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 struct bio_vec
*bvec
;
199 unsigned int nr_bvecs
;
200 unsigned long acct_pages
;
206 struct list_head list
;
213 struct fixed_rsrc_table
{
217 struct fixed_rsrc_ref_node
{
218 struct percpu_ref refs
;
219 struct list_head node
;
220 struct list_head rsrc_list
;
221 struct fixed_rsrc_data
*rsrc_data
;
222 void (*rsrc_put
)(struct io_ring_ctx
*ctx
,
223 struct io_rsrc_put
*prsrc
);
224 struct llist_node llist
;
228 struct fixed_rsrc_data
{
229 struct fixed_rsrc_table
*table
;
230 struct io_ring_ctx
*ctx
;
232 struct fixed_rsrc_ref_node
*node
;
233 struct percpu_ref refs
;
234 struct completion done
;
239 struct list_head list
;
245 struct io_restriction
{
246 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
247 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
248 u8 sqe_flags_allowed
;
249 u8 sqe_flags_required
;
254 IO_SQ_THREAD_SHOULD_STOP
= 0,
255 IO_SQ_THREAD_SHOULD_PARK
,
260 atomic_t park_pending
;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list
;
266 struct task_struct
*thread
;
267 struct wait_queue_head wait
;
269 unsigned sq_thread_idle
;
275 struct completion exited
;
276 struct callback_head
*park_task_work
;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state
{
285 struct io_kiocb
*reqs
[IO_COMPL_BATCH
];
287 unsigned int locked_free_nr
;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list
;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list
;
294 struct io_submit_link
{
295 struct io_kiocb
*head
;
296 struct io_kiocb
*last
;
299 struct io_submit_state
{
300 struct blk_plug plug
;
301 struct io_submit_link link
;
304 * io_kiocb alloc cache
306 void *reqs
[IO_REQ_CACHE_SIZE
];
307 unsigned int free_reqs
;
312 * Batch completion logic
314 struct io_comp_state comp
;
317 * File reference cache
321 unsigned int file_refs
;
322 unsigned int ios_left
;
327 struct percpu_ref refs
;
328 } ____cacheline_aligned_in_smp
;
332 unsigned int compat
: 1;
333 unsigned int cq_overflow_flushed
: 1;
334 unsigned int drain_next
: 1;
335 unsigned int eventfd_async
: 1;
336 unsigned int restricted
: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head
;
353 unsigned sq_thread_idle
;
354 unsigned cached_sq_dropped
;
355 unsigned cached_cq_overflow
;
356 unsigned long sq_check_overflow
;
358 /* hashed buffered write serialization */
359 struct io_wq_hash
*hash_map
;
361 struct list_head defer_list
;
362 struct list_head timeout_list
;
363 struct list_head cq_overflow_list
;
365 struct io_uring_sqe
*sq_sqes
;
366 } ____cacheline_aligned_in_smp
;
369 struct mutex uring_lock
;
370 wait_queue_head_t wait
;
371 } ____cacheline_aligned_in_smp
;
373 struct io_submit_state submit_state
;
375 struct io_rings
*rings
;
377 /* Only used for accounting purposes */
378 struct mm_struct
*mm_account
;
380 const struct cred
*sq_creds
; /* cred used for __io_sq_thread() */
381 struct io_sq_data
*sq_data
; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait
;
384 struct list_head sqd_list
;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data
*file_data
;
392 unsigned nr_user_files
;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs
;
396 struct io_mapped_ubuf
*user_bufs
;
398 struct user_struct
*user
;
400 struct completion ref_comp
;
402 #if defined(CONFIG_UNIX)
403 struct socket
*ring_sock
;
406 struct xarray io_buffers
;
408 struct xarray personalities
;
412 unsigned cached_cq_tail
;
415 atomic_t cq_timeouts
;
416 unsigned cq_last_tm_flush
;
417 unsigned long cq_check_overflow
;
418 struct wait_queue_head cq_wait
;
419 struct fasync_struct
*cq_fasync
;
420 struct eventfd_ctx
*cq_ev_fd
;
421 } ____cacheline_aligned_in_smp
;
424 spinlock_t completion_lock
;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list
;
433 struct hlist_head
*cancel_hash
;
434 unsigned cancel_hash_bits
;
435 bool poll_multi_file
;
437 spinlock_t inflight_lock
;
438 struct list_head inflight_list
;
439 } ____cacheline_aligned_in_smp
;
441 struct delayed_work rsrc_put_work
;
442 struct llist_head rsrc_put_llist
;
443 struct list_head rsrc_ref_list
;
444 spinlock_t rsrc_ref_lock
;
446 struct io_restriction restrictions
;
449 struct callback_head
*exit_task_work
;
451 struct wait_queue_head hash_wait
;
453 /* Keep this last, we don't need it for the fast path */
454 struct work_struct exit_work
;
455 struct list_head tctx_list
;
458 struct io_uring_task
{
459 /* submission side */
461 struct wait_queue_head wait
;
462 const struct io_ring_ctx
*last
;
464 struct percpu_counter inflight
;
468 spinlock_t task_lock
;
469 struct io_wq_work_list task_list
;
470 unsigned long task_state
;
471 struct callback_head task_work
;
475 * First field must be the file pointer in all the
476 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
478 struct io_poll_iocb
{
480 struct wait_queue_head
*head
;
484 struct wait_queue_entry wait
;
487 struct io_poll_remove
{
497 struct io_timeout_data
{
498 struct io_kiocb
*req
;
499 struct hrtimer timer
;
500 struct timespec64 ts
;
501 enum hrtimer_mode mode
;
506 struct sockaddr __user
*addr
;
507 int __user
*addr_len
;
509 unsigned long nofile
;
529 struct list_head list
;
530 /* head of the link, used by linked timeouts only */
531 struct io_kiocb
*head
;
534 struct io_timeout_rem
{
539 struct timespec64 ts
;
544 /* NOTE: kiocb has the file as the first member, so don't do it here */
552 struct sockaddr __user
*addr
;
559 struct user_msghdr __user
*umsg
;
565 struct io_buffer
*kbuf
;
571 struct filename
*filename
;
573 unsigned long nofile
;
576 struct io_rsrc_update
{
602 struct epoll_event event
;
606 struct file
*file_out
;
607 struct file
*file_in
;
614 struct io_provide_buf
{
628 const char __user
*filename
;
629 struct statx __user
*buffer
;
641 struct filename
*oldpath
;
642 struct filename
*newpath
;
650 struct filename
*filename
;
653 struct io_completion
{
655 struct list_head list
;
659 struct io_async_connect
{
660 struct sockaddr_storage address
;
663 struct io_async_msghdr
{
664 struct iovec fast_iov
[UIO_FASTIOV
];
665 /* points to an allocated iov, if NULL we use fast_iov instead */
666 struct iovec
*free_iov
;
667 struct sockaddr __user
*uaddr
;
669 struct sockaddr_storage addr
;
673 struct iovec fast_iov
[UIO_FASTIOV
];
674 const struct iovec
*free_iovec
;
675 struct iov_iter iter
;
677 struct wait_page_queue wpq
;
681 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
682 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
683 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
684 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
685 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
686 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
692 REQ_F_LINK_TIMEOUT_BIT
,
694 REQ_F_NEED_CLEANUP_BIT
,
696 REQ_F_BUFFER_SELECTED_BIT
,
697 REQ_F_NO_FILE_TABLE_BIT
,
698 REQ_F_LTIMEOUT_ACTIVE_BIT
,
699 REQ_F_COMPLETE_INLINE_BIT
,
702 /* not a real bit, just to check we're not overflowing the space */
708 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
709 /* drain existing IO first */
710 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
712 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
713 /* doesn't sever on completion < 0 */
714 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
716 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
717 /* IOSQE_BUFFER_SELECT */
718 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
720 /* fail rest of links */
721 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
722 /* on inflight list, should be cancelled and waited on exit reliably */
723 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
724 /* read/write uses file position */
725 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
726 /* must not punt to workers */
727 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
728 /* has or had linked timeout */
729 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
731 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
733 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
734 /* already went through poll handler */
735 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
736 /* buffer already selected */
737 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
738 /* doesn't need file table for this request */
739 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
740 /* linked timeout is active, i.e. prepared by link's head */
741 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
742 /* completion is deferred through io_comp_state */
743 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
744 /* caller should reissue async */
745 REQ_F_REISSUE
= BIT(REQ_F_REISSUE_BIT
),
749 struct io_poll_iocb poll
;
750 struct io_poll_iocb
*double_poll
;
753 struct io_task_work
{
754 struct io_wq_work_node node
;
755 task_work_func_t func
;
759 * NOTE! Each of the iocb union members has the file pointer
760 * as the first entry in their struct definition. So you can
761 * access the file pointer through any of the sub-structs,
762 * or directly as just 'ki_filp' in this struct.
768 struct io_poll_iocb poll
;
769 struct io_poll_remove poll_remove
;
770 struct io_accept accept
;
772 struct io_cancel cancel
;
773 struct io_timeout timeout
;
774 struct io_timeout_rem timeout_rem
;
775 struct io_connect connect
;
776 struct io_sr_msg sr_msg
;
778 struct io_close close
;
779 struct io_rsrc_update rsrc_update
;
780 struct io_fadvise fadvise
;
781 struct io_madvise madvise
;
782 struct io_epoll epoll
;
783 struct io_splice splice
;
784 struct io_provide_buf pbuf
;
785 struct io_statx statx
;
786 struct io_shutdown shutdown
;
787 struct io_rename rename
;
788 struct io_unlink unlink
;
789 /* use only after cleaning per-op data, see io_clean_op() */
790 struct io_completion
compl;
793 /* opcode allocated if it needs to store data for async defer */
796 /* polled IO has completed */
802 struct io_ring_ctx
*ctx
;
805 struct task_struct
*task
;
808 struct io_kiocb
*link
;
809 struct percpu_ref
*fixed_rsrc_refs
;
812 * 1. used with ctx->iopoll_list with reads/writes
813 * 2. to track reqs with ->files (see io_op_def::file_table)
815 struct list_head inflight_entry
;
817 struct io_task_work io_task_work
;
818 struct callback_head task_work
;
820 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
821 struct hlist_node hash_node
;
822 struct async_poll
*apoll
;
823 struct io_wq_work work
;
826 struct io_tctx_node
{
827 struct list_head ctx_node
;
828 struct task_struct
*task
;
829 struct io_ring_ctx
*ctx
;
832 struct io_defer_entry
{
833 struct list_head list
;
834 struct io_kiocb
*req
;
839 /* needs req->file assigned */
840 unsigned needs_file
: 1;
841 /* hash wq insertion if file is a regular file */
842 unsigned hash_reg_file
: 1;
843 /* unbound wq insertion if file is a non-regular file */
844 unsigned unbound_nonreg_file
: 1;
845 /* opcode is not supported by this kernel */
846 unsigned not_supported
: 1;
847 /* set if opcode supports polled "wait" */
849 unsigned pollout
: 1;
850 /* op supports buffer selection */
851 unsigned buffer_select
: 1;
852 /* must always have async data allocated */
853 unsigned needs_async_data
: 1;
854 /* should block plug */
856 /* size of async data needed, if any */
857 unsigned short async_size
;
860 static const struct io_op_def io_op_defs
[] = {
861 [IORING_OP_NOP
] = {},
862 [IORING_OP_READV
] = {
864 .unbound_nonreg_file
= 1,
867 .needs_async_data
= 1,
869 .async_size
= sizeof(struct io_async_rw
),
871 [IORING_OP_WRITEV
] = {
874 .unbound_nonreg_file
= 1,
876 .needs_async_data
= 1,
878 .async_size
= sizeof(struct io_async_rw
),
880 [IORING_OP_FSYNC
] = {
883 [IORING_OP_READ_FIXED
] = {
885 .unbound_nonreg_file
= 1,
888 .async_size
= sizeof(struct io_async_rw
),
890 [IORING_OP_WRITE_FIXED
] = {
893 .unbound_nonreg_file
= 1,
896 .async_size
= sizeof(struct io_async_rw
),
898 [IORING_OP_POLL_ADD
] = {
900 .unbound_nonreg_file
= 1,
902 [IORING_OP_POLL_REMOVE
] = {},
903 [IORING_OP_SYNC_FILE_RANGE
] = {
906 [IORING_OP_SENDMSG
] = {
908 .unbound_nonreg_file
= 1,
910 .needs_async_data
= 1,
911 .async_size
= sizeof(struct io_async_msghdr
),
913 [IORING_OP_RECVMSG
] = {
915 .unbound_nonreg_file
= 1,
918 .needs_async_data
= 1,
919 .async_size
= sizeof(struct io_async_msghdr
),
921 [IORING_OP_TIMEOUT
] = {
922 .needs_async_data
= 1,
923 .async_size
= sizeof(struct io_timeout_data
),
925 [IORING_OP_TIMEOUT_REMOVE
] = {
926 /* used by timeout updates' prep() */
928 [IORING_OP_ACCEPT
] = {
930 .unbound_nonreg_file
= 1,
933 [IORING_OP_ASYNC_CANCEL
] = {},
934 [IORING_OP_LINK_TIMEOUT
] = {
935 .needs_async_data
= 1,
936 .async_size
= sizeof(struct io_timeout_data
),
938 [IORING_OP_CONNECT
] = {
940 .unbound_nonreg_file
= 1,
942 .needs_async_data
= 1,
943 .async_size
= sizeof(struct io_async_connect
),
945 [IORING_OP_FALLOCATE
] = {
948 [IORING_OP_OPENAT
] = {},
949 [IORING_OP_CLOSE
] = {},
950 [IORING_OP_FILES_UPDATE
] = {},
951 [IORING_OP_STATX
] = {},
954 .unbound_nonreg_file
= 1,
958 .async_size
= sizeof(struct io_async_rw
),
960 [IORING_OP_WRITE
] = {
962 .unbound_nonreg_file
= 1,
965 .async_size
= sizeof(struct io_async_rw
),
967 [IORING_OP_FADVISE
] = {
970 [IORING_OP_MADVISE
] = {},
973 .unbound_nonreg_file
= 1,
978 .unbound_nonreg_file
= 1,
982 [IORING_OP_OPENAT2
] = {
984 [IORING_OP_EPOLL_CTL
] = {
985 .unbound_nonreg_file
= 1,
987 [IORING_OP_SPLICE
] = {
990 .unbound_nonreg_file
= 1,
992 [IORING_OP_PROVIDE_BUFFERS
] = {},
993 [IORING_OP_REMOVE_BUFFERS
] = {},
997 .unbound_nonreg_file
= 1,
999 [IORING_OP_SHUTDOWN
] = {
1002 [IORING_OP_RENAMEAT
] = {},
1003 [IORING_OP_UNLINKAT
] = {},
1006 static bool io_disarm_next(struct io_kiocb
*req
);
1007 static void io_uring_del_task_file(unsigned long index
);
1008 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1009 struct task_struct
*task
,
1010 struct files_struct
*files
);
1011 static void io_uring_cancel_sqpoll(struct io_ring_ctx
*ctx
);
1012 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node
*ref_node
);
1013 static struct fixed_rsrc_ref_node
*alloc_fixed_rsrc_ref_node(
1014 struct io_ring_ctx
*ctx
);
1015 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
);
1017 static bool io_rw_reissue(struct io_kiocb
*req
);
1018 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
1019 static void io_put_req(struct io_kiocb
*req
);
1020 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1021 static void io_double_put_req(struct io_kiocb
*req
);
1022 static void io_dismantle_req(struct io_kiocb
*req
);
1023 static void io_put_task(struct task_struct
*task
, int nr
);
1024 static void io_queue_next(struct io_kiocb
*req
);
1025 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1026 static void __io_queue_linked_timeout(struct io_kiocb
*req
);
1027 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1028 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
1029 struct io_uring_rsrc_update
*ip
,
1031 static void __io_clean_op(struct io_kiocb
*req
);
1032 static struct file
*io_file_get(struct io_submit_state
*state
,
1033 struct io_kiocb
*req
, int fd
, bool fixed
);
1034 static void __io_queue_sqe(struct io_kiocb
*req
);
1035 static void io_rsrc_put_work(struct work_struct
*work
);
1037 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
1038 struct iov_iter
*iter
, bool needs_lock
);
1039 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
1040 const struct iovec
*fast_iov
,
1041 struct iov_iter
*iter
, bool force
);
1042 static void io_req_task_queue(struct io_kiocb
*req
);
1043 static void io_submit_flush_completions(struct io_comp_state
*cs
,
1044 struct io_ring_ctx
*ctx
);
1046 static struct kmem_cache
*req_cachep
;
1048 static const struct file_operations io_uring_fops
;
1050 struct sock
*io_uring_get_socket(struct file
*file
)
1052 #if defined(CONFIG_UNIX)
1053 if (file
->f_op
== &io_uring_fops
) {
1054 struct io_ring_ctx
*ctx
= file
->private_data
;
1056 return ctx
->ring_sock
->sk
;
1061 EXPORT_SYMBOL(io_uring_get_socket
);
1063 #define io_for_each_link(pos, head) \
1064 for (pos = (head); pos; pos = pos->link)
1066 static inline void io_clean_op(struct io_kiocb
*req
)
1068 if (req
->flags
& (REQ_F_NEED_CLEANUP
| REQ_F_BUFFER_SELECTED
))
1072 static inline void io_set_resource_node(struct io_kiocb
*req
)
1074 struct io_ring_ctx
*ctx
= req
->ctx
;
1076 if (!req
->fixed_rsrc_refs
) {
1077 req
->fixed_rsrc_refs
= &ctx
->file_data
->node
->refs
;
1078 percpu_ref_get(req
->fixed_rsrc_refs
);
1082 static bool io_match_task(struct io_kiocb
*head
,
1083 struct task_struct
*task
,
1084 struct files_struct
*files
)
1086 struct io_kiocb
*req
;
1088 if (task
&& head
->task
!= task
) {
1089 /* in terms of cancelation, always match if req task is dead */
1090 if (head
->task
->flags
& PF_EXITING
)
1097 io_for_each_link(req
, head
) {
1098 if (req
->flags
& REQ_F_INFLIGHT
)
1104 static inline void req_set_fail_links(struct io_kiocb
*req
)
1106 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1107 req
->flags
|= REQ_F_FAIL_LINK
;
1110 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1112 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1114 complete(&ctx
->ref_comp
);
1117 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1119 return !req
->timeout
.off
;
1122 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1124 struct io_ring_ctx
*ctx
;
1127 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1132 * Use 5 bits less than the max cq entries, that should give us around
1133 * 32 entries per hash list if totally full and uniformly spread.
1135 hash_bits
= ilog2(p
->cq_entries
);
1139 ctx
->cancel_hash_bits
= hash_bits
;
1140 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1142 if (!ctx
->cancel_hash
)
1144 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1146 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1147 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1150 ctx
->flags
= p
->flags
;
1151 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1152 INIT_LIST_HEAD(&ctx
->sqd_list
);
1153 init_waitqueue_head(&ctx
->cq_wait
);
1154 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1155 init_completion(&ctx
->ref_comp
);
1156 xa_init_flags(&ctx
->io_buffers
, XA_FLAGS_ALLOC1
);
1157 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1158 mutex_init(&ctx
->uring_lock
);
1159 init_waitqueue_head(&ctx
->wait
);
1160 spin_lock_init(&ctx
->completion_lock
);
1161 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1162 INIT_LIST_HEAD(&ctx
->defer_list
);
1163 INIT_LIST_HEAD(&ctx
->timeout_list
);
1164 spin_lock_init(&ctx
->inflight_lock
);
1165 INIT_LIST_HEAD(&ctx
->inflight_list
);
1166 spin_lock_init(&ctx
->rsrc_ref_lock
);
1167 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1168 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1169 init_llist_head(&ctx
->rsrc_put_llist
);
1170 INIT_LIST_HEAD(&ctx
->tctx_list
);
1171 INIT_LIST_HEAD(&ctx
->submit_state
.comp
.free_list
);
1172 INIT_LIST_HEAD(&ctx
->submit_state
.comp
.locked_free_list
);
1175 kfree(ctx
->cancel_hash
);
1180 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1182 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1183 struct io_ring_ctx
*ctx
= req
->ctx
;
1185 return seq
!= ctx
->cached_cq_tail
1186 + READ_ONCE(ctx
->cached_cq_overflow
);
1192 static void io_req_track_inflight(struct io_kiocb
*req
)
1194 struct io_ring_ctx
*ctx
= req
->ctx
;
1196 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1197 req
->flags
|= REQ_F_INFLIGHT
;
1199 spin_lock_irq(&ctx
->inflight_lock
);
1200 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
1201 spin_unlock_irq(&ctx
->inflight_lock
);
1205 static void io_prep_async_work(struct io_kiocb
*req
)
1207 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1208 struct io_ring_ctx
*ctx
= req
->ctx
;
1210 if (!req
->work
.creds
)
1211 req
->work
.creds
= get_current_cred();
1213 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1214 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1216 if (req
->flags
& REQ_F_ISREG
) {
1217 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1218 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1219 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
1220 if (def
->unbound_nonreg_file
)
1221 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1225 static void io_prep_async_link(struct io_kiocb
*req
)
1227 struct io_kiocb
*cur
;
1229 io_for_each_link(cur
, req
)
1230 io_prep_async_work(cur
);
1233 static void io_queue_async_work(struct io_kiocb
*req
)
1235 struct io_ring_ctx
*ctx
= req
->ctx
;
1236 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1237 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1240 BUG_ON(!tctx
->io_wq
);
1242 /* init ->work of the whole link before punting */
1243 io_prep_async_link(req
);
1244 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1245 &req
->work
, req
->flags
);
1246 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
1248 io_queue_linked_timeout(link
);
1251 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1253 struct io_timeout_data
*io
= req
->async_data
;
1256 ret
= hrtimer_try_to_cancel(&io
->timer
);
1258 atomic_set(&req
->ctx
->cq_timeouts
,
1259 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1260 list_del_init(&req
->timeout
.list
);
1261 io_cqring_fill_event(req
, status
);
1262 io_put_req_deferred(req
, 1);
1266 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1269 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1270 struct io_defer_entry
, list
);
1272 if (req_need_defer(de
->req
, de
->seq
))
1274 list_del_init(&de
->list
);
1275 io_req_task_queue(de
->req
);
1277 } while (!list_empty(&ctx
->defer_list
));
1280 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1284 if (list_empty(&ctx
->timeout_list
))
1287 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1290 u32 events_needed
, events_got
;
1291 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1292 struct io_kiocb
, timeout
.list
);
1294 if (io_is_timeout_noseq(req
))
1298 * Since seq can easily wrap around over time, subtract
1299 * the last seq at which timeouts were flushed before comparing.
1300 * Assuming not more than 2^31-1 events have happened since,
1301 * these subtractions won't have wrapped, so we can check if
1302 * target is in [last_seq, current_seq] by comparing the two.
