list
);
3992 return i
? i
: -ENOMEM
;
3995 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
3997 struct io_provide_buf
*p
= &req
->pbuf
;
3998 struct io_ring_ctx
*ctx
= req
->ctx
;
3999 struct io_buffer
*head
, *list
;
4001 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4003 io_ring_submit_lock(ctx
, !force_nonblock
);
4005 lockdep_assert_held(&ctx
->uring_lock
);
4007 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4009 ret
= io_add_buffers(p
, &head
);
4010 if (ret
>= 0 && !list
) {
4011 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4013 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4016 req_set_fail_links(req
);
4017 /* complete before unlock, IOPOLL may need the lock */
4018 __io_req_complete(req
, issue_flags
, ret
, 0);
4019 io_ring_submit_unlock(ctx
, !force_nonblock
);
4023 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4024 const struct io_uring_sqe
*sqe
)
4026 #if defined(CONFIG_EPOLL)
4027 if (sqe
->ioprio
|| sqe
->buf_index
)
4029 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4032 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4033 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4034 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4036 if (ep_op_has_event(req
->epoll
.op
)) {
4037 struct epoll_event __user
*ev
;
4039 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4040 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4050 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4052 #if defined(CONFIG_EPOLL)
4053 struct io_epoll
*ie
= &req
->epoll
;
4055 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4057 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4058 if (force_nonblock
&& ret
== -EAGAIN
)
4062 req_set_fail_links(req
);
4063 __io_req_complete(req
, issue_flags
, ret
, 0);
4070 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4072 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4073 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4075 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4078 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4079 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4080 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4087 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4089 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4090 struct io_madvise
*ma
= &req
->madvise
;
4093 if (issue_flags
& IO_URING_F_NONBLOCK
)
4096 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4098 req_set_fail_links(req
);
4099 io_req_complete(req
, ret
);
4106 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4108 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4110 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4113 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4114 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4115 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4119 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4121 struct io_fadvise
*fa
= &req
->fadvise
;
4124 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4125 switch (fa
->advice
) {
4126 case POSIX_FADV_NORMAL
:
4127 case POSIX_FADV_RANDOM
:
4128 case POSIX_FADV_SEQUENTIAL
:
4135 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4137 req_set_fail_links(req
);
4138 __io_req_complete(req
, issue_flags
, ret
, 0);
4142 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4144 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4146 if (sqe
->ioprio
|| sqe
->buf_index
)
4148 if (req
->flags
& REQ_F_FIXED_FILE
)
4151 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4152 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4153 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4154 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4155 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4160 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4162 struct io_statx
*ctx
= &req
->statx
;
4165 if (issue_flags
& IO_URING_F_NONBLOCK
)
4168 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4172 req_set_fail_links(req
);
4173 io_req_complete(req
, ret
);
4177 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4179 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4181 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4182 sqe
->rw_flags
|| sqe
->buf_index
)
4184 if (req
->flags
& REQ_F_FIXED_FILE
)
4187 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4191 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4193 struct files_struct
*files
= current
->files
;
4194 struct io_close
*close
= &req
->close
;
4195 struct fdtable
*fdt
;
4196 struct file
*file
= NULL
;
4199 spin_lock(&files
->file_lock
);
4200 fdt
= files_fdtable(files
);
4201 if (close
->fd
>= fdt
->max_fds
) {
4202 spin_unlock(&files
->file_lock
);
4205 file
= fdt
->fd
[close
->fd
];
4206 if (!file
|| file
->f_op
== &io_uring_fops
) {
4207 spin_unlock(&files
->file_lock
);
4212 /* if the file has a flush method, be safe and punt to async */
4213 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4214 spin_unlock(&files
->file_lock
);
4218 ret
= __close_fd_get_file(close
->fd
, &file
);
4219 spin_unlock(&files
->file_lock
);
4226 /* No ->flush() or already async, safely close from here */
4227 ret
= filp_close(file
, current
->files
);
4230 req_set_fail_links(req
);
4233 __io_req_complete(req
, issue_flags
, ret
, 0);
4237 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4239 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
, unsigned int issue_flags
)
4256 /* sync_file_range always requires a blocking context */
4257 if (issue_flags
& IO_URING_F_NONBLOCK
)
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 kfree(kmsg
->free_iov
);
4280 async_msg
= req
->async_data
;
4281 req
->flags
|= REQ_F_NEED_CLEANUP
;
4282 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4283 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4284 /* if were using fast_iov, set it to the new one */
4285 if (!async_msg
->free_iov
)
4286 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4291 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4292 struct io_async_msghdr
*iomsg
)
4294 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4295 iomsg
->free_iov
= iomsg
->fast_iov
;
4296 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4297 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4300 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4304 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4306 req
->flags
|= REQ_F_NEED_CLEANUP
;
4310 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4312 struct io_sr_msg
*sr
= &req
->sr_msg
;
4314 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4317 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4318 sr
->len
= READ_ONCE(sqe
->len
);
4319 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4320 if (sr
->msg_flags
& MSG_DONTWAIT
)
4321 req
->flags
|= REQ_F_NOWAIT
;
4323 #ifdef CONFIG_COMPAT
4324 if (req
->ctx
->compat
)
4325 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4330 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4332 struct io_async_msghdr iomsg
, *kmsg
;
4333 struct socket
*sock
;
4338 sock
= sock_from_file(req
->file
);
4339 if (unlikely(!sock
))
4342 kmsg
= req
->async_data
;
4344 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4350 flags
= req
->sr_msg
.msg_flags
;
4351 if (issue_flags
& IO_URING_F_NONBLOCK
)
4352 flags
|= MSG_DONTWAIT
;
4353 if (flags
& MSG_WAITALL
)
4354 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4356 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4357 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4358 return io_setup_async_msg(req
, kmsg
);
4359 if (ret
== -ERESTARTSYS
)
4362 /* fast path, check for non-NULL to avoid function call */
4364 kfree(kmsg
->free_iov
);
4365 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4367 req_set_fail_links(req
);
4368 __io_req_complete(req
, issue_flags
, ret
, 0);
4372 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4374 struct io_sr_msg
*sr
= &req
->sr_msg
;
4377 struct socket
*sock
;
4382 sock
= sock_from_file(req
->file
);
4383 if (unlikely(!sock
))
4386 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4390 msg
.msg_name
= NULL
;
4391 msg
.msg_control
= NULL
;
4392 msg
.msg_controllen
= 0;
4393 msg
.msg_namelen
= 0;
4395 flags
= req
->sr_msg
.msg_flags
;
4396 if (issue_flags
& IO_URING_F_NONBLOCK
)
4397 flags
|= MSG_DONTWAIT
;
4398 if (flags
& MSG_WAITALL
)
4399 min_ret
= iov_iter_count(&msg
.msg_iter
);
4401 msg
.msg_flags
= flags
;
4402 ret
= sock_sendmsg(sock
, &msg
);
4403 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4405 if (ret
== -ERESTARTSYS
)
4409 req_set_fail_links(req
);
4410 __io_req_complete(req
, issue_flags
, ret
, 0);
4414 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4415 struct io_async_msghdr
*iomsg
)
4417 struct io_sr_msg
*sr
= &req
->sr_msg
;
4418 struct iovec __user
*uiov
;
4422 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4423 &iomsg
->uaddr
, &uiov
, &iov_len
);
4427 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4430 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4432 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4433 iomsg
->free_iov
= NULL
;
4435 iomsg
->free_iov
= iomsg
->fast_iov
;
4436 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4437 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4446 #ifdef CONFIG_COMPAT
4447 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4448 struct io_async_msghdr
*iomsg
)
4450 struct io_sr_msg
*sr
= &req
->sr_msg
;
4451 struct compat_iovec __user
*uiov
;
4456 ret
= __get_compat_msghdr(&iomsg
->msg
, sr
->umsg_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
))
4474 iomsg
->free_iov
= NULL
;
4476 iomsg
->free_iov
= iomsg
->fast_iov
;
4477 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4478 UIO_FASTIOV
, &iomsg
->free_iov
,
4479 &iomsg
->msg
.msg_iter
, true);
4488 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4489 struct io_async_msghdr
*iomsg
)
4491 iomsg
->msg
.msg_name
= &iomsg
->addr
;
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_async(struct io_kiocb
*req
)
4525 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
4527 req
->flags
|= REQ_F_NEED_CLEANUP
;
4531 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4533 struct io_sr_msg
*sr
= &req
->sr_msg
;
4535 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4538 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4539 sr
->len
= READ_ONCE(sqe
->len
);
4540 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4541 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4542 if (sr
->msg_flags
& MSG_DONTWAIT
)
4543 req
->flags
|= REQ_F_NOWAIT
;
4545 #ifdef CONFIG_COMPAT
4546 if (req
->ctx
->compat
)
4547 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4552 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4554 struct io_async_msghdr iomsg
, *kmsg
;
4555 struct socket
*sock
;
4556 struct io_buffer
*kbuf
;
4559 int ret
, cflags
= 0;
4560 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4562 sock
= sock_from_file(req
->file
);
4563 if (unlikely(!sock
))
4566 kmsg
= req
->async_data
;
4568 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4574 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4575 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4577 return PTR_ERR(kbuf
);
4578 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4579 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
4580 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
4581 1, req
->sr_msg
.len
);
4584 flags
= req
->sr_msg
.msg_flags
;
4586 flags
|= MSG_DONTWAIT
;
4587 if (flags
& MSG_WAITALL
)
4588 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4590 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4591 kmsg
->uaddr
, flags
);
4592 if (force_nonblock
&& ret
== -EAGAIN
)
4593 return io_setup_async_msg(req
, kmsg
);
4594 if (ret
== -ERESTARTSYS
)
4597 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4598 cflags
= io_put_recv_kbuf(req
);
4599 /* fast path, check for non-NULL to avoid function call */
4601 kfree(kmsg
->free_iov
);
4602 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4603 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4604 req_set_fail_links(req
);
4605 __io_req_complete(req
, issue_flags
, ret
, cflags
);
4609 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
4611 struct io_buffer
*kbuf
;
4612 struct io_sr_msg
*sr
= &req
->sr_msg
;
4614 void __user
*buf
= sr
->buf
;
4615 struct socket
*sock
;
4619 int ret
, cflags
= 0;
4620 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4622 sock
= sock_from_file(req
->file
);
4623 if (unlikely(!sock
))
4626 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4627 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4629 return PTR_ERR(kbuf
);
4630 buf
= u64_to_user_ptr(kbuf
->addr
);
4633 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4637 msg
.msg_name
= NULL
;
4638 msg
.msg_control
= NULL
;
4639 msg
.msg_controllen
= 0;
4640 msg
.msg_namelen
= 0;
4641 msg
.msg_iocb
= NULL
;
4644 flags
= req
->sr_msg
.msg_flags
;
4646 flags
|= MSG_DONTWAIT
;
4647 if (flags
& MSG_WAITALL
)
4648 min_ret
= iov_iter_count(&msg
.msg_iter
);
4650 ret
= sock_recvmsg(sock
, &msg
, flags
);
4651 if (force_nonblock
&& ret
== -EAGAIN
)
4653 if (ret
== -ERESTARTSYS
)
4656 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4657 cflags
= io_put_recv_kbuf(req
);
4658 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4659 req_set_fail_links(req
);
4660 __io_req_complete(req
, issue_flags
, ret
, cflags
);
4664 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4666 struct io_accept
*accept
= &req
->accept
;
4668 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4670 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4673 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4674 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4675 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4676 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4680 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
4682 struct io_accept
*accept
= &req
->accept
;
4683 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4684 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4687 if (req
->file
->f_flags
& O_NONBLOCK
)
4688 req
->flags
|= REQ_F_NOWAIT
;
4690 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4691 accept
->addr_len
, accept
->flags
,
4693 if (ret
== -EAGAIN
&& force_nonblock
)
4696 if (ret
== -ERESTARTSYS
)
4698 req_set_fail_links(req
);
4700 __io_req_complete(req
, issue_flags
, ret
, 0);
4704 static int io_connect_prep_async(struct io_kiocb
*req
)
4706 struct io_async_connect
*io
= req
->async_data
;
4707 struct io_connect
*conn
= &req
->connect
;
4709 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
4712 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4714 struct io_connect
*conn
= &req
->connect
;
4716 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4718 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4721 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4722 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4726 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
4728 struct io_async_connect __io
, *io
;
4729 unsigned file_flags
;
4731 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4733 if (req
->async_data
) {
4734 io
= req
->async_data
;
4736 ret
= move_addr_to_kernel(req
->connect
.addr
,
4737 req
->connect
.addr_len
,
4744 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4746 ret
= __sys_connect_file(req
->file
, &io
->address
,
4747 req
->connect
.addr_len
, file_flags
);
4748 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4749 if (req
->async_data
)
4751 if (io_alloc_async_data(req
)) {
4755 memcpy(req
->async_data
, &__io
, sizeof(__io
));
4758 if (ret
== -ERESTARTSYS
)
4762 req_set_fail_links(req
);
4763 __io_req_complete(req
, issue_flags
, ret
, 0);
4766 #else /* !CONFIG_NET */
4767 #define IO_NETOP_FN(op) \
4768 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4770 return -EOPNOTSUPP; \
4773 #define IO_NETOP_PREP(op) \
4775 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4777 return -EOPNOTSUPP; \
4780 #define IO_NETOP_PREP_ASYNC(op) \
4782 static int io_##op##_prep_async(struct io_kiocb *req) \
4784 return -EOPNOTSUPP; \
4787 IO_NETOP_PREP_ASYNC(sendmsg
);
4788 IO_NETOP_PREP_ASYNC(recvmsg
);
4789 IO_NETOP_PREP_ASYNC(connect
);
4790 IO_NETOP_PREP(accept
);
4793 #endif /* CONFIG_NET */
4795 struct io_poll_table
{
4796 struct poll_table_struct pt
;
4797 struct io_kiocb
*req
;
4801 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
4802 __poll_t mask
, task_work_func_t func
)
4806 /* for instances that support it check for an event match first: */
4807 if (mask
&& !(mask
& poll
->events
))
4810 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
4812 list_del_init(&poll
->wait
.entry
);
4815 req
->task_work
.func
= func
;
4818 * If this fails, then the task is exiting. When a task exits, the
4819 * work gets canceled, so just cancel this request as well instead
4820 * of executing it. We can't safely execute it anyway, as we may not
4821 * have the needed state needed for it anyway.