1304 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1305 events_got
= seq
- ctx
->cq_last_tm_flush
;
1306 if (events_got
< events_needed
)
1309 list_del_init(&req
->timeout
.list
);
1310 io_kill_timeout(req
, 0);
1311 } while (!list_empty(&ctx
->timeout_list
));
1313 ctx
->cq_last_tm_flush
= seq
;
1316 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1318 io_flush_timeouts(ctx
);
1320 /* order cqe stores with ring update */
1321 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1323 if (unlikely(!list_empty(&ctx
->defer_list
)))
1324 __io_queue_deferred(ctx
);
1327 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1329 struct io_rings
*r
= ctx
->rings
;
1331 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== r
->sq_ring_entries
;
1334 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1336 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1339 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1341 struct io_rings
*rings
= ctx
->rings
;
1345 * writes to the cq entry need to come after reading head; the
1346 * control dependency is enough as we're using WRITE_ONCE to
1349 if (__io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1352 tail
= ctx
->cached_cq_tail
++;
1353 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1356 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1360 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1362 if (!ctx
->eventfd_async
)
1364 return io_wq_current_is_worker();
1367 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1369 /* see waitqueue_active() comment */
1372 if (waitqueue_active(&ctx
->wait
))
1373 wake_up(&ctx
->wait
);
1374 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1375 wake_up(&ctx
->sq_data
->wait
);
1376 if (io_should_trigger_evfd(ctx
))
1377 eventfd_signal(ctx
->cq_ev_fd
, 1);
1378 if (waitqueue_active(&ctx
->cq_wait
)) {
1379 wake_up_interruptible(&ctx
->cq_wait
);
1380 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1384 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1386 /* see waitqueue_active() comment */
1389 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1390 if (waitqueue_active(&ctx
->wait
))
1391 wake_up(&ctx
->wait
);
1393 if (io_should_trigger_evfd(ctx
))
1394 eventfd_signal(ctx
->cq_ev_fd
, 1);
1395 if (waitqueue_active(&ctx
->cq_wait
)) {
1396 wake_up_interruptible(&ctx
->cq_wait
);
1397 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1401 /* Returns true if there are no backlogged entries after the flush */
1402 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1403 struct task_struct
*tsk
,
1404 struct files_struct
*files
)
1406 struct io_rings
*rings
= ctx
->rings
;
1407 struct io_kiocb
*req
, *tmp
;
1408 struct io_uring_cqe
*cqe
;
1409 unsigned long flags
;
1410 bool all_flushed
, posted
;
1413 if (!force
&& __io_cqring_events(ctx
) == rings
->cq_ring_entries
)
1417 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1418 list_for_each_entry_safe(req
, tmp
, &ctx
->cq_overflow_list
, compl.list
) {
1419 if (!io_match_task(req
, tsk
, files
))
1422 cqe
= io_get_cqring(ctx
);
1426 list_move(&req
->compl.list
, &list
);
1428 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1429 WRITE_ONCE(cqe
->res
, req
->result
);
1430 WRITE_ONCE(cqe
->flags
, req
->compl.cflags
);
1432 ctx
->cached_cq_overflow
++;
1433 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1434 ctx
->cached_cq_overflow
);
1439 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1441 clear_bit(0, &ctx
->sq_check_overflow
);
1442 clear_bit(0, &ctx
->cq_check_overflow
);
1443 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1447 io_commit_cqring(ctx
);
1448 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1450 io_cqring_ev_posted(ctx
);
1452 while (!list_empty(&list
)) {
1453 req
= list_first_entry(&list
, struct io_kiocb
, compl.list
);
1454 list_del(&req
->compl.list
);
1461 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
,
1462 struct task_struct
*tsk
,
1463 struct files_struct
*files
)
1467 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1468 /* iopoll syncs against uring_lock, not completion_lock */
1469 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1470 mutex_lock(&ctx
->uring_lock
);
1471 ret
= __io_cqring_overflow_flush(ctx
, force
, tsk
, files
);
1472 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1473 mutex_unlock(&ctx
->uring_lock
);
1479 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1481 struct io_ring_ctx
*ctx
= req
->ctx
;
1482 struct io_uring_cqe
*cqe
;
1484 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1487 * If we can't get a cq entry, userspace overflowed the
1488 * submission (by quite a lot). Increment the overflow count in
1491 cqe
= io_get_cqring(ctx
);
1493 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1494 WRITE_ONCE(cqe
->res
, res
);
1495 WRITE_ONCE(cqe
->flags
, cflags
);
1496 } else if (ctx
->cq_overflow_flushed
||
1497 atomic_read(&req
->task
->io_uring
->in_idle
)) {
1499 * If we're in ring overflow flush mode, or in task cancel mode,
1500 * then we cannot store the request for later flushing, we need
1501 * to drop it on the floor.
1503 ctx
->cached_cq_overflow
++;
1504 WRITE_ONCE(ctx
->rings
->cq_overflow
, ctx
->cached_cq_overflow
);
1506 if (list_empty(&ctx
->cq_overflow_list
)) {
1507 set_bit(0, &ctx
->sq_check_overflow
);
1508 set_bit(0, &ctx
->cq_check_overflow
);
1509 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1513 req
->compl.cflags
= cflags
;
1514 refcount_inc(&req
->refs
);
1515 list_add_tail(&req
->compl.list
, &ctx
->cq_overflow_list
);
1519 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1521 __io_cqring_fill_event(req
, res
, 0);
1524 static void io_req_complete_post(struct io_kiocb
*req
, long res
,
1525 unsigned int cflags
)
1527 struct io_ring_ctx
*ctx
= req
->ctx
;
1528 unsigned long flags
;
1530 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1531 __io_cqring_fill_event(req
, res
, cflags
);
1533 * If we're the last reference to this request, add to our locked
1536 if (refcount_dec_and_test(&req
->refs
)) {
1537 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
1539 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1540 if (req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_FAIL_LINK
))
1541 io_disarm_next(req
);
1543 io_req_task_queue(req
->link
);
1547 io_dismantle_req(req
);
1548 io_put_task(req
->task
, 1);
1549 list_add(&req
->compl.list
, &cs
->locked_free_list
);
1550 cs
->locked_free_nr
++;
1552 if (!percpu_ref_tryget(&ctx
->refs
))
1555 io_commit_cqring(ctx
);
1556 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1559 io_cqring_ev_posted(ctx
);
1560 percpu_ref_put(&ctx
->refs
);
1564 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1565 unsigned int cflags
)
1569 req
->compl.cflags
= cflags
;
1570 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1573 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1574 long res
, unsigned cflags
)
1576 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1577 io_req_complete_state(req
, res
, cflags
);
1579 io_req_complete_post(req
, res
, cflags
);
1582 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1584 __io_req_complete(req
, 0, res
, 0);
1587 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1589 struct io_submit_state
*state
= &ctx
->submit_state
;
1590 struct io_comp_state
*cs
= &state
->comp
;
1591 struct io_kiocb
*req
= NULL
;
1594 * If we have more than a batch's worth of requests in our IRQ side
1595 * locked cache, grab the lock and move them over to our submission
1598 if (READ_ONCE(cs
->locked_free_nr
) > IO_COMPL_BATCH
) {
1599 spin_lock_irq(&ctx
->completion_lock
);
1600 list_splice_init(&cs
->locked_free_list
, &cs
->free_list
);
1601 cs
->locked_free_nr
= 0;
1602 spin_unlock_irq(&ctx
->completion_lock
);
1605 while (!list_empty(&cs
->free_list
)) {
1606 req
= list_first_entry(&cs
->free_list
, struct io_kiocb
,
1608 list_del(&req
->compl.list
);
1609 state
->reqs
[state
->free_reqs
++] = req
;
1610 if (state
->free_reqs
== ARRAY_SIZE(state
->reqs
))
1617 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1619 struct io_submit_state
*state
= &ctx
->submit_state
;
1621 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH
> ARRAY_SIZE(state
->reqs
));
1623 if (!state
->free_reqs
) {
1624 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1627 if (io_flush_cached_reqs(ctx
))
1630 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1634 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1635 * retry single alloc to be on the safe side.
1637 if (unlikely(ret
<= 0)) {
1638 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1639 if (!state
->reqs
[0])
1643 state
->free_reqs
= ret
;
1647 return state
->reqs
[state
->free_reqs
];
1650 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
1657 static void io_dismantle_req(struct io_kiocb
*req
)
1661 if (req
->async_data
)
1662 kfree(req
->async_data
);
1664 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
1665 if (req
->fixed_rsrc_refs
)
1666 percpu_ref_put(req
->fixed_rsrc_refs
);
1667 if (req
->work
.creds
) {
1668 put_cred(req
->work
.creds
);
1669 req
->work
.creds
= NULL
;
1672 if (req
->flags
& REQ_F_INFLIGHT
) {
1673 struct io_ring_ctx
*ctx
= req
->ctx
;
1674 unsigned long flags
;
1676 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1677 list_del(&req
->inflight_entry
);
1678 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1679 req
->flags
&= ~REQ_F_INFLIGHT
;
1683 /* must to be called somewhat shortly after putting a request */
1684 static inline void io_put_task(struct task_struct
*task
, int nr
)
1686 struct io_uring_task
*tctx
= task
->io_uring
;
1688 percpu_counter_sub(&tctx
->inflight
, nr
);
1689 if (unlikely(atomic_read(&tctx
->in_idle
)))
1690 wake_up(&tctx
->wait
);
1691 put_task_struct_many(task
, nr
);
1694 static void __io_free_req(struct io_kiocb
*req
)
1696 struct io_ring_ctx
*ctx
= req
->ctx
;
1698 io_dismantle_req(req
);
1699 io_put_task(req
->task
, 1);
1701 kmem_cache_free(req_cachep
, req
);
1702 percpu_ref_put(&ctx
->refs
);
1705 static inline void io_remove_next_linked(struct io_kiocb
*req
)
1707 struct io_kiocb
*nxt
= req
->link
;
1709 req
->link
= nxt
->link
;
1713 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
1714 __must_hold(&req
->ctx
->completion_lock
)
1716 struct io_kiocb
*link
= req
->link
;
1717 bool cancelled
= false;
1720 * Can happen if a linked timeout fired and link had been like
1721 * req -> link t-out -> link t-out [-> ...]
1723 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
1724 struct io_timeout_data
*io
= link
->async_data
;
1727 io_remove_next_linked(req
);
1728 link
->timeout
.head
= NULL
;
1729 ret
= hrtimer_try_to_cancel(&io
->timer
);
1731 io_cqring_fill_event(link
, -ECANCELED
);
1732 io_put_req_deferred(link
, 1);
1736 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1740 static void io_fail_links(struct io_kiocb
*req
)
1741 __must_hold(&req
->ctx
->completion_lock
)
1743 struct io_kiocb
*nxt
, *link
= req
->link
;
1750 trace_io_uring_fail_link(req
, link
);
1751 io_cqring_fill_event(link
, -ECANCELED
);
1752 io_put_req_deferred(link
, 2);
1757 static bool io_disarm_next(struct io_kiocb
*req
)
1758 __must_hold(&req
->ctx
->completion_lock
)
1760 bool posted
= false;
1762 if (likely(req
->flags
& REQ_F_LINK_TIMEOUT
))
1763 posted
= io_kill_linked_timeout(req
);
1764 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
1765 posted
|= (req
->link
!= NULL
);
1771 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
1773 struct io_kiocb
*nxt
;
1776 * If LINK is set, we have dependent requests in this chain. If we
1777 * didn't fail this request, queue the first one up, moving any other
1778 * dependencies to the next request. In case of failure, fail the rest
1781 if (req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_FAIL_LINK
)) {
1782 struct io_ring_ctx
*ctx
= req
->ctx
;
1783 unsigned long flags
;
1786 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1787 posted
= io_disarm_next(req
);
1789 io_commit_cqring(req
->ctx
);
1790 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1792 io_cqring_ev_posted(ctx
);
1799 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1801 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
1803 return __io_req_find_next(req
);
1806 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
)
1810 if (ctx
->submit_state
.comp
.nr
) {
1811 mutex_lock(&ctx
->uring_lock
);
1812 io_submit_flush_completions(&ctx
->submit_state
.comp
, ctx
);
1813 mutex_unlock(&ctx
->uring_lock
);
1815 percpu_ref_put(&ctx
->refs
);
1818 static bool __tctx_task_work(struct io_uring_task
*tctx
)
1820 struct io_ring_ctx
*ctx
= NULL
;
1821 struct io_wq_work_list list
;
1822 struct io_wq_work_node
*node
;
1824 if (wq_list_empty(&tctx
->task_list
))
1827 spin_lock_irq(&tctx
->task_lock
);
1828 list
= tctx
->task_list
;
1829 INIT_WQ_LIST(&tctx
->task_list
);
1830 spin_unlock_irq(&tctx
->task_lock
);
1834 struct io_wq_work_node
*next
= node
->next
;
1835 struct io_kiocb
*req
;
1837 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1838 if (req
->ctx
!= ctx
) {
1839 ctx_flush_and_put(ctx
);
1841 percpu_ref_get(&ctx
->refs
);
1844 req
->task_work
.func(&req
->task_work
);
1848 ctx_flush_and_put(ctx
);
1849 return list
.first
!= NULL
;
1852 static void tctx_task_work(struct callback_head
*cb
)
1854 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
, task_work
);
1856 clear_bit(0, &tctx
->task_state
);
1858 while (__tctx_task_work(tctx
))
1862 static int io_task_work_add(struct task_struct
*tsk
, struct io_kiocb
*req
,
1863 enum task_work_notify_mode notify
)
1865 struct io_uring_task
*tctx
= tsk
->io_uring
;
1866 struct io_wq_work_node
*node
, *prev
;
1867 unsigned long flags
;
1870 WARN_ON_ONCE(!tctx
);
1872 spin_lock_irqsave(&tctx
->task_lock
, flags
);
1873 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
1874 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
1876 /* task_work already pending, we're done */
1877 if (test_bit(0, &tctx
->task_state
) ||
1878 test_and_set_bit(0, &tctx
->task_state
))
1881 if (!task_work_add(tsk
, &tctx
->task_work
, notify
))
1885 * Slow path - we failed, find and delete work. if the work is not
1886 * in the list, it got run and we're fine.
1889 spin_lock_irqsave(&tctx
->task_lock
, flags
);
1890 wq_list_for_each(node
, prev
, &tctx
->task_list
) {
1891 if (&req
->io_task_work
.node
== node
) {
1892 wq_list_del(&tctx
->task_list
, node
, prev
);
1897 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
1898 clear_bit(0, &tctx
->task_state
);
1902 static int io_req_task_work_add(struct io_kiocb
*req
)
1904 struct task_struct
*tsk
= req
->task
;
1905 struct io_ring_ctx
*ctx
= req
->ctx
;
1906 enum task_work_notify_mode notify
;
1909 if (tsk
->flags
& PF_EXITING
)
1913 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1914 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1915 * processing task_work. There's no reliable way to tell if TWA_RESUME
1919 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
))
1920 notify
= TWA_SIGNAL
;
1922 ret
= io_task_work_add(tsk
, req
, notify
);
1924 wake_up_process(tsk
);
1929 static bool io_run_task_work_head(struct callback_head
**work_head
)
1931 struct callback_head
*work
, *next
;
1932 bool executed
= false;
1935 work
= xchg(work_head
, NULL
);
1951 static void io_task_work_add_head(struct callback_head
**work_head
,
1952 struct callback_head
*task_work
)
1954 struct callback_head
*head
;
1957 head
= READ_ONCE(*work_head
);
1958 task_work
->next
= head
;
1959 } while (cmpxchg(work_head
, head
, task_work
) != head
);
1962 static void io_req_task_work_add_fallback(struct io_kiocb
*req
,
1963 task_work_func_t cb
)
1965 init_task_work(&req
->task_work
, cb
);
1966 io_task_work_add_head(&req
->ctx
->exit_task_work
, &req
->task_work
);
1969 static void __io_req_task_cancel(struct io_kiocb
*req
, int error
)
1971 struct io_ring_ctx
*ctx
= req
->ctx
;
1973 spin_lock_irq(&ctx
->completion_lock
);
1974 io_cqring_fill_event(req
, error
);
1975 io_commit_cqring(ctx
);
1976 spin_unlock_irq(&ctx
->completion_lock
);
1978 io_cqring_ev_posted(ctx
);
1979 req_set_fail_links(req
);
1980 io_double_put_req(req
);
1983 static void io_req_task_cancel(struct callback_head
*cb
)
1985 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
1986 struct io_ring_ctx
*ctx
= req
->ctx
;
1988 mutex_lock(&ctx
->uring_lock
);
1989 __io_req_task_cancel(req
, req
->result
);
1990 mutex_unlock(&ctx
->uring_lock
);
1991 percpu_ref_put(&ctx
->refs
);
1994 static void __io_req_task_submit(struct io_kiocb
*req
)
1996 struct io_ring_ctx
*ctx
= req
->ctx
;
1998 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1999 mutex_lock(&ctx
->uring_lock
);
2000 if (!(current
->flags
& PF_EXITING
) && !current
->in_execve
)
2001 __io_queue_sqe(req
);
2003 __io_req_task_cancel(req
, -EFAULT
);
2004 mutex_unlock(&ctx
->uring_lock
);
2007 static void io_req_task_submit(struct callback_head
*cb
)
2009 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2011 __io_req_task_submit(req
);
2014 static void io_req_task_queue(struct io_kiocb
*req
)
2018 req
->task_work
.func
= io_req_task_submit
;
2019 ret
= io_req_task_work_add(req
);
2020 if (unlikely(ret
)) {
2021 req
->result
= -ECANCELED
;
2022 percpu_ref_get(&req
->ctx
->refs
);
2023 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2027 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2029 percpu_ref_get(&req
->ctx
->refs
);
2031 req
->task_work
.func
= io_req_task_cancel
;
2033 if (unlikely(io_req_task_work_add(req
)))
2034 io_req_task_work_add_fallback(req
, io_req_task_cancel
);
2037 static inline void io_queue_next(struct io_kiocb
*req
)
2039 struct io_kiocb
*nxt
= io_req_find_next(req
);
2042 io_req_task_queue(nxt
);
2045 static void io_free_req(struct io_kiocb
*req
)
2052 struct task_struct
*task
;
2057 static inline void io_init_req_batch(struct req_batch
*rb
)
2064 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2065 struct req_batch
*rb
)
2068 io_put_task(rb
->task
, rb
->task_refs
);
2070 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2073 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2074 struct io_submit_state
*state
)
2078 if (req
->task
!= rb
->task
) {
2080 io_put_task(rb
->task
, rb
->task_refs
);
2081 rb
->task
= req
->task
;
2087 io_dismantle_req(req
);
2088 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2089 state
->reqs
[state
->free_reqs
++] = req
;
2091 list_add(&req
->compl.list
, &state
->comp
.free_list
);
2094 static void io_submit_flush_completions(struct io_comp_state
*cs
,
2095 struct io_ring_ctx
*ctx
)
2098 struct io_kiocb
*req
;
2099 struct req_batch rb
;
2101 io_init_req_batch(&rb
);
2102 spin_lock_irq(&ctx
->completion_lock
);
2103 for (i
= 0; i
< nr
; i
++) {
2105 __io_cqring_fill_event(req
, req
->result
, req
->compl.cflags
);
2107 io_commit_cqring(ctx
);
2108 spin_unlock_irq(&ctx
->completion_lock
);
2110 io_cqring_ev_posted(ctx
);
2111 for (i
= 0; i
< nr
; i
++) {
2114 /* submission and completion refs */
2115 if (refcount_sub_and_test(2, &req
->refs
))
2116 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2119 io_req_free_batch_finish(ctx
, &rb
);
2124 * Drop reference to request, return next in chain (if there is one) if this
2125 * was the last reference to this request.
2127 static struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2129 struct io_kiocb
*nxt
= NULL
;
2131 if (refcount_dec_and_test(&req
->refs
)) {
2132 nxt
= io_req_find_next(req
);
2138 static void io_put_req(struct io_kiocb
*req
)
2140 if (refcount_dec_and_test(&req
->refs
))
2144 static void io_put_req_deferred_cb(struct callback_head
*cb
)
2146 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2151 static void io_free_req_deferred(struct io_kiocb
*req
)
2155 req
->task_work
.func
= io_put_req_deferred_cb
;
2156 ret
= io_req_task_work_add(req
);
2158 io_req_task_work_add_fallback(req
, io_put_req_deferred_cb
);
2161 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2163 if (refcount_sub_and_test(refs
, &req
->refs
))
2164 io_free_req_deferred(req
);
2167 static void io_double_put_req(struct io_kiocb
*req
)
2169 /* drop both submit and complete references */
2170 if (refcount_sub_and_test(2, &req
->refs
))
2174 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2176 /* See comment at the top of this file */
2178 return __io_cqring_events(ctx
);
2181 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2183 struct io_rings
*rings
= ctx
->rings
;
2185 /* make sure SQ entry isn't read before tail */
2186 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2189 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2191 unsigned int cflags
;
2193 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2194 cflags
|= IORING_CQE_F_BUFFER
;
2195 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2200 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2202 struct io_buffer
*kbuf
;
2204 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2205 return io_put_kbuf(req
, kbuf
);
2208 static inline bool io_run_task_work(void)
2211 * Not safe to run on exiting task, and the task_work handling will
2212 * not add work to such a task.
2214 if (unlikely(current
->flags
& PF_EXITING
))
2216 if (current
->task_works
) {
2217 __set_current_state(TASK_RUNNING
);
2226 * Find and free completed poll iocbs
2228 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2229 struct list_head
*done
)
2231 struct req_batch rb
;
2232 struct io_kiocb
*req
;
2234 /* order with ->result store in io_complete_rw_iopoll() */
2237 io_init_req_batch(&rb
);
2238 while (!list_empty(done
)) {
2241 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2242 list_del(&req
->inflight_entry
);
2244 if (READ_ONCE(req
->result
) == -EAGAIN
) {
2245 req
->iopoll_completed
= 0;
2246 if (io_rw_reissue(req
))
2250 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2251 cflags
= io_put_rw_kbuf(req
);
2253 __io_cqring_fill_event(req
, req
->result
, cflags
);
2256 if (refcount_dec_and_test(&req
->refs
))
2257 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2260 io_commit_cqring(ctx
);
2261 io_cqring_ev_posted_iopoll(ctx
);
2262 io_req_free_batch_finish(ctx
, &rb
);
2265 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2268 struct io_kiocb
*req
, *tmp
;
2274 * Only spin for completions if we don't have multiple devices hanging
2275 * off our complete list, and we're under the requested amount.
2277 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
2280 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2281 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2284 * Move completed and retryable entries to our local lists.
2285 * If we find a request that requires polling, break out
2286 * and complete those lists first, if we have entries there.
2288 if (READ_ONCE(req
->iopoll_completed
)) {
2289 list_move_tail(&req
->inflight_entry
, &done
);
2292 if (!list_empty(&done
))
2295 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2299 /* iopoll may have completed current req */
2300 if (READ_ONCE(req
->iopoll_completed
))
2301 list_move_tail(&req
->inflight_entry
, &done
);
2308 if (!list_empty(&done
))
2309 io_iopoll_complete(ctx
, nr_events
, &done
);
2315 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2316 * non-spinning poll check - we'll still enter the driver poll loop, but only
2317 * as a non-spinning completion check.
2319 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2322 while (!list_empty(&ctx
->iopoll_list
) && !need_resched()) {
2325 ret
= io_do_iopoll(ctx
, nr_events
, min
);
2328 if (*nr_events
>= min
)
2336 * We can't just wait for polled events to come to us, we have to actively
2337 * find and complete them.
2339 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2341 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2344 mutex_lock(&ctx
->uring_lock
);
2345 while (!list_empty(&ctx
->iopoll_list
)) {
2346 unsigned int nr_events
= 0;
2348 io_do_iopoll(ctx
, &nr_events
, 0);
2350 /* let it sleep and repeat later if can't complete a request */
2354 * Ensure we allow local-to-the-cpu processing to take place,
2355 * in this case we need to ensure that we reap all events.
2356 * Also let task_work, etc. to progress by releasing the mutex
2358 if (need_resched()) {
2359 mutex_unlock(&ctx
->uring_lock
);
2361 mutex_lock(&ctx
->uring_lock
);
2364 mutex_unlock(&ctx
->uring_lock
);
2367 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2369 unsigned int nr_events
= 0;
2370 int iters
= 0, ret
= 0;
2373 * We disallow the app entering submit/complete with polling, but we
2374 * still need to lock the ring to prevent racing with polled issue
2375 * that got punted to a workqueue.