4823 ret
= io_req_task_work_add(req
);
4824 if (unlikely(ret
)) {
4825 WRITE_ONCE(poll
->canceled
, true);
4826 io_req_task_work_add_fallback(req
, func
);
4831 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
4832 __acquires(&req
->ctx
->completion_lock
)
4834 struct io_ring_ctx
*ctx
= req
->ctx
;
4836 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4837 struct poll_table_struct pt
= { ._key
= poll
->events
};
4839 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
4842 spin_lock_irq(&ctx
->completion_lock
);
4843 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4844 add_wait_queue(poll
->head
, &poll
->wait
);
4851 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
4853 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4854 if (req
->opcode
== IORING_OP_POLL_ADD
)
4855 return req
->async_data
;
4856 return req
->apoll
->double_poll
;
4859 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
4861 if (req
->opcode
== IORING_OP_POLL_ADD
)
4863 return &req
->apoll
->poll
;
4866 static void io_poll_remove_double(struct io_kiocb
*req
)
4867 __must_hold(&req
->ctx
->completion_lock
)
4869 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
4871 lockdep_assert_held(&req
->ctx
->completion_lock
);
4873 if (poll
&& poll
->head
) {
4874 struct wait_queue_head
*head
= poll
->head
;
4876 spin_lock(&head
->lock
);
4877 list_del_init(&poll
->wait
.entry
);
4878 if (poll
->wait
.private)
4881 spin_unlock(&head
->lock
);
4885 static bool io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
4886 __must_hold(&req
->ctx
->completion_lock
)
4888 struct io_ring_ctx
*ctx
= req
->ctx
;
4889 unsigned flags
= IORING_CQE_F_MORE
;
4892 if (READ_ONCE(req
->poll
.canceled
)) {
4894 req
->poll
.events
|= EPOLLONESHOT
;
4896 error
= mangle_poll(mask
);
4898 if (req
->poll
.events
& EPOLLONESHOT
)
4900 if (!io_cqring_fill_event(ctx
, req
->user_data
, error
, flags
)) {
4901 io_poll_remove_waitqs(req
);
4902 req
->poll
.done
= true;
4905 if (flags
& IORING_CQE_F_MORE
)
4908 io_commit_cqring(ctx
);
4909 return !(flags
& IORING_CQE_F_MORE
);
4912 static void io_poll_task_func(struct callback_head
*cb
)
4914 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4915 struct io_ring_ctx
*ctx
= req
->ctx
;
4916 struct io_kiocb
*nxt
;
4918 if (io_poll_rewait(req
, &req
->poll
)) {
4919 spin_unlock_irq(&ctx
->completion_lock
);
4923 done
= io_poll_complete(req
, req
->result
);
4925 hash_del(&req
->hash_node
);
4928 add_wait_queue(req
->poll
.head
, &req
->poll
.wait
);
4930 spin_unlock_irq(&ctx
->completion_lock
);
4931 io_cqring_ev_posted(ctx
);
4934 nxt
= io_put_req_find_next(req
);
4936 __io_req_task_submit(nxt
);
4941 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
4942 int sync
, void *key
)
4944 struct io_kiocb
*req
= wait
->private;
4945 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
4946 __poll_t mask
= key_to_poll(key
);
4948 /* for instances that support it check for an event match first: */
4949 if (mask
&& !(mask
& poll
->events
))
4951 if (!(poll
->events
& EPOLLONESHOT
))
4952 return poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
4954 list_del_init(&wait
->entry
);
4956 if (poll
&& poll
->head
) {
4959 spin_lock(&poll
->head
->lock
);
4960 done
= list_empty(&poll
->wait
.entry
);
4962 list_del_init(&poll
->wait
.entry
);
4963 /* make sure double remove sees this as being gone */
4964 wait
->private = NULL
;
4965 spin_unlock(&poll
->head
->lock
);
4967 /* use wait func handler, so it matches the rq type */
4968 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
4975 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
4976 wait_queue_func_t wake_func
)
4980 poll
->canceled
= false;
4981 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4982 /* mask in events that we always want/need */
4983 poll
->events
= events
| IO_POLL_UNMASK
;
4984 INIT_LIST_HEAD(&poll
->wait
.entry
);
4985 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
4988 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
4989 struct wait_queue_head
*head
,
4990 struct io_poll_iocb
**poll_ptr
)
4992 struct io_kiocb
*req
= pt
->req
;
4995 * If poll->head is already set, it's because the file being polled
4996 * uses multiple waitqueues for poll handling (eg one for read, one
4997 * for write). Setup a separate io_poll_iocb if this happens.
4999 if (unlikely(poll
->head
)) {
5000 struct io_poll_iocb
*poll_one
= poll
;
5002 /* already have a 2nd entry, fail a third attempt */
5004 pt
->error
= -EINVAL
;
5008 * Can't handle multishot for double wait for now, turn it
5009 * into one-shot mode.
5011 if (!(req
->poll
.events
& EPOLLONESHOT
))
5012 req
->poll
.events
|= EPOLLONESHOT
;
5013 /* double add on the same waitqueue head, ignore */
5014 if (poll
->head
== head
)
5016 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5018 pt
->error
= -ENOMEM
;
5021 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5023 poll
->wait
.private = req
;
5030 if (poll
->events
& EPOLLEXCLUSIVE
)
5031 add_wait_queue_exclusive(head
, &poll
->wait
);
5033 add_wait_queue(head
, &poll
->wait
);
5036 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5037 struct poll_table_struct
*p
)
5039 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5040 struct async_poll
*apoll
= pt
->req
->apoll
;
5042 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5045 static void io_async_task_func(struct callback_head
*cb
)
5047 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5048 struct async_poll
*apoll
= req
->apoll
;
5049 struct io_ring_ctx
*ctx
= req
->ctx
;
5051 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5053 if (io_poll_rewait(req
, &apoll
->poll
)) {
5054 spin_unlock_irq(&ctx
->completion_lock
);
5058 hash_del(&req
->hash_node
);
5059 io_poll_remove_double(req
);
5060 spin_unlock_irq(&ctx
->completion_lock
);
5062 if (!READ_ONCE(apoll
->poll
.canceled
))
5063 __io_req_task_submit(req
);
5065 io_req_complete_failed(req
, -ECANCELED
);
5068 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5071 struct io_kiocb
*req
= wait
->private;
5072 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5074 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5077 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5080 static void io_poll_req_insert(struct io_kiocb
*req
)
5082 struct io_ring_ctx
*ctx
= req
->ctx
;
5083 struct hlist_head
*list
;
5085 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5086 hlist_add_head(&req
->hash_node
, list
);
5089 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5090 struct io_poll_iocb
*poll
,
5091 struct io_poll_table
*ipt
, __poll_t mask
,
5092 wait_queue_func_t wake_func
)
5093 __acquires(&ctx
->completion_lock
)
5095 struct io_ring_ctx
*ctx
= req
->ctx
;
5096 bool cancel
= false;
5098 INIT_HLIST_NODE(&req
->hash_node
);
5099 io_init_poll_iocb(poll
, mask
, wake_func
);
5100 poll
->file
= req
->file
;
5101 poll
->wait
.private = req
;
5103 ipt
->pt
._key
= mask
;
5105 ipt
->error
= -EINVAL
;
5107 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5109 spin_lock_irq(&ctx
->completion_lock
);
5110 if (likely(poll
->head
)) {
5111 spin_lock(&poll
->head
->lock
);
5112 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5118 if ((mask
&& (poll
->events
& EPOLLONESHOT
)) || ipt
->error
)
5119 list_del_init(&poll
->wait
.entry
);
5121 WRITE_ONCE(poll
->canceled
, true);
5122 else if (!poll
->done
) /* actually waiting for an event */
5123 io_poll_req_insert(req
);
5124 spin_unlock(&poll
->head
->lock
);
5130 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5132 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5133 struct io_ring_ctx
*ctx
= req
->ctx
;
5134 struct async_poll
*apoll
;
5135 struct io_poll_table ipt
;
5139 if (!req
->file
|| !file_can_poll(req
->file
))
5141 if (req
->flags
& REQ_F_POLLED
)
5145 else if (def
->pollout
)
5149 /* if we can't nonblock try, then no point in arming a poll handler */
5150 if (!io_file_supports_async(req
, rw
))
5153 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5154 if (unlikely(!apoll
))
5156 apoll
->double_poll
= NULL
;
5158 req
->flags
|= REQ_F_POLLED
;
5161 mask
= EPOLLONESHOT
;
5163 mask
|= POLLIN
| POLLRDNORM
;
5165 mask
|= POLLOUT
| POLLWRNORM
;
5167 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5168 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5169 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5172 mask
|= POLLERR
| POLLPRI
;
5174 ipt
.pt
._qproc
= io_async_queue_proc
;
5176 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5178 if (ret
|| ipt
.error
) {
5179 io_poll_remove_double(req
);
5180 spin_unlock_irq(&ctx
->completion_lock
);
5183 spin_unlock_irq(&ctx
->completion_lock
);
5184 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5185 apoll
->poll
.events
);
5189 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5190 struct io_poll_iocb
*poll
, bool do_cancel
)
5191 __must_hold(&req
->ctx
->completion_lock
)
5193 bool do_complete
= false;
5197 spin_lock(&poll
->head
->lock
);
5199 WRITE_ONCE(poll
->canceled
, true);
5200 if (!list_empty(&poll
->wait
.entry
)) {
5201 list_del_init(&poll
->wait
.entry
);
5204 spin_unlock(&poll
->head
->lock
);
5205 hash_del(&req
->hash_node
);
5209 static bool io_poll_remove_waitqs(struct io_kiocb
*req
)
5210 __must_hold(&req
->ctx
->completion_lock
)
5214 io_poll_remove_double(req
);
5215 do_complete
= __io_poll_remove_one(req
, io_poll_get_single(req
), true);
5217 if (req
->opcode
!= IORING_OP_POLL_ADD
&& do_complete
) {
5218 /* non-poll requests have submit ref still */
5224 static bool io_poll_remove_one(struct io_kiocb
*req
)
5225 __must_hold(&req
->ctx
->completion_lock
)
5229 do_complete
= io_poll_remove_waitqs(req
);
5231 io_cqring_fill_event(req
->ctx
, req
->user_data
, -ECANCELED
, 0);
5232 io_commit_cqring(req
->ctx
);
5233 req_set_fail_links(req
);
5234 io_put_req_deferred(req
, 1);
5241 * Returns true if we found and killed one or more poll requests
5243 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5244 struct files_struct
*files
)
5246 struct hlist_node
*tmp
;
5247 struct io_kiocb
*req
;
5250 spin_lock_irq(&ctx
->completion_lock
);
5251 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5252 struct hlist_head
*list
;
5254 list
= &ctx
->cancel_hash
[i
];
5255 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5256 if (io_match_task(req
, tsk
, files
))
5257 posted
+= io_poll_remove_one(req
);
5260 spin_unlock_irq(&ctx
->completion_lock
);
5263 io_cqring_ev_posted(ctx
);
5268 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5270 __must_hold(&ctx
->completion_lock
)
5272 struct hlist_head
*list
;
5273 struct io_kiocb
*req
;
5275 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5276 hlist_for_each_entry(req
, list
, hash_node
) {
5277 if (sqe_addr
!= req
->user_data
)
5279 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
5286 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5288 __must_hold(&ctx
->completion_lock
)
5290 struct io_kiocb
*req
;
5292 req
= io_poll_find(ctx
, sqe_addr
, poll_only
);
5295 if (io_poll_remove_one(req
))
5301 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
5306 events
= READ_ONCE(sqe
->poll32_events
);
5308 events
= swahw32(events
);
5310 if (!(flags
& IORING_POLL_ADD_MULTI
))
5311 events
|= EPOLLONESHOT
;
5312 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
5315 static int io_poll_update_prep(struct io_kiocb
*req
,
5316 const struct io_uring_sqe
*sqe
)
5318 struct io_poll_update
*upd
= &req
->poll_update
;
5321 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5323 if (sqe
->ioprio
|| sqe
->buf_index
)
5325 flags
= READ_ONCE(sqe
->len
);
5326 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
5327 IORING_POLL_ADD_MULTI
))
5329 /* meaningless without update */
5330 if (flags
== IORING_POLL_ADD_MULTI
)
5333 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
5334 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
5335 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
5337 upd
->new_user_data
= READ_ONCE(sqe
->off
);
5338 if (!upd
->update_user_data
&& upd
->new_user_data
)
5340 if (upd
->update_events
)
5341 upd
->events
= io_poll_parse_events(sqe
, flags
);
5342 else if (sqe
->poll32_events
)
5348 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5351 struct io_kiocb
*req
= wait
->private;
5352 struct io_poll_iocb
*poll
= &req
->poll
;
5354 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5357 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5358 struct poll_table_struct
*p
)
5360 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5362 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5365 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5367 struct io_poll_iocb
*poll
= &req
->poll
;
5370 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5372 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
5374 flags
= READ_ONCE(sqe
->len
);
5375 if (flags
& ~IORING_POLL_ADD_MULTI
)
5378 poll
->events
= io_poll_parse_events(sqe
, flags
);
5382 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5384 struct io_poll_iocb
*poll
= &req
->poll
;
5385 struct io_ring_ctx
*ctx
= req
->ctx
;
5386 struct io_poll_table ipt
;
5389 ipt
.pt
._qproc
= io_poll_queue_proc
;
5391 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5394 if (mask
) { /* no async, we'd stolen it */
5396 io_poll_complete(req
, mask
);
5398 spin_unlock_irq(&ctx
->completion_lock
);
5401 io_cqring_ev_posted(ctx
);
5402 if (poll
->events
& EPOLLONESHOT
)
5408 static int io_poll_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5410 struct io_ring_ctx
*ctx
= req
->ctx
;
5411 struct io_kiocb
*preq
;
5415 spin_lock_irq(&ctx
->completion_lock
);
5416 preq
= io_poll_find(ctx
, req
->poll_update
.old_user_data
, true);
5422 if (!req
->poll_update
.update_events
&& !req
->poll_update
.update_user_data
) {
5424 ret
= io_poll_remove_one(preq
) ? 0 : -EALREADY
;
5429 * Don't allow racy completion with singleshot, as we cannot safely
5430 * update those. For multishot, if we're racing with completion, just
5431 * let completion re-add it.
5433 completing
= !__io_poll_remove_one(preq
, &preq
->poll
, false);
5434 if (completing
&& (preq
->poll
.events
& EPOLLONESHOT
)) {
5438 /* we now have a detached poll request. reissue. */
5442 spin_unlock_irq(&ctx
->completion_lock
);
5443 req_set_fail_links(req
);
5444 io_req_complete(req
, ret
);
5447 /* only mask one event flags, keep behavior flags */
5448 if (req
->poll_update
.update_events
) {
5449 preq
->poll
.events
&= ~0xffff;
5450 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
5451 preq
->poll
.events
|= IO_POLL_UNMASK
;
5453 if (req
->poll_update
.update_user_data
)
5454 preq
->user_data
= req
->poll_update
.new_user_data
;
5455 spin_unlock_irq(&ctx
->completion_lock
);
5457 /* complete update request, we're done with it */
5458 io_req_complete(req
, ret
);
5461 ret
= io_poll_add(preq
, issue_flags
);
5463 req_set_fail_links(preq
);
5464 io_req_complete(preq
, ret
);
5470 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5472 struct io_timeout_data
*data
= container_of(timer
,
5473 struct io_timeout_data
, timer
);
5474 struct io_kiocb
*req
= data
->req
;
5475 struct io_ring_ctx
*ctx
= req
->ctx
;
5476 unsigned long flags
;
5478 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5479 list_del_init(&req
->timeout
.list
);
5480 atomic_set(&req
->ctx
->cq_timeouts
,
5481 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5483 io_cqring_fill_event(ctx
, req
->user_data
, -ETIME
, 0);
5484 io_commit_cqring(ctx
);
5485 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5487 io_cqring_ev_posted(ctx
);
5488 req_set_fail_links(req
);
5490 return HRTIMER_NORESTART
;
5493 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5495 __must_hold(&ctx
->completion_lock
)
5497 struct io_timeout_data
*io
;
5498 struct io_kiocb
*req
;
5501 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5502 found
= user_data
== req
->user_data
;
5507 return ERR_PTR(-ENOENT
);
5509 io
= req
->async_data
;
5510 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
5511 return ERR_PTR(-EALREADY
);
5512 list_del_init(&req
->timeout
.list
);
5516 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5517 __must_hold(&ctx
->completion_lock
)
5519 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5522 return PTR_ERR(req
);
5524 req_set_fail_links(req
);
5525 io_cqring_fill_event(ctx
, req
->user_data
, -ECANCELED
, 0);
5526 io_put_req_deferred(req
, 1);
5530 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5531 struct timespec64
*ts
, enum hrtimer_mode mode
)
5532 __must_hold(&ctx
->completion_lock
)
5534 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5535 struct io_timeout_data
*data
;
5538 return PTR_ERR(req
);
5540 req
->timeout
.off
= 0; /* noseq */
5541 data
= req
->async_data
;
5542 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5543 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5544 data
->timer
.function
= io_timeout_fn
;
5545 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5549 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5550 const struct io_uring_sqe
*sqe
)
5552 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5554 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5556 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5558 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
5561 tr
->addr
= READ_ONCE(sqe
->addr
);
5562 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5563 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5564 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5566 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5568 } else if (tr
->flags
) {
5569 /* timeout removal doesn't support flags */
5576 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
5578 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
5583 * Remove or update an existing timeout command
5585 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
5587 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5588 struct io_ring_ctx
*ctx
= req
->ctx
;
5591 spin_lock_irq(&ctx
->completion_lock
);
5592 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
))
5593 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5595 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
,
5596 io_translate_timeout_mode(tr
->flags
));
5598 io_cqring_fill_event(ctx
, req
->user_data
, ret
, 0);
5599 io_commit_cqring(ctx
);
5600 spin_unlock_irq(&ctx
->completion_lock
);
5601 io_cqring_ev_posted(ctx
);
5603 req_set_fail_links(req
);
5608 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5609 bool is_timeout_link
)
5611 struct io_timeout_data
*data
;
5613 u32 off
= READ_ONCE(sqe
->off
);
5615 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5617 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5619 if (off
&& is_timeout_link
)
5621 flags
= READ_ONCE(sqe
->timeout_flags
);
5622 if (flags
& ~IORING_TIMEOUT_ABS
)
5625 req
->timeout
.off
= off
;
5627 if (!req
->async_data
&& io_alloc_async_data(req
))
5630 data
= req
->async_data
;
5633 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5636 data
->mode
= io_translate_timeout_mode(flags
);
5637 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5638 if (is_timeout_link
)
5639 io_req_track_inflight(req
);
5643 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
5645 struct io_ring_ctx
*ctx
= req
->ctx
;
5646 struct io_timeout_data
*data
= req
->async_data
;
5647 struct list_head
*entry
;
5648 u32 tail
, off
= req
->timeout
.off
;
5650 spin_lock_irq(&ctx
->completion_lock
);
5653 * sqe->off holds how many events that need to occur for this
5654 * timeout event to be satisfied. If it isn't set, then this is
5655 * a pure timeout request, sequence isn't used.
5657 if (io_is_timeout_noseq(req
)) {
5658 entry
= ctx
->timeout_list
.prev
;
5662 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5663 req
->timeout
.target_seq
= tail
+ off
;
5665 /* Update the last seq here in case io_flush_timeouts() hasn't.
5666 * This is safe because ->completion_lock is held, and submissions
5667 * and completions are never mixed in the same ->completion_lock section.
5669 ctx
->cq_last_tm_flush
= tail
;
5672 * Insertion sort, ensuring the first entry in the list is always
5673 * the one we need first.
5675 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5676 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5679 if (io_is_timeout_noseq(nxt
))
5681 /* nxt.seq is behind @tail, otherwise would've been completed */
5682 if (off
>= nxt
->timeout
.target_seq
- tail
)
5686 list_add(&req
->timeout
.list
, entry
);
5687 data
->timer
.function
= io_timeout_fn
;
5688 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5689 spin_unlock_irq(&ctx
->completion_lock
);
5693 struct io_cancel_data
{
5694 struct io_ring_ctx
*ctx
;
5698 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5700 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5701 struct io_cancel_data
*cd
= data
;
5703 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
5706 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
5707 struct io_ring_ctx
*ctx
)
5709 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
5710 enum io_wq_cancel cancel_ret
;
5713 if (!tctx
|| !tctx
->io_wq
)
5716 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
5717 switch (cancel_ret
) {
5718 case IO_WQ_CANCEL_OK
:
5721 case IO_WQ_CANCEL_RUNNING
:
5724 case IO_WQ_CANCEL_NOTFOUND
:
5732 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5733 struct io_kiocb
*req
, __u64 sqe_addr
,
5736 unsigned long flags
;
5739 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
5740 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5743 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5746 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
5750 io_cqring_fill_event(ctx
, req
->user_data
, ret
, 0);
5751 io_commit_cqring(ctx
);
5752 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5753 io_cqring_ev_posted(ctx
);
5756 req_set_fail_links(req
);
5759 static int io_async_cancel_prep(struct io_kiocb
*req
,
5760 const struct io_uring_sqe
*sqe
)
5762 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5764 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5766 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5769 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5773 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
5775 struct io_ring_ctx
*ctx
= req
->ctx
;
5776 u64 sqe_addr
= req
->cancel
.addr
;
5777 struct io_tctx_node
*node
;
5780 /* tasks should wait for their io-wq threads, so safe w/o sync */
5781 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
5782 spin_lock_irq(&ctx
->completion_lock
);
5785 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5788 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
5791 spin_unlock_irq(&ctx
->completion_lock
);
5793 /* slow path, try all io-wq's */
5794 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
5796 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
5797 struct io_uring_task
*tctx
= node
->task
->io_uring
;
5799 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
5803 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
5805 spin_lock_irq(&ctx
->completion_lock
);
5807 io_cqring_fill_event(ctx
, req
->user_data
, ret
, 0);
5808 io_commit_cqring(ctx
);
5809 spin_unlock_irq(&ctx
->completion_lock
);
5810 io_cqring_ev_posted(ctx
);
5813 req_set_fail_links(req
);
5818 static int io_rsrc_update_prep(struct io_kiocb
*req
,
5819 const struct io_uring_sqe
*sqe
)
5821 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
5823 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5825 if (sqe
->ioprio
|| sqe
->rw_flags
)
5828 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
5829 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
5830 if (!req
->rsrc_update
.nr_args
)
5832 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
5836 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5838 struct io_ring_ctx
*ctx
= req
->ctx
;
5839 struct io_uring_rsrc_update2 up
;
5842 if (issue_flags
& IO_URING_F_NONBLOCK
)
5845 up
.offset
= req
->rsrc_update
.offset
;
5846 up
.data
= req
->rsrc_update
.arg
;
5851 mutex_lock(&ctx
->uring_lock
);
5852 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
5853 &up
, req
->rsrc_update
.nr_args
);
5854 mutex_unlock(&ctx
->uring_lock
);
5857 req_set_fail_links(req
);
5858 __io_req_complete(req
, issue_flags
, ret
, 0);
5862 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5864 switch (req
->opcode
) {
5867 case IORING_OP_READV
:
5868 case IORING_OP_READ_FIXED
:
5869 case IORING_OP_READ
:
5870 return io_read_prep(req
, sqe
);
5871 case IORING_OP_WRITEV
:
5872 case IORING_OP_WRITE_FIXED
:
5873 case IORING_OP_WRITE
:
5874 return io_write_prep(req
, sqe
);
5875 case IORING_OP_POLL_ADD
:
5876 return io_poll_add_prep(req
, sqe
);
5877 case IORING_OP_POLL_REMOVE
:
5878 return io_poll_update_prep(req
, sqe
);
5879 case IORING_OP_FSYNC
:
5880 return io_fsync_prep(req
, sqe
);
5881 case IORING_OP_SYNC_FILE_RANGE
:
5882 return io_sfr_prep(req
, sqe
);
5883 case IORING_OP_SENDMSG
:
5884 case IORING_OP_SEND
:
5885 return io_sendmsg_prep(req
, sqe
);
5886 case IORING_OP_RECVMSG
:
5887 case IORING_OP_RECV
:
5888 return io_recvmsg_prep(req
, sqe
);
5889 case IORING_OP_CONNECT
:
5890 return io_connect_prep(req
, sqe
);
5891 case IORING_OP_TIMEOUT
:
5892 return io_timeout_prep(req
, sqe
, false);
5893 case IORING_OP_TIMEOUT_REMOVE
:
5894 return io_timeout_remove_prep(req
, sqe
);
5895 case IORING_OP_ASYNC_CANCEL
:
5896 return io_async_cancel_prep(req
, sqe
);
5897 case IORING_OP_LINK_TIMEOUT
:
5898 return io_timeout_prep(req
, sqe
, true);
5899 case IORING_OP_ACCEPT
:
5900 return io_accept_prep(req
, sqe
);
5901 case IORING_OP_FALLOCATE
:
5902 return io_fallocate_prep(req
, sqe
);
5903 case IORING_OP_OPENAT
:
5904 return io_openat_prep(req
, sqe
);
5905 case IORING_OP_CLOSE
:
5906 return io_close_prep(req
, sqe
);
5907 case IORING_OP_FILES_UPDATE
:
5908 return io_rsrc_update_prep(req
, sqe
);
5909 case IORING_OP_STATX
:
5910 return io_statx_prep(req
, sqe
);
5911 case IORING_OP_FADVISE
:
5912 return io_fadvise_prep(req
, sqe
);
5913 case IORING_OP_MADVISE
:
5914 return io_madvise_prep(req
, sqe
);
5915 case IORING_OP_OPENAT2
:
5916 return io_openat2_prep(req
, sqe
);
5917 case IORING_OP_EPOLL_CTL
:
5918 return io_epoll_ctl_prep(req
, sqe
);
5919 case IORING_OP_SPLICE
:
5920 return io_splice_prep(req
, sqe
);
5921 case IORING_OP_PROVIDE_BUFFERS
:
5922 return io_provide_buffers_prep(req
, sqe
);
5923 case IORING_OP_REMOVE_BUFFERS
:
5924 return io_remove_buffers_prep(req
, sqe
);
5926 return io_tee_prep(req
, sqe
);
5927 case IORING_OP_SHUTDOWN
:
5928 return io_shutdown_prep(req
, sqe
);
5929 case IORING_OP_RENAMEAT
:
5930 return io_renameat_prep(req
, sqe
);
5931 case IORING_OP_UNLINKAT
:
5932 return io_unlinkat_prep(req
, sqe
);
5935 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5940 static int io_req_prep_async(struct io_kiocb
*req
)
5942 if (!io_op_defs
[req
->opcode
].needs_async_setup
)
5944 if (WARN_ON_ONCE(req
->async_data
))
5946 if (io_alloc_async_data(req
))
5949 switch (req
->opcode
) {
5950 case IORING_OP_READV
:
5951 return io_rw_prep_async(req
, READ
);
5952 case IORING_OP_WRITEV
:
5953 return io_rw_prep_async(req
, WRITE
);
5954 case IORING_OP_SENDMSG
:
5955 return io_sendmsg_prep_async(req
);
5956 case IORING_OP_RECVMSG
:
5957 return io_recvmsg_prep_async(req
);
5958 case IORING_OP_CONNECT
:
5959 return io_connect_prep_async(req
);
5961 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
5966 static u32
io_get_sequence(struct io_kiocb
*req
)
5968 struct io_kiocb
*pos
;
5969 struct io_ring_ctx
*ctx
= req
->ctx
;
5970 u32 total_submitted
, nr_reqs
= 0;
5972 io_for_each_link(pos
, req
)
5975 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
5976 return total_submitted
- nr_reqs
;
5979 static int io_req_defer(struct io_kiocb
*req
)
5981 struct io_ring_ctx
*ctx
= req
->ctx
;
5982 struct io_defer_entry
*de
;
5986 /* Still need defer if there is pending req in defer list. */
5987 if (likely(list_empty_careful(&ctx
->defer_list
) &&
5988 !(req
->flags
& REQ_F_IO_DRAIN
)))
5991 seq
= io_get_sequence(req
);
5992 /* Still a chance to pass the sequence check */
5993 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
5996 ret
= io_req_prep_async(req
);
5999 io_prep_async_link(req
);
6000 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6004 spin_lock_irq(&ctx
->completion_lock
);
6005 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6006 spin_unlock_irq(&ctx
->completion_lock
);
6008 io_queue_async_work(req
);
6009 return -EIOCBQUEUED
;
6012 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6015 list_add_tail(&de
->list
, &ctx
->defer_list
);
6016 spin_unlock_irq(&ctx
->completion_lock
);
6017 return -EIOCBQUEUED
;
6020 static void io_clean_op(struct io_kiocb
*req
)
6022 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6023 switch (req
->opcode
) {
6024 case IORING_OP_READV
:
6025 case IORING_OP_READ_FIXED
:
6026 case IORING_OP_READ
:
6027 kfree((void *)(unsigned long)req
->rw
.addr
);
6029 case IORING_OP_RECVMSG
:
6030 case IORING_OP_RECV
:
6031 kfree(req
->sr_msg
.kbuf
);
6034 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
6037 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6038 switch (req
->opcode
) {
6039 case IORING_OP_READV
:
6040 case IORING_OP_READ_FIXED
:
6041 case IORING_OP_READ
:
6042 case IORING_OP_WRITEV
:
6043 case IORING_OP_WRITE_FIXED
:
6044 case IORING_OP_WRITE
: {
6045 struct io_async_rw
*io
= req
->async_data
;
6047 kfree(io
->free_iovec
);
6050 case IORING_OP_RECVMSG
:
6051 case IORING_OP_SENDMSG
: {
6052 struct io_async_msghdr
*io
= req
->async_data
;
6054 kfree(io
->free_iov
);
6057 case IORING_OP_SPLICE
:
6059 if (!(req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
))
6060 io_put_file(req
->splice
.file_in
);
6062 case IORING_OP_OPENAT
:
6063 case IORING_OP_OPENAT2
:
6064 if (req
->open
.filename
)
6065 putname(req
->open
.filename
);
6067 case IORING_OP_RENAMEAT
:
6068 putname(req
->rename
.oldpath
);
6069 putname(req
->rename
.newpath
);
6071 case IORING_OP_UNLINKAT
:
6072 putname(req
->unlink
.filename
);
6075 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6077 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
6078 kfree(req
->apoll
->double_poll
);
6082 if (req
->flags
& REQ_F_INFLIGHT
) {
6083 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6085 atomic_dec(&tctx
->inflight_tracked
);
6086 req
->flags
&= ~REQ_F_INFLIGHT
;
6090 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6092 struct io_ring_ctx
*ctx
= req
->ctx
;
6093 const struct cred
*creds
= NULL
;
6096 if (req
->work
.creds
&& req
->work
.creds
!= current_cred())
6097 creds
= override_creds(req
->work
.creds
);
6099 switch (req
->opcode
) {
6101 ret
= io_nop(req
, issue_flags
);
6103 case IORING_OP_READV
:
6104 case IORING_OP_READ_FIXED
:
6105 case IORING_OP_READ
:
6106 ret
= io_read(req
, issue_flags
);
6108 case IORING_OP_WRITEV
:
6109 case IORING_OP_WRITE_FIXED
:
6110 case IORING_OP_WRITE
:
6111 ret
= io_write(req
, issue_flags
);
6113 case IORING_OP_FSYNC
:
6114 ret
= io_fsync(req
, issue_flags
);
6116 case IORING_OP_POLL_ADD
:
6117 ret
= io_poll_add(req
, issue_flags
);
6119 case IORING_OP_POLL_REMOVE
:
6120 ret
= io_poll_update(req
, issue_flags
);
6122 case IORING_OP_SYNC_FILE_RANGE
:
6123 ret
= io_sync_file_range(req
, issue_flags
);
6125 case IORING_OP_SENDMSG
:
6126 ret
= io_sendmsg(req
, issue_flags
);
6128 case IORING_OP_SEND
:
6129 ret
= io_send(req
, issue_flags
);
6131 case IORING_OP_RECVMSG
:
6132 ret
= io_recvmsg(req
, issue_flags
);
6134 case IORING_OP_RECV
:
6135 ret
= io_recv(req
, issue_flags
);
6137 case IORING_OP_TIMEOUT
:
6138 ret
= io_timeout(req
, issue_flags
);
6140 case IORING_OP_TIMEOUT_REMOVE
:
6141 ret
= io_timeout_remove(req
, issue_flags
);
6143 case IORING_OP_ACCEPT
:
6144 ret
= io_accept(req
, issue_flags
);
6146 case IORING_OP_CONNECT
:
6147 ret
= io_connect(req
, issue_flags
);
6149 case IORING_OP_ASYNC_CANCEL
:
6150 ret
= io_async_cancel(req
, issue_flags
);
6152 case IORING_OP_FALLOCATE
:
6153 ret
= io_fallocate(req
, issue_flags
);
6155 case IORING_OP_OPENAT
:
6156 ret
= io_openat(req
, issue_flags
);
6158 case IORING_OP_CLOSE
:
6159 ret
= io_close(req
, issue_flags
);
6161 case IORING_OP_FILES_UPDATE
:
6162 ret
= io_files_update(req
, issue_flags
);
6164 case IORING_OP_STATX
:
6165 ret
= io_statx(req
, issue_flags
);
6167 case IORING_OP_FADVISE
:
6168 ret
= io_fadvise(req
, issue_flags
);
6170 case IORING_OP_MADVISE
:
6171 ret
= io_madvise(req
, issue_flags
);
6173 case IORING_OP_OPENAT2
:
6174 ret
= io_openat2(req
, issue_flags
);
6176 case IORING_OP_EPOLL_CTL
:
6177 ret
= io_epoll_ctl(req
, issue_flags
);
6179 case IORING_OP_SPLICE
:
6180 ret
= io_splice(req
, issue_flags
);
6182 case IORING_OP_PROVIDE_BUFFERS
:
6183 ret
= io_provide_buffers(req
, issue_flags
);
6185 case IORING_OP_REMOVE_BUFFERS
:
6186 ret
= io_remove_buffers(req
, issue_flags
);
6189 ret
= io_tee(req
, issue_flags
);
6191 case IORING_OP_SHUTDOWN
:
6192 ret
= io_shutdown(req
, issue_flags
);
6194 case IORING_OP_RENAMEAT
:
6195 ret
= io_renameat(req
, issue_flags
);
6197 case IORING_OP_UNLINKAT
:
6198 ret
= io_unlinkat(req
, issue_flags
);
6206 revert_creds(creds
);
6211 /* If the op doesn't have a file, we're not polling for it */
6212 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6213 const bool in_async
= io_wq_current_is_worker();
6215 /* workqueue context doesn't hold uring_lock, grab it now */
6217 mutex_lock(&ctx
->uring_lock
);
6219 io_iopoll_req_issued(req
, in_async
);
6222 mutex_unlock(&ctx
->uring_lock
);
6228 static void io_wq_submit_work(struct io_wq_work
*work
)
6230 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6231 struct io_kiocb
*timeout
;
6234 timeout
= io_prep_linked_timeout(req
);
6236 io_queue_linked_timeout(timeout
);
6238 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6243 ret
= io_issue_sqe(req
, 0);
6245 * We can get EAGAIN for polled IO even though we're
6246 * forcing a sync submission from here, since we can't
6247 * wait for request slots on the block side.