2377 mutex_lock(&ctx
->uring_lock
);
2380 * Don't enter poll loop if we already have events pending.
2381 * If we do, we can potentially be spinning for commands that
2382 * already triggered a CQE (eg in error).
2384 if (test_bit(0, &ctx
->cq_check_overflow
))
2385 __io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
2386 if (io_cqring_events(ctx
))
2390 * If a submit got punted to a workqueue, we can have the
2391 * application entering polling for a command before it gets
2392 * issued. That app will hold the uring_lock for the duration
2393 * of the poll right here, so we need to take a breather every
2394 * now and then to ensure that the issue has a chance to add
2395 * the poll to the issued list. Otherwise we can spin here
2396 * forever, while the workqueue is stuck trying to acquire the
2399 if (!(++iters
& 7)) {
2400 mutex_unlock(&ctx
->uring_lock
);
2402 mutex_lock(&ctx
->uring_lock
);
2405 ret
= io_iopoll_getevents(ctx
, &nr_events
, min
);
2409 } while (min
&& !nr_events
&& !need_resched());
2411 mutex_unlock(&ctx
->uring_lock
);
2415 static void kiocb_end_write(struct io_kiocb
*req
)
2418 * Tell lockdep we inherited freeze protection from submission
2421 if (req
->flags
& REQ_F_ISREG
) {
2422 struct inode
*inode
= file_inode(req
->file
);
2424 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
2426 file_end_write(req
->file
);
2430 static bool io_resubmit_prep(struct io_kiocb
*req
)
2432 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2434 struct iov_iter iter
;
2436 /* already prepared */
2437 if (req
->async_data
)
2440 switch (req
->opcode
) {
2441 case IORING_OP_READV
:
2442 case IORING_OP_READ_FIXED
:
2443 case IORING_OP_READ
:
2446 case IORING_OP_WRITEV
:
2447 case IORING_OP_WRITE_FIXED
:
2448 case IORING_OP_WRITE
:
2452 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2457 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2460 return !io_setup_async_rw(req
, iovec
, inline_vecs
, &iter
, false);
2463 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2465 umode_t mode
= file_inode(req
->file
)->i_mode
;
2466 struct io_ring_ctx
*ctx
= req
->ctx
;
2468 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2470 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2471 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2474 * If ref is dying, we might be running poll reap from the exit work.
2475 * Don't attempt to reissue from that path, just let it fail with
2478 if (percpu_ref_is_dying(&ctx
->refs
))
2483 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2489 static bool io_rw_reissue(struct io_kiocb
*req
)
2492 if (!io_rw_should_reissue(req
))
2495 lockdep_assert_held(&req
->ctx
->uring_lock
);
2497 if (io_resubmit_prep(req
)) {
2498 refcount_inc(&req
->refs
);
2499 io_queue_async_work(req
);
2502 req_set_fail_links(req
);
2507 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2508 unsigned int issue_flags
)
2512 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
)
2513 kiocb_end_write(req
);
2514 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) && io_rw_should_reissue(req
)) {
2515 req
->flags
|= REQ_F_REISSUE
;
2518 if (res
!= req
->result
)
2519 req_set_fail_links(req
);
2520 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2521 cflags
= io_put_rw_kbuf(req
);
2522 __io_req_complete(req
, issue_flags
, res
, cflags
);
2525 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2527 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2529 __io_complete_rw(req
, res
, res2
, 0);
2532 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2534 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2537 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2538 if (res
== -EAGAIN
&& io_rw_should_reissue(req
)) {
2539 struct io_async_rw
*rw
= req
->async_data
;
2542 iov_iter_revert(&rw
->iter
,
2543 req
->result
- iov_iter_count(&rw
->iter
));
2544 else if (!io_resubmit_prep(req
))
2549 if (kiocb
->ki_flags
& IOCB_WRITE
)
2550 kiocb_end_write(req
);
2552 if (res
!= -EAGAIN
&& res
!= req
->result
)
2553 req_set_fail_links(req
);
2555 WRITE_ONCE(req
->result
, res
);
2556 /* order with io_poll_complete() checking ->result */
2558 WRITE_ONCE(req
->iopoll_completed
, 1);
2562 * After the iocb has been issued, it's safe to be found on the poll list.
2563 * Adding the kiocb to the list AFTER submission ensures that we don't
2564 * find it from a io_iopoll_getevents() thread before the issuer is done
2565 * accessing the kiocb cookie.
2567 static void io_iopoll_req_issued(struct io_kiocb
*req
, bool in_async
)
2569 struct io_ring_ctx
*ctx
= req
->ctx
;
2572 * Track whether we have multiple files in our lists. This will impact
2573 * how we do polling eventually, not spinning if we're on potentially
2574 * different devices.
2576 if (list_empty(&ctx
->iopoll_list
)) {
2577 ctx
->poll_multi_file
= false;
2578 } else if (!ctx
->poll_multi_file
) {
2579 struct io_kiocb
*list_req
;
2581 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2583 if (list_req
->file
!= req
->file
)
2584 ctx
->poll_multi_file
= true;
2588 * For fast devices, IO may have already completed. If it has, add
2589 * it to the front so we find it first.
2591 if (READ_ONCE(req
->iopoll_completed
))
2592 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2594 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2597 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2598 * task context or in io worker task context. If current task context is
2599 * sq thread, we don't need to check whether should wake up sq thread.
2601 if (in_async
&& (ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2602 wq_has_sleeper(&ctx
->sq_data
->wait
))
2603 wake_up(&ctx
->sq_data
->wait
);
2606 static inline void io_state_file_put(struct io_submit_state
*state
)
2608 if (state
->file_refs
) {
2609 fput_many(state
->file
, state
->file_refs
);
2610 state
->file_refs
= 0;
2615 * Get as many references to a file as we have IOs left in this submission,
2616 * assuming most submissions are for one file, or at least that each file
2617 * has more than one submission.
2619 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2624 if (state
->file_refs
) {
2625 if (state
->fd
== fd
) {
2629 io_state_file_put(state
);
2631 state
->file
= fget_many(fd
, state
->ios_left
);
2632 if (unlikely(!state
->file
))
2636 state
->file_refs
= state
->ios_left
- 1;
2640 static bool io_bdev_nowait(struct block_device
*bdev
)
2642 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2646 * If we tracked the file through the SCM inflight mechanism, we could support
2647 * any file. For now, just ensure that anything potentially problematic is done
2650 static bool io_file_supports_async(struct file
*file
, int rw
)
2652 umode_t mode
= file_inode(file
)->i_mode
;
2654 if (S_ISBLK(mode
)) {
2655 if (IS_ENABLED(CONFIG_BLOCK
) &&
2656 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2660 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2662 if (S_ISREG(mode
)) {
2663 if (IS_ENABLED(CONFIG_BLOCK
) &&
2664 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2665 file
->f_op
!= &io_uring_fops
)
2670 /* any ->read/write should understand O_NONBLOCK */
2671 if (file
->f_flags
& O_NONBLOCK
)
2674 if (!(file
->f_mode
& FMODE_NOWAIT
))
2678 return file
->f_op
->read_iter
!= NULL
;
2680 return file
->f_op
->write_iter
!= NULL
;
2683 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2685 struct io_ring_ctx
*ctx
= req
->ctx
;
2686 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2687 struct file
*file
= req
->file
;
2691 if (S_ISREG(file_inode(file
)->i_mode
))
2692 req
->flags
|= REQ_F_ISREG
;
2694 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2695 if (kiocb
->ki_pos
== -1 && !(file
->f_mode
& FMODE_STREAM
)) {
2696 req
->flags
|= REQ_F_CUR_POS
;
2697 kiocb
->ki_pos
= file
->f_pos
;
2699 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2700 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2701 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2705 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2706 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) || (file
->f_flags
& O_NONBLOCK
))
2707 req
->flags
|= REQ_F_NOWAIT
;
2709 ioprio
= READ_ONCE(sqe
->ioprio
);
2711 ret
= ioprio_check_cap(ioprio
);
2715 kiocb
->ki_ioprio
= ioprio
;
2717 kiocb
->ki_ioprio
= get_current_ioprio();
2719 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2720 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2721 !kiocb
->ki_filp
->f_op
->iopoll
)
2724 kiocb
->ki_flags
|= IOCB_HIPRI
;
2725 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2726 req
->iopoll_completed
= 0;
2728 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2730 kiocb
->ki_complete
= io_complete_rw
;
2733 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2734 req
->rw
.len
= READ_ONCE(sqe
->len
);
2735 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2739 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2745 case -ERESTARTNOINTR
:
2746 case -ERESTARTNOHAND
:
2747 case -ERESTART_RESTARTBLOCK
:
2749 * We can't just restart the syscall, since previously
2750 * submitted sqes may already be in progress. Just fail this
2756 kiocb
->ki_complete(kiocb
, ret
, 0);
2760 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2761 unsigned int issue_flags
)
2763 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2764 struct io_async_rw
*io
= req
->async_data
;
2765 bool check_reissue
= kiocb
->ki_complete
== io_complete_rw
;
2767 /* add previously done IO, if any */
2768 if (io
&& io
->bytes_done
> 0) {
2770 ret
= io
->bytes_done
;
2772 ret
+= io
->bytes_done
;
2775 if (req
->flags
& REQ_F_CUR_POS
)
2776 req
->file
->f_pos
= kiocb
->ki_pos
;
2777 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2778 __io_complete_rw(req
, ret
, 0, issue_flags
);
2780 io_rw_done(kiocb
, ret
);
2782 if (check_reissue
&& req
->flags
& REQ_F_REISSUE
) {
2783 req
->flags
&= ~REQ_F_REISSUE
;
2784 if (!io_rw_reissue(req
)) {
2787 req_set_fail_links(req
);
2788 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2789 cflags
= io_put_rw_kbuf(req
);
2790 __io_req_complete(req
, issue_flags
, ret
, cflags
);
2795 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
2797 struct io_ring_ctx
*ctx
= req
->ctx
;
2798 size_t len
= req
->rw
.len
;
2799 struct io_mapped_ubuf
*imu
;
2800 u16 index
, buf_index
= req
->buf_index
;
2804 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
2806 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
2807 imu
= &ctx
->user_bufs
[index
];
2808 buf_addr
= req
->rw
.addr
;
2811 if (buf_addr
+ len
< buf_addr
)
2813 /* not inside the mapped region */
2814 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
2818 * May not be a start of buffer, set size appropriately
2819 * and advance us to the beginning.
2821 offset
= buf_addr
- imu
->ubuf
;
2822 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
2826 * Don't use iov_iter_advance() here, as it's really slow for
2827 * using the latter parts of a big fixed buffer - it iterates
2828 * over each segment manually. We can cheat a bit here, because
2831 * 1) it's a BVEC iter, we set it up
2832 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2833 * first and last bvec
2835 * So just find our index, and adjust the iterator afterwards.
2836 * If the offset is within the first bvec (or the whole first
2837 * bvec, just use iov_iter_advance(). This makes it easier
2838 * since we can just skip the first segment, which may not
2839 * be PAGE_SIZE aligned.
2841 const struct bio_vec
*bvec
= imu
->bvec
;
2843 if (offset
<= bvec
->bv_len
) {
2844 iov_iter_advance(iter
, offset
);
2846 unsigned long seg_skip
;
2848 /* skip first vec */
2849 offset
-= bvec
->bv_len
;
2850 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
2852 iter
->bvec
= bvec
+ seg_skip
;
2853 iter
->nr_segs
-= seg_skip
;
2854 iter
->count
-= bvec
->bv_len
+ offset
;
2855 iter
->iov_offset
= offset
& ~PAGE_MASK
;
2862 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2865 mutex_unlock(&ctx
->uring_lock
);
2868 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2871 * "Normal" inline submissions always hold the uring_lock, since we
2872 * grab it from the system call. Same is true for the SQPOLL offload.
2873 * The only exception is when we've detached the request and issue it
2874 * from an async worker thread, grab the lock for that case.
2877 mutex_lock(&ctx
->uring_lock
);
2880 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
2881 int bgid
, struct io_buffer
*kbuf
,
2884 struct io_buffer
*head
;
2886 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2889 io_ring_submit_lock(req
->ctx
, needs_lock
);
2891 lockdep_assert_held(&req
->ctx
->uring_lock
);
2893 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
2895 if (!list_empty(&head
->list
)) {
2896 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
2898 list_del(&kbuf
->list
);
2901 xa_erase(&req
->ctx
->io_buffers
, bgid
);
2903 if (*len
> kbuf
->len
)
2906 kbuf
= ERR_PTR(-ENOBUFS
);
2909 io_ring_submit_unlock(req
->ctx
, needs_lock
);
2914 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
2917 struct io_buffer
*kbuf
;
2920 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2921 bgid
= req
->buf_index
;
2922 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
2925 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
2926 req
->flags
|= REQ_F_BUFFER_SELECTED
;
2927 return u64_to_user_ptr(kbuf
->addr
);
2930 #ifdef CONFIG_COMPAT
2931 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
2934 struct compat_iovec __user
*uiov
;
2935 compat_ssize_t clen
;
2939 uiov
= u64_to_user_ptr(req
->rw
.addr
);
2940 if (!access_ok(uiov
, sizeof(*uiov
)))
2942 if (__get_user(clen
, &uiov
->iov_len
))
2948 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2950 return PTR_ERR(buf
);
2951 iov
[0].iov_base
= buf
;
2952 iov
[0].iov_len
= (compat_size_t
) len
;
2957 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2960 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
2964 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
2967 len
= iov
[0].iov_len
;
2970 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2972 return PTR_ERR(buf
);
2973 iov
[0].iov_base
= buf
;
2974 iov
[0].iov_len
= len
;
2978 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2981 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
2982 struct io_buffer
*kbuf
;
2984 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2985 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
2986 iov
[0].iov_len
= kbuf
->len
;
2989 if (req
->rw
.len
!= 1)
2992 #ifdef CONFIG_COMPAT
2993 if (req
->ctx
->compat
)
2994 return io_compat_import(req
, iov
, needs_lock
);
2997 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3000 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
3001 struct iov_iter
*iter
, bool needs_lock
)
3003 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3004 size_t sqe_len
= req
->rw
.len
;
3005 u8 opcode
= req
->opcode
;
3008 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3010 return io_import_fixed(req
, rw
, iter
);
3013 /* buffer index only valid with fixed read/write, or buffer select */
3014 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3017 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3018 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3019 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3021 return PTR_ERR(buf
);
3022 req
->rw
.len
= sqe_len
;
3025 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3030 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3031 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3033 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3038 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3042 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3044 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3048 * For files that don't have ->read_iter() and ->write_iter(), handle them
3049 * by looping over ->read() or ->write() manually.
3051 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3053 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3054 struct file
*file
= req
->file
;
3058 * Don't support polled IO through this interface, and we can't
3059 * support non-blocking either. For the latter, this just causes
3060 * the kiocb to be handled from an async context.
3062 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3064 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3067 while (iov_iter_count(iter
)) {
3071 if (!iov_iter_is_bvec(iter
)) {
3072 iovec
= iov_iter_iovec(iter
);
3074 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3075 iovec
.iov_len
= req
->rw
.len
;
3079 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3080 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3082 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3083 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3092 if (nr
!= iovec
.iov_len
)
3096 iov_iter_advance(iter
, nr
);
3102 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3103 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3105 struct io_async_rw
*rw
= req
->async_data
;
3107 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3108 rw
->free_iovec
= iovec
;
3110 /* can only be fixed buffers, no need to do anything */
3111 if (iov_iter_is_bvec(iter
))
3114 unsigned iov_off
= 0;
3116 rw
->iter
.iov
= rw
->fast_iov
;
3117 if (iter
->iov
!= fast_iov
) {
3118 iov_off
= iter
->iov
- fast_iov
;
3119 rw
->iter
.iov
+= iov_off
;
3121 if (rw
->fast_iov
!= fast_iov
)
3122 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3123 sizeof(struct iovec
) * iter
->nr_segs
);
3125 req
->flags
|= REQ_F_NEED_CLEANUP
;
3129 static inline int __io_alloc_async_data(struct io_kiocb
*req
)
3131 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3132 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3133 return req
->async_data
== NULL
;
3136 static int io_alloc_async_data(struct io_kiocb
*req
)
3138 if (!io_op_defs
[req
->opcode
].needs_async_data
)
3141 return __io_alloc_async_data(req
);
3144 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3145 const struct iovec
*fast_iov
,
3146 struct iov_iter
*iter
, bool force
)
3148 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_data
)
3150 if (!req
->async_data
) {
3151 if (__io_alloc_async_data(req
)) {
3156 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3161 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3163 struct io_async_rw
*iorw
= req
->async_data
;
3164 struct iovec
*iov
= iorw
->fast_iov
;
3167 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3168 if (unlikely(ret
< 0))
3171 iorw
->bytes_done
= 0;
3172 iorw
->free_iovec
= iov
;
3174 req
->flags
|= REQ_F_NEED_CLEANUP
;
3178 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3180 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3182 return io_prep_rw(req
, sqe
);
3186 * This is our waitqueue callback handler, registered through lock_page_async()
3187 * when we initially tried to do the IO with the iocb armed our waitqueue.
3188 * This gets called when the page is unlocked, and we generally expect that to
3189 * happen when the page IO is completed and the page is now uptodate. This will
3190 * queue a task_work based retry of the operation, attempting to copy the data
3191 * again. If the latter fails because the page was NOT uptodate, then we will
3192 * do a thread based blocking retry of the operation. That's the unexpected
3195 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3196 int sync
, void *arg
)
3198 struct wait_page_queue
*wpq
;
3199 struct io_kiocb
*req
= wait
->private;
3200 struct wait_page_key
*key
= arg
;
3202 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3204 if (!wake_page_match(wpq
, key
))
3207 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3208 list_del_init(&wait
->entry
);
3210 /* submit ref gets dropped, acquire a new one */
3211 refcount_inc(&req
->refs
);
3212 io_req_task_queue(req
);
3217 * This controls whether a given IO request should be armed for async page
3218 * based retry. If we return false here, the request is handed to the async
3219 * worker threads for retry. If we're doing buffered reads on a regular file,
3220 * we prepare a private wait_page_queue entry and retry the operation. This
3221 * will either succeed because the page is now uptodate and unlocked, or it
3222 * will register a callback when the page is unlocked at IO completion. Through
3223 * that callback, io_uring uses task_work to setup a retry of the operation.
3224 * That retry will attempt the buffered read again. The retry will generally
3225 * succeed, or in rare cases where it fails, we then fall back to using the
3226 * async worker threads for a blocking retry.
3228 static bool io_rw_should_retry(struct io_kiocb
*req
)
3230 struct io_async_rw
*rw
= req
->async_data
;
3231 struct wait_page_queue
*wait
= &rw
->wpq
;
3232 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3234 /* never retry for NOWAIT, we just complete with -EAGAIN */
3235 if (req
->flags
& REQ_F_NOWAIT
)
3238 /* Only for buffered IO */
3239 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3243 * just use poll if we can, and don't attempt if the fs doesn't
3244 * support callback based unlocks
3246 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3249 wait
->wait
.func
= io_async_buf_func
;
3250 wait
->wait
.private = req
;
3251 wait
->wait
.flags
= 0;
3252 INIT_LIST_HEAD(&wait
->wait
.entry
);
3253 kiocb
->ki_flags
|= IOCB_WAITQ
;
3254 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3255 kiocb
->ki_waitq
= wait
;
3259 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3261 if (req
->file
->f_op
->read_iter
)
3262 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3263 else if (req
->file
->f_op
->read
)
3264 return loop_rw_iter(READ
, req
, iter
);
3269 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3271 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3272 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3273 struct iov_iter __iter
, *iter
= &__iter
;
3274 struct io_async_rw
*rw
= req
->async_data
;
3275 ssize_t io_size
, ret
, ret2
;
3276 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3282 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3286 io_size
= iov_iter_count(iter
);
3287 req
->result
= io_size
;
3289 /* Ensure we clear previously set non-block flag */
3290 if (!force_nonblock
)
3291 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3293 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3295 /* If the file doesn't support async, just async punt */
3296 if (force_nonblock
&& !io_file_supports_async(req
->file
, READ
)) {
3297 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3298 return ret
?: -EAGAIN
;
3301 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3302 if (unlikely(ret
)) {
3307 ret
= io_iter_do_read(req
, iter
);
3309 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
3310 req
->flags
&= ~REQ_F_REISSUE
;
3311 /* IOPOLL retry should happen for io-wq threads */
3312 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3314 /* no retry on NONBLOCK nor RWF_NOWAIT */
3315 if (req
->flags
& REQ_F_NOWAIT
)
3317 /* some cases will consume bytes even on error returns */
3318 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3320 } else if (ret
== -EIOCBQUEUED
) {
3322 } else if (ret
<= 0 || ret
== io_size
|| !force_nonblock
||
3323 (req
->flags
& REQ_F_NOWAIT
) || !(req
->flags
& REQ_F_ISREG
)) {
3324 /* read all, failed, already did sync or don't want to retry */
3328 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3333 rw
= req
->async_data
;
3334 /* now use our persistent iterator, if we aren't already */
3339 rw
->bytes_done
+= ret
;
3340 /* if we can retry, do so with the callbacks armed */
3341 if (!io_rw_should_retry(req
)) {
3342 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3347 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3348 * we get -EIOCBQUEUED, then we'll get a notification when the
3349 * desired page gets unlocked. We can also get a partial read
3350 * here, and if we do, then just retry at the new offset.
3352 ret
= io_iter_do_read(req
, iter
);
3353 if (ret
== -EIOCBQUEUED
)
3355 /* we got some bytes, but not all. retry. */
3356 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3357 } while (ret
> 0 && ret
< io_size
);
3359 kiocb_done(kiocb
, ret
, issue_flags
);
3361 /* it's faster to check here then delegate to kfree */
3367 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3369 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3371 return io_prep_rw(req
, sqe
);
3374 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3376 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3377 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3378 struct iov_iter __iter
, *iter
= &__iter
;
3379 struct io_async_rw
*rw
= req
->async_data
;
3380 ssize_t ret
, ret2
, io_size
;
3381 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3387 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3391 io_size
= iov_iter_count(iter
);
3392 req
->result
= io_size
;
3394 /* Ensure we clear previously set non-block flag */
3395 if (!force_nonblock
)
3396 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3398 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3400 /* If the file doesn't support async, just async punt */
3401 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3404 /* file path doesn't support NOWAIT for non-direct_IO */
3405 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3406 (req
->flags
& REQ_F_ISREG
))
3409 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3414 * Open-code file_start_write here to grab freeze protection,
3415 * which will be released by another thread in
3416 * io_complete_rw(). Fool lockdep by telling it the lock got
3417 * released so that it doesn't complain about the held lock when
3418 * we return to userspace.
3420 if (req
->flags
& REQ_F_ISREG
) {
3421 sb_start_write(file_inode(req
->file
)->i_sb
);
3422 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3425 kiocb
->ki_flags
|= IOCB_WRITE
;
3427 if (req
->file
->f_op
->write_iter
)
3428 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3429 else if (req
->file
->f_op
->write
)
3430 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3434 if (req
->flags
& REQ_F_REISSUE
) {
3435 req
->flags
&= ~REQ_F_REISSUE
;
3440 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3441 * retry them without IOCB_NOWAIT.