6255 /* avoid locking problems by failing it from a clean context */
6257 /* io-wq is going to take one down */
6259 io_req_task_queue_fail(req
, ret
);
6263 #define FFS_ASYNC_READ 0x1UL
6264 #define FFS_ASYNC_WRITE 0x2UL
6266 #define FFS_ISREG 0x4UL
6268 #define FFS_ISREG 0x0UL
6270 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6272 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
6275 struct io_fixed_file
*table_l2
;
6277 table_l2
= table
->files
[i
>> IORING_FILE_TABLE_SHIFT
];
6278 return &table_l2
[i
& IORING_FILE_TABLE_MASK
];
6281 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6284 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
6286 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
6289 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
6291 unsigned long file_ptr
= (unsigned long) file
;
6293 if (__io_file_supports_async(file
, READ
))
6294 file_ptr
|= FFS_ASYNC_READ
;
6295 if (__io_file_supports_async(file
, WRITE
))
6296 file_ptr
|= FFS_ASYNC_WRITE
;
6297 if (S_ISREG(file_inode(file
)->i_mode
))
6298 file_ptr
|= FFS_ISREG
;
6299 file_slot
->file_ptr
= file_ptr
;
6302 static struct file
*io_file_get(struct io_submit_state
*state
,
6303 struct io_kiocb
*req
, int fd
, bool fixed
)
6305 struct io_ring_ctx
*ctx
= req
->ctx
;
6309 unsigned long file_ptr
;
6311 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6313 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6314 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
6315 file
= (struct file
*) (file_ptr
& FFS_MASK
);
6316 file_ptr
&= ~FFS_MASK
;
6317 /* mask in overlapping REQ_F and FFS bits */
6318 req
->flags
|= (file_ptr
<< REQ_F_ASYNC_READ_BIT
);
6319 io_req_set_rsrc_node(req
);
6321 trace_io_uring_file_get(ctx
, fd
);
6322 file
= __io_file_get(state
, fd
);
6324 /* we don't allow fixed io_uring files */
6325 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6326 io_req_track_inflight(req
);
6332 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6334 struct io_timeout_data
*data
= container_of(timer
,
6335 struct io_timeout_data
, timer
);
6336 struct io_kiocb
*prev
, *req
= data
->req
;
6337 struct io_ring_ctx
*ctx
= req
->ctx
;
6338 unsigned long flags
;
6340 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6341 prev
= req
->timeout
.head
;
6342 req
->timeout
.head
= NULL
;
6345 * We don't expect the list to be empty, that will only happen if we
6346 * race with the completion of the linked work.
6348 if (prev
&& req_ref_inc_not_zero(prev
))
6349 io_remove_next_linked(prev
);
6352 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6355 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6356 io_put_req_deferred(prev
, 1);
6358 io_req_complete_post(req
, -ETIME
, 0);
6360 io_put_req_deferred(req
, 1);
6361 return HRTIMER_NORESTART
;
6364 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6366 struct io_ring_ctx
*ctx
= req
->ctx
;
6368 spin_lock_irq(&ctx
->completion_lock
);
6370 * If the back reference is NULL, then our linked request finished
6371 * before we got a chance to setup the timer
6373 if (req
->timeout
.head
) {
6374 struct io_timeout_data
*data
= req
->async_data
;
6376 data
->timer
.function
= io_link_timeout_fn
;
6377 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6380 spin_unlock_irq(&ctx
->completion_lock
);
6381 /* drop submission reference */
6385 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6387 struct io_kiocb
*nxt
= req
->link
;
6389 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6390 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6393 nxt
->timeout
.head
= req
;
6394 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6395 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6399 static void __io_queue_sqe(struct io_kiocb
*req
)
6401 struct io_kiocb
*linked_timeout
= io_prep_linked_timeout(req
);
6404 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
6407 * We async punt it if the file wasn't marked NOWAIT, or if the file
6408 * doesn't support non-blocking read/write attempts
6411 /* drop submission reference */
6412 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
6413 struct io_ring_ctx
*ctx
= req
->ctx
;
6414 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
6416 cs
->reqs
[cs
->nr
++] = req
;
6417 if (cs
->nr
== ARRAY_SIZE(cs
->reqs
))
6418 io_submit_flush_completions(cs
, ctx
);
6422 } else if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6423 if (!io_arm_poll_handler(req
)) {
6425 * Queued up for async execution, worker will release
6426 * submit reference when the iocb is actually submitted.
6428 io_queue_async_work(req
);
6431 io_req_complete_failed(req
, ret
);
6434 io_queue_linked_timeout(linked_timeout
);
6437 static void io_queue_sqe(struct io_kiocb
*req
)
6441 ret
= io_req_defer(req
);
6443 if (ret
!= -EIOCBQUEUED
) {
6445 io_req_complete_failed(req
, ret
);
6447 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6448 ret
= io_req_prep_async(req
);
6451 io_queue_async_work(req
);
6453 __io_queue_sqe(req
);
6458 * Check SQE restrictions (opcode and flags).
6460 * Returns 'true' if SQE is allowed, 'false' otherwise.
6462 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6463 struct io_kiocb
*req
,
6464 unsigned int sqe_flags
)
6466 if (!ctx
->restricted
)
6469 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6472 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6473 ctx
->restrictions
.sqe_flags_required
)
6476 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6477 ctx
->restrictions
.sqe_flags_required
))
6483 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6484 const struct io_uring_sqe
*sqe
)
6486 struct io_submit_state
*state
;
6487 unsigned int sqe_flags
;
6488 int personality
, ret
= 0;
6490 req
->opcode
= READ_ONCE(sqe
->opcode
);
6491 /* same numerical values with corresponding REQ_F_*, safe to copy */
6492 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
6493 req
->user_data
= READ_ONCE(sqe
->user_data
);
6494 req
->async_data
= NULL
;
6498 req
->fixed_rsrc_refs
= NULL
;
6499 /* one is dropped after submission, the other at completion */
6500 atomic_set(&req
->refs
, 2);
6501 req
->task
= current
;
6503 req
->work
.creds
= NULL
;
6505 /* enforce forwards compatibility on users */
6506 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
)) {
6511 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6514 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6517 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6518 !io_op_defs
[req
->opcode
].buffer_select
)
6521 personality
= READ_ONCE(sqe
->personality
);
6523 req
->work
.creds
= xa_load(&ctx
->personalities
, personality
);
6524 if (!req
->work
.creds
)
6526 get_cred(req
->work
.creds
);
6528 state
= &ctx
->submit_state
;
6531 * Plug now if we have more than 1 IO left after this, and the target
6532 * is potentially a read/write to block based storage.
6534 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6535 io_op_defs
[req
->opcode
].plug
) {
6536 blk_start_plug(&state
->plug
);
6537 state
->plug_started
= true;
6540 if (io_op_defs
[req
->opcode
].needs_file
) {
6541 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6543 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6544 if (unlikely(!req
->file
))
6552 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6553 const struct io_uring_sqe
*sqe
)
6555 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
6558 ret
= io_init_req(ctx
, req
, sqe
);
6559 if (unlikely(ret
)) {
6562 /* fail even hard links since we don't submit */
6563 link
->head
->flags
|= REQ_F_FAIL_LINK
;
6564 io_req_complete_failed(link
->head
, -ECANCELED
);
6567 io_req_complete_failed(req
, ret
);
6570 ret
= io_req_prep(req
, sqe
);
6574 /* don't need @sqe from now on */
6575 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6576 true, ctx
->flags
& IORING_SETUP_SQPOLL
);
6579 * If we already have a head request, queue this one for async
6580 * submittal once the head completes. If we don't have a head but
6581 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6582 * submitted sync once the chain is complete. If none of those
6583 * conditions are true (normal request), then just queue it.
6586 struct io_kiocb
*head
= link
->head
;
6589 * Taking sequential execution of a link, draining both sides
6590 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6591 * requests in the link. So, it drains the head and the
6592 * next after the link request. The last one is done via
6593 * drain_next flag to persist the effect across calls.
6595 if (req
->flags
& REQ_F_IO_DRAIN
) {
6596 head
->flags
|= REQ_F_IO_DRAIN
;
6597 ctx
->drain_next
= 1;
6599 ret
= io_req_prep_async(req
);
6602 trace_io_uring_link(ctx
, req
, head
);
6603 link
->last
->link
= req
;
6606 /* last request of a link, enqueue the link */
6607 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6612 if (unlikely(ctx
->drain_next
)) {
6613 req
->flags
|= REQ_F_IO_DRAIN
;
6614 ctx
->drain_next
= 0;
6616 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6628 * Batched submission is done, ensure local IO is flushed out.
6630 static void io_submit_state_end(struct io_submit_state
*state
,
6631 struct io_ring_ctx
*ctx
)
6633 if (state
->link
.head
)
6634 io_queue_sqe(state
->link
.head
);
6636 io_submit_flush_completions(&state
->comp
, ctx
);
6637 if (state
->plug_started
)
6638 blk_finish_plug(&state
->plug
);
6639 io_state_file_put(state
);
6643 * Start submission side cache.
6645 static void io_submit_state_start(struct io_submit_state
*state
,
6646 unsigned int max_ios
)
6648 state
->plug_started
= false;
6649 state
->ios_left
= max_ios
;
6650 /* set only head, no need to init link_last in advance */
6651 state
->link
.head
= NULL
;
6654 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6656 struct io_rings
*rings
= ctx
->rings
;
6659 * Ensure any loads from the SQEs are done at this point,
6660 * since once we write the new head, the application could
6661 * write new data to them.
6663 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6667 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6668 * that is mapped by userspace. This means that care needs to be taken to
6669 * ensure that reads are stable, as we cannot rely on userspace always
6670 * being a good citizen. If members of the sqe are validated and then later
6671 * used, it's important that those reads are done through READ_ONCE() to
6672 * prevent a re-load down the line.
6674 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6676 u32
*sq_array
= ctx
->sq_array
;
6680 * The cached sq head (or cq tail) serves two purposes:
6682 * 1) allows us to batch the cost of updating the user visible
6684 * 2) allows the kernel side to track the head on its own, even
6685 * though the application is the one updating it.
6687 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
++ & ctx
->sq_mask
]);
6688 if (likely(head
< ctx
->sq_entries
))
6689 return &ctx
->sq_sqes
[head
];
6691 /* drop invalid entries */
6692 ctx
->cached_sq_dropped
++;
6693 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6697 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6701 /* make sure SQ entry isn't read before tail */
6702 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6704 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6707 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6708 refcount_add(nr
, ¤t
->usage
);
6709 io_submit_state_start(&ctx
->submit_state
, nr
);
6711 while (submitted
< nr
) {
6712 const struct io_uring_sqe
*sqe
;
6713 struct io_kiocb
*req
;
6715 req
= io_alloc_req(ctx
);
6716 if (unlikely(!req
)) {
6718 submitted
= -EAGAIN
;
6721 sqe
= io_get_sqe(ctx
);
6722 if (unlikely(!sqe
)) {
6723 kmem_cache_free(req_cachep
, req
);
6726 /* will complete beyond this point, count as submitted */
6728 if (io_submit_sqe(ctx
, req
, sqe
))
6732 if (unlikely(submitted
!= nr
)) {
6733 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6734 struct io_uring_task
*tctx
= current
->io_uring
;
6735 int unused
= nr
- ref_used
;
6737 percpu_ref_put_many(&ctx
->refs
, unused
);
6738 percpu_counter_sub(&tctx
->inflight
, unused
);
6739 put_task_struct_many(current
, unused
);
6742 io_submit_state_end(&ctx
->submit_state
, ctx
);
6743 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6744 io_commit_sqring(ctx
);
6749 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6751 /* Tell userspace we may need a wakeup call */
6752 spin_lock_irq(&ctx
->completion_lock
);
6753 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6754 spin_unlock_irq(&ctx
->completion_lock
);
6757 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6759 spin_lock_irq(&ctx
->completion_lock
);
6760 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6761 spin_unlock_irq(&ctx
->completion_lock
);
6764 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
6766 unsigned int to_submit
;
6769 to_submit
= io_sqring_entries(ctx
);
6770 /* if we're handling multiple rings, cap submit size for fairness */
6771 if (cap_entries
&& to_submit
> 8)
6774 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
6775 unsigned nr_events
= 0;
6777 mutex_lock(&ctx
->uring_lock
);
6778 if (!list_empty(&ctx
->iopoll_list
))
6779 io_do_iopoll(ctx
, &nr_events
, 0);
6782 * Don't submit if refs are dying, good for io_uring_register(),
6783 * but also it is relied upon by io_ring_exit_work()
6785 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
6786 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
6787 ret
= io_submit_sqes(ctx
, to_submit
);
6788 mutex_unlock(&ctx
->uring_lock
);
6791 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
6792 wake_up(&ctx
->sqo_sq_wait
);
6797 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
6799 struct io_ring_ctx
*ctx
;
6800 unsigned sq_thread_idle
= 0;
6802 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6803 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
6804 sqd
->sq_thread_idle
= sq_thread_idle
;
6807 static int io_sq_thread(void *data
)
6809 struct io_sq_data
*sqd
= data
;
6810 struct io_ring_ctx
*ctx
;
6811 unsigned long timeout
= 0;
6812 char buf
[TASK_COMM_LEN
];
6815 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
6816 set_task_comm(current
, buf
);
6818 if (sqd
->sq_cpu
!= -1)
6819 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
6821 set_cpus_allowed_ptr(current
, cpu_online_mask
);
6822 current
->flags
|= PF_NO_SETAFFINITY
;
6824 mutex_lock(&sqd
->lock
);
6825 /* a user may had exited before the thread started */
6826 io_run_task_work_head(&sqd
->park_task_work
);
6828 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
)) {
6830 bool cap_entries
, sqt_spin
, needs_sched
;
6832 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
6833 signal_pending(current
)) {
6834 bool did_sig
= false;
6836 mutex_unlock(&sqd
->lock
);
6837 if (signal_pending(current
)) {
6838 struct ksignal ksig
;
6840 did_sig
= get_signal(&ksig
);
6843 mutex_lock(&sqd
->lock
);
6845 io_run_task_work_head(&sqd
->park_task_work
);
6848 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6852 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
6853 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6854 const struct cred
*creds
= NULL
;
6856 if (ctx
->sq_creds
!= current_cred())
6857 creds
= override_creds(ctx
->sq_creds
);
6858 ret
= __io_sq_thread(ctx
, cap_entries
);
6860 revert_creds(creds
);
6861 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
6865 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
6869 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6873 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
6874 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
)) {
6875 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6876 io_ring_set_wakeup_flag(ctx
);
6879 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6880 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6881 !list_empty_careful(&ctx
->iopoll_list
)) {
6882 needs_sched
= false;
6885 if (io_sqring_entries(ctx
)) {
6886 needs_sched
= false;
6892 mutex_unlock(&sqd
->lock
);
6894 mutex_lock(&sqd
->lock
);
6896 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6897 io_ring_clear_wakeup_flag(ctx
);
6900 finish_wait(&sqd
->wait
, &wait
);
6901 io_run_task_work_head(&sqd
->park_task_work
);
6902 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6905 io_uring_cancel_sqpoll(sqd
);
6907 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6908 io_ring_set_wakeup_flag(ctx
);
6910 io_run_task_work_head(&sqd
->park_task_work
);
6911 mutex_unlock(&sqd
->lock
);
6913 complete(&sqd
->exited
);
6917 struct io_wait_queue
{
6918 struct wait_queue_entry wq
;
6919 struct io_ring_ctx
*ctx
;
6921 unsigned nr_timeouts
;
6924 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
6926 struct io_ring_ctx
*ctx
= iowq
->ctx
;
6929 * Wake up if we have enough events, or if a timeout occurred since we
6930 * started waiting. For timeouts, we always want to return to userspace,
6931 * regardless of event count.
6933 return io_cqring_events(ctx
) >= iowq
->to_wait
||
6934 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
6937 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
6938 int wake_flags
, void *key
)
6940 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
6944 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6945 * the task, and the next invocation will do it.
6947 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->cq_check_overflow
))
6948 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
6952 static int io_run_task_work_sig(void)
6954 if (io_run_task_work())
6956 if (!signal_pending(current
))
6958 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
6959 return -ERESTARTSYS
;
6963 /* when returns >0, the caller should retry */
6964 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
6965 struct io_wait_queue
*iowq
,
6966 signed long *timeout
)
6970 /* make sure we run task_work before checking for signals */
6971 ret
= io_run_task_work_sig();
6972 if (ret
|| io_should_wake(iowq
))
6974 /* let the caller flush overflows, retry */
6975 if (test_bit(0, &ctx
->cq_check_overflow
))
6978 *timeout
= schedule_timeout(*timeout
);
6979 return !*timeout
? -ETIME
: 1;
6983 * Wait until events become available, if we don't already have some. The
6984 * application must reap them itself, as they reside on the shared cq ring.
6986 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
6987 const sigset_t __user
*sig
, size_t sigsz
,
6988 struct __kernel_timespec __user
*uts
)
6990 struct io_wait_queue iowq
= {
6993 .func
= io_wake_function
,
6994 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
6997 .to_wait
= min_events
,
6999 struct io_rings
*rings
= ctx
->rings
;
7000 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
7004 io_cqring_overflow_flush(ctx
, false);
7005 if (io_cqring_events(ctx
) >= min_events
)
7007 if (!io_run_task_work())
7012 #ifdef CONFIG_COMPAT
7013 if (in_compat_syscall())
7014 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7018 ret
= set_user_sigmask(sig
, sigsz
);
7025 struct timespec64 ts
;
7027 if (get_timespec64(&ts
, uts
))
7029 timeout
= timespec64_to_jiffies(&ts
);
7032 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7033 trace_io_uring_cqring_wait(ctx
, min_events
);
7035 /* if we can't even flush overflow, don't wait for more */
7036 if (!io_cqring_overflow_flush(ctx
, false)) {
7040 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
7041 TASK_INTERRUPTIBLE
);
7042 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
7043 finish_wait(&ctx
->wait
, &iowq
.wq
);
7047 restore_saved_sigmask_unless(ret
== -EINTR
);
7049 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7052 static void io_free_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7054 unsigned i
, nr_tables
= DIV_ROUND_UP(nr_files
, IORING_MAX_FILES_TABLE
);
7056 for (i
= 0; i
< nr_tables
; i
++)
7057 kfree(table
->files
[i
]);
7058 kfree(table
->files
);
7059 table
->files
= NULL
;
7062 static inline void io_rsrc_ref_lock(struct io_ring_ctx
*ctx
)
7064 spin_lock_bh(&ctx
->rsrc_ref_lock
);
7067 static inline void io_rsrc_ref_unlock(struct io_ring_ctx
*ctx
)
7069 spin_unlock_bh(&ctx
->rsrc_ref_lock
);
7072 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
7074 percpu_ref_exit(&ref_node
->refs
);
7078 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
7079 struct io_rsrc_data
*data_to_kill
)
7081 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
7082 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
7085 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
7087 rsrc_node
->rsrc_data
= data_to_kill
;
7088 io_rsrc_ref_lock(ctx
);
7089 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
7090 io_rsrc_ref_unlock(ctx
);
7092 atomic_inc(&data_to_kill
->refs
);
7093 percpu_ref_kill(&rsrc_node
->refs
);
7094 ctx
->rsrc_node
= NULL
;
7097 if (!ctx
->rsrc_node
) {
7098 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
7099 ctx
->rsrc_backup_node
= NULL
;
7103 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
7105 if (ctx
->rsrc_backup_node
)
7107 ctx
->rsrc_backup_node
= io_rsrc_node_alloc(ctx
);
7108 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
7111 static int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
, struct io_ring_ctx
*ctx
)
7115 /* As we may drop ->uring_lock, other task may have started quiesce */
7119 data
->quiesce
= true;
7121 ret
= io_rsrc_node_switch_start(ctx
);
7124 io_rsrc_node_switch(ctx
, data
);
7126 /* kill initial ref, already quiesced if zero */
7127 if (atomic_dec_and_test(&data
->refs
))
7129 flush_delayed_work(&ctx
->rsrc_put_work
);
7130 ret
= wait_for_completion_interruptible(&data
->done
);
7134 atomic_inc(&data
->refs
);
7135 /* wait for all works potentially completing data->done */
7136 flush_delayed_work(&ctx
->rsrc_put_work
);
7137 reinit_completion(&data
->done
);
7139 mutex_unlock(&ctx
->uring_lock
);
7140 ret
= io_run_task_work_sig();
7141 mutex_lock(&ctx
->uring_lock
);
7143 data
->quiesce
= false;
7148 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
7154 static struct io_rsrc_data
*io_rsrc_data_alloc(struct io_ring_ctx
*ctx
,
7155 rsrc_put_fn
*do_put
,
7158 struct io_rsrc_data
*data
;
7160 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7164 data
->tags
= kvcalloc(nr
, sizeof(*data
->tags
), GFP_KERNEL
);
7170 atomic_set(&data
->refs
, 1);
7172 data
->do_put
= do_put
;
7173 init_completion(&data
->done
);
7177 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7179 #if defined(CONFIG_UNIX)
7180 if (ctx
->ring_sock
) {
7181 struct sock
*sock
= ctx
->ring_sock
->sk
;
7182 struct sk_buff
*skb
;
7184 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7190 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7193 file
= io_file_from_index(ctx
, i
);
7198 io_free_file_tables(&ctx
->file_table
, ctx
->nr_user_files
);
7199 io_rsrc_data_free(ctx
->file_data
);
7200 ctx
->file_data
= NULL
;
7201 ctx
->nr_user_files
= 0;
7204 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7208 if (!ctx
->file_data
)
7210 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
7212 __io_sqe_files_unregister(ctx
);
7216 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7217 __releases(&sqd
->lock
)
7219 WARN_ON_ONCE(sqd
->thread
== current
);
7222 * Do the dance but not conditional clear_bit() because it'd race with
7223 * other threads incrementing park_pending and setting the bit.
7225 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7226 if (atomic_dec_return(&sqd
->park_pending
))
7227 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7228 mutex_unlock(&sqd
->lock
);
7231 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7232 __acquires(&sqd
->lock
)
7234 WARN_ON_ONCE(sqd
->thread
== current
);
7236 atomic_inc(&sqd
->park_pending
);
7237 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7238 mutex_lock(&sqd
->lock
);
7240 wake_up_process(sqd
->thread
);
7243 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7245 WARN_ON_ONCE(sqd
->thread
== current
);
7246 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
7248 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7249 mutex_lock(&sqd
->lock
);
7251 wake_up_process(sqd
->thread
);
7252 mutex_unlock(&sqd
->lock
);
7253 wait_for_completion(&sqd
->exited
);
7256 static void io_put_sq_data(struct io_sq_data
*sqd
)
7258 if (refcount_dec_and_test(&sqd
->refs
)) {
7259 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
7261 io_sq_thread_stop(sqd
);
7266 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
7268 struct io_sq_data
*sqd
= ctx
->sq_data
;
7271 io_sq_thread_park(sqd
);
7272 list_del_init(&ctx
->sqd_list
);
7273 io_sqd_update_thread_idle(sqd
);
7274 io_sq_thread_unpark(sqd
);
7276 io_put_sq_data(sqd
);
7277 ctx
->sq_data
= NULL
;
7281 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7283 struct io_ring_ctx
*ctx_attach
;
7284 struct io_sq_data
*sqd
;
7287 f
= fdget(p
->wq_fd
);
7289 return ERR_PTR(-ENXIO
);
7290 if (f
.file
->f_op
!= &io_uring_fops
) {
7292 return ERR_PTR(-EINVAL
);
7295 ctx_attach
= f
.file
->private_data
;
7296 sqd
= ctx_attach
->sq_data
;
7299 return ERR_PTR(-EINVAL
);
7301 if (sqd
->task_tgid
!= current
->tgid
) {
7303 return ERR_PTR(-EPERM
);
7306 refcount_inc(&sqd
->refs
);
7311 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
7314 struct io_sq_data
*sqd
;
7317 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
7318 sqd
= io_attach_sq_data(p
);
7323 /* fall through for EPERM case, setup new sqd/task */
7324 if (PTR_ERR(sqd
) != -EPERM
)
7328 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7330 return ERR_PTR(-ENOMEM
);
7332 atomic_set(&sqd
->park_pending
, 0);
7333 refcount_set(&sqd
->refs
, 1);
7334 INIT_LIST_HEAD(&sqd
->ctx_list
);
7335 mutex_init(&sqd
->lock
);
7336 init_waitqueue_head(&sqd
->wait
);
7337 init_completion(&sqd
->exited
);
7341 #if defined(CONFIG_UNIX)
7343 * Ensure the UNIX gc is aware of our file set, so we are certain that
7344 * the io_uring can be safely unregistered on process exit, even if we have
7345 * loops in the file referencing.
7347 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7349 struct sock
*sk
= ctx
->ring_sock
->sk
;
7350 struct scm_fp_list
*fpl
;
7351 struct sk_buff
*skb
;
7354 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7358 skb
= alloc_skb(0, GFP_KERNEL
);
7367 fpl
->user
= get_uid(current_user());
7368 for (i
= 0; i
< nr
; i
++) {
7369 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7373 fpl
->fp
[nr_files
] = get_file(file
);
7374 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7379 fpl
->max
= SCM_MAX_FD
;
7380 fpl
->count
= nr_files
;
7381 UNIXCB(skb
).fp
= fpl
;
7382 skb
->destructor
= unix_destruct_scm
;
7383 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7384 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7386 for (i
= 0; i
< nr_files
; i
++)
7397 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7398 * causes regular reference counting to break down. We rely on the UNIX
7399 * garbage collection to take care of this problem for us.
7401 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7403 unsigned left
, total
;
7407 left
= ctx
->nr_user_files
;
7409 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7411 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7415 total
+= this_files
;
7421 while (total
< ctx
->nr_user_files
) {
7422 struct file
*file
= io_file_from_index(ctx
, total
);
7432 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7438 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7440 unsigned i
, nr_tables
= DIV_ROUND_UP(nr_files
, IORING_MAX_FILES_TABLE
);
7442 table
->files
= kcalloc(nr_tables
, sizeof(*table
->files
), GFP_KERNEL
);
7446 for (i
= 0; i
< nr_tables
; i
++) {
7447 unsigned int this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7449 table
->files
[i
] = kcalloc(this_files
, sizeof(*table
->files
[i
]),
7451 if (!table
->files
[i
])
7453 nr_files
-= this_files
;
7459 io_free_file_tables(table
, nr_tables
* IORING_MAX_FILES_TABLE
);
7463 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
7465 struct file
*file
= prsrc
->file
;
7466 #if defined(CONFIG_UNIX)
7467 struct sock
*sock
= ctx
->ring_sock
->sk
;
7468 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7469 struct sk_buff
*skb
;
7472 __skb_queue_head_init(&list
);
7475 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7476 * remove this entry and rearrange the file array.
7478 skb
= skb_dequeue(head
);
7480 struct scm_fp_list
*fp
;
7482 fp
= UNIXCB(skb
).fp
;
7483 for (i
= 0; i
< fp
->count
; i
++) {
7486 if (fp
->fp
[i
] != file
)
7489 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7490 left
= fp
->count
- 1 - i
;
7492 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7493 left
* sizeof(struct file
*));
7500 __skb_queue_tail(&list
, skb
);
7510 __skb_queue_tail(&list
, skb
);
7512 skb
= skb_dequeue(head
);
7515 if (skb_peek(&list
)) {
7516 spin_lock_irq(&head
->lock
);
7517 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7518 __skb_queue_tail(head
, skb
);
7519 spin_unlock_irq(&head
->lock
);
7526 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
7528 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
7529 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
7530 struct io_rsrc_put
*prsrc
, *tmp
;
7532 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
7533 list_del(&prsrc
->list
);
7536 bool lock_ring
= ctx
->flags
& IORING_SETUP_IOPOLL
;
7537 unsigned long flags
;
7539 io_ring_submit_lock(ctx
, lock_ring
);
7540 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
7541 io_cqring_fill_event(ctx
, prsrc
->tag
, 0, 0);
7542 io_commit_cqring(ctx
);
7543 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
7544 io_cqring_ev_posted(ctx
);
7545 io_ring_submit_unlock(ctx
, lock_ring
);
7548 rsrc_data
->do_put(ctx
, prsrc
);
7552 io_rsrc_node_destroy(ref_node
);
7553 if (atomic_dec_and_test(&rsrc_data
->refs
))
7554 complete(&rsrc_data
->done
);
7557 static void io_rsrc_put_work(struct work_struct
*work
)
7559 struct io_ring_ctx
*ctx
;
7560 struct llist_node
*node
;
7562 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
7563 node
= llist_del_all(&ctx
->rsrc_put_llist
);
7566 struct io_rsrc_node
*ref_node
;
7567 struct llist_node
*next
= node
->next
;
7569 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
7570 __io_rsrc_put_work(ref_node
);
7575 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7577 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
7578 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
7579 bool first_add
= false;
7581 io_rsrc_ref_lock(ctx
);
7584 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7585 node
= list_first_entry(&ctx
->rsrc_ref_list
,
7586 struct io_rsrc_node
, node
);
7587 /* recycle ref nodes in order */
7590 list_del(&node
->node
);
7591 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
7593 io_rsrc_ref_unlock(ctx
);
7596 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, HZ
);
7599 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
)
7601 struct io_rsrc_node
*ref_node
;
7603 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7607 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7612 INIT_LIST_HEAD(&ref_node
->node
);
7613 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7614 ref_node
->done
= false;
7618 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7619 unsigned nr_args
, u64 __user
*tags
)
7621 __s32 __user
*fds
= (__s32 __user
*) arg
;
7625 struct io_rsrc_data
*file_data
;
7631 if (nr_args
> IORING_MAX_FIXED_FILES
)
7633 ret
= io_rsrc_node_switch_start(ctx
);
7637 file_data
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, nr_args
);
7640 ctx
->file_data
= file_data
;
7642 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
))
7645 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7648 if ((tags
&& copy_from_user(&tag
, &tags
[i
], sizeof(tag
))) ||
7649 copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7653 /* allow sparse sets */
7663 if (unlikely(!file
))
7667 * Don't allow io_uring instances to be registered. If UNIX
7668 * isn't enabled, then this causes a reference cycle and this
7669 * instance can never get freed. If UNIX is enabled we'll
7670 * handle it just fine, but there's still no point in allowing
7671 * a ring fd as it doesn't support regular read/write anyway.
7673 if (file
->f_op
== &io_uring_fops
) {
7677 ctx
->file_data
->tags
[i
] = tag
;
7678 io_fixed_file_set(io_fixed_file_slot(&ctx
->file_table
, i
), file
);
7681 ret
= io_sqe_files_scm(ctx
);
7683 __io_sqe_files_unregister(ctx
);
7687 io_rsrc_node_switch(ctx
, NULL
);
7690 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7691 file
= io_file_from_index(ctx
, i
);
7695 io_free_file_tables(&ctx
->file_table
, nr_args
);
7696 ctx
->nr_user_files
= 0;
7698 io_rsrc_data_free(ctx
->file_data
);
7699 ctx
->file_data
= NULL
;
7703 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7706 #if defined(CONFIG_UNIX)
7707 struct sock
*sock
= ctx
->ring_sock
->sk
;
7708 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7709 struct sk_buff
*skb
;
7712 * See if we can merge this file into an existing skb SCM_RIGHTS
7713 * file set. If there's no room, fall back to allocating a new skb
7714 * and filling it in.