3443 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3445 /* no retry on NONBLOCK nor RWF_NOWAIT */
3446 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3448 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3449 /* IOPOLL retry should happen for io-wq threads */
3450 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3453 kiocb_done(kiocb
, ret2
, issue_flags
);
3456 /* some cases will consume bytes even on error returns */
3457 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3458 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3459 return ret
?: -EAGAIN
;
3462 /* it's reportedly faster than delegating the null check to kfree() */
3468 static int io_renameat_prep(struct io_kiocb
*req
,
3469 const struct io_uring_sqe
*sqe
)
3471 struct io_rename
*ren
= &req
->rename
;
3472 const char __user
*oldf
, *newf
;
3474 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3477 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3478 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3479 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3480 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3481 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3483 ren
->oldpath
= getname(oldf
);
3484 if (IS_ERR(ren
->oldpath
))
3485 return PTR_ERR(ren
->oldpath
);
3487 ren
->newpath
= getname(newf
);
3488 if (IS_ERR(ren
->newpath
)) {
3489 putname(ren
->oldpath
);
3490 return PTR_ERR(ren
->newpath
);
3493 req
->flags
|= REQ_F_NEED_CLEANUP
;
3497 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3499 struct io_rename
*ren
= &req
->rename
;
3502 if (issue_flags
& IO_URING_F_NONBLOCK
)
3505 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3506 ren
->newpath
, ren
->flags
);
3508 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3510 req_set_fail_links(req
);
3511 io_req_complete(req
, ret
);
3515 static int io_unlinkat_prep(struct io_kiocb
*req
,
3516 const struct io_uring_sqe
*sqe
)
3518 struct io_unlink
*un
= &req
->unlink
;
3519 const char __user
*fname
;
3521 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3524 un
->dfd
= READ_ONCE(sqe
->fd
);
3526 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3527 if (un
->flags
& ~AT_REMOVEDIR
)
3530 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3531 un
->filename
= getname(fname
);
3532 if (IS_ERR(un
->filename
))
3533 return PTR_ERR(un
->filename
);
3535 req
->flags
|= REQ_F_NEED_CLEANUP
;
3539 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3541 struct io_unlink
*un
= &req
->unlink
;
3544 if (issue_flags
& IO_URING_F_NONBLOCK
)
3547 if (un
->flags
& AT_REMOVEDIR
)
3548 ret
= do_rmdir(un
->dfd
, un
->filename
);
3550 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3552 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3554 req_set_fail_links(req
);
3555 io_req_complete(req
, ret
);
3559 static int io_shutdown_prep(struct io_kiocb
*req
,
3560 const struct io_uring_sqe
*sqe
)
3562 #if defined(CONFIG_NET)
3563 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3565 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3569 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3576 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3578 #if defined(CONFIG_NET)
3579 struct socket
*sock
;
3582 if (issue_flags
& IO_URING_F_NONBLOCK
)
3585 sock
= sock_from_file(req
->file
);
3586 if (unlikely(!sock
))
3589 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3591 req_set_fail_links(req
);
3592 io_req_complete(req
, ret
);
3599 static int __io_splice_prep(struct io_kiocb
*req
,
3600 const struct io_uring_sqe
*sqe
)
3602 struct io_splice
* sp
= &req
->splice
;
3603 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3605 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3609 sp
->len
= READ_ONCE(sqe
->len
);
3610 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3612 if (unlikely(sp
->flags
& ~valid_flags
))
3615 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3616 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3619 req
->flags
|= REQ_F_NEED_CLEANUP
;
3621 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3623 * Splice operation will be punted aync, and here need to
3624 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3626 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3632 static int io_tee_prep(struct io_kiocb
*req
,
3633 const struct io_uring_sqe
*sqe
)
3635 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3637 return __io_splice_prep(req
, sqe
);
3640 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
3642 struct io_splice
*sp
= &req
->splice
;
3643 struct file
*in
= sp
->file_in
;
3644 struct file
*out
= sp
->file_out
;
3645 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3648 if (issue_flags
& IO_URING_F_NONBLOCK
)
3651 ret
= do_tee(in
, out
, sp
->len
, flags
);
3653 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3654 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3657 req_set_fail_links(req
);
3658 io_req_complete(req
, ret
);
3662 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3664 struct io_splice
* sp
= &req
->splice
;
3666 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3667 sp
->off_out
= READ_ONCE(sqe
->off
);
3668 return __io_splice_prep(req
, sqe
);
3671 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
3673 struct io_splice
*sp
= &req
->splice
;
3674 struct file
*in
= sp
->file_in
;
3675 struct file
*out
= sp
->file_out
;
3676 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3677 loff_t
*poff_in
, *poff_out
;
3680 if (issue_flags
& IO_URING_F_NONBLOCK
)
3683 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3684 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3687 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3689 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3690 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3693 req_set_fail_links(req
);
3694 io_req_complete(req
, ret
);
3699 * IORING_OP_NOP just posts a completion event, nothing else.
3701 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
3703 struct io_ring_ctx
*ctx
= req
->ctx
;
3705 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3708 __io_req_complete(req
, issue_flags
, 0, 0);
3712 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3714 struct io_ring_ctx
*ctx
= req
->ctx
;
3719 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3721 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3724 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3725 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3728 req
->sync
.off
= READ_ONCE(sqe
->off
);
3729 req
->sync
.len
= READ_ONCE(sqe
->len
);
3733 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
3735 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3738 /* fsync always requires a blocking context */
3739 if (issue_flags
& IO_URING_F_NONBLOCK
)
3742 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3743 end
> 0 ? end
: LLONG_MAX
,
3744 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3746 req_set_fail_links(req
);
3747 io_req_complete(req
, ret
);
3751 static int io_fallocate_prep(struct io_kiocb
*req
,
3752 const struct io_uring_sqe
*sqe
)
3754 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3756 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3759 req
->sync
.off
= READ_ONCE(sqe
->off
);
3760 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3761 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3765 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
3769 /* fallocate always requiring blocking context */
3770 if (issue_flags
& IO_URING_F_NONBLOCK
)
3772 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3775 req_set_fail_links(req
);
3776 io_req_complete(req
, ret
);
3780 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3782 const char __user
*fname
;
3785 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3787 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3790 /* open.how should be already initialised */
3791 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
3792 req
->open
.how
.flags
|= O_LARGEFILE
;
3794 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3795 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3796 req
->open
.filename
= getname(fname
);
3797 if (IS_ERR(req
->open
.filename
)) {
3798 ret
= PTR_ERR(req
->open
.filename
);
3799 req
->open
.filename
= NULL
;
3802 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
3803 req
->flags
|= REQ_F_NEED_CLEANUP
;
3807 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3811 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3813 mode
= READ_ONCE(sqe
->len
);
3814 flags
= READ_ONCE(sqe
->open_flags
);
3815 req
->open
.how
= build_open_how(flags
, mode
);
3816 return __io_openat_prep(req
, sqe
);
3819 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3821 struct open_how __user
*how
;
3825 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3827 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3828 len
= READ_ONCE(sqe
->len
);
3829 if (len
< OPEN_HOW_SIZE_VER0
)
3832 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
3837 return __io_openat_prep(req
, sqe
);
3840 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
3842 struct open_flags op
;
3845 bool resolve_nonblock
;
3848 ret
= build_open_flags(&req
->open
.how
, &op
);
3851 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
3852 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
3853 if (issue_flags
& IO_URING_F_NONBLOCK
) {
3855 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3856 * it'll always -EAGAIN
3858 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
3860 op
.lookup_flags
|= LOOKUP_CACHED
;
3861 op
.open_flag
|= O_NONBLOCK
;
3864 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
3868 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
3869 /* only retry if RESOLVE_CACHED wasn't already set by application */
3870 if ((!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)) &&
3871 file
== ERR_PTR(-EAGAIN
)) {
3873 * We could hang on to this 'fd', but seems like marginal
3874 * gain for something that is now known to be a slower path.
3875 * So just put it, and we'll get a new one when we retry.
3883 ret
= PTR_ERR(file
);
3885 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
3886 file
->f_flags
&= ~O_NONBLOCK
;
3887 fsnotify_open(file
);
3888 fd_install(ret
, file
);
3891 putname(req
->open
.filename
);
3892 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3894 req_set_fail_links(req
);
3895 io_req_complete(req
, ret
);
3899 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
3901 return io_openat2(req
, issue_flags
);
3904 static int io_remove_buffers_prep(struct io_kiocb
*req
,
3905 const struct io_uring_sqe
*sqe
)
3907 struct io_provide_buf
*p
= &req
->pbuf
;
3910 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
3913 tmp
= READ_ONCE(sqe
->fd
);
3914 if (!tmp
|| tmp
> USHRT_MAX
)
3917 memset(p
, 0, sizeof(*p
));
3919 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3923 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
3924 int bgid
, unsigned nbufs
)
3928 /* shouldn't happen */
3932 /* the head kbuf is the list itself */
3933 while (!list_empty(&buf
->list
)) {
3934 struct io_buffer
*nxt
;
3936 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
3937 list_del(&nxt
->list
);
3944 xa_erase(&ctx
->io_buffers
, bgid
);
3949 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
3951 struct io_provide_buf
*p
= &req
->pbuf
;
3952 struct io_ring_ctx
*ctx
= req
->ctx
;
3953 struct io_buffer
*head
;
3955 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3957 io_ring_submit_lock(ctx
, !force_nonblock
);
3959 lockdep_assert_held(&ctx
->uring_lock
);
3962 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
3964 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
3966 req_set_fail_links(req
);
3968 /* need to hold the lock to complete IOPOLL requests */
3969 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
3970 __io_req_complete(req
, issue_flags
, ret
, 0);
3971 io_ring_submit_unlock(ctx
, !force_nonblock
);
3973 io_ring_submit_unlock(ctx
, !force_nonblock
);
3974 __io_req_complete(req
, issue_flags
, ret
, 0);
3979 static int io_provide_buffers_prep(struct io_kiocb
*req
,
3980 const struct io_uring_sqe
*sqe
)
3983 struct io_provide_buf
*p
= &req
->pbuf
;
3986 if (sqe
->ioprio
|| sqe
->rw_flags
)
3989 tmp
= READ_ONCE(sqe
->fd
);
3990 if (!tmp
|| tmp
> USHRT_MAX
)
3993 p
->addr
= READ_ONCE(sqe
->addr
);
3994 p
->len
= READ_ONCE(sqe
->len
);
3996 size
= (unsigned long)p
->len
* p
->nbufs
;
3997 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4000 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4001 tmp
= READ_ONCE(sqe
->off
);
4002 if (tmp
> USHRT_MAX
)
4008 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4010 struct io_buffer
*buf
;
4011 u64 addr
= pbuf
->addr
;
4012 int i
, bid
= pbuf
->bid
;
4014 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4015 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
4020 buf
->len
= pbuf
->len
;
4025 INIT_LIST_HEAD(&buf
->list
);
4028 list_add_tail(&buf
->list
, &(*head
)->list
);
4032 return i
? i
: -ENOMEM
;
4035 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4037 struct io_provide_buf
*p
= &req
->pbuf
;
4038 struct io_ring_ctx
*ctx
= req
->ctx
;
4039 struct io_buffer
*head
, *list
;
4041 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4043 io_ring_submit_lock(ctx
, !force_nonblock
);
4045 lockdep_assert_held(&ctx
->uring_lock
);
4047 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4049 ret
= io_add_buffers(p
, &head
);
4050 if (ret
>= 0 && !list
) {
4051 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4053 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4056 req_set_fail_links(req
);
4058 /* need to hold the lock to complete IOPOLL requests */
4059 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
4060 __io_req_complete(req
, issue_flags
, ret
, 0);
4061 io_ring_submit_unlock(ctx
, !force_nonblock
);
4063 io_ring_submit_unlock(ctx
, !force_nonblock
);
4064 __io_req_complete(req
, issue_flags
, ret
, 0);
4069 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4070 const struct io_uring_sqe
*sqe
)
4072 #if defined(CONFIG_EPOLL)
4073 if (sqe
->ioprio
|| sqe
->buf_index
)
4075 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4078 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4079 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4080 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4082 if (ep_op_has_event(req
->epoll
.op
)) {
4083 struct epoll_event __user
*ev
;
4085 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4086 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4096 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4098 #if defined(CONFIG_EPOLL)
4099 struct io_epoll
*ie
= &req
->epoll
;
4101 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4103 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4104 if (force_nonblock
&& ret
== -EAGAIN
)
4108 req_set_fail_links(req
);
4109 __io_req_complete(req
, issue_flags
, ret
, 0);
4116 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4118 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4119 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4121 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4124 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4125 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4126 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4133 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4135 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4136 struct io_madvise
*ma
= &req
->madvise
;
4139 if (issue_flags
& IO_URING_F_NONBLOCK
)
4142 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4144 req_set_fail_links(req
);
4145 io_req_complete(req
, ret
);
4152 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4154 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4156 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4159 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4160 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4161 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4165 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4167 struct io_fadvise
*fa
= &req
->fadvise
;
4170 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4171 switch (fa
->advice
) {
4172 case POSIX_FADV_NORMAL
:
4173 case POSIX_FADV_RANDOM
:
4174 case POSIX_FADV_SEQUENTIAL
:
4181 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4183 req_set_fail_links(req
);
4184 io_req_complete(req
, ret
);
4188 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4190 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
)))
4192 if (sqe
->ioprio
|| sqe
->buf_index
)
4194 if (req
->flags
& REQ_F_FIXED_FILE
)
4197 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4198 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4199 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4200 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4201 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4206 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4208 struct io_statx
*ctx
= &req
->statx
;
4211 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4212 /* only need file table for an actual valid fd */
4213 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
4214 req
->flags
|= REQ_F_NO_FILE_TABLE
;
4218 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4222 req_set_fail_links(req
);
4223 io_req_complete(req
, ret
);
4227 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4229 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4231 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4232 sqe
->rw_flags
|| sqe
->buf_index
)
4234 if (req
->flags
& REQ_F_FIXED_FILE
)
4237 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4241 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4243 struct files_struct
*files
= current
->files
;
4244 struct io_close
*close
= &req
->close
;
4245 struct fdtable
*fdt
;
4251 spin_lock(&files
->file_lock
);
4252 fdt
= files_fdtable(files
);
4253 if (close
->fd
>= fdt
->max_fds
) {
4254 spin_unlock(&files
->file_lock
);
4257 file
= fdt
->fd
[close
->fd
];
4259 spin_unlock(&files
->file_lock
);
4263 if (file
->f_op
== &io_uring_fops
) {
4264 spin_unlock(&files
->file_lock
);
4269 /* if the file has a flush method, be safe and punt to async */
4270 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4271 spin_unlock(&files
->file_lock
);
4275 ret
= __close_fd_get_file(close
->fd
, &file
);
4276 spin_unlock(&files
->file_lock
);
4283 /* No ->flush() or already async, safely close from here */
4284 ret
= filp_close(file
, current
->files
);
4287 req_set_fail_links(req
);
4290 __io_req_complete(req
, issue_flags
, ret
, 0);
4294 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4296 struct io_ring_ctx
*ctx
= req
->ctx
;
4298 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4300 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4303 req
->sync
.off
= READ_ONCE(sqe
->off
);
4304 req
->sync
.len
= READ_ONCE(sqe
->len
);
4305 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4309 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4313 /* sync_file_range always requires a blocking context */
4314 if (issue_flags
& IO_URING_F_NONBLOCK
)
4317 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4320 req_set_fail_links(req
);
4321 io_req_complete(req
, ret
);
4325 #if defined(CONFIG_NET)
4326 static int io_setup_async_msg(struct io_kiocb
*req
,
4327 struct io_async_msghdr
*kmsg
)
4329 struct io_async_msghdr
*async_msg
= req
->async_data
;
4333 if (io_alloc_async_data(req
)) {
4334 kfree(kmsg
->free_iov
);
4337 async_msg
= req
->async_data
;
4338 req
->flags
|= REQ_F_NEED_CLEANUP
;
4339 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4340 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4341 /* if were using fast_iov, set it to the new one */
4342 if (!async_msg
->free_iov
)
4343 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4348 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4349 struct io_async_msghdr
*iomsg
)
4351 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4352 iomsg
->free_iov
= iomsg
->fast_iov
;
4353 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4354 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4357 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4361 if (!io_op_defs
[req
->opcode
].needs_async_data
)
4363 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4365 req
->flags
|= REQ_F_NEED_CLEANUP
;
4369 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4371 struct io_sr_msg
*sr
= &req
->sr_msg
;
4373 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4376 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4377 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4378 sr
->len
= READ_ONCE(sqe
->len
);
4380 #ifdef CONFIG_COMPAT
4381 if (req
->ctx
->compat
)
4382 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4387 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4389 struct io_async_msghdr iomsg
, *kmsg
;
4390 struct socket
*sock
;
4395 sock
= sock_from_file(req
->file
);
4396 if (unlikely(!sock
))
4399 kmsg
= req
->async_data
;
4401 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4407 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4408 if (flags
& MSG_DONTWAIT
)
4409 req
->flags
|= REQ_F_NOWAIT
;
4410 else if (issue_flags
& IO_URING_F_NONBLOCK
)
4411 flags
|= MSG_DONTWAIT
;
4413 if (flags
& MSG_WAITALL
)
4414 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4416 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4417 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4418 return io_setup_async_msg(req
, kmsg
);
4419 if (ret
== -ERESTARTSYS
)
4422 /* fast path, check for non-NULL to avoid function call */
4424 kfree(kmsg
->free_iov
);
4425 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4427 req_set_fail_links(req
);
4428 __io_req_complete(req
, issue_flags
, ret
, 0);
4432 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4434 struct io_sr_msg
*sr
= &req
->sr_msg
;
4437 struct socket
*sock
;
4442 sock
= sock_from_file(req
->file
);
4443 if (unlikely(!sock
))
4446 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4450 msg
.msg_name
= NULL
;
4451 msg
.msg_control
= NULL
;
4452 msg
.msg_controllen
= 0;
4453 msg
.msg_namelen
= 0;
4455 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4456 if (flags
& MSG_DONTWAIT
)
4457 req
->flags
|= REQ_F_NOWAIT
;
4458 else if (issue_flags
& IO_URING_F_NONBLOCK
)
4459 flags
|= MSG_DONTWAIT
;
4461 if (flags
& MSG_WAITALL
)
4462 min_ret
= iov_iter_count(&msg
.msg_iter
);
4464 msg
.msg_flags
= flags
;
4465 ret
= sock_sendmsg(sock
, &msg
);
4466 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4468 if (ret
== -ERESTARTSYS
)
4472 req_set_fail_links(req
);
4473 __io_req_complete(req
, issue_flags
, ret
, 0);
4477 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4478 struct io_async_msghdr
*iomsg
)
4480 struct io_sr_msg
*sr
= &req
->sr_msg
;
4481 struct iovec __user
*uiov
;
4485 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4486 &iomsg
->uaddr
, &uiov
, &iov_len
);
4490 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4493 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4495 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4496 iomsg
->free_iov
= NULL
;
4498 iomsg
->free_iov
= iomsg
->fast_iov
;
4499 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4500 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4509 #ifdef CONFIG_COMPAT
4510 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4511 struct io_async_msghdr
*iomsg
)
4513 struct compat_msghdr __user
*msg_compat
;
4514 struct io_sr_msg
*sr
= &req
->sr_msg
;
4515 struct compat_iovec __user
*uiov
;
4520 msg_compat
= (struct compat_msghdr __user
*) sr
->umsg
;
4521 ret
= __get_compat_msghdr(&iomsg
->msg
, msg_compat
, &iomsg
->uaddr
,
4526 uiov
= compat_ptr(ptr
);
4527 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4528 compat_ssize_t clen
;
4532 if (!access_ok(uiov
, sizeof(*uiov
)))
4534 if (__get_user(clen
, &uiov
->iov_len
))
4539 iomsg
->free_iov
= NULL
;
4541 iomsg
->free_iov
= iomsg
->fast_iov
;
4542 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4543 UIO_FASTIOV
, &iomsg
->free_iov
,
4544 &iomsg
->msg
.msg_iter
, true);
4553 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4554 struct io_async_msghdr
*iomsg
)
4556 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4558 #ifdef CONFIG_COMPAT
4559 if (req
->ctx
->compat
)
4560 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4563 return __io_recvmsg_copy_hdr(req
, iomsg
);
4566 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4569 struct io_sr_msg
*sr
= &req
->sr_msg
;
4570 struct io_buffer
*kbuf
;
4572 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4577 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4581 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4583 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4586 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
4590 if (!io_op_defs
[req
->opcode
].needs_async_data
)
4592 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
4594 req
->flags
|= REQ_F_NEED_CLEANUP
;
4598 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4600 struct io_sr_msg
*sr
= &req
->sr_msg
;
4602 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4605 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4606 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4607 sr
->len
= READ_ONCE(sqe
->len
);
4608 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4610 #ifdef CONFIG_COMPAT
4611 if (req
->ctx
->compat
)
4612 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4617 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4619 struct io_async_msghdr iomsg
, *kmsg
;
4620 struct socket
*sock
;
4621 struct io_buffer
*kbuf
;
4624 int ret
, cflags
= 0;
4625 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4627 sock
= sock_from_file(req
->file
);
4628 if (unlikely(!sock
))
4631 kmsg
= req
->async_data
;
4633 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4639 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4640 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4642 return PTR_ERR(kbuf
);
4643 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4644 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
4645 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
4646 1, req
->sr_msg
.len
);
4649 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4650 if (flags
& MSG_DONTWAIT
)
4651 req
->flags
|= REQ_F_NOWAIT
;
4652 else if (force_nonblock
)
4653 flags
|= MSG_DONTWAIT
;
4655 if (flags
& MSG_WAITALL
)
4656 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4658 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4659 kmsg
->uaddr
, flags
);
4660 if (force_nonblock
&& ret
== -EAGAIN
)
4661 return io_setup_async_msg(req
, kmsg
);
4662 if (ret
== -ERESTARTSYS
)
4665 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4666 cflags
= io_put_recv_kbuf(req
);
4667 /* fast path, check for non-NULL to avoid function call */
4669 kfree(kmsg
->free_iov
);
4670 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4671 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4672 req_set_fail_links(req
);
4673 __io_req_complete(req
, issue_flags
, ret
, cflags
);
4677 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
4679 struct io_buffer
*kbuf
;
4680 struct io_sr_msg
*sr
= &req
->sr_msg
;
4682 void __user
*buf
= sr
->buf
;
4683 struct socket
*sock
;
4687 int ret
, cflags
= 0;
4688 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4690 sock
= sock_from_file(req
->file
);
4691 if (unlikely(!sock
))
4694 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4695 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4697 return PTR_ERR(kbuf
);
4698 buf
= u64_to_user_ptr(kbuf
->addr
);
4701 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4705 msg
.msg_name
= NULL
;
4706 msg
.msg_control
= NULL
;
4707 msg
.msg_controllen
= 0;
4708 msg
.msg_namelen
= 0;
4709 msg
.msg_iocb
= NULL
;
4712 flags
= req
->sr_msg
.msg_flags
| MSG_NOSIGNAL
;
4713 if (flags
& MSG_DONTWAIT
)
4714 req
->flags
|= REQ_F_NOWAIT
;
4715 else if (force_nonblock
)
4716 flags
|= MSG_DONTWAIT
;
4718 if (flags
& MSG_WAITALL
)
4719 min_ret
= iov_iter_count(&msg
.msg_iter
);
4721 ret
= sock_recvmsg(sock
, &msg
, flags
);
4722 if (force_nonblock
&& ret
== -EAGAIN
)
4724 if (ret
== -ERESTARTSYS
)
4727 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4728 cflags
= io_put_recv_kbuf(req
);
4729 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4730 req_set_fail_links(req
);
4731 __io_req_complete(req
, issue_flags
, ret
, cflags
);
4735 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4737 struct io_accept
*accept
= &req
->accept
;
4739 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4741 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4744 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4745 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4746 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4747 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4751 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
4753 struct io_accept
*accept
= &req
->accept
;
4754 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4755 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4758 if (req
->file
->f_flags
& O_NONBLOCK
)
4759 req
->flags
|= REQ_F_NOWAIT
;
4761 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4762 accept
->addr_len
, accept
->flags
,
4764 if (ret
== -EAGAIN
&& force_nonblock
)
4767 if (ret
== -ERESTARTSYS
)
4769 req_set_fail_links(req
);
4771 __io_req_complete(req
, issue_flags
, ret
, 0);
4775 static int io_connect_prep_async(struct io_kiocb
*req
)
4777 struct io_async_connect
*io
= req
->async_data
;
4778 struct io_connect
*conn
= &req
->connect
;
4780 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
4783 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4785 struct io_connect
*conn
= &req
->connect
;
4787 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4789 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4792 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4793 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4797 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
4799 struct io_async_connect __io
, *io
;
4800 unsigned file_flags
;
4802 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4804 if (req
->async_data
) {
4805 io
= req
->async_data
;
4807 ret
= move_addr_to_kernel(req
->connect
.addr
,
4808 req
->connect
.addr_len
,
4815 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4817 ret
= __sys_connect_file(req
->file
, &io
->address
,
4818 req
->connect
.addr_len
, file_flags
);
4819 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4820 if (req
->async_data
)
4822 if (io_alloc_async_data(req
)) {
4826 memcpy(req
->async_data
, &__io
, sizeof(__io
));
4829 if (ret
== -ERESTARTSYS
)
4833 req_set_fail_links(req
);
4834 __io_req_complete(req
, issue_flags
, ret
, 0);
4837 #else /* !CONFIG_NET */
4838 #define IO_NETOP_FN(op) \
4839 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4841 return -EOPNOTSUPP; \
4844 #define IO_NETOP_PREP(op) \
4846 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4848 return -EOPNOTSUPP; \
4851 #define IO_NETOP_PREP_ASYNC(op) \
4853 static int io_##op##_prep_async(struct io_kiocb *req) \
4855 return -EOPNOTSUPP; \
4858 IO_NETOP_PREP_ASYNC(sendmsg
);
4859 IO_NETOP_PREP_ASYNC(recvmsg
);
4860 IO_NETOP_PREP_ASYNC(connect
);
4861 IO_NETOP_PREP(accept
);
4864 #endif /* CONFIG_NET */
4866 struct io_poll_table
{
4867 struct poll_table_struct pt
;
4868 struct io_kiocb
*req
;
4872 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
4873 __poll_t mask
, task_work_func_t func
)
4877 /* for instances that support it check for an event match first: */
4878 if (mask
&& !(mask
& poll
->events
))
4881 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
4883 list_del_init(&poll
->wait
.entry
);
4886 req
->task_work
.func
= func
;
4887 percpu_ref_get(&req
->ctx
->refs
);
4890 * If this fails, then the task is exiting. When a task exits, the
4891 * work gets canceled, so just cancel this request as well instead
4892 * of executing it. We can't safely execute it anyway, as we may not
4893 * have the needed state needed for it anyway.
4895 ret
= io_req_task_work_add(req
);
4896 if (unlikely(ret
)) {
4897 WRITE_ONCE(poll
->canceled
, true);
4898 io_req_task_work_add_fallback(req
, func
);
4903 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
4904 __acquires(&req
->ctx
->completion_lock
)
4906 struct io_ring_ctx
*ctx
= req
->ctx
;
4908 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4909 struct poll_table_struct pt
= { ._key
= poll
->events
};
4911 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
4914 spin_lock_irq(&ctx
->completion_lock
);
4915 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4916 add_wait_queue(poll
->head
, &poll
->wait
);
4923 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
4925 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4926 if (req
->opcode
== IORING_OP_POLL_ADD
)
4927 return req
->async_data
;
4928 return req
->apoll
->double_poll
;
4931 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
4933 if (req
->opcode
== IORING_OP_POLL_ADD
)
4935 return &req
->apoll
->poll
;
4938 static void io_poll_remove_double(struct io_kiocb
*req
)
4940 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
4942 lockdep_assert_held(&req
->ctx
->completion_lock
);
4944 if (poll
&& poll
->head
) {
4945 struct wait_queue_head
*head
= poll
->head
;
4947 spin_lock(&head
->lock
);
4948 list_del_init(&poll
->wait
.entry
);
4949 if (poll
->wait
.private)
4950 refcount_dec(&req
->refs
);
4952 spin_unlock(&head
->lock
);
4956 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
4958 struct io_ring_ctx
*ctx
= req
->ctx
;
4960 io_poll_remove_double(req
);
4961 req
->poll
.done
= true;
4962 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
4963 io_commit_cqring(ctx
);
4966 static void io_poll_task_func(struct callback_head
*cb
)
4968 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4969 struct io_ring_ctx
*ctx
= req
->ctx
;
4970 struct io_kiocb
*nxt
;
4972 if (io_poll_rewait(req
, &req
->poll
)) {
4973 spin_unlock_irq(&ctx
->completion_lock
);
4975 hash_del(&req
->hash_node
);
4976 io_poll_complete(req
, req
->result
, 0);
4977 spin_unlock_irq(&ctx
->completion_lock
);
4979 nxt
= io_put_req_find_next(req
);
4980 io_cqring_ev_posted(ctx
);
4982 __io_req_task_submit(nxt
);
4985 percpu_ref_put(&ctx
->refs
);
4988 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
4989 int sync
, void *key
)
4991 struct io_kiocb
*req
= wait
->private;
4992 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
4993 __poll_t mask
= key_to_poll(key
);
4995 /* for instances that support it check for an event match first: */
4996 if (mask
&& !(mask
& poll
->events
))
4999 list_del_init(&wait
->entry
);
5001 if (poll
&& poll
->head
) {
5004 spin_lock(&poll
->head
->lock
);
5005 done
= list_empty(&poll
->wait
.entry
);
5007 list_del_init(&poll
->wait
.entry
);
5008 /* make sure double remove sees this as being gone */
5009 wait
->private = NULL
;
5010 spin_unlock(&poll
->head
->lock
);
5012 /* use wait func handler, so it matches the rq type */
5013 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5016 refcount_dec(&req
->refs
);
5020 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5021 wait_queue_func_t wake_func
)
5025 poll
->canceled
= false;
5026 poll
->events
= events
;
5027 INIT_LIST_HEAD(&poll
->wait
.entry
);
5028 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5031 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5032 struct wait_queue_head
*head
,
5033 struct io_poll_iocb
**poll_ptr
)
5035 struct io_kiocb
*req
= pt
->req
;
5038 * If poll->head is already set, it's because the file being polled
5039 * uses multiple waitqueues for poll handling (eg one for read, one
5040 * for write). Setup a separate io_poll_iocb if this happens.
5042 if (unlikely(poll
->head
)) {
5043 struct io_poll_iocb
*poll_one
= poll
;
5045 /* already have a 2nd entry, fail a third attempt */
5047 pt
->error
= -EINVAL
;
5050 /* double add on the same waitqueue head, ignore */
5051 if (poll
->head
== head
)
5053 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5055 pt
->error
= -ENOMEM
;
5058 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5059 refcount_inc(&req
->refs
);
5060 poll
->wait
.private = req
;
5067 if (poll
->events
& EPOLLEXCLUSIVE
)
5068 add_wait_queue_exclusive(head
, &poll
->wait
);
5070 add_wait_queue(head
, &poll
->wait
);
5073 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5074 struct poll_table_struct
*p
)
5076 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5077 struct async_poll
*apoll
= pt
->req
->apoll
;
5079 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5082 static void io_async_task_func(struct callback_head
*cb
)
5084 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
5085 struct async_poll
*apoll
= req
->apoll
;
5086 struct io_ring_ctx
*ctx
= req
->ctx
;
5088 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
5090 if (io_poll_rewait(req
, &apoll
->poll
)) {
5091 spin_unlock_irq(&ctx
->completion_lock
);
5092 percpu_ref_put(&ctx
->refs
);
5096 /* If req is still hashed, it cannot have been canceled. Don't check. */
5097 if (hash_hashed(&req
->hash_node
))
5098 hash_del(&req
->hash_node
);
5100 io_poll_remove_double(req
);
5101 spin_unlock_irq(&ctx
->completion_lock
);
5103 if (!READ_ONCE(apoll
->poll
.canceled
))
5104 __io_req_task_submit(req
);
5106 __io_req_task_cancel(req
, -ECANCELED
);
5108 percpu_ref_put(&ctx
->refs
);
5109 kfree(apoll
->double_poll
);
5113 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5116 struct io_kiocb
*req
= wait
->private;
5117 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5119 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5122 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5125 static void io_poll_req_insert(struct io_kiocb
*req
)
5127 struct io_ring_ctx
*ctx
= req
->ctx
;
5128 struct hlist_head
*list
;
5130 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5131 hlist_add_head(&req
->hash_node
, list
);
5134 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5135 struct io_poll_iocb
*poll
,
5136 struct io_poll_table
*ipt
, __poll_t mask
,
5137 wait_queue_func_t wake_func
)
5138 __acquires(&ctx
->completion_lock
)
5140 struct io_ring_ctx
*ctx
= req
->ctx
;
5141 bool cancel
= false;
5143 INIT_HLIST_NODE(&req
->hash_node
);
5144 io_init_poll_iocb(poll
, mask
, wake_func
);
5145 poll
->file
= req
->file
;
5146 poll
->wait
.private = req
;
5148 ipt
->pt
._key
= mask
;
5150 ipt
->error
= -EINVAL
;
5152 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5154 spin_lock_irq(&ctx
->completion_lock
);
5155 if (likely(poll
->head
)) {
5156 spin_lock(&poll
->head
->lock
);
5157 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5163 if (mask
|| ipt
->error
)
5164 list_del_init(&poll
->wait
.entry
);
5166 WRITE_ONCE(poll
->canceled
, true);
5167 else if (!poll
->done
) /* actually waiting for an event */
5168 io_poll_req_insert(req
);
5169 spin_unlock(&poll
->head
->lock
);
5175 static bool io_arm_poll_handler(struct io_kiocb
*req
)
5177 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5178 struct io_ring_ctx
*ctx
= req
->ctx
;
5179 struct async_poll
*apoll
;
5180 struct io_poll_table ipt
;
5184 if (!req
->file
|| !file_can_poll(req
->file
))
5186 if (req
->flags
& REQ_F_POLLED
)
5190 else if (def
->pollout
)
5194 /* if we can't nonblock try, then no point in arming a poll handler */
5195 if (!io_file_supports_async(req
->file
, rw
))
5198 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5199 if (unlikely(!apoll
))
5201 apoll
->double_poll
= NULL
;
5203 req
->flags
|= REQ_F_POLLED
;
5208 mask
|= POLLIN
| POLLRDNORM
;
5210 mask
|= POLLOUT
| POLLWRNORM
;
5212 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5213 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5214 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5217 mask
|= POLLERR
| POLLPRI
;
5219 ipt
.pt
._qproc
= io_async_queue_proc
;
5221 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5223 if (ret
|| ipt
.error
) {
5224 io_poll_remove_double(req
);
5225 spin_unlock_irq(&ctx
->completion_lock
);
5226 kfree(apoll
->double_poll
);
5230 spin_unlock_irq(&ctx
->completion_lock
);
5231 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
5232 apoll
->poll
.events
);
5236 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5237 struct io_poll_iocb
*poll
)
5239 bool do_complete
= false;
5241 spin_lock(&poll
->head
->lock
);
5242 WRITE_ONCE(poll
->canceled
, true);
5243 if (!list_empty(&poll
->wait
.entry
)) {
5244 list_del_init(&poll
->wait
.entry
);
5247 spin_unlock(&poll
->head
->lock
);
5248 hash_del(&req
->hash_node
);
5252 static bool io_poll_remove_one(struct io_kiocb
*req
)
5256 io_poll_remove_double(req
);
5258 if (req
->opcode
== IORING_OP_POLL_ADD
) {
5259 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
5261 struct async_poll
*apoll
= req
->apoll
;
5263 /* non-poll requests have submit ref still */
5264 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
5267 kfree(apoll
->double_poll
);
5273 io_cqring_fill_event(req
, -ECANCELED
);
5274 io_commit_cqring(req
->ctx
);
5275 req_set_fail_links(req
);
5276 io_put_req_deferred(req
, 1);
5283 * Returns true if we found and killed one or more poll requests
5285 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5286 struct files_struct
*files
)
5288 struct hlist_node
*tmp
;
5289 struct io_kiocb
*req
;
5292 spin_lock_irq(&ctx
->completion_lock
);
5293 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5294 struct hlist_head
*list
;
5296 list
= &ctx
->cancel_hash
[i
];
5297 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5298 if (io_match_task(req
, tsk
, files
))
5299 posted
+= io_poll_remove_one(req
);
5302 spin_unlock_irq(&ctx
->completion_lock
);
5305 io_cqring_ev_posted(ctx
);
5310 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
5312 struct hlist_head
*list
;
5313 struct io_kiocb
*req
;
5315 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5316 hlist_for_each_entry(req
, list
, hash_node
) {
5317 if (sqe_addr
!= req
->user_data
)
5319 if (io_poll_remove_one(req
))
5327 static int io_poll_remove_prep(struct io_kiocb
*req
,
5328 const struct io_uring_sqe
*sqe
)
5330 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5332 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
5336 req
->poll_remove
.addr
= READ_ONCE(sqe
->addr
);
5341 * Find a running poll command that matches one specified in sqe->addr,
5342 * and remove it if found.
5344 static int io_poll_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
5346 struct io_ring_ctx
*ctx
= req
->ctx
;
5349 spin_lock_irq(&ctx
->completion_lock
);
5350 ret
= io_poll_cancel(ctx
, req
->poll_remove
.addr
);
5351 spin_unlock_irq(&ctx
->completion_lock
);
5354 req_set_fail_links(req
);
5355 io_req_complete(req
, ret
);
5359 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5362 struct io_kiocb
*req
= wait
->private;
5363 struct io_poll_iocb
*poll
= &req
->poll
;
5365 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5368 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5369 struct poll_table_struct
*p
)
5371 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5373 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5376 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5378 struct io_poll_iocb
*poll
= &req
->poll
;
5381 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5383 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
5386 events
= READ_ONCE(sqe
->poll32_events
);
5388 events
= swahw32(events
);
5390 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
5391 (events
& EPOLLEXCLUSIVE
);
5395 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5397 struct io_poll_iocb
*poll
= &req
->poll
;
5398 struct io_ring_ctx
*ctx
= req
->ctx
;
5399 struct io_poll_table ipt
;
5402 ipt
.pt
._qproc
= io_poll_queue_proc
;
5404 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5407 if (mask
) { /* no async, we'd stolen it */
5409 io_poll_complete(req
, mask
, 0);
5411 spin_unlock_irq(&ctx
->completion_lock
);
5414 io_cqring_ev_posted(ctx
);
5420 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5422 struct io_timeout_data
*data
= container_of(timer
,
5423 struct io_timeout_data
, timer
);
5424 struct io_kiocb
*req
= data
->req
;
5425 struct io_ring_ctx
*ctx
= req
->ctx
;
5426 unsigned long flags
;
5428 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5429 list_del_init(&req
->timeout
.list
);
5430 atomic_set(&req
->ctx
->cq_timeouts
,
5431 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5433 io_cqring_fill_event(req
, -ETIME
);
5434 io_commit_cqring(ctx
);
5435 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5437 io_cqring_ev_posted(ctx
);
5438 req_set_fail_links(req
);
5440 return HRTIMER_NORESTART
;
5443 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5446 struct io_timeout_data
*io
;
5447 struct io_kiocb
*req
;
5450 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5451 if (user_data
== req
->user_data
) {
5458 return ERR_PTR(ret
);
5460 io
= req
->async_data
;
5461 ret
= hrtimer_try_to_cancel(&io
->timer
);
5463 return ERR_PTR(-EALREADY
);
5464 list_del_init(&req
->timeout
.list
);
5468 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5470 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5473 return PTR_ERR(req
);
5475 req_set_fail_links(req
);
5476 io_cqring_fill_event(req
, -ECANCELED
);
5477 io_put_req_deferred(req
, 1);
5481 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5482 struct timespec64
*ts
, enum hrtimer_mode mode
)
5484 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5485 struct io_timeout_data
*data
;
5488 return PTR_ERR(req
);
5490 req
->timeout
.off
= 0; /* noseq */
5491 data
= req
->async_data
;
5492 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5493 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5494 data
->timer
.function
= io_timeout_fn
;
5495 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5499 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5500 const struct io_uring_sqe
*sqe
)
5502 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5504 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5506 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5508 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
5511 tr
->addr
= READ_ONCE(sqe
->addr
);
5512 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5513 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5514 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5516 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5518 } else if (tr
->flags
) {
5519 /* timeout removal doesn't support flags */
5526 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
5528 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
5533 * Remove or update an existing timeout command
5535 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
5537 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5538 struct io_ring_ctx
*ctx
= req
->ctx
;
5541 spin_lock_irq(&ctx
->completion_lock
);
5542 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
))
5543 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5545 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
,
5546 io_translate_timeout_mode(tr
->flags
));
5548 io_cqring_fill_event(req
, ret
);
5549 io_commit_cqring(ctx
);
5550 spin_unlock_irq(&ctx
->completion_lock
);
5551 io_cqring_ev_posted(ctx
);
5553 req_set_fail_links(req
);
5558 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5559 bool is_timeout_link
)
5561 struct io_timeout_data
*data
;
5563 u32 off
= READ_ONCE(sqe
->off
);
5565 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5567 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5569 if (off
&& is_timeout_link
)
5571 flags
= READ_ONCE(sqe
->timeout_flags
);
5572 if (flags
& ~IORING_TIMEOUT_ABS
)
5575 req
->timeout
.off
= off
;
5577 if (!req
->async_data
&& io_alloc_async_data(req
))
5580 data
= req
->async_data
;
5583 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5586 data
->mode
= io_translate_timeout_mode(flags
);
5587 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5588 if (is_timeout_link
)
5589 io_req_track_inflight(req
);
5593 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
5595 struct io_ring_ctx
*ctx
= req
->ctx
;
5596 struct io_timeout_data
*data
= req
->async_data
;
5597 struct list_head
*entry
;
5598 u32 tail
, off
= req
->timeout
.off
;
5600 spin_lock_irq(&ctx
->completion_lock
);
5603 * sqe->off holds how many events that need to occur for this
5604 * timeout event to be satisfied. If it isn't set, then this is
5605 * a pure timeout request, sequence isn't used.
5607 if (io_is_timeout_noseq(req
)) {
5608 entry
= ctx
->timeout_list
.prev
;
5612 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5613 req
->timeout
.target_seq
= tail
+ off
;
5615 /* Update the last seq here in case io_flush_timeouts() hasn't.
5616 * This is safe because ->completion_lock is held, and submissions
5617 * and completions are never mixed in the same ->completion_lock section.
5619 ctx
->cq_last_tm_flush
= tail
;
5622 * Insertion sort, ensuring the first entry in the list is always
5623 * the one we need first.