7716 spin_lock_irq(&head
->lock
);
7717 skb
= skb_peek(head
);
7719 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7721 if (fpl
->count
< SCM_MAX_FD
) {
7722 __skb_unlink(skb
, head
);
7723 spin_unlock_irq(&head
->lock
);
7724 fpl
->fp
[fpl
->count
] = get_file(file
);
7725 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7727 spin_lock_irq(&head
->lock
);
7728 __skb_queue_head(head
, skb
);
7733 spin_unlock_irq(&head
->lock
);
7740 return __io_sqe_files_scm(ctx
, 1, index
);
7746 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
7747 struct io_rsrc_node
*node
, void *rsrc
)
7749 struct io_rsrc_put
*prsrc
;
7751 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
7755 prsrc
->tag
= data
->tags
[idx
];
7757 list_add(&prsrc
->list
, &node
->rsrc_list
);
7761 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7762 struct io_uring_rsrc_update2
*up
,
7765 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
7766 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
7767 struct io_rsrc_data
*data
= ctx
->file_data
;
7768 struct io_fixed_file
*file_slot
;
7772 bool needs_switch
= false;
7774 if (!ctx
->file_data
)
7776 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
7779 for (done
= 0; done
< nr_args
; done
++) {
7782 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
7783 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7787 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
7791 if (fd
== IORING_REGISTER_FILES_SKIP
)
7794 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
7795 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
7797 if (file_slot
->file_ptr
) {
7798 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
7799 err
= io_queue_rsrc_removal(data
, up
->offset
+ done
,
7800 ctx
->rsrc_node
, file
);
7803 file_slot
->file_ptr
= 0;
7804 needs_switch
= true;
7813 * Don't allow io_uring instances to be registered. If
7814 * UNIX isn't enabled, then this causes a reference
7815 * cycle and this instance can never get freed. If UNIX
7816 * is enabled we'll handle it just fine, but there's
7817 * still no point in allowing a ring fd as it doesn't
7818 * support regular read/write anyway.
7820 if (file
->f_op
== &io_uring_fops
) {
7825 data
->tags
[up
->offset
+ done
] = tag
;
7826 io_fixed_file_set(file_slot
, file
);
7827 err
= io_sqe_file_register(ctx
, file
, i
);
7829 file_slot
->file_ptr
= 0;
7837 io_rsrc_node_switch(ctx
, data
);
7838 return done
? done
: err
;
7841 static struct io_wq_work
*io_free_work(struct io_wq_work
*work
)
7843 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7845 req
= io_put_req_find_next(req
);
7846 return req
? &req
->work
: NULL
;
7849 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
7850 struct task_struct
*task
)
7852 struct io_wq_hash
*hash
;
7853 struct io_wq_data data
;
7854 unsigned int concurrency
;
7856 hash
= ctx
->hash_map
;
7858 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
7860 return ERR_PTR(-ENOMEM
);
7861 refcount_set(&hash
->refs
, 1);
7862 init_waitqueue_head(&hash
->wait
);
7863 ctx
->hash_map
= hash
;
7868 data
.free_work
= io_free_work
;
7869 data
.do_work
= io_wq_submit_work
;
7871 /* Do QD, or 4 * CPUS, whatever is smallest */
7872 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
7874 return io_wq_create(concurrency
, &data
);
7877 static int io_uring_alloc_task_context(struct task_struct
*task
,
7878 struct io_ring_ctx
*ctx
)
7880 struct io_uring_task
*tctx
;
7883 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
7884 if (unlikely(!tctx
))
7887 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
7888 if (unlikely(ret
)) {
7893 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
7894 if (IS_ERR(tctx
->io_wq
)) {
7895 ret
= PTR_ERR(tctx
->io_wq
);
7896 percpu_counter_destroy(&tctx
->inflight
);
7902 init_waitqueue_head(&tctx
->wait
);
7904 atomic_set(&tctx
->in_idle
, 0);
7905 atomic_set(&tctx
->inflight_tracked
, 0);
7906 task
->io_uring
= tctx
;
7907 spin_lock_init(&tctx
->task_lock
);
7908 INIT_WQ_LIST(&tctx
->task_list
);
7909 tctx
->task_state
= 0;
7910 init_task_work(&tctx
->task_work
, tctx_task_work
);
7914 void __io_uring_free(struct task_struct
*tsk
)
7916 struct io_uring_task
*tctx
= tsk
->io_uring
;
7918 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
7919 WARN_ON_ONCE(tctx
->io_wq
);
7921 percpu_counter_destroy(&tctx
->inflight
);
7923 tsk
->io_uring
= NULL
;
7926 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
7927 struct io_uring_params
*p
)
7931 /* Retain compatibility with failing for an invalid attach attempt */
7932 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
7933 IORING_SETUP_ATTACH_WQ
) {
7936 f
= fdget(p
->wq_fd
);
7940 if (f
.file
->f_op
!= &io_uring_fops
)
7943 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7944 struct task_struct
*tsk
;
7945 struct io_sq_data
*sqd
;
7948 sqd
= io_get_sq_data(p
, &attached
);
7954 ctx
->sq_creds
= get_current_cred();
7956 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
7957 if (!ctx
->sq_thread_idle
)
7958 ctx
->sq_thread_idle
= HZ
;
7960 io_sq_thread_park(sqd
);
7961 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
7962 io_sqd_update_thread_idle(sqd
);
7963 /* don't attach to a dying SQPOLL thread, would be racy */
7964 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
7965 io_sq_thread_unpark(sqd
);
7972 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7973 int cpu
= p
->sq_thread_cpu
;
7976 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
7983 sqd
->task_pid
= current
->pid
;
7984 sqd
->task_tgid
= current
->tgid
;
7985 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
7992 ret
= io_uring_alloc_task_context(tsk
, ctx
);
7993 wake_up_new_task(tsk
);
7996 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7997 /* Can't have SQ_AFF without SQPOLL */
8004 complete(&ctx
->sq_data
->exited
);
8006 io_sq_thread_finish(ctx
);
8010 static inline void __io_unaccount_mem(struct user_struct
*user
,
8011 unsigned long nr_pages
)
8013 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8016 static inline int __io_account_mem(struct user_struct
*user
,
8017 unsigned long nr_pages
)
8019 unsigned long page_limit
, cur_pages
, new_pages
;
8021 /* Don't allow more pages than we can safely lock */
8022 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8025 cur_pages
= atomic_long_read(&user
->locked_vm
);
8026 new_pages
= cur_pages
+ nr_pages
;
8027 if (new_pages
> page_limit
)
8029 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8030 new_pages
) != cur_pages
);
8035 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8038 __io_unaccount_mem(ctx
->user
, nr_pages
);
8040 if (ctx
->mm_account
)
8041 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8044 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8049 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8054 if (ctx
->mm_account
)
8055 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8060 static void io_mem_free(void *ptr
)
8067 page
= virt_to_head_page(ptr
);
8068 if (put_page_testzero(page
))
8069 free_compound_page(page
);
8072 static void *io_mem_alloc(size_t size
)
8074 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8075 __GFP_NORETRY
| __GFP_ACCOUNT
;
8077 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8080 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8083 struct io_rings
*rings
;
8084 size_t off
, sq_array_size
;
8086 off
= struct_size(rings
, cqes
, cq_entries
);
8087 if (off
== SIZE_MAX
)
8091 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8099 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8100 if (sq_array_size
== SIZE_MAX
)
8103 if (check_add_overflow(off
, sq_array_size
, &off
))
8109 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
8111 struct io_mapped_ubuf
*imu
= *slot
;
8114 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
8115 unpin_user_page(imu
->bvec
[i
].bv_page
);
8116 if (imu
->acct_pages
)
8117 io_unaccount_mem(ctx
, imu
->acct_pages
);
8122 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8124 io_buffer_unmap(ctx
, &prsrc
->buf
);
8128 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8132 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
8133 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
8134 kfree(ctx
->user_bufs
);
8135 kfree(ctx
->buf_data
);
8136 ctx
->user_bufs
= NULL
;
8137 ctx
->buf_data
= NULL
;
8138 ctx
->nr_user_bufs
= 0;
8141 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8148 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
8150 __io_sqe_buffers_unregister(ctx
);
8154 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8155 void __user
*arg
, unsigned index
)
8157 struct iovec __user
*src
;
8159 #ifdef CONFIG_COMPAT
8161 struct compat_iovec __user
*ciovs
;
8162 struct compat_iovec ciov
;
8164 ciovs
= (struct compat_iovec __user
*) arg
;
8165 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8168 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8169 dst
->iov_len
= ciov
.iov_len
;
8173 src
= (struct iovec __user
*) arg
;
8174 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8180 * Not super efficient, but this is just a registration time. And we do cache
8181 * the last compound head, so generally we'll only do a full search if we don't
8184 * We check if the given compound head page has already been accounted, to
8185 * avoid double accounting it. This allows us to account the full size of the
8186 * page, not just the constituent pages of a huge page.
8188 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8189 int nr_pages
, struct page
*hpage
)
8193 /* check current page array */
8194 for (i
= 0; i
< nr_pages
; i
++) {
8195 if (!PageCompound(pages
[i
]))
8197 if (compound_head(pages
[i
]) == hpage
)
8201 /* check previously registered pages */
8202 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8203 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
8205 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8206 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8208 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8216 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8217 int nr_pages
, struct io_mapped_ubuf
*imu
,
8218 struct page
**last_hpage
)
8222 for (i
= 0; i
< nr_pages
; i
++) {
8223 if (!PageCompound(pages
[i
])) {
8228 hpage
= compound_head(pages
[i
]);
8229 if (hpage
== *last_hpage
)
8231 *last_hpage
= hpage
;
8232 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8234 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8238 if (!imu
->acct_pages
)
8241 ret
= io_account_mem(ctx
, imu
->acct_pages
);
8243 imu
->acct_pages
= 0;
8247 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
8248 struct io_mapped_ubuf
**pimu
,
8249 struct page
**last_hpage
)
8251 struct io_mapped_ubuf
*imu
= NULL
;
8252 struct vm_area_struct
**vmas
= NULL
;
8253 struct page
**pages
= NULL
;
8254 unsigned long off
, start
, end
, ubuf
;
8256 int ret
, pret
, nr_pages
, i
;
8258 ubuf
= (unsigned long) iov
->iov_base
;
8259 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8260 start
= ubuf
>> PAGE_SHIFT
;
8261 nr_pages
= end
- start
;
8266 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
8270 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
8275 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
8280 mmap_read_lock(current
->mm
);
8281 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
8283 if (pret
== nr_pages
) {
8284 /* don't support file backed memory */
8285 for (i
= 0; i
< nr_pages
; i
++) {
8286 struct vm_area_struct
*vma
= vmas
[i
];
8289 !is_file_hugepages(vma
->vm_file
)) {
8295 ret
= pret
< 0 ? pret
: -EFAULT
;
8297 mmap_read_unlock(current
->mm
);
8300 * if we did partial map, or found file backed vmas,
8301 * release any pages we did get
8304 unpin_user_pages(pages
, pret
);
8308 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
8310 unpin_user_pages(pages
, pret
);
8314 off
= ubuf
& ~PAGE_MASK
;
8315 size
= iov
->iov_len
;
8316 for (i
= 0; i
< nr_pages
; i
++) {
8319 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8320 imu
->bvec
[i
].bv_page
= pages
[i
];
8321 imu
->bvec
[i
].bv_len
= vec_len
;
8322 imu
->bvec
[i
].bv_offset
= off
;
8326 /* store original address for later verification */
8328 imu
->ubuf_end
= ubuf
+ iov
->iov_len
;
8329 imu
->nr_bvecs
= nr_pages
;
8340 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
8342 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
8343 return ctx
->user_bufs
? 0 : -ENOMEM
;
8346 static int io_buffer_validate(struct iovec
*iov
)
8348 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
8351 * Don't impose further limits on the size and buffer
8352 * constraints here, we'll -EINVAL later when IO is
8353 * submitted if they are wrong.
8355 if (!iov
->iov_base
|| !iov
->iov_len
)
8358 /* arbitrary limit, but we need something */
8359 if (iov
->iov_len
> SZ_1G
)
8362 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
8368 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8369 unsigned int nr_args
, u64 __user
*tags
)
8371 struct page
*last_hpage
= NULL
;
8372 struct io_rsrc_data
*data
;
8378 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8380 ret
= io_rsrc_node_switch_start(ctx
);
8383 data
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, nr_args
);
8386 ret
= io_buffers_map_alloc(ctx
, nr_args
);
8392 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
8395 if (tags
&& copy_from_user(&tag
, &tags
[i
], sizeof(tag
))) {
8399 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8402 ret
= io_buffer_validate(&iov
);
8406 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
8410 data
->tags
[i
] = tag
;
8413 WARN_ON_ONCE(ctx
->buf_data
);
8415 ctx
->buf_data
= data
;
8417 __io_sqe_buffers_unregister(ctx
);
8419 io_rsrc_node_switch(ctx
, NULL
);
8423 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
8424 struct io_uring_rsrc_update2
*up
,
8425 unsigned int nr_args
)
8427 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
8428 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
8429 struct page
*last_hpage
= NULL
;
8430 bool needs_switch
= false;
8436 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
8439 for (done
= 0; done
< nr_args
; done
++) {
8440 struct io_mapped_ubuf
*imu
;
8441 int offset
= up
->offset
+ done
;
8444 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
8447 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
8451 err
= io_buffer_validate(&iov
);
8454 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
8458 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
8459 if (ctx
->user_bufs
[i
]) {
8460 err
= io_queue_rsrc_removal(ctx
->buf_data
, offset
,
8461 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
8462 if (unlikely(err
)) {
8463 io_buffer_unmap(ctx
, &imu
);
8466 ctx
->user_bufs
[i
] = NULL
;
8467 needs_switch
= true;
8470 ctx
->user_bufs
[i
] = imu
;
8471 ctx
->buf_data
->tags
[offset
] = tag
;
8475 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
8476 return done
? done
: err
;
8479 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8481 __s32 __user
*fds
= arg
;
8487 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8490 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8491 if (IS_ERR(ctx
->cq_ev_fd
)) {
8492 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8493 ctx
->cq_ev_fd
= NULL
;
8500 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8502 if (ctx
->cq_ev_fd
) {
8503 eventfd_ctx_put(ctx
->cq_ev_fd
);
8504 ctx
->cq_ev_fd
= NULL
;
8511 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8513 struct io_buffer
*buf
;
8514 unsigned long index
;
8516 xa_for_each(&ctx
->io_buffers
, index
, buf
)
8517 __io_remove_buffers(ctx
, buf
, index
, -1U);
8520 static void io_req_cache_free(struct list_head
*list
, struct task_struct
*tsk
)
8522 struct io_kiocb
*req
, *nxt
;
8524 list_for_each_entry_safe(req
, nxt
, list
, compl.list
) {
8525 if (tsk
&& req
->task
!= tsk
)
8527 list_del(&req
->compl.list
);
8528 kmem_cache_free(req_cachep
, req
);
8532 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
8534 struct io_submit_state
*submit_state
= &ctx
->submit_state
;
8535 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
8537 mutex_lock(&ctx
->uring_lock
);
8539 if (submit_state
->free_reqs
) {
8540 kmem_cache_free_bulk(req_cachep
, submit_state
->free_reqs
,
8541 submit_state
->reqs
);
8542 submit_state
->free_reqs
= 0;
8545 io_flush_cached_locked_reqs(ctx
, cs
);
8546 io_req_cache_free(&cs
->free_list
, NULL
);
8547 mutex_unlock(&ctx
->uring_lock
);
8550 static bool io_wait_rsrc_data(struct io_rsrc_data
*data
)
8554 if (!atomic_dec_and_test(&data
->refs
))
8555 wait_for_completion(&data
->done
);
8559 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8561 io_sq_thread_finish(ctx
);
8563 if (ctx
->mm_account
) {
8564 mmdrop(ctx
->mm_account
);
8565 ctx
->mm_account
= NULL
;
8568 mutex_lock(&ctx
->uring_lock
);
8569 if (io_wait_rsrc_data(ctx
->buf_data
))
8570 __io_sqe_buffers_unregister(ctx
);
8571 if (io_wait_rsrc_data(ctx
->file_data
))
8572 __io_sqe_files_unregister(ctx
);
8574 __io_cqring_overflow_flush(ctx
, true);
8575 mutex_unlock(&ctx
->uring_lock
);
8576 io_eventfd_unregister(ctx
);
8577 io_destroy_buffers(ctx
);
8579 put_cred(ctx
->sq_creds
);
8581 /* there are no registered resources left, nobody uses it */
8583 io_rsrc_node_destroy(ctx
->rsrc_node
);
8584 if (ctx
->rsrc_backup_node
)
8585 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
8586 flush_delayed_work(&ctx
->rsrc_put_work
);
8588 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
8589 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
8591 #if defined(CONFIG_UNIX)
8592 if (ctx
->ring_sock
) {
8593 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8594 sock_release(ctx
->ring_sock
);
8598 io_mem_free(ctx
->rings
);
8599 io_mem_free(ctx
->sq_sqes
);
8601 percpu_ref_exit(&ctx
->refs
);
8602 free_uid(ctx
->user
);
8603 io_req_caches_free(ctx
);
8605 io_wq_put_hash(ctx
->hash_map
);
8606 kfree(ctx
->cancel_hash
);
8610 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8612 struct io_ring_ctx
*ctx
= file
->private_data
;
8615 poll_wait(file
, &ctx
->cq_wait
, wait
);
8617 * synchronizes with barrier from wq_has_sleeper call in
8621 if (!io_sqring_full(ctx
))
8622 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8625 * Don't flush cqring overflow list here, just do a simple check.
8626 * Otherwise there could possible be ABBA deadlock:
8629 * lock(&ctx->uring_lock);
8631 * lock(&ctx->uring_lock);
8634 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8635 * pushs them to do the flush.
8637 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->cq_check_overflow
))
8638 mask
|= EPOLLIN
| EPOLLRDNORM
;
8643 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8645 struct io_ring_ctx
*ctx
= file
->private_data
;
8647 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8650 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8652 const struct cred
*creds
;
8654 creds
= xa_erase(&ctx
->personalities
, id
);
8663 static inline bool io_run_ctx_fallback(struct io_ring_ctx
*ctx
)
8665 return io_run_task_work_head(&ctx
->exit_task_work
);
8668 struct io_tctx_exit
{
8669 struct callback_head task_work
;
8670 struct completion completion
;
8671 struct io_ring_ctx
*ctx
;
8674 static void io_tctx_exit_cb(struct callback_head
*cb
)
8676 struct io_uring_task
*tctx
= current
->io_uring
;
8677 struct io_tctx_exit
*work
;
8679 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
8681 * When @in_idle, we're in cancellation and it's racy to remove the
8682 * node. It'll be removed by the end of cancellation, just ignore it.
8684 if (!atomic_read(&tctx
->in_idle
))
8685 io_uring_del_task_file((unsigned long)work
->ctx
);
8686 complete(&work
->completion
);
8689 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8691 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8693 return req
->ctx
== data
;
8696 static void io_ring_exit_work(struct work_struct
*work
)
8698 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
8699 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
8700 struct io_tctx_exit exit
;
8701 struct io_tctx_node
*node
;
8705 * If we're doing polled IO and end up having requests being
8706 * submitted async (out-of-line), then completions can come in while
8707 * we're waiting for refs to drop. We need to reap these manually,
8708 * as nobody else will be looking for them.
8711 io_uring_try_cancel_requests(ctx
, NULL
, NULL
);
8713 struct io_sq_data
*sqd
= ctx
->sq_data
;
8714 struct task_struct
*tsk
;
8716 io_sq_thread_park(sqd
);
8718 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
8719 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
8720 io_cancel_ctx_cb
, ctx
, true);
8721 io_sq_thread_unpark(sqd
);
8724 WARN_ON_ONCE(time_after(jiffies
, timeout
));
8725 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8727 init_completion(&exit
.completion
);
8728 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
8731 * Some may use context even when all refs and requests have been put,
8732 * and they are free to do so while still holding uring_lock or
8733 * completion_lock, see __io_req_task_submit(). Apart from other work,
8734 * this lock/unlock section also waits them to finish.
8736 mutex_lock(&ctx
->uring_lock
);
8737 while (!list_empty(&ctx
->tctx_list
)) {
8738 WARN_ON_ONCE(time_after(jiffies
, timeout
));
8740 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
8742 /* don't spin on a single task if cancellation failed */
8743 list_rotate_left(&ctx
->tctx_list
);
8744 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
8745 if (WARN_ON_ONCE(ret
))
8747 wake_up_process(node
->task
);
8749 mutex_unlock(&ctx
->uring_lock
);
8750 wait_for_completion(&exit
.completion
);
8751 mutex_lock(&ctx
->uring_lock
);
8753 mutex_unlock(&ctx
->uring_lock
);
8754 spin_lock_irq(&ctx
->completion_lock
);
8755 spin_unlock_irq(&ctx
->completion_lock
);
8757 io_ring_ctx_free(ctx
);
8760 /* Returns true if we found and killed one or more timeouts */
8761 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
8762 struct files_struct
*files
)
8764 struct io_kiocb
*req
, *tmp
;
8767 spin_lock_irq(&ctx
->completion_lock
);
8768 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
8769 if (io_match_task(req
, tsk
, files
)) {
8770 io_kill_timeout(req
, -ECANCELED
);
8775 io_commit_cqring(ctx
);
8776 spin_unlock_irq(&ctx
->completion_lock
);
8778 io_cqring_ev_posted(ctx
);
8779 return canceled
!= 0;
8782 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8784 unsigned long index
;
8785 struct creds
*creds
;
8787 mutex_lock(&ctx
->uring_lock
);
8788 percpu_ref_kill(&ctx
->refs
);
8790 __io_cqring_overflow_flush(ctx
, true);
8791 xa_for_each(&ctx
->personalities
, index
, creds
)
8792 io_unregister_personality(ctx
, index
);
8793 mutex_unlock(&ctx
->uring_lock
);
8795 io_kill_timeouts(ctx
, NULL
, NULL
);
8796 io_poll_remove_all(ctx
, NULL
, NULL
);
8798 /* if we failed setting up the ctx, we might not have any rings */
8799 io_iopoll_try_reap_events(ctx
);
8801 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8803 * Use system_unbound_wq to avoid spawning tons of event kworkers
8804 * if we're exiting a ton of rings at the same time. It just adds
8805 * noise and overhead, there's no discernable change in runtime
8806 * over using system_wq.
8808 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8811 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8813 struct io_ring_ctx
*ctx
= file
->private_data
;
8815 file
->private_data
= NULL
;
8816 io_ring_ctx_wait_and_kill(ctx
);
8820 struct io_task_cancel
{
8821 struct task_struct
*task
;
8822 struct files_struct
*files
;
8825 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8827 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8828 struct io_task_cancel
*cancel
= data
;
8831 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8832 unsigned long flags
;
8833 struct io_ring_ctx
*ctx
= req
->ctx
;
8835 /* protect against races with linked timeouts */
8836 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8837 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8838 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8840 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8845 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8846 struct task_struct
*task
,
8847 struct files_struct
*files
)
8849 struct io_defer_entry
*de
;
8852 spin_lock_irq(&ctx
->completion_lock
);
8853 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8854 if (io_match_task(de
->req
, task
, files
)) {
8855 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8859 spin_unlock_irq(&ctx
->completion_lock
);
8860 if (list_empty(&list
))
8863 while (!list_empty(&list
)) {
8864 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8865 list_del_init(&de
->list
);
8866 io_req_complete_failed(de
->req
, -ECANCELED
);
8872 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
8874 struct io_tctx_node
*node
;
8875 enum io_wq_cancel cret
;
8878 mutex_lock(&ctx
->uring_lock
);
8879 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
8880 struct io_uring_task
*tctx
= node
->task
->io_uring
;
8883 * io_wq will stay alive while we hold uring_lock, because it's
8884 * killed after ctx nodes, which requires to take the lock.
8886 if (!tctx
|| !tctx
->io_wq
)
8888 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8889 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8891 mutex_unlock(&ctx
->uring_lock
);
8896 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
8897 struct task_struct
*task
,
8898 struct files_struct
*files
)
8900 struct io_task_cancel cancel
= { .task
= task
, .files
= files
, };
8901 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
8904 enum io_wq_cancel cret
;
8908 ret
|= io_uring_try_cancel_iowq(ctx
);
8909 } else if (tctx
&& tctx
->io_wq
) {
8911 * Cancels requests of all rings, not only @ctx, but
8912 * it's fine as the task is in exit/exec.
8914 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
8916 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8919 /* SQPOLL thread does its own polling */
8920 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && !files
) ||
8921 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
8922 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8923 io_iopoll_try_reap_events(ctx
);
8928 ret
|= io_cancel_defer_files(ctx
, task
, files
);
8929 ret
|= io_poll_remove_all(ctx
, task
, files
);
8930 ret
|= io_kill_timeouts(ctx
, task
, files
);
8931 ret
|= io_run_task_work();
8932 ret
|= io_run_ctx_fallback(ctx
);
8939 static int __io_uring_add_task_file(struct io_ring_ctx
*ctx
)
8941 struct io_uring_task
*tctx
= current
->io_uring
;
8942 struct io_tctx_node
*node
;
8945 if (unlikely(!tctx
)) {
8946 ret
= io_uring_alloc_task_context(current
, ctx
);
8949 tctx
= current
->io_uring
;
8951 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
8952 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
8956 node
->task
= current
;
8958 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
8965 mutex_lock(&ctx
->uring_lock
);
8966 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
8967 mutex_unlock(&ctx
->uring_lock
);
8974 * Note that this task has used io_uring. We use it for cancelation purposes.
8976 static inline int io_uring_add_task_file(struct io_ring_ctx
*ctx
)
8978 struct io_uring_task
*tctx
= current
->io_uring
;
8980 if (likely(tctx
&& tctx
->last
== ctx
))
8982 return __io_uring_add_task_file(ctx
);
8986 * Remove this io_uring_file -> task mapping.
8988 static void io_uring_del_task_file(unsigned long index
)
8990 struct io_uring_task
*tctx
= current
->io_uring
;
8991 struct io_tctx_node
*node
;
8995 node
= xa_erase(&tctx
->xa
, index
);
8999 WARN_ON_ONCE(current
!= node
->task
);
9000 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
9002 mutex_lock(&node
->ctx
->uring_lock
);
9003 list_del(&node
->ctx_node
);
9004 mutex_unlock(&node
->ctx
->uring_lock
);
9006 if (tctx
->last
== node
->ctx
)
9011 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
9013 struct io_tctx_node
*node
;
9014 unsigned long index
;
9016 xa_for_each(&tctx
->xa
, index
, node
)
9017 io_uring_del_task_file(index
);
9019 io_wq_put_and_exit(tctx
->io_wq
);
9024 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
9027 return atomic_read(&tctx
->inflight_tracked
);
9028 return percpu_counter_sum(&tctx
->inflight
);
9031 static void io_uring_try_cancel(struct files_struct
*files
)
9033 struct io_uring_task
*tctx
= current
->io_uring
;
9034 struct io_tctx_node
*node
;
9035 unsigned long index
;
9037 xa_for_each(&tctx
->xa
, index
, node
) {
9038 struct io_ring_ctx
*ctx
= node
->ctx
;
9040 /* sqpoll task will cancel all its requests */
9042 io_uring_try_cancel_requests(ctx
, current
, files
);
9046 /* should only be called by SQPOLL task */
9047 static void io_uring_cancel_sqpoll(struct io_sq_data
*sqd
)
9049 struct io_uring_task
*tctx
= current
->io_uring
;
9050 struct io_ring_ctx
*ctx
;
9054 if (!current
->io_uring
)
9056 WARN_ON_ONCE(!sqd
|| sqd
->thread
!= current
);
9058 atomic_inc(&tctx
->in_idle
);
9060 /* read completions before cancelations */
9061 inflight
= tctx_inflight(tctx
, false);
9064 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9065 io_uring_try_cancel_requests(ctx
, current
, NULL
);
9067 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9069 * If we've seen completions, retry without waiting. This
9070 * avoids a race where a completion comes in before we did
9071 * prepare_to_wait().
9073 if (inflight
== tctx_inflight(tctx
, false))
9075 finish_wait(&tctx
->wait
, &wait
);
9077 atomic_dec(&tctx
->in_idle
);
9081 * Find any io_uring fd that this task has registered or done IO on, and cancel
9084 void __io_uring_cancel(struct files_struct
*files
)
9086 struct io_uring_task
*tctx
= current
->io_uring
;
9090 /* make sure overflow events are dropped */
9091 atomic_inc(&tctx
->in_idle
);
9093 /* read completions before cancelations */
9094 inflight
= tctx_inflight(tctx
, !!files
);
9097 io_uring_try_cancel(files
);
9098 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9101 * If we've seen completions, retry without waiting. This
9102 * avoids a race where a completion comes in before we did
9103 * prepare_to_wait().
9105 if (inflight
== tctx_inflight(tctx
, !!files
))
9107 finish_wait(&tctx
->wait
, &wait
);
9109 atomic_dec(&tctx
->in_idle
);
9111 io_uring_clean_tctx(tctx
);
9113 /* for exec all current's requests should be gone, kill tctx */
9114 __io_uring_free(current
);
9118 static void *io_uring_validate_mmap_request(struct file
*file
,
9119 loff_t pgoff
, size_t sz
)
9121 struct io_ring_ctx
*ctx
= file
->private_data
;
9122 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9127 case IORING_OFF_SQ_RING
:
9128 case IORING_OFF_CQ_RING
:
9131 case IORING_OFF_SQES
:
9135 return ERR_PTR(-EINVAL
);
9138 page
= virt_to_head_page(ptr
);
9139 if (sz
> page_size(page
))
9140 return ERR_PTR(-EINVAL
);
9147 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9149 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9153 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9155 return PTR_ERR(ptr
);
9157 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9158 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9161 #else /* !CONFIG_MMU */
9163 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9165 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9168 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9170 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9173 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9174 unsigned long addr
, unsigned long len
,
9175 unsigned long pgoff
, unsigned long flags
)
9179 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9181 return PTR_ERR(ptr
);
9183 return (unsigned long) ptr
;
9186 #endif /* !CONFIG_MMU */
9188 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9193 if (!io_sqring_full(ctx
))
9195 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9197 if (!io_sqring_full(ctx
))
9200 } while (!signal_pending(current
));
9202 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9206 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9207 struct __kernel_timespec __user
**ts
,
9208 const sigset_t __user
**sig
)
9210 struct io_uring_getevents_arg arg
;
9213 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9214 * is just a pointer to the sigset_t.
9216 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9217 *sig
= (const sigset_t __user
*) argp
;
9223 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9224 * timespec and sigset_t pointers if good.
9226 if (*argsz
!= sizeof(arg
))
9228 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9230 *sig
= u64_to_user_ptr(arg
.sigmask
);
9231 *argsz
= arg
.sigmask_sz
;
9232 *ts
= u64_to_user_ptr(arg
.ts
);
9236 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9237 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9240 struct io_ring_ctx
*ctx
;
9247 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9248 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
)))
9252 if (unlikely(!f
.file
))
9256 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
9260 ctx
= f
.file
->private_data
;
9261 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
9265 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9269 * For SQ polling, the thread will do all submissions and completions.
9270 * Just return the requested submit count, and wake the thread if
9274 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9275 io_cqring_overflow_flush(ctx
, false);
9278 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
9281 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9282 wake_up(&ctx
->sq_data
->wait
);
9283 if (flags
& IORING_ENTER_SQ_WAIT
) {
9284 ret
= io_sqpoll_wait_sq(ctx
);
9288 submitted
= to_submit
;
9289 } else if (to_submit
) {
9290 ret
= io_uring_add_task_file(ctx
);
9293 mutex_lock(&ctx
->uring_lock
);
9294 submitted
= io_submit_sqes(ctx
, to_submit
);
9295 mutex_unlock(&ctx
->uring_lock
);
9297 if (submitted
!= to_submit
)
9300 if (flags
& IORING_ENTER_GETEVENTS
) {
9301 const sigset_t __user
*sig
;
9302 struct __kernel_timespec __user
*ts
;
9304 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9308 min_complete
= min(min_complete
, ctx
->cq_entries
);
9311 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9312 * space applications don't need to do io completion events
9313 * polling again, they can rely on io_sq_thread to do polling
9314 * work, which can reduce cpu usage and uring_lock contention.