5625 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5626 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5629 if (io_is_timeout_noseq(nxt
))
5631 /* nxt.seq is behind @tail, otherwise would've been completed */
5632 if (off
>= nxt
->timeout
.target_seq
- tail
)
5636 list_add(&req
->timeout
.list
, entry
);
5637 data
->timer
.function
= io_timeout_fn
;
5638 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5639 spin_unlock_irq(&ctx
->completion_lock
);
5643 struct io_cancel_data
{
5644 struct io_ring_ctx
*ctx
;
5648 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5650 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5651 struct io_cancel_data
*cd
= data
;
5653 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
5656 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
5657 struct io_ring_ctx
*ctx
)
5659 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
5660 enum io_wq_cancel cancel_ret
;
5663 if (!tctx
|| !tctx
->io_wq
)
5666 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
5667 switch (cancel_ret
) {
5668 case IO_WQ_CANCEL_OK
:
5671 case IO_WQ_CANCEL_RUNNING
:
5674 case IO_WQ_CANCEL_NOTFOUND
:
5682 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5683 struct io_kiocb
*req
, __u64 sqe_addr
,
5686 unsigned long flags
;
5689 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
5690 if (ret
!= -ENOENT
) {
5691 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5695 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5696 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5699 ret
= io_poll_cancel(ctx
, sqe_addr
);
5703 io_cqring_fill_event(req
, ret
);
5704 io_commit_cqring(ctx
);
5705 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5706 io_cqring_ev_posted(ctx
);
5709 req_set_fail_links(req
);
5713 static int io_async_cancel_prep(struct io_kiocb
*req
,
5714 const struct io_uring_sqe
*sqe
)
5716 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5718 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5720 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5723 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5727 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
5729 struct io_ring_ctx
*ctx
= req
->ctx
;
5730 u64 sqe_addr
= req
->cancel
.addr
;
5731 struct io_tctx_node
*node
;
5734 /* tasks should wait for their io-wq threads, so safe w/o sync */
5735 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
5736 spin_lock_irq(&ctx
->completion_lock
);
5739 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5742 ret
= io_poll_cancel(ctx
, sqe_addr
);
5745 spin_unlock_irq(&ctx
->completion_lock
);
5747 /* slow path, try all io-wq's */
5748 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
5750 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
5751 struct io_uring_task
*tctx
= node
->task
->io_uring
;
5753 if (!tctx
|| !tctx
->io_wq
)
5755 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
5759 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
5761 spin_lock_irq(&ctx
->completion_lock
);
5763 io_cqring_fill_event(req
, ret
);
5764 io_commit_cqring(ctx
);
5765 spin_unlock_irq(&ctx
->completion_lock
);
5766 io_cqring_ev_posted(ctx
);
5769 req_set_fail_links(req
);
5774 static int io_rsrc_update_prep(struct io_kiocb
*req
,
5775 const struct io_uring_sqe
*sqe
)
5777 if (unlikely(req
->ctx
->flags
& IORING_SETUP_SQPOLL
))
5779 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5781 if (sqe
->ioprio
|| sqe
->rw_flags
)
5784 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
5785 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
5786 if (!req
->rsrc_update
.nr_args
)
5788 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
5792 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5794 struct io_ring_ctx
*ctx
= req
->ctx
;
5795 struct io_uring_rsrc_update up
;
5798 if (issue_flags
& IO_URING_F_NONBLOCK
)
5801 up
.offset
= req
->rsrc_update
.offset
;
5802 up
.data
= req
->rsrc_update
.arg
;
5804 mutex_lock(&ctx
->uring_lock
);
5805 ret
= __io_sqe_files_update(ctx
, &up
, req
->rsrc_update
.nr_args
);
5806 mutex_unlock(&ctx
->uring_lock
);
5809 req_set_fail_links(req
);
5810 __io_req_complete(req
, issue_flags
, ret
, 0);
5814 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5816 switch (req
->opcode
) {
5819 case IORING_OP_READV
:
5820 case IORING_OP_READ_FIXED
:
5821 case IORING_OP_READ
:
5822 return io_read_prep(req
, sqe
);
5823 case IORING_OP_WRITEV
:
5824 case IORING_OP_WRITE_FIXED
:
5825 case IORING_OP_WRITE
:
5826 return io_write_prep(req
, sqe
);
5827 case IORING_OP_POLL_ADD
:
5828 return io_poll_add_prep(req
, sqe
);
5829 case IORING_OP_POLL_REMOVE
:
5830 return io_poll_remove_prep(req
, sqe
);
5831 case IORING_OP_FSYNC
:
5832 return io_fsync_prep(req
, sqe
);
5833 case IORING_OP_SYNC_FILE_RANGE
:
5834 return io_sfr_prep(req
, sqe
);
5835 case IORING_OP_SENDMSG
:
5836 case IORING_OP_SEND
:
5837 return io_sendmsg_prep(req
, sqe
);
5838 case IORING_OP_RECVMSG
:
5839 case IORING_OP_RECV
:
5840 return io_recvmsg_prep(req
, sqe
);
5841 case IORING_OP_CONNECT
:
5842 return io_connect_prep(req
, sqe
);
5843 case IORING_OP_TIMEOUT
:
5844 return io_timeout_prep(req
, sqe
, false);
5845 case IORING_OP_TIMEOUT_REMOVE
:
5846 return io_timeout_remove_prep(req
, sqe
);
5847 case IORING_OP_ASYNC_CANCEL
:
5848 return io_async_cancel_prep(req
, sqe
);
5849 case IORING_OP_LINK_TIMEOUT
:
5850 return io_timeout_prep(req
, sqe
, true);
5851 case IORING_OP_ACCEPT
:
5852 return io_accept_prep(req
, sqe
);
5853 case IORING_OP_FALLOCATE
:
5854 return io_fallocate_prep(req
, sqe
);
5855 case IORING_OP_OPENAT
:
5856 return io_openat_prep(req
, sqe
);
5857 case IORING_OP_CLOSE
:
5858 return io_close_prep(req
, sqe
);
5859 case IORING_OP_FILES_UPDATE
:
5860 return io_rsrc_update_prep(req
, sqe
);
5861 case IORING_OP_STATX
:
5862 return io_statx_prep(req
, sqe
);
5863 case IORING_OP_FADVISE
:
5864 return io_fadvise_prep(req
, sqe
);
5865 case IORING_OP_MADVISE
:
5866 return io_madvise_prep(req
, sqe
);
5867 case IORING_OP_OPENAT2
:
5868 return io_openat2_prep(req
, sqe
);
5869 case IORING_OP_EPOLL_CTL
:
5870 return io_epoll_ctl_prep(req
, sqe
);
5871 case IORING_OP_SPLICE
:
5872 return io_splice_prep(req
, sqe
);
5873 case IORING_OP_PROVIDE_BUFFERS
:
5874 return io_provide_buffers_prep(req
, sqe
);
5875 case IORING_OP_REMOVE_BUFFERS
:
5876 return io_remove_buffers_prep(req
, sqe
);
5878 return io_tee_prep(req
, sqe
);
5879 case IORING_OP_SHUTDOWN
:
5880 return io_shutdown_prep(req
, sqe
);
5881 case IORING_OP_RENAMEAT
:
5882 return io_renameat_prep(req
, sqe
);
5883 case IORING_OP_UNLINKAT
:
5884 return io_unlinkat_prep(req
, sqe
);
5887 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5892 static int io_req_prep_async(struct io_kiocb
*req
)
5894 switch (req
->opcode
) {
5895 case IORING_OP_READV
:
5896 case IORING_OP_READ_FIXED
:
5897 case IORING_OP_READ
:
5898 return io_rw_prep_async(req
, READ
);
5899 case IORING_OP_WRITEV
:
5900 case IORING_OP_WRITE_FIXED
:
5901 case IORING_OP_WRITE
:
5902 return io_rw_prep_async(req
, WRITE
);
5903 case IORING_OP_SENDMSG
:
5904 case IORING_OP_SEND
:
5905 return io_sendmsg_prep_async(req
);
5906 case IORING_OP_RECVMSG
:
5907 case IORING_OP_RECV
:
5908 return io_recvmsg_prep_async(req
);
5909 case IORING_OP_CONNECT
:
5910 return io_connect_prep_async(req
);
5915 static int io_req_defer_prep(struct io_kiocb
*req
)
5917 if (!io_op_defs
[req
->opcode
].needs_async_data
)
5919 /* some opcodes init it during the inital prep */
5920 if (req
->async_data
)
5922 if (__io_alloc_async_data(req
))
5924 return io_req_prep_async(req
);
5927 static u32
io_get_sequence(struct io_kiocb
*req
)
5929 struct io_kiocb
*pos
;
5930 struct io_ring_ctx
*ctx
= req
->ctx
;
5931 u32 total_submitted
, nr_reqs
= 0;
5933 io_for_each_link(pos
, req
)
5936 total_submitted
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
5937 return total_submitted
- nr_reqs
;
5940 static int io_req_defer(struct io_kiocb
*req
)
5942 struct io_ring_ctx
*ctx
= req
->ctx
;
5943 struct io_defer_entry
*de
;
5947 /* Still need defer if there is pending req in defer list. */
5948 if (likely(list_empty_careful(&ctx
->defer_list
) &&
5949 !(req
->flags
& REQ_F_IO_DRAIN
)))
5952 seq
= io_get_sequence(req
);
5953 /* Still a chance to pass the sequence check */
5954 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
5957 ret
= io_req_defer_prep(req
);
5960 io_prep_async_link(req
);
5961 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
5965 spin_lock_irq(&ctx
->completion_lock
);
5966 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
5967 spin_unlock_irq(&ctx
->completion_lock
);
5969 io_queue_async_work(req
);
5970 return -EIOCBQUEUED
;
5973 trace_io_uring_defer(ctx
, req
, req
->user_data
);
5976 list_add_tail(&de
->list
, &ctx
->defer_list
);
5977 spin_unlock_irq(&ctx
->completion_lock
);
5978 return -EIOCBQUEUED
;
5981 static void __io_clean_op(struct io_kiocb
*req
)
5983 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
5984 switch (req
->opcode
) {
5985 case IORING_OP_READV
:
5986 case IORING_OP_READ_FIXED
:
5987 case IORING_OP_READ
:
5988 kfree((void *)(unsigned long)req
->rw
.addr
);
5990 case IORING_OP_RECVMSG
:
5991 case IORING_OP_RECV
:
5992 kfree(req
->sr_msg
.kbuf
);
5995 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
5998 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
5999 switch (req
->opcode
) {
6000 case IORING_OP_READV
:
6001 case IORING_OP_READ_FIXED
:
6002 case IORING_OP_READ
:
6003 case IORING_OP_WRITEV
:
6004 case IORING_OP_WRITE_FIXED
:
6005 case IORING_OP_WRITE
: {
6006 struct io_async_rw
*io
= req
->async_data
;
6008 kfree(io
->free_iovec
);
6011 case IORING_OP_RECVMSG
:
6012 case IORING_OP_SENDMSG
: {
6013 struct io_async_msghdr
*io
= req
->async_data
;
6015 kfree(io
->free_iov
);
6018 case IORING_OP_SPLICE
:
6020 io_put_file(req
, req
->splice
.file_in
,
6021 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
6023 case IORING_OP_OPENAT
:
6024 case IORING_OP_OPENAT2
:
6025 if (req
->open
.filename
)
6026 putname(req
->open
.filename
);
6028 case IORING_OP_RENAMEAT
:
6029 putname(req
->rename
.oldpath
);
6030 putname(req
->rename
.newpath
);
6032 case IORING_OP_UNLINKAT
:
6033 putname(req
->unlink
.filename
);
6036 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
6040 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6042 struct io_ring_ctx
*ctx
= req
->ctx
;
6043 const struct cred
*creds
= NULL
;
6046 if (req
->work
.creds
&& req
->work
.creds
!= current_cred())
6047 creds
= override_creds(req
->work
.creds
);
6049 switch (req
->opcode
) {
6051 ret
= io_nop(req
, issue_flags
);
6053 case IORING_OP_READV
:
6054 case IORING_OP_READ_FIXED
:
6055 case IORING_OP_READ
:
6056 ret
= io_read(req
, issue_flags
);
6058 case IORING_OP_WRITEV
:
6059 case IORING_OP_WRITE_FIXED
:
6060 case IORING_OP_WRITE
:
6061 ret
= io_write(req
, issue_flags
);
6063 case IORING_OP_FSYNC
:
6064 ret
= io_fsync(req
, issue_flags
);
6066 case IORING_OP_POLL_ADD
:
6067 ret
= io_poll_add(req
, issue_flags
);
6069 case IORING_OP_POLL_REMOVE
:
6070 ret
= io_poll_remove(req
, issue_flags
);
6072 case IORING_OP_SYNC_FILE_RANGE
:
6073 ret
= io_sync_file_range(req
, issue_flags
);
6075 case IORING_OP_SENDMSG
:
6076 ret
= io_sendmsg(req
, issue_flags
);
6078 case IORING_OP_SEND
:
6079 ret
= io_send(req
, issue_flags
);
6081 case IORING_OP_RECVMSG
:
6082 ret
= io_recvmsg(req
, issue_flags
);
6084 case IORING_OP_RECV
:
6085 ret
= io_recv(req
, issue_flags
);
6087 case IORING_OP_TIMEOUT
:
6088 ret
= io_timeout(req
, issue_flags
);
6090 case IORING_OP_TIMEOUT_REMOVE
:
6091 ret
= io_timeout_remove(req
, issue_flags
);
6093 case IORING_OP_ACCEPT
:
6094 ret
= io_accept(req
, issue_flags
);
6096 case IORING_OP_CONNECT
:
6097 ret
= io_connect(req
, issue_flags
);
6099 case IORING_OP_ASYNC_CANCEL
:
6100 ret
= io_async_cancel(req
, issue_flags
);
6102 case IORING_OP_FALLOCATE
:
6103 ret
= io_fallocate(req
, issue_flags
);
6105 case IORING_OP_OPENAT
:
6106 ret
= io_openat(req
, issue_flags
);
6108 case IORING_OP_CLOSE
:
6109 ret
= io_close(req
, issue_flags
);
6111 case IORING_OP_FILES_UPDATE
:
6112 ret
= io_files_update(req
, issue_flags
);
6114 case IORING_OP_STATX
:
6115 ret
= io_statx(req
, issue_flags
);
6117 case IORING_OP_FADVISE
:
6118 ret
= io_fadvise(req
, issue_flags
);
6120 case IORING_OP_MADVISE
:
6121 ret
= io_madvise(req
, issue_flags
);
6123 case IORING_OP_OPENAT2
:
6124 ret
= io_openat2(req
, issue_flags
);
6126 case IORING_OP_EPOLL_CTL
:
6127 ret
= io_epoll_ctl(req
, issue_flags
);
6129 case IORING_OP_SPLICE
:
6130 ret
= io_splice(req
, issue_flags
);
6132 case IORING_OP_PROVIDE_BUFFERS
:
6133 ret
= io_provide_buffers(req
, issue_flags
);
6135 case IORING_OP_REMOVE_BUFFERS
:
6136 ret
= io_remove_buffers(req
, issue_flags
);
6139 ret
= io_tee(req
, issue_flags
);
6141 case IORING_OP_SHUTDOWN
:
6142 ret
= io_shutdown(req
, issue_flags
);
6144 case IORING_OP_RENAMEAT
:
6145 ret
= io_renameat(req
, issue_flags
);
6147 case IORING_OP_UNLINKAT
:
6148 ret
= io_unlinkat(req
, issue_flags
);
6156 revert_creds(creds
);
6161 /* If the op doesn't have a file, we're not polling for it */
6162 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
6163 const bool in_async
= io_wq_current_is_worker();
6165 /* workqueue context doesn't hold uring_lock, grab it now */
6167 mutex_lock(&ctx
->uring_lock
);
6169 io_iopoll_req_issued(req
, in_async
);
6172 mutex_unlock(&ctx
->uring_lock
);
6178 static void io_wq_submit_work(struct io_wq_work
*work
)
6180 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6181 struct io_kiocb
*timeout
;
6184 timeout
= io_prep_linked_timeout(req
);
6186 io_queue_linked_timeout(timeout
);
6188 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6193 ret
= io_issue_sqe(req
, 0);
6195 * We can get EAGAIN for polled IO even though we're
6196 * forcing a sync submission from here, since we can't
6197 * wait for request slots on the block side.
6205 /* avoid locking problems by failing it from a clean context */
6207 /* io-wq is going to take one down */
6208 refcount_inc(&req
->refs
);
6209 io_req_task_queue_fail(req
, ret
);
6213 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6216 struct fixed_rsrc_table
*table
;
6218 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
6219 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
6222 static struct file
*io_file_get(struct io_submit_state
*state
,
6223 struct io_kiocb
*req
, int fd
, bool fixed
)
6225 struct io_ring_ctx
*ctx
= req
->ctx
;
6229 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6231 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6232 file
= io_file_from_index(ctx
, fd
);
6233 io_set_resource_node(req
);
6235 trace_io_uring_file_get(ctx
, fd
);
6236 file
= __io_file_get(state
, fd
);
6239 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6240 io_req_track_inflight(req
);
6244 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6246 struct io_timeout_data
*data
= container_of(timer
,
6247 struct io_timeout_data
, timer
);
6248 struct io_kiocb
*prev
, *req
= data
->req
;
6249 struct io_ring_ctx
*ctx
= req
->ctx
;
6250 unsigned long flags
;
6252 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6253 prev
= req
->timeout
.head
;
6254 req
->timeout
.head
= NULL
;
6257 * We don't expect the list to be empty, that will only happen if we
6258 * race with the completion of the linked work.
6260 if (prev
&& refcount_inc_not_zero(&prev
->refs
))
6261 io_remove_next_linked(prev
);
6264 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6267 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6268 io_put_req_deferred(prev
, 1);
6270 io_req_complete_post(req
, -ETIME
, 0);
6271 io_put_req_deferred(req
, 1);
6273 return HRTIMER_NORESTART
;
6276 static void __io_queue_linked_timeout(struct io_kiocb
*req
)
6279 * If the back reference is NULL, then our linked request finished
6280 * before we got a chance to setup the timer
6282 if (req
->timeout
.head
) {
6283 struct io_timeout_data
*data
= req
->async_data
;
6285 data
->timer
.function
= io_link_timeout_fn
;
6286 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6291 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6293 struct io_ring_ctx
*ctx
= req
->ctx
;
6295 spin_lock_irq(&ctx
->completion_lock
);
6296 __io_queue_linked_timeout(req
);
6297 spin_unlock_irq(&ctx
->completion_lock
);
6299 /* drop submission reference */
6303 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6305 struct io_kiocb
*nxt
= req
->link
;
6307 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6308 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6311 nxt
->timeout
.head
= req
;
6312 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6313 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6317 static void __io_queue_sqe(struct io_kiocb
*req
)
6319 struct io_kiocb
*linked_timeout
= io_prep_linked_timeout(req
);
6322 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
6325 * We async punt it if the file wasn't marked NOWAIT, or if the file
6326 * doesn't support non-blocking read/write attempts
6328 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6329 if (!io_arm_poll_handler(req
)) {
6331 * Queued up for async execution, worker will release
6332 * submit reference when the iocb is actually submitted.
6334 io_queue_async_work(req
);
6336 } else if (likely(!ret
)) {
6337 /* drop submission reference */
6338 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
6339 struct io_ring_ctx
*ctx
= req
->ctx
;
6340 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
6342 cs
->reqs
[cs
->nr
++] = req
;
6343 if (cs
->nr
== ARRAY_SIZE(cs
->reqs
))
6344 io_submit_flush_completions(cs
, ctx
);
6349 req_set_fail_links(req
);
6351 io_req_complete(req
, ret
);
6354 io_queue_linked_timeout(linked_timeout
);
6357 static void io_queue_sqe(struct io_kiocb
*req
)
6361 ret
= io_req_defer(req
);
6363 if (ret
!= -EIOCBQUEUED
) {
6365 req_set_fail_links(req
);
6367 io_req_complete(req
, ret
);
6369 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
6370 ret
= io_req_defer_prep(req
);
6373 io_queue_async_work(req
);
6375 __io_queue_sqe(req
);
6380 * Check SQE restrictions (opcode and flags).
6382 * Returns 'true' if SQE is allowed, 'false' otherwise.
6384 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6385 struct io_kiocb
*req
,
6386 unsigned int sqe_flags
)
6388 if (!ctx
->restricted
)
6391 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6394 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6395 ctx
->restrictions
.sqe_flags_required
)
6398 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6399 ctx
->restrictions
.sqe_flags_required
))
6405 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6406 const struct io_uring_sqe
*sqe
)
6408 struct io_submit_state
*state
;
6409 unsigned int sqe_flags
;
6410 int personality
, ret
= 0;
6412 req
->opcode
= READ_ONCE(sqe
->opcode
);
6413 /* same numerical values with corresponding REQ_F_*, safe to copy */
6414 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
6415 req
->user_data
= READ_ONCE(sqe
->user_data
);
6416 req
->async_data
= NULL
;
6420 req
->fixed_rsrc_refs
= NULL
;
6421 /* one is dropped after submission, the other at completion */
6422 refcount_set(&req
->refs
, 2);
6423 req
->task
= current
;
6425 req
->work
.list
.next
= NULL
;
6426 req
->work
.creds
= NULL
;
6427 req
->work
.flags
= 0;
6429 /* enforce forwards compatibility on users */
6430 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
)) {
6435 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6438 if (unlikely(!io_check_restriction(ctx
, req
, sqe_flags
)))
6441 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6442 !io_op_defs
[req
->opcode
].buffer_select
)
6445 personality
= READ_ONCE(sqe
->personality
);
6447 req
->work
.creds
= xa_load(&ctx
->personalities
, personality
);
6448 if (!req
->work
.creds
)
6450 get_cred(req
->work
.creds
);
6452 state
= &ctx
->submit_state
;
6455 * Plug now if we have more than 1 IO left after this, and the target
6456 * is potentially a read/write to block based storage.
6458 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6459 io_op_defs
[req
->opcode
].plug
) {
6460 blk_start_plug(&state
->plug
);
6461 state
->plug_started
= true;
6464 if (io_op_defs
[req
->opcode
].needs_file
) {
6465 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6467 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6468 if (unlikely(!req
->file
))
6476 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6477 const struct io_uring_sqe
*sqe
)
6479 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
6482 ret
= io_init_req(ctx
, req
, sqe
);
6483 if (unlikely(ret
)) {
6486 /* fail even hard links since we don't submit */
6487 link
->head
->flags
|= REQ_F_FAIL_LINK
;
6488 io_put_req(link
->head
);
6489 io_req_complete(link
->head
, -ECANCELED
);
6493 io_req_complete(req
, ret
);
6496 ret
= io_req_prep(req
, sqe
);
6500 /* don't need @sqe from now on */
6501 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6502 true, ctx
->flags
& IORING_SETUP_SQPOLL
);
6505 * If we already have a head request, queue this one for async
6506 * submittal once the head completes. If we don't have a head but
6507 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6508 * submitted sync once the chain is complete. If none of those
6509 * conditions are true (normal request), then just queue it.
6512 struct io_kiocb
*head
= link
->head
;
6515 * Taking sequential execution of a link, draining both sides
6516 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6517 * requests in the link. So, it drains the head and the
6518 * next after the link request. The last one is done via
6519 * drain_next flag to persist the effect across calls.
6521 if (req
->flags
& REQ_F_IO_DRAIN
) {
6522 head
->flags
|= REQ_F_IO_DRAIN
;
6523 ctx
->drain_next
= 1;
6525 ret
= io_req_defer_prep(req
);
6528 trace_io_uring_link(ctx
, req
, head
);
6529 link
->last
->link
= req
;
6532 /* last request of a link, enqueue the link */
6533 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6538 if (unlikely(ctx
->drain_next
)) {
6539 req
->flags
|= REQ_F_IO_DRAIN
;
6540 ctx
->drain_next
= 0;
6542 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6554 * Batched submission is done, ensure local IO is flushed out.
6556 static void io_submit_state_end(struct io_submit_state
*state
,
6557 struct io_ring_ctx
*ctx
)
6559 if (state
->link
.head
)
6560 io_queue_sqe(state
->link
.head
);
6562 io_submit_flush_completions(&state
->comp
, ctx
);
6563 if (state
->plug_started
)
6564 blk_finish_plug(&state
->plug
);
6565 io_state_file_put(state
);
6569 * Start submission side cache.
6571 static void io_submit_state_start(struct io_submit_state
*state
,
6572 unsigned int max_ios
)
6574 state
->plug_started
= false;
6575 state
->ios_left
= max_ios
;
6576 /* set only head, no need to init link_last in advance */
6577 state
->link
.head
= NULL
;
6580 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6582 struct io_rings
*rings
= ctx
->rings
;
6585 * Ensure any loads from the SQEs are done at this point,
6586 * since once we write the new head, the application could
6587 * write new data to them.
6589 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6593 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6594 * that is mapped by userspace. This means that care needs to be taken to
6595 * ensure that reads are stable, as we cannot rely on userspace always
6596 * being a good citizen. If members of the sqe are validated and then later
6597 * used, it's important that those reads are done through READ_ONCE() to
6598 * prevent a re-load down the line.
6600 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6602 u32
*sq_array
= ctx
->sq_array
;
6606 * The cached sq head (or cq tail) serves two purposes:
6608 * 1) allows us to batch the cost of updating the user visible
6610 * 2) allows the kernel side to track the head on its own, even
6611 * though the application is the one updating it.
6613 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
++ & ctx
->sq_mask
]);
6614 if (likely(head
< ctx
->sq_entries
))
6615 return &ctx
->sq_sqes
[head
];
6617 /* drop invalid entries */
6618 ctx
->cached_sq_dropped
++;
6619 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6623 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6627 /* if we have a backlog and couldn't flush it all, return BUSY */
6628 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6629 if (!__io_cqring_overflow_flush(ctx
, false, NULL
, NULL
))
6633 /* make sure SQ entry isn't read before tail */
6634 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6636 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6639 percpu_counter_add(¤t
->io_uring
->inflight
, nr
);
6640 refcount_add(nr
, ¤t
->usage
);
6641 io_submit_state_start(&ctx
->submit_state
, nr
);
6643 while (submitted
< nr
) {
6644 const struct io_uring_sqe
*sqe
;
6645 struct io_kiocb
*req
;
6647 req
= io_alloc_req(ctx
);
6648 if (unlikely(!req
)) {
6650 submitted
= -EAGAIN
;
6653 sqe
= io_get_sqe(ctx
);
6654 if (unlikely(!sqe
)) {
6655 kmem_cache_free(req_cachep
, req
);
6658 /* will complete beyond this point, count as submitted */
6660 if (io_submit_sqe(ctx
, req
, sqe
))
6664 if (unlikely(submitted
!= nr
)) {
6665 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6666 struct io_uring_task
*tctx
= current
->io_uring
;
6667 int unused
= nr
- ref_used
;
6669 percpu_ref_put_many(&ctx
->refs
, unused
);
6670 percpu_counter_sub(&tctx
->inflight
, unused
);
6671 put_task_struct_many(current
, unused
);
6674 io_submit_state_end(&ctx
->submit_state
, ctx
);
6675 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6676 io_commit_sqring(ctx
);
6681 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6683 /* Tell userspace we may need a wakeup call */
6684 spin_lock_irq(&ctx
->completion_lock
);
6685 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6686 spin_unlock_irq(&ctx
->completion_lock
);
6689 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6691 spin_lock_irq(&ctx
->completion_lock
);
6692 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6693 spin_unlock_irq(&ctx
->completion_lock
);
6696 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
6698 unsigned int to_submit
;
6701 to_submit
= io_sqring_entries(ctx
);
6702 /* if we're handling multiple rings, cap submit size for fairness */
6703 if (cap_entries
&& to_submit
> 8)
6706 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
6707 unsigned nr_events
= 0;
6709 mutex_lock(&ctx
->uring_lock
);
6710 if (!list_empty(&ctx
->iopoll_list
))
6711 io_do_iopoll(ctx
, &nr_events
, 0);
6713 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
6714 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
6715 ret
= io_submit_sqes(ctx
, to_submit
);
6716 mutex_unlock(&ctx
->uring_lock
);
6719 if (!io_sqring_full(ctx
) && wq_has_sleeper(&ctx
->sqo_sq_wait
))
6720 wake_up(&ctx
->sqo_sq_wait
);
6725 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
6727 struct io_ring_ctx
*ctx
;
6728 unsigned sq_thread_idle
= 0;
6730 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6731 if (sq_thread_idle
< ctx
->sq_thread_idle
)
6732 sq_thread_idle
= ctx
->sq_thread_idle
;
6735 sqd
->sq_thread_idle
= sq_thread_idle
;
6738 static int io_sq_thread(void *data
)
6740 struct io_sq_data
*sqd
= data
;
6741 struct io_ring_ctx
*ctx
;
6742 unsigned long timeout
= 0;
6743 char buf
[TASK_COMM_LEN
];
6746 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
6747 set_task_comm(current
, buf
);
6748 current
->pf_io_worker
= NULL
;
6750 if (sqd
->sq_cpu
!= -1)
6751 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
6753 set_cpus_allowed_ptr(current
, cpu_online_mask
);
6754 current
->flags
|= PF_NO_SETAFFINITY
;
6756 mutex_lock(&sqd
->lock
);
6757 /* a user may had exited before the thread started */
6758 io_run_task_work_head(&sqd
->park_task_work
);
6760 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
)) {
6762 bool cap_entries
, sqt_spin
, needs_sched
;
6764 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
6765 signal_pending(current
)) {
6766 bool did_sig
= false;
6768 mutex_unlock(&sqd
->lock
);
6769 if (signal_pending(current
)) {
6770 struct ksignal ksig
;
6772 did_sig
= get_signal(&ksig
);
6775 mutex_lock(&sqd
->lock
);
6777 io_run_task_work_head(&sqd
->park_task_work
);
6780 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6784 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
6785 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6786 const struct cred
*creds
= NULL
;
6788 if (ctx
->sq_creds
!= current_cred())
6789 creds
= override_creds(ctx
->sq_creds
);
6790 ret
= __io_sq_thread(ctx
, cap_entries
);
6792 revert_creds(creds
);
6793 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
6797 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
6801 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6806 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
6807 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6808 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6809 !list_empty_careful(&ctx
->iopoll_list
)) {
6810 needs_sched
= false;
6813 if (io_sqring_entries(ctx
)) {
6814 needs_sched
= false;
6819 if (needs_sched
&& !test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
)) {
6820 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6821 io_ring_set_wakeup_flag(ctx
);
6823 mutex_unlock(&sqd
->lock
);
6825 mutex_lock(&sqd
->lock
);
6826 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6827 io_ring_clear_wakeup_flag(ctx
);
6830 finish_wait(&sqd
->wait
, &wait
);
6831 io_run_task_work_head(&sqd
->park_task_work
);
6832 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6835 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6836 io_uring_cancel_sqpoll(ctx
);
6838 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6839 io_ring_set_wakeup_flag(ctx
);
6840 mutex_unlock(&sqd
->lock
);
6843 io_run_task_work_head(&sqd
->park_task_work
);
6844 complete(&sqd
->exited
);
6848 struct io_wait_queue
{
6849 struct wait_queue_entry wq
;
6850 struct io_ring_ctx
*ctx
;
6852 unsigned nr_timeouts
;
6855 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
6857 struct io_ring_ctx
*ctx
= iowq
->ctx
;
6860 * Wake up if we have enough events, or if a timeout occurred since we
6861 * started waiting. For timeouts, we always want to return to userspace,
6862 * regardless of event count.
6864 return io_cqring_events(ctx
) >= iowq
->to_wait
||
6865 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
6868 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
6869 int wake_flags
, void *key
)
6871 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
6875 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6876 * the task, and the next invocation will do it.
6878 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->cq_check_overflow
))
6879 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
6883 static int io_run_task_work_sig(void)
6885 if (io_run_task_work())
6887 if (!signal_pending(current
))
6889 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
6890 return -ERESTARTSYS
;
6894 /* when returns >0, the caller should retry */
6895 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
6896 struct io_wait_queue
*iowq
,
6897 signed long *timeout
)
6901 /* make sure we run task_work before checking for signals */
6902 ret
= io_run_task_work_sig();
6903 if (ret
|| io_should_wake(iowq
))
6905 /* let the caller flush overflows, retry */
6906 if (test_bit(0, &ctx
->cq_check_overflow
))
6909 *timeout
= schedule_timeout(*timeout
);
6910 return !*timeout
? -ETIME
: 1;
6914 * Wait until events become available, if we don't already have some. The
6915 * application must reap them itself, as they reside on the shared cq ring.
6917 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
6918 const sigset_t __user
*sig
, size_t sigsz
,
6919 struct __kernel_timespec __user
*uts
)
6921 struct io_wait_queue iowq
= {
6924 .func
= io_wake_function
,
6925 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
6928 .to_wait
= min_events
,
6930 struct io_rings
*rings
= ctx
->rings
;
6931 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
6935 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
6936 if (io_cqring_events(ctx
) >= min_events
)
6938 if (!io_run_task_work())
6943 #ifdef CONFIG_COMPAT
6944 if (in_compat_syscall())
6945 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
6949 ret
= set_user_sigmask(sig
, sigsz
);
6956 struct timespec64 ts
;
6958 if (get_timespec64(&ts
, uts
))
6960 timeout
= timespec64_to_jiffies(&ts
);
6963 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
6964 trace_io_uring_cqring_wait(ctx
, min_events
);
6966 /* if we can't even flush overflow, don't wait for more */
6967 if (!io_cqring_overflow_flush(ctx
, false, NULL
, NULL
)) {
6971 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
6972 TASK_INTERRUPTIBLE
);
6973 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
6974 finish_wait(&ctx
->wait
, &iowq
.wq
);
6978 restore_saved_sigmask_unless(ret
== -EINTR
);
6980 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
6983 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6985 #if defined(CONFIG_UNIX)
6986 if (ctx
->ring_sock
) {
6987 struct sock
*sock
= ctx
->ring_sock
->sk
;
6988 struct sk_buff
*skb
;
6990 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
6996 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6999 file
= io_file_from_index(ctx
, i
);
7006 static void io_rsrc_data_ref_zero(struct percpu_ref
*ref
)
7008 struct fixed_rsrc_data
*data
;
7010 data
= container_of(ref
, struct fixed_rsrc_data
, refs
);
7011 complete(&data
->done
);
7014 static inline void io_rsrc_ref_lock(struct io_ring_ctx
*ctx
)
7016 spin_lock_bh(&ctx
->rsrc_ref_lock
);
7019 static inline void io_rsrc_ref_unlock(struct io_ring_ctx
*ctx
)
7021 spin_unlock_bh(&ctx
->rsrc_ref_lock
);
7024 static void io_sqe_rsrc_set_node(struct io_ring_ctx
*ctx
,
7025 struct fixed_rsrc_data
*rsrc_data
,
7026 struct fixed_rsrc_ref_node
*ref_node
)
7028 io_rsrc_ref_lock(ctx
);
7029 rsrc_data
->node
= ref_node
;
7030 list_add_tail(&ref_node
->node
, &ctx
->rsrc_ref_list
);
7031 io_rsrc_ref_unlock(ctx
);
7032 percpu_ref_get(&rsrc_data
->refs
);
7035 static void io_sqe_rsrc_kill_node(struct io_ring_ctx
*ctx
, struct fixed_rsrc_data
*data
)
7037 struct fixed_rsrc_ref_node
*ref_node
= NULL
;
7039 io_rsrc_ref_lock(ctx
);
7040 ref_node
= data
->node
;
7042 io_rsrc_ref_unlock(ctx
);
7044 percpu_ref_kill(&ref_node
->refs
);
7047 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data
*data
,
7048 struct io_ring_ctx
*ctx
,
7049 void (*rsrc_put
)(struct io_ring_ctx
*ctx
,
7050 struct io_rsrc_put
*prsrc
))
7052 struct fixed_rsrc_ref_node
*backup_node
;
7058 data
->quiesce
= true;
7061 backup_node
= alloc_fixed_rsrc_ref_node(ctx
);
7064 backup_node
->rsrc_data
= data
;
7065 backup_node
->rsrc_put
= rsrc_put
;
7067 io_sqe_rsrc_kill_node(ctx
, data
);
7068 percpu_ref_kill(&data
->refs
);
7069 flush_delayed_work(&ctx
->rsrc_put_work
);
7071 ret
= wait_for_completion_interruptible(&data
->done
);
7075 percpu_ref_resurrect(&data
->refs
);
7076 io_sqe_rsrc_set_node(ctx
, data
, backup_node
);
7078 reinit_completion(&data
->done
);
7079 mutex_unlock(&ctx
->uring_lock
);
7080 ret
= io_run_task_work_sig();
7081 mutex_lock(&ctx
->uring_lock
);
7083 data
->quiesce
= false;
7086 destroy_fixed_rsrc_ref_node(backup_node
);
7090 static struct fixed_rsrc_data
*alloc_fixed_rsrc_data(struct io_ring_ctx
*ctx
)
7092 struct fixed_rsrc_data
*data
;
7094 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7098 if (percpu_ref_init(&data
->refs
, io_rsrc_data_ref_zero
,
7099 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
)) {
7104 init_completion(&data
->done
);
7108 static void free_fixed_rsrc_data(struct fixed_rsrc_data
*data
)
7110 percpu_ref_exit(&data
->refs
);
7115 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7117 struct fixed_rsrc_data
*data
= ctx
->file_data
;
7118 unsigned nr_tables
, i
;
7122 * percpu_ref_is_dying() is to stop parallel files unregister
7123 * Since we possibly drop uring lock later in this function to
7126 if (!data
|| percpu_ref_is_dying(&data
->refs
))
7128 ret
= io_rsrc_ref_quiesce(data
, ctx
, io_ring_file_put
);
7132 __io_sqe_files_unregister(ctx
);
7133 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
7134 for (i
= 0; i
< nr_tables
; i
++)
7135 kfree(data
->table
[i
].files
);
7136 free_fixed_rsrc_data(data
);
7137 ctx
->file_data
= NULL
;
7138 ctx
->nr_user_files
= 0;
7142 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7143 __releases(&sqd
->lock
)
7145 WARN_ON_ONCE(sqd
->thread
== current
);
7148 * Do the dance but not conditional clear_bit() because it'd race with
7149 * other threads incrementing park_pending and setting the bit.
7151 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7152 if (atomic_dec_return(&sqd
->park_pending
))
7153 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7154 mutex_unlock(&sqd
->lock
);
7157 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7158 __acquires(&sqd
->lock
)
7160 WARN_ON_ONCE(sqd
->thread
== current
);
7162 atomic_inc(&sqd
->park_pending
);
7163 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7164 mutex_lock(&sqd
->lock
);
7166 wake_up_process(sqd
->thread
);
7169 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7171 WARN_ON_ONCE(sqd
->thread
== current
);
7173 mutex_lock(&sqd
->lock
);
7174 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7176 wake_up_process(sqd
->thread
);
7177 mutex_unlock(&sqd
->lock
);
7178 wait_for_completion(&sqd
->exited
);
7181 static void io_put_sq_data(struct io_sq_data
*sqd
)
7183 if (refcount_dec_and_test(&sqd
->refs
)) {
7184 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
7186 io_sq_thread_stop(sqd
);
7191 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
7193 struct io_sq_data
*sqd
= ctx
->sq_data
;
7196 io_sq_thread_park(sqd
);
7197 list_del_init(&ctx
->sqd_list
);
7198 io_sqd_update_thread_idle(sqd
);
7199 io_sq_thread_unpark(sqd
);
7201 io_put_sq_data(sqd
);
7202 ctx
->sq_data
= NULL
;
7204 put_cred(ctx
->sq_creds
);
7208 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7210 struct io_ring_ctx
*ctx_attach
;
7211 struct io_sq_data
*sqd
;
7214 f
= fdget(p
->wq_fd
);
7216 return ERR_PTR(-ENXIO
);
7217 if (f
.file
->f_op
!= &io_uring_fops
) {
7219 return ERR_PTR(-EINVAL
);
7222 ctx_attach
= f
.file
->private_data
;
7223 sqd
= ctx_attach
->sq_data
;
7226 return ERR_PTR(-EINVAL
);
7228 if (sqd
->task_tgid
!= current
->tgid
) {
7230 return ERR_PTR(-EPERM
);
7233 refcount_inc(&sqd
->refs
);
7238 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
7241 struct io_sq_data
*sqd
;
7244 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
7245 sqd
= io_attach_sq_data(p
);
7250 /* fall through for EPERM case, setup new sqd/task */
7251 if (PTR_ERR(sqd
) != -EPERM
)
7255 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7257 return ERR_PTR(-ENOMEM
);
7259 atomic_set(&sqd
->park_pending
, 0);
7260 refcount_set(&sqd
->refs
, 1);
7261 INIT_LIST_HEAD(&sqd
->ctx_list
);
7262 mutex_init(&sqd
->lock
);
7263 init_waitqueue_head(&sqd
->wait
);
7264 init_completion(&sqd
->exited
);
7268 #if defined(CONFIG_UNIX)
7270 * Ensure the UNIX gc is aware of our file set, so we are certain that
7271 * the io_uring can be safely unregistered on process exit, even if we have
7272 * loops in the file referencing.
7274 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7276 struct sock
*sk
= ctx
->ring_sock
->sk
;
7277 struct scm_fp_list
*fpl
;
7278 struct sk_buff
*skb
;
7281 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7285 skb
= alloc_skb(0, GFP_KERNEL
);
7294 fpl
->user
= get_uid(current_user());
7295 for (i
= 0; i
< nr
; i
++) {
7296 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7300 fpl
->fp
[nr_files
] = get_file(file
);
7301 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7306 fpl
->max
= SCM_MAX_FD
;
7307 fpl
->count
= nr_files
;
7308 UNIXCB(skb
).fp
= fpl
;
7309 skb
->destructor
= unix_destruct_scm
;
7310 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7311 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7313 for (i
= 0; i
< nr_files
; i
++)
7324 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7325 * causes regular reference counting to break down. We rely on the UNIX
7326 * garbage collection to take care of this problem for us.
7328 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7330 unsigned left
, total
;
7334 left
= ctx
->nr_user_files
;
7336 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7338 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7342 total
+= this_files
;
7348 while (total
< ctx
->nr_user_files
) {
7349 struct file
*file
= io_file_from_index(ctx
, total
);
7359 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7365 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data
*file_data
,
7366 unsigned nr_tables
, unsigned nr_files
)
7370 for (i
= 0; i
< nr_tables
; i
++) {
7371 struct fixed_rsrc_table
*table
= &file_data
->table
[i
];
7372 unsigned this_files
;
7374 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
7375 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
7379 nr_files
-= this_files
;
7385 for (i
= 0; i
< nr_tables
; i
++) {
7386 struct fixed_rsrc_table
*table
= &file_data
->table
[i
];
7387 kfree(table
->files
);
7392 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
7394 struct file
*file
= prsrc
->file
;
7395 #if defined(CONFIG_UNIX)
7396 struct sock
*sock
= ctx
->ring_sock
->sk
;
7397 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7398 struct sk_buff
*skb
;
7401 __skb_queue_head_init(&list
);
7404 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7405 * remove this entry and rearrange the file array.
7407 skb
= skb_dequeue(head
);
7409 struct scm_fp_list
*fp
;
7411 fp
= UNIXCB(skb
).fp
;
7412 for (i
= 0; i
< fp
->count
; i
++) {
7415 if (fp
->fp
[i
] != file
)
7418 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7419 left
= fp
->count
- 1 - i
;
7421 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7422 left
* sizeof(struct file
*));
7429 __skb_queue_tail(&list
, skb
);
7439 __skb_queue_tail(&list
, skb
);
7441 skb
= skb_dequeue(head
);
7444 if (skb_peek(&list
)) {
7445 spin_lock_irq(&head
->lock
);
7446 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7447 __skb_queue_tail(head
, skb
);
7448 spin_unlock_irq(&head
->lock
);
7455 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node
*ref_node
)
7457 struct fixed_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
7458 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
7459 struct io_rsrc_put
*prsrc
, *tmp
;
7461 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
7462 list_del(&prsrc
->list
);
7463 ref_node
->rsrc_put(ctx
, prsrc
);
7467 percpu_ref_exit(&ref_node
->refs
);
7469 percpu_ref_put(&rsrc_data
->refs
);
7472 static void io_rsrc_put_work(struct work_struct
*work
)
7474 struct io_ring_ctx
*ctx
;
7475 struct llist_node
*node
;
7477 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
7478 node
= llist_del_all(&ctx
->rsrc_put_llist
);
7481 struct fixed_rsrc_ref_node
*ref_node
;
7482 struct llist_node
*next
= node
->next
;
7484 ref_node
= llist_entry(node
, struct fixed_rsrc_ref_node
, llist
);
7485 __io_rsrc_put_work(ref_node
);
7490 static struct file
**io_fixed_file_slot(struct fixed_rsrc_data
*file_data
,
7493 struct fixed_rsrc_table
*table
;
7495 table
= &file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7496 return &table
->files
[i
& IORING_FILE_TABLE_MASK
];
7499 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7501 struct fixed_rsrc_ref_node
*ref_node
;
7502 struct fixed_rsrc_data
*data
;
7503 struct io_ring_ctx
*ctx
;
7504 bool first_add
= false;
7507 ref_node
= container_of(ref
, struct fixed_rsrc_ref_node
, refs
);
7508 data
= ref_node
->rsrc_data
;
7511 io_rsrc_ref_lock(ctx
);
7512 ref_node
->done
= true;
7514 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7515 ref_node
= list_first_entry(&ctx
->rsrc_ref_list
,
7516 struct fixed_rsrc_ref_node
, node
);
7517 /* recycle ref nodes in order */
7518 if (!ref_node
->done
)
7520 list_del(&ref_node
->node
);
7521 first_add
|= llist_add(&ref_node
->llist
, &ctx
->rsrc_put_llist
);
7523 io_rsrc_ref_unlock(ctx
);
7525 if (percpu_ref_is_dying(&data
->refs
))
7529 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, 0);
7531 queue_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
7534 static struct fixed_rsrc_ref_node
*alloc_fixed_rsrc_ref_node(
7535 struct io_ring_ctx
*ctx
)
7537 struct fixed_rsrc_ref_node
*ref_node
;
7539 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7543 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7548 INIT_LIST_HEAD(&ref_node
->node
);
7549 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7550 ref_node
->done
= false;
7554 static void init_fixed_file_ref_node(struct io_ring_ctx
*ctx
,
7555 struct fixed_rsrc_ref_node
*ref_node
)
7557 ref_node
->rsrc_data
= ctx
->file_data
;
7558 ref_node
->rsrc_put
= io_ring_file_put
;
7561 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node
*ref_node
)
7563 percpu_ref_exit(&ref_node
->refs
);
7568 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7571 __s32 __user
*fds
= (__s32 __user
*) arg
;
7572 unsigned nr_tables
, i
;
7574 int fd
, ret
= -ENOMEM
;
7575 struct fixed_rsrc_ref_node
*ref_node
;
7576 struct fixed_rsrc_data
*file_data
;
7582 if (nr_args
> IORING_MAX_FIXED_FILES
)
7585 file_data
= alloc_fixed_rsrc_data(ctx
);
7588 ctx
->file_data
= file_data
;
7590 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
7591 file_data
->table
= kcalloc(nr_tables
, sizeof(*file_data
->table
),
7593 if (!file_data
->table
)
7596 if (io_sqe_alloc_file_tables(file_data
, nr_tables
, nr_args
))
7599 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7600 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7604 /* allow sparse sets */
7614 * Don't allow io_uring instances to be registered. If UNIX
7615 * isn't enabled, then this causes a reference cycle and this
7616 * instance can never get freed. If UNIX is enabled we'll
7617 * handle it just fine, but there's still no point in allowing
7618 * a ring fd as it doesn't support regular read/write anyway.
7620 if (file
->f_op
== &io_uring_fops
) {
7624 *io_fixed_file_slot(file_data
, i
) = file
;
7627 ret
= io_sqe_files_scm(ctx
);
7629 io_sqe_files_unregister(ctx
);
7633 ref_node
= alloc_fixed_rsrc_ref_node(ctx
);
7635 io_sqe_files_unregister(ctx
);
7638 init_fixed_file_ref_node(ctx
, ref_node
);
7640 io_sqe_rsrc_set_node(ctx
, file_data
, ref_node
);
7643 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7644 file
= io_file_from_index(ctx
, i
);
7648 for (i
= 0; i
< nr_tables
; i
++)
7649 kfree(file_data
->table
[i
].files
);
7650 ctx
->nr_user_files
= 0;
7652 free_fixed_rsrc_data(ctx
->file_data
);
7653 ctx
->file_data
= NULL
;
7657 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7660 #if defined(CONFIG_UNIX)
7661 struct sock
*sock
= ctx
->ring_sock
->sk
;
7662 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7663 struct sk_buff
*skb
;
7666 * See if we can merge this file into an existing skb SCM_RIGHTS
7667 * file set. If there's no room, fall back to allocating a new skb
7668 * and filling it in.
7670 spin_lock_irq(&head
->lock
);
7671 skb
= skb_peek(head
);
7673 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7675 if (fpl
->count
< SCM_MAX_FD
) {
7676 __skb_unlink(skb
, head
);
7677 spin_unlock_irq(&head
->lock
);
7678 fpl
->fp
[fpl
->count
] = get_file(file
);
7679 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7681 spin_lock_irq(&head
->lock
);
7682 __skb_queue_head(head
, skb
);
7687 spin_unlock_irq(&head
->lock
);
7694 return __io_sqe_files_scm(ctx
, 1, index
);
7700 static int io_queue_rsrc_removal(struct fixed_rsrc_data
*data
, void *rsrc
)
7702 struct io_rsrc_put
*prsrc
;
7703 struct fixed_rsrc_ref_node
*ref_node
= data
->node
;
7705 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
7710 list_add(&prsrc
->list
, &ref_node
->rsrc_list
);
7715 static inline int io_queue_file_removal(struct fixed_rsrc_data
*data
,
7718 return io_queue_rsrc_removal(data
, (void *)file
);
7721 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7722 struct io_uring_rsrc_update
*up
,
7725 struct fixed_rsrc_data
*data
= ctx
->file_data
;
7726 struct fixed_rsrc_ref_node
*ref_node
;
7727 struct file
*file
, **file_slot
;
7731 bool needs_switch
= false;
7733 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7735 if (done
> ctx
->nr_user_files
)
7738 ref_node
= alloc_fixed_rsrc_ref_node(ctx
);
7741 init_fixed_file_ref_node(ctx
, ref_node
);
7743 fds
= u64_to_user_ptr(up
->data
);
7744 for (done
= 0; done
< nr_args
; done
++) {
7746 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7750 if (fd
== IORING_REGISTER_FILES_SKIP
)
7753 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
7754 file_slot
= io_fixed_file_slot(ctx
->file_data
, i
);
7757 err
= io_queue_file_removal(data
, *file_slot
);
7761 needs_switch
= true;
7770 * Don't allow io_uring instances to be registered. If
7771 * UNIX isn't enabled, then this causes a reference
7772 * cycle and this instance can never get freed. If UNIX
7773 * is enabled we'll handle it just fine, but there's
7774 * still no point in allowing a ring fd as it doesn't
7775 * support regular read/write anyway.