9316 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9317 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9318 ret
= io_iopoll_check(ctx
, min_complete
);
9320 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9325 percpu_ref_put(&ctx
->refs
);
9328 return submitted
? submitted
: ret
;
9331 #ifdef CONFIG_PROC_FS
9332 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
9333 const struct cred
*cred
)
9335 struct user_namespace
*uns
= seq_user_ns(m
);
9336 struct group_info
*gi
;
9341 seq_printf(m
, "%5d\n", id
);
9342 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9343 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9344 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9345 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9346 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9347 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9348 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9349 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9350 seq_puts(m
, "\n\tGroups:\t");
9351 gi
= cred
->group_info
;
9352 for (g
= 0; g
< gi
->ngroups
; g
++) {
9353 seq_put_decimal_ull(m
, g
? " " : "",
9354 from_kgid_munged(uns
, gi
->gid
[g
]));
9356 seq_puts(m
, "\n\tCapEff:\t");
9357 cap
= cred
->cap_effective
;
9358 CAP_FOR_EACH_U32(__capi
)
9359 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9364 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9366 struct io_sq_data
*sq
= NULL
;
9371 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9372 * since fdinfo case grabs it in the opposite direction of normal use
9373 * cases. If we fail to get the lock, we just don't iterate any
9374 * structures that could be going away outside the io_uring mutex.
9376 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9378 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9384 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9385 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9386 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9387 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9388 struct file
*f
= io_file_from_index(ctx
, i
);
9391 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9393 seq_printf(m
, "%5u: <none>\n", i
);
9395 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9396 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9397 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
9398 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
9400 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
9402 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
9403 unsigned long index
;
9404 const struct cred
*cred
;
9406 seq_printf(m
, "Personalities:\n");
9407 xa_for_each(&ctx
->personalities
, index
, cred
)
9408 io_uring_show_cred(m
, index
, cred
);
9410 seq_printf(m
, "PollList:\n");
9411 spin_lock_irq(&ctx
->completion_lock
);
9412 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9413 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9414 struct io_kiocb
*req
;
9416 hlist_for_each_entry(req
, list
, hash_node
)
9417 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9418 req
->task
->task_works
!= NULL
);
9420 spin_unlock_irq(&ctx
->completion_lock
);
9422 mutex_unlock(&ctx
->uring_lock
);
9425 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9427 struct io_ring_ctx
*ctx
= f
->private_data
;
9429 if (percpu_ref_tryget(&ctx
->refs
)) {
9430 __io_uring_show_fdinfo(ctx
, m
);
9431 percpu_ref_put(&ctx
->refs
);
9436 static const struct file_operations io_uring_fops
= {
9437 .release
= io_uring_release
,
9438 .mmap
= io_uring_mmap
,
9440 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9441 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9443 .poll
= io_uring_poll
,
9444 .fasync
= io_uring_fasync
,
9445 #ifdef CONFIG_PROC_FS
9446 .show_fdinfo
= io_uring_show_fdinfo
,
9450 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9451 struct io_uring_params
*p
)
9453 struct io_rings
*rings
;
9454 size_t size
, sq_array_offset
;
9456 /* make sure these are sane, as we already accounted them */
9457 ctx
->sq_entries
= p
->sq_entries
;
9458 ctx
->cq_entries
= p
->cq_entries
;
9460 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9461 if (size
== SIZE_MAX
)
9464 rings
= io_mem_alloc(size
);
9469 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9470 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9471 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9472 rings
->sq_ring_entries
= p
->sq_entries
;
9473 rings
->cq_ring_entries
= p
->cq_entries
;
9474 ctx
->sq_mask
= rings
->sq_ring_mask
;
9475 ctx
->cq_mask
= rings
->cq_ring_mask
;
9477 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9478 if (size
== SIZE_MAX
) {
9479 io_mem_free(ctx
->rings
);
9484 ctx
->sq_sqes
= io_mem_alloc(size
);
9485 if (!ctx
->sq_sqes
) {
9486 io_mem_free(ctx
->rings
);
9494 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9498 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9502 ret
= io_uring_add_task_file(ctx
);
9507 fd_install(fd
, file
);
9512 * Allocate an anonymous fd, this is what constitutes the application
9513 * visible backing of an io_uring instance. The application mmaps this
9514 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9515 * we have to tie this fd to a socket for file garbage collection purposes.
9517 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9520 #if defined(CONFIG_UNIX)
9523 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9526 return ERR_PTR(ret
);
9529 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9530 O_RDWR
| O_CLOEXEC
);
9531 #if defined(CONFIG_UNIX)
9533 sock_release(ctx
->ring_sock
);
9534 ctx
->ring_sock
= NULL
;
9536 ctx
->ring_sock
->file
= file
;
9542 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9543 struct io_uring_params __user
*params
)
9545 struct io_ring_ctx
*ctx
;
9551 if (entries
> IORING_MAX_ENTRIES
) {
9552 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9554 entries
= IORING_MAX_ENTRIES
;
9558 * Use twice as many entries for the CQ ring. It's possible for the
9559 * application to drive a higher depth than the size of the SQ ring,
9560 * since the sqes are only used at submission time. This allows for
9561 * some flexibility in overcommitting a bit. If the application has
9562 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9563 * of CQ ring entries manually.
9565 p
->sq_entries
= roundup_pow_of_two(entries
);
9566 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9568 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9569 * to a power-of-two, if it isn't already. We do NOT impose
9570 * any cq vs sq ring sizing.
9574 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9575 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9577 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9579 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9580 if (p
->cq_entries
< p
->sq_entries
)
9583 p
->cq_entries
= 2 * p
->sq_entries
;
9586 ctx
= io_ring_ctx_alloc(p
);
9589 ctx
->compat
= in_compat_syscall();
9590 if (!capable(CAP_IPC_LOCK
))
9591 ctx
->user
= get_uid(current_user());
9594 * This is just grabbed for accounting purposes. When a process exits,
9595 * the mm is exited and dropped before the files, hence we need to hang
9596 * on to this mm purely for the purposes of being able to unaccount
9597 * memory (locked/pinned vm). It's not used for anything else.
9599 mmgrab(current
->mm
);
9600 ctx
->mm_account
= current
->mm
;
9602 ret
= io_allocate_scq_urings(ctx
, p
);
9606 ret
= io_sq_offload_create(ctx
, p
);
9609 /* always set a rsrc node */
9610 io_rsrc_node_switch_start(ctx
);
9611 io_rsrc_node_switch(ctx
, NULL
);
9613 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9614 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9615 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9616 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9617 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9618 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9619 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9620 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9622 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9623 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9624 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9625 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9626 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9627 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9628 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9629 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9631 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9632 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9633 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9634 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9635 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
;
9637 if (copy_to_user(params
, p
, sizeof(*p
))) {
9642 file
= io_uring_get_file(ctx
);
9644 ret
= PTR_ERR(file
);
9649 * Install ring fd as the very last thing, so we don't risk someone
9650 * having closed it before we finish setup
9652 ret
= io_uring_install_fd(ctx
, file
);
9654 /* fput will clean it up */
9659 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9662 io_ring_ctx_wait_and_kill(ctx
);
9667 * Sets up an aio uring context, and returns the fd. Applications asks for a
9668 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9669 * params structure passed in.
9671 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9673 struct io_uring_params p
;
9676 if (copy_from_user(&p
, params
, sizeof(p
)))
9678 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9683 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9684 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9685 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9686 IORING_SETUP_R_DISABLED
))
9689 return io_uring_create(entries
, &p
, params
);
9692 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9693 struct io_uring_params __user
*, params
)
9695 return io_uring_setup(entries
, params
);
9698 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9700 struct io_uring_probe
*p
;
9704 size
= struct_size(p
, ops
, nr_args
);
9705 if (size
== SIZE_MAX
)
9707 p
= kzalloc(size
, GFP_KERNEL
);
9712 if (copy_from_user(p
, arg
, size
))
9715 if (memchr_inv(p
, 0, size
))
9718 p
->last_op
= IORING_OP_LAST
- 1;
9719 if (nr_args
> IORING_OP_LAST
)
9720 nr_args
= IORING_OP_LAST
;
9722 for (i
= 0; i
< nr_args
; i
++) {
9724 if (!io_op_defs
[i
].not_supported
)
9725 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9730 if (copy_to_user(arg
, p
, size
))
9737 static int io_register_personality(struct io_ring_ctx
*ctx
)
9739 const struct cred
*creds
;
9743 creds
= get_current_cred();
9745 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
9746 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
9753 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9754 unsigned int nr_args
)
9756 struct io_uring_restriction
*res
;
9760 /* Restrictions allowed only if rings started disabled */
9761 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9764 /* We allow only a single restrictions registration */
9765 if (ctx
->restrictions
.registered
)
9768 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9771 size
= array_size(nr_args
, sizeof(*res
));
9772 if (size
== SIZE_MAX
)
9775 res
= memdup_user(arg
, size
);
9777 return PTR_ERR(res
);
9781 for (i
= 0; i
< nr_args
; i
++) {
9782 switch (res
[i
].opcode
) {
9783 case IORING_RESTRICTION_REGISTER_OP
:
9784 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9789 __set_bit(res
[i
].register_op
,
9790 ctx
->restrictions
.register_op
);
9792 case IORING_RESTRICTION_SQE_OP
:
9793 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9798 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9800 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9801 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9803 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9804 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9813 /* Reset all restrictions if an error happened */
9815 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9817 ctx
->restrictions
.registered
= true;
9823 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
9825 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9828 if (ctx
->restrictions
.registered
)
9829 ctx
->restricted
= 1;
9831 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
9832 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
9833 wake_up(&ctx
->sq_data
->wait
);
9837 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
9838 struct io_uring_rsrc_update2
*up
,
9846 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
9848 err
= io_rsrc_node_switch_start(ctx
);
9853 case IORING_RSRC_FILE
:
9854 return __io_sqe_files_update(ctx
, up
, nr_args
);
9855 case IORING_RSRC_BUFFER
:
9856 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
9861 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
9864 struct io_uring_rsrc_update2 up
;
9868 memset(&up
, 0, sizeof(up
));
9869 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
9871 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
9874 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
9877 struct io_uring_rsrc_update2 up
;
9879 if (size
!= sizeof(up
))
9881 if (copy_from_user(&up
, arg
, sizeof(up
)))
9885 return __io_register_rsrc_update(ctx
, up
.type
, &up
, up
.nr
);
9888 static int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
9891 struct io_uring_rsrc_register rr
;
9893 /* keep it extendible */
9894 if (size
!= sizeof(rr
))
9897 memset(&rr
, 0, sizeof(rr
));
9898 if (copy_from_user(&rr
, arg
, size
))
9904 case IORING_RSRC_FILE
:
9905 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
9906 rr
.nr
, u64_to_user_ptr(rr
.tags
));
9907 case IORING_RSRC_BUFFER
:
9908 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
9909 rr
.nr
, u64_to_user_ptr(rr
.tags
));
9914 static bool io_register_op_must_quiesce(int op
)
9917 case IORING_REGISTER_BUFFERS
:
9918 case IORING_UNREGISTER_BUFFERS
:
9919 case IORING_REGISTER_FILES
:
9920 case IORING_UNREGISTER_FILES
:
9921 case IORING_REGISTER_FILES_UPDATE
:
9922 case IORING_REGISTER_PROBE
:
9923 case IORING_REGISTER_PERSONALITY
:
9924 case IORING_UNREGISTER_PERSONALITY
:
9925 case IORING_REGISTER_RSRC
:
9926 case IORING_REGISTER_RSRC_UPDATE
:
9933 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
9934 void __user
*arg
, unsigned nr_args
)
9935 __releases(ctx
->uring_lock
)
9936 __acquires(ctx
->uring_lock
)
9941 * We're inside the ring mutex, if the ref is already dying, then
9942 * someone else killed the ctx or is already going through
9943 * io_uring_register().
9945 if (percpu_ref_is_dying(&ctx
->refs
))
9948 if (ctx
->restricted
) {
9949 if (opcode
>= IORING_REGISTER_LAST
)
9951 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
9952 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
9956 if (io_register_op_must_quiesce(opcode
)) {
9957 percpu_ref_kill(&ctx
->refs
);
9960 * Drop uring mutex before waiting for references to exit. If
9961 * another thread is currently inside io_uring_enter() it might
9962 * need to grab the uring_lock to make progress. If we hold it
9963 * here across the drain wait, then we can deadlock. It's safe
9964 * to drop the mutex here, since no new references will come in
9965 * after we've killed the percpu ref.
9967 mutex_unlock(&ctx
->uring_lock
);
9969 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
9972 ret
= io_run_task_work_sig();
9976 mutex_lock(&ctx
->uring_lock
);
9979 io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
);
9985 case IORING_REGISTER_BUFFERS
:
9986 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
9988 case IORING_UNREGISTER_BUFFERS
:
9992 ret
= io_sqe_buffers_unregister(ctx
);
9994 case IORING_REGISTER_FILES
:
9995 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
9997 case IORING_UNREGISTER_FILES
:
10001 ret
= io_sqe_files_unregister(ctx
);
10003 case IORING_REGISTER_FILES_UPDATE
:
10004 ret
= io_register_files_update(ctx
, arg
, nr_args
);
10006 case IORING_REGISTER_EVENTFD
:
10007 case IORING_REGISTER_EVENTFD_ASYNC
:
10011 ret
= io_eventfd_register(ctx
, arg
);
10014 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10015 ctx
->eventfd_async
= 1;
10017 ctx
->eventfd_async
= 0;
10019 case IORING_UNREGISTER_EVENTFD
:
10021 if (arg
|| nr_args
)
10023 ret
= io_eventfd_unregister(ctx
);
10025 case IORING_REGISTER_PROBE
:
10027 if (!arg
|| nr_args
> 256)
10029 ret
= io_probe(ctx
, arg
, nr_args
);
10031 case IORING_REGISTER_PERSONALITY
:
10033 if (arg
|| nr_args
)
10035 ret
= io_register_personality(ctx
);
10037 case IORING_UNREGISTER_PERSONALITY
:
10041 ret
= io_unregister_personality(ctx
, nr_args
);
10043 case IORING_REGISTER_ENABLE_RINGS
:
10045 if (arg
|| nr_args
)
10047 ret
= io_register_enable_rings(ctx
);
10049 case IORING_REGISTER_RESTRICTIONS
:
10050 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10052 case IORING_REGISTER_RSRC
:
10053 ret
= io_register_rsrc(ctx
, arg
, nr_args
);
10055 case IORING_REGISTER_RSRC_UPDATE
:
10056 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
);
10063 if (io_register_op_must_quiesce(opcode
)) {
10064 /* bring the ctx back to life */
10065 percpu_ref_reinit(&ctx
->refs
);
10066 reinit_completion(&ctx
->ref_comp
);
10071 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10072 void __user
*, arg
, unsigned int, nr_args
)
10074 struct io_ring_ctx
*ctx
;
10083 if (f
.file
->f_op
!= &io_uring_fops
)
10086 ctx
= f
.file
->private_data
;
10088 io_run_task_work();
10090 mutex_lock(&ctx
->uring_lock
);
10091 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10092 mutex_unlock(&ctx
->uring_lock
);
10093 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10094 ctx
->cq_ev_fd
!= NULL
, ret
);
10100 static int __init
io_uring_init(void)
10102 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10103 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10104 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10107 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10108 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10109 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10110 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10111 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10112 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10113 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10114 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10115 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10116 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10117 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10118 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10119 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10120 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10121 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10122 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10123 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10124 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10125 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10126 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10127 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10128 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10129 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10130 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10131 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10132 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10133 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10134 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10135 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10136 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10137 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10139 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10140 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
10141 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
10145 __initcall(io_uring_init
);