7777 if (file
->f_op
== &io_uring_fops
) {
7783 err
= io_sqe_file_register(ctx
, file
, i
);
7793 percpu_ref_kill(&data
->node
->refs
);
7794 io_sqe_rsrc_set_node(ctx
, data
, ref_node
);
7796 destroy_fixed_rsrc_ref_node(ref_node
);
7798 return done
? done
: err
;
7801 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
7804 struct io_uring_rsrc_update up
;
7806 if (!ctx
->file_data
)
7810 if (copy_from_user(&up
, arg
, sizeof(up
)))
7815 return __io_sqe_files_update(ctx
, &up
, nr_args
);
7818 static struct io_wq_work
*io_free_work(struct io_wq_work
*work
)
7820 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7822 req
= io_put_req_find_next(req
);
7823 return req
? &req
->work
: NULL
;
7826 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
)
7828 struct io_wq_hash
*hash
;
7829 struct io_wq_data data
;
7830 unsigned int concurrency
;
7832 hash
= ctx
->hash_map
;
7834 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
7836 return ERR_PTR(-ENOMEM
);
7837 refcount_set(&hash
->refs
, 1);
7838 init_waitqueue_head(&hash
->wait
);
7839 ctx
->hash_map
= hash
;
7843 data
.free_work
= io_free_work
;
7844 data
.do_work
= io_wq_submit_work
;
7846 /* Do QD, or 4 * CPUS, whatever is smallest */
7847 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
7849 return io_wq_create(concurrency
, &data
);
7852 static int io_uring_alloc_task_context(struct task_struct
*task
,
7853 struct io_ring_ctx
*ctx
)
7855 struct io_uring_task
*tctx
;
7858 tctx
= kmalloc(sizeof(*tctx
), GFP_KERNEL
);
7859 if (unlikely(!tctx
))
7862 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
7863 if (unlikely(ret
)) {
7868 tctx
->io_wq
= io_init_wq_offload(ctx
);
7869 if (IS_ERR(tctx
->io_wq
)) {
7870 ret
= PTR_ERR(tctx
->io_wq
);
7871 percpu_counter_destroy(&tctx
->inflight
);
7877 init_waitqueue_head(&tctx
->wait
);
7879 atomic_set(&tctx
->in_idle
, 0);
7880 task
->io_uring
= tctx
;
7881 spin_lock_init(&tctx
->task_lock
);
7882 INIT_WQ_LIST(&tctx
->task_list
);
7883 tctx
->task_state
= 0;
7884 init_task_work(&tctx
->task_work
, tctx_task_work
);
7888 void __io_uring_free(struct task_struct
*tsk
)
7890 struct io_uring_task
*tctx
= tsk
->io_uring
;
7892 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
7893 WARN_ON_ONCE(tctx
->io_wq
);
7895 percpu_counter_destroy(&tctx
->inflight
);
7897 tsk
->io_uring
= NULL
;
7900 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
7901 struct io_uring_params
*p
)
7905 /* Retain compatibility with failing for an invalid attach attempt */
7906 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
7907 IORING_SETUP_ATTACH_WQ
) {
7910 f
= fdget(p
->wq_fd
);
7913 if (f
.file
->f_op
!= &io_uring_fops
) {
7919 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7920 struct task_struct
*tsk
;
7921 struct io_sq_data
*sqd
;
7925 if (!capable(CAP_SYS_ADMIN
) && !capable(CAP_SYS_NICE
))
7928 sqd
= io_get_sq_data(p
, &attached
);
7934 ctx
->sq_creds
= get_current_cred();
7936 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
7937 if (!ctx
->sq_thread_idle
)
7938 ctx
->sq_thread_idle
= HZ
;
7941 io_sq_thread_park(sqd
);
7942 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
7943 io_sqd_update_thread_idle(sqd
);
7944 /* don't attach to a dying SQPOLL thread, would be racy */
7945 if (attached
&& !sqd
->thread
)
7947 io_sq_thread_unpark(sqd
);
7954 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7955 int cpu
= p
->sq_thread_cpu
;
7958 if (cpu
>= nr_cpu_ids
)
7960 if (!cpu_online(cpu
))
7968 sqd
->task_pid
= current
->pid
;
7969 sqd
->task_tgid
= current
->tgid
;
7970 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
7977 ret
= io_uring_alloc_task_context(tsk
, ctx
);
7978 wake_up_new_task(tsk
);
7981 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7982 /* Can't have SQ_AFF without SQPOLL */
7989 io_sq_thread_finish(ctx
);
7992 complete(&ctx
->sq_data
->exited
);
7996 static inline void __io_unaccount_mem(struct user_struct
*user
,
7997 unsigned long nr_pages
)
7999 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8002 static inline int __io_account_mem(struct user_struct
*user
,
8003 unsigned long nr_pages
)
8005 unsigned long page_limit
, cur_pages
, new_pages
;
8007 /* Don't allow more pages than we can safely lock */
8008 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8011 cur_pages
= atomic_long_read(&user
->locked_vm
);
8012 new_pages
= cur_pages
+ nr_pages
;
8013 if (new_pages
> page_limit
)
8015 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8016 new_pages
) != cur_pages
);
8021 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8024 __io_unaccount_mem(ctx
->user
, nr_pages
);
8026 if (ctx
->mm_account
)
8027 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8030 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8035 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8040 if (ctx
->mm_account
)
8041 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8046 static void io_mem_free(void *ptr
)
8053 page
= virt_to_head_page(ptr
);
8054 if (put_page_testzero(page
))
8055 free_compound_page(page
);
8058 static void *io_mem_alloc(size_t size
)
8060 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8061 __GFP_NORETRY
| __GFP_ACCOUNT
;
8063 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8066 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8069 struct io_rings
*rings
;
8070 size_t off
, sq_array_size
;
8072 off
= struct_size(rings
, cqes
, cq_entries
);
8073 if (off
== SIZE_MAX
)
8077 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8085 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8086 if (sq_array_size
== SIZE_MAX
)
8089 if (check_add_overflow(off
, sq_array_size
, &off
))
8095 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8099 if (!ctx
->user_bufs
)
8102 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8103 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8105 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
8106 unpin_user_page(imu
->bvec
[j
].bv_page
);
8108 if (imu
->acct_pages
)
8109 io_unaccount_mem(ctx
, imu
->acct_pages
);
8114 kfree(ctx
->user_bufs
);
8115 ctx
->user_bufs
= NULL
;
8116 ctx
->nr_user_bufs
= 0;
8120 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8121 void __user
*arg
, unsigned index
)
8123 struct iovec __user
*src
;
8125 #ifdef CONFIG_COMPAT
8127 struct compat_iovec __user
*ciovs
;
8128 struct compat_iovec ciov
;
8130 ciovs
= (struct compat_iovec __user
*) arg
;
8131 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8134 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8135 dst
->iov_len
= ciov
.iov_len
;
8139 src
= (struct iovec __user
*) arg
;
8140 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8146 * Not super efficient, but this is just a registration time. And we do cache
8147 * the last compound head, so generally we'll only do a full search if we don't
8150 * We check if the given compound head page has already been accounted, to
8151 * avoid double accounting it. This allows us to account the full size of the
8152 * page, not just the constituent pages of a huge page.
8154 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8155 int nr_pages
, struct page
*hpage
)
8159 /* check current page array */
8160 for (i
= 0; i
< nr_pages
; i
++) {
8161 if (!PageCompound(pages
[i
]))
8163 if (compound_head(pages
[i
]) == hpage
)
8167 /* check previously registered pages */
8168 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8169 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8171 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8172 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8174 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8182 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8183 int nr_pages
, struct io_mapped_ubuf
*imu
,
8184 struct page
**last_hpage
)
8188 for (i
= 0; i
< nr_pages
; i
++) {
8189 if (!PageCompound(pages
[i
])) {
8194 hpage
= compound_head(pages
[i
]);
8195 if (hpage
== *last_hpage
)
8197 *last_hpage
= hpage
;
8198 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8200 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8204 if (!imu
->acct_pages
)
8207 ret
= io_account_mem(ctx
, imu
->acct_pages
);
8209 imu
->acct_pages
= 0;
8213 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
8214 struct io_mapped_ubuf
*imu
,
8215 struct page
**last_hpage
)
8217 struct vm_area_struct
**vmas
= NULL
;
8218 struct page
**pages
= NULL
;
8219 unsigned long off
, start
, end
, ubuf
;
8221 int ret
, pret
, nr_pages
, i
;
8223 ubuf
= (unsigned long) iov
->iov_base
;
8224 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8225 start
= ubuf
>> PAGE_SHIFT
;
8226 nr_pages
= end
- start
;
8230 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
8234 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
8239 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
8245 mmap_read_lock(current
->mm
);
8246 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
8248 if (pret
== nr_pages
) {
8249 /* don't support file backed memory */
8250 for (i
= 0; i
< nr_pages
; i
++) {
8251 struct vm_area_struct
*vma
= vmas
[i
];
8254 !is_file_hugepages(vma
->vm_file
)) {
8260 ret
= pret
< 0 ? pret
: -EFAULT
;
8262 mmap_read_unlock(current
->mm
);
8265 * if we did partial map, or found file backed vmas,
8266 * release any pages we did get
8269 unpin_user_pages(pages
, pret
);
8274 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
8276 unpin_user_pages(pages
, pret
);
8281 off
= ubuf
& ~PAGE_MASK
;
8282 size
= iov
->iov_len
;
8283 for (i
= 0; i
< nr_pages
; i
++) {
8286 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8287 imu
->bvec
[i
].bv_page
= pages
[i
];
8288 imu
->bvec
[i
].bv_len
= vec_len
;
8289 imu
->bvec
[i
].bv_offset
= off
;
8293 /* store original address for later verification */
8295 imu
->len
= iov
->iov_len
;
8296 imu
->nr_bvecs
= nr_pages
;
8304 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
8308 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
8311 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
8313 if (!ctx
->user_bufs
)
8319 static int io_buffer_validate(struct iovec
*iov
)
8322 * Don't impose further limits on the size and buffer
8323 * constraints here, we'll -EINVAL later when IO is
8324 * submitted if they are wrong.
8326 if (!iov
->iov_base
|| !iov
->iov_len
)
8329 /* arbitrary limit, but we need something */
8330 if (iov
->iov_len
> SZ_1G
)
8336 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8337 unsigned int nr_args
)
8341 struct page
*last_hpage
= NULL
;
8343 ret
= io_buffers_map_alloc(ctx
, nr_args
);
8347 for (i
= 0; i
< nr_args
; i
++) {
8348 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
8350 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8354 ret
= io_buffer_validate(&iov
);
8358 ret
= io_sqe_buffer_register(ctx
, &iov
, imu
, &last_hpage
);
8362 ctx
->nr_user_bufs
++;
8366 io_sqe_buffers_unregister(ctx
);
8371 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8373 __s32 __user
*fds
= arg
;
8379 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8382 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8383 if (IS_ERR(ctx
->cq_ev_fd
)) {
8384 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8385 ctx
->cq_ev_fd
= NULL
;
8392 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8394 if (ctx
->cq_ev_fd
) {
8395 eventfd_ctx_put(ctx
->cq_ev_fd
);
8396 ctx
->cq_ev_fd
= NULL
;
8403 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8405 struct io_buffer
*buf
;
8406 unsigned long index
;
8408 xa_for_each(&ctx
->io_buffers
, index
, buf
)
8409 __io_remove_buffers(ctx
, buf
, index
, -1U);
8412 static void io_req_cache_free(struct list_head
*list
, struct task_struct
*tsk
)
8414 struct io_kiocb
*req
, *nxt
;
8416 list_for_each_entry_safe(req
, nxt
, list
, compl.list
) {
8417 if (tsk
&& req
->task
!= tsk
)
8419 list_del(&req
->compl.list
);
8420 kmem_cache_free(req_cachep
, req
);
8424 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
8426 struct io_submit_state
*submit_state
= &ctx
->submit_state
;
8427 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
8429 mutex_lock(&ctx
->uring_lock
);
8431 if (submit_state
->free_reqs
) {
8432 kmem_cache_free_bulk(req_cachep
, submit_state
->free_reqs
,
8433 submit_state
->reqs
);
8434 submit_state
->free_reqs
= 0;
8437 spin_lock_irq(&ctx
->completion_lock
);
8438 list_splice_init(&cs
->locked_free_list
, &cs
->free_list
);
8439 cs
->locked_free_nr
= 0;
8440 spin_unlock_irq(&ctx
->completion_lock
);
8442 io_req_cache_free(&cs
->free_list
, NULL
);
8444 mutex_unlock(&ctx
->uring_lock
);
8447 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8450 * Some may use context even when all refs and requests have been put,
8451 * and they are free to do so while still holding uring_lock or
8452 * completion_lock, see __io_req_task_submit(). Wait for them to finish.
8454 mutex_lock(&ctx
->uring_lock
);
8455 mutex_unlock(&ctx
->uring_lock
);
8456 spin_lock_irq(&ctx
->completion_lock
);
8457 spin_unlock_irq(&ctx
->completion_lock
);
8459 io_sq_thread_finish(ctx
);
8460 io_sqe_buffers_unregister(ctx
);
8462 if (ctx
->mm_account
) {
8463 mmdrop(ctx
->mm_account
);
8464 ctx
->mm_account
= NULL
;
8467 mutex_lock(&ctx
->uring_lock
);
8468 io_sqe_files_unregister(ctx
);
8469 mutex_unlock(&ctx
->uring_lock
);
8470 io_eventfd_unregister(ctx
);
8471 io_destroy_buffers(ctx
);
8473 #if defined(CONFIG_UNIX)
8474 if (ctx
->ring_sock
) {
8475 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8476 sock_release(ctx
->ring_sock
);
8480 io_mem_free(ctx
->rings
);
8481 io_mem_free(ctx
->sq_sqes
);
8483 percpu_ref_exit(&ctx
->refs
);
8484 free_uid(ctx
->user
);
8485 io_req_caches_free(ctx
);
8487 io_wq_put_hash(ctx
->hash_map
);
8488 kfree(ctx
->cancel_hash
);
8492 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8494 struct io_ring_ctx
*ctx
= file
->private_data
;
8497 poll_wait(file
, &ctx
->cq_wait
, wait
);
8499 * synchronizes with barrier from wq_has_sleeper call in
8503 if (!io_sqring_full(ctx
))
8504 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8507 * Don't flush cqring overflow list here, just do a simple check.
8508 * Otherwise there could possible be ABBA deadlock:
8511 * lock(&ctx->uring_lock);
8513 * lock(&ctx->uring_lock);
8516 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8517 * pushs them to do the flush.
8519 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->cq_check_overflow
))
8520 mask
|= EPOLLIN
| EPOLLRDNORM
;
8525 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8527 struct io_ring_ctx
*ctx
= file
->private_data
;
8529 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8532 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8534 const struct cred
*creds
;
8536 creds
= xa_erase(&ctx
->personalities
, id
);
8545 static inline bool io_run_ctx_fallback(struct io_ring_ctx
*ctx
)
8547 return io_run_task_work_head(&ctx
->exit_task_work
);
8550 struct io_tctx_exit
{
8551 struct callback_head task_work
;
8552 struct completion completion
;
8553 struct io_ring_ctx
*ctx
;
8556 static void io_tctx_exit_cb(struct callback_head
*cb
)
8558 struct io_uring_task
*tctx
= current
->io_uring
;
8559 struct io_tctx_exit
*work
;
8561 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
8563 * When @in_idle, we're in cancellation and it's racy to remove the
8564 * node. It'll be removed by the end of cancellation, just ignore it.
8566 if (!atomic_read(&tctx
->in_idle
))
8567 io_uring_del_task_file((unsigned long)work
->ctx
);
8568 complete(&work
->completion
);
8571 static void io_ring_exit_work(struct work_struct
*work
)
8573 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
8574 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
8575 struct io_tctx_exit exit
;
8576 struct io_tctx_node
*node
;
8579 /* prevent SQPOLL from submitting new requests */
8581 io_sq_thread_park(ctx
->sq_data
);
8582 list_del_init(&ctx
->sqd_list
);
8583 io_sqd_update_thread_idle(ctx
->sq_data
);
8584 io_sq_thread_unpark(ctx
->sq_data
);
8588 * If we're doing polled IO and end up having requests being
8589 * submitted async (out-of-line), then completions can come in while
8590 * we're waiting for refs to drop. We need to reap these manually,
8591 * as nobody else will be looking for them.
8594 io_uring_try_cancel_requests(ctx
, NULL
, NULL
);
8596 WARN_ON_ONCE(time_after(jiffies
, timeout
));
8597 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8599 mutex_lock(&ctx
->uring_lock
);
8600 while (!list_empty(&ctx
->tctx_list
)) {
8601 WARN_ON_ONCE(time_after(jiffies
, timeout
));
8603 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
8606 init_completion(&exit
.completion
);
8607 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
8608 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
8609 if (WARN_ON_ONCE(ret
))
8611 wake_up_process(node
->task
);
8613 mutex_unlock(&ctx
->uring_lock
);
8614 wait_for_completion(&exit
.completion
);
8616 mutex_lock(&ctx
->uring_lock
);
8618 mutex_unlock(&ctx
->uring_lock
);
8620 io_ring_ctx_free(ctx
);
8623 /* Returns true if we found and killed one or more timeouts */
8624 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
8625 struct files_struct
*files
)
8627 struct io_kiocb
*req
, *tmp
;
8630 spin_lock_irq(&ctx
->completion_lock
);
8631 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
8632 if (io_match_task(req
, tsk
, files
)) {
8633 io_kill_timeout(req
, -ECANCELED
);
8638 io_commit_cqring(ctx
);
8639 spin_unlock_irq(&ctx
->completion_lock
);
8641 io_cqring_ev_posted(ctx
);
8642 return canceled
!= 0;
8645 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8647 unsigned long index
;
8648 struct creds
*creds
;
8650 mutex_lock(&ctx
->uring_lock
);
8651 percpu_ref_kill(&ctx
->refs
);
8652 /* if force is set, the ring is going away. always drop after that */
8653 ctx
->cq_overflow_flushed
= 1;
8655 __io_cqring_overflow_flush(ctx
, true, NULL
, NULL
);
8656 xa_for_each(&ctx
->personalities
, index
, creds
)
8657 io_unregister_personality(ctx
, index
);
8658 mutex_unlock(&ctx
->uring_lock
);
8660 io_kill_timeouts(ctx
, NULL
, NULL
);
8661 io_poll_remove_all(ctx
, NULL
, NULL
);
8663 /* if we failed setting up the ctx, we might not have any rings */
8664 io_iopoll_try_reap_events(ctx
);
8666 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8668 * Use system_unbound_wq to avoid spawning tons of event kworkers
8669 * if we're exiting a ton of rings at the same time. It just adds
8670 * noise and overhead, there's no discernable change in runtime
8671 * over using system_wq.
8673 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8676 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8678 struct io_ring_ctx
*ctx
= file
->private_data
;
8680 file
->private_data
= NULL
;
8681 io_ring_ctx_wait_and_kill(ctx
);
8685 struct io_task_cancel
{
8686 struct task_struct
*task
;
8687 struct files_struct
*files
;
8690 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8692 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8693 struct io_task_cancel
*cancel
= data
;
8696 if (cancel
->files
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8697 unsigned long flags
;
8698 struct io_ring_ctx
*ctx
= req
->ctx
;
8700 /* protect against races with linked timeouts */
8701 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8702 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8703 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8705 ret
= io_match_task(req
, cancel
->task
, cancel
->files
);
8710 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8711 struct task_struct
*task
,
8712 struct files_struct
*files
)
8714 struct io_defer_entry
*de
;
8717 spin_lock_irq(&ctx
->completion_lock
);
8718 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8719 if (io_match_task(de
->req
, task
, files
)) {
8720 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8724 spin_unlock_irq(&ctx
->completion_lock
);
8725 if (list_empty(&list
))
8728 while (!list_empty(&list
)) {
8729 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8730 list_del_init(&de
->list
);
8731 req_set_fail_links(de
->req
);
8732 io_put_req(de
->req
);
8733 io_req_complete(de
->req
, -ECANCELED
);
8739 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8741 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8743 return req
->ctx
== data
;
8746 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
8748 struct io_tctx_node
*node
;
8749 enum io_wq_cancel cret
;
8752 mutex_lock(&ctx
->uring_lock
);
8753 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
8754 struct io_uring_task
*tctx
= node
->task
->io_uring
;
8757 * io_wq will stay alive while we hold uring_lock, because it's
8758 * killed after ctx nodes, which requires to take the lock.
8760 if (!tctx
|| !tctx
->io_wq
)
8762 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8763 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8765 mutex_unlock(&ctx
->uring_lock
);
8770 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
8771 struct task_struct
*task
,
8772 struct files_struct
*files
)
8774 struct io_task_cancel cancel
= { .task
= task
, .files
= files
, };
8775 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
8778 enum io_wq_cancel cret
;
8782 ret
|= io_uring_try_cancel_iowq(ctx
);
8783 } else if (tctx
&& tctx
->io_wq
) {
8785 * Cancels requests of all rings, not only @ctx, but
8786 * it's fine as the task is in exit/exec.
8788 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
8790 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8793 /* SQPOLL thread does its own polling */
8794 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && !files
) ||
8795 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
8796 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8797 io_iopoll_try_reap_events(ctx
);
8802 ret
|= io_cancel_defer_files(ctx
, task
, files
);
8803 ret
|= io_poll_remove_all(ctx
, task
, files
);
8804 ret
|= io_kill_timeouts(ctx
, task
, files
);
8805 ret
|= io_run_task_work();
8806 ret
|= io_run_ctx_fallback(ctx
);
8807 io_cqring_overflow_flush(ctx
, true, task
, files
);
8814 static int io_uring_count_inflight(struct io_ring_ctx
*ctx
,
8815 struct task_struct
*task
,
8816 struct files_struct
*files
)
8818 struct io_kiocb
*req
;
8821 spin_lock_irq(&ctx
->inflight_lock
);
8822 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
)
8823 cnt
+= io_match_task(req
, task
, files
);
8824 spin_unlock_irq(&ctx
->inflight_lock
);
8828 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
8829 struct task_struct
*task
,
8830 struct files_struct
*files
)
8832 while (!list_empty_careful(&ctx
->inflight_list
)) {
8836 inflight
= io_uring_count_inflight(ctx
, task
, files
);
8840 io_uring_try_cancel_requests(ctx
, task
, files
);
8842 prepare_to_wait(&task
->io_uring
->wait
, &wait
,
8843 TASK_UNINTERRUPTIBLE
);
8844 if (inflight
== io_uring_count_inflight(ctx
, task
, files
))
8846 finish_wait(&task
->io_uring
->wait
, &wait
);
8851 * Note that this task has used io_uring. We use it for cancelation purposes.
8853 static int io_uring_add_task_file(struct io_ring_ctx
*ctx
)
8855 struct io_uring_task
*tctx
= current
->io_uring
;
8856 struct io_tctx_node
*node
;
8859 if (unlikely(!tctx
)) {
8860 ret
= io_uring_alloc_task_context(current
, ctx
);
8863 tctx
= current
->io_uring
;
8865 if (tctx
->last
!= ctx
) {
8866 void *old
= xa_load(&tctx
->xa
, (unsigned long)ctx
);
8869 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
8873 node
->task
= current
;
8875 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
8882 mutex_lock(&ctx
->uring_lock
);
8883 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
8884 mutex_unlock(&ctx
->uring_lock
);
8892 * Remove this io_uring_file -> task mapping.
8894 static void io_uring_del_task_file(unsigned long index
)
8896 struct io_uring_task
*tctx
= current
->io_uring
;
8897 struct io_tctx_node
*node
;
8901 node
= xa_erase(&tctx
->xa
, index
);
8905 WARN_ON_ONCE(current
!= node
->task
);
8906 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
8908 mutex_lock(&node
->ctx
->uring_lock
);
8909 list_del(&node
->ctx_node
);
8910 mutex_unlock(&node
->ctx
->uring_lock
);
8912 if (tctx
->last
== node
->ctx
)
8917 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
8919 struct io_tctx_node
*node
;
8920 unsigned long index
;
8922 xa_for_each(&tctx
->xa
, index
, node
)
8923 io_uring_del_task_file(index
);
8925 io_wq_put_and_exit(tctx
->io_wq
);
8930 static s64
tctx_inflight(struct io_uring_task
*tctx
)
8932 return percpu_counter_sum(&tctx
->inflight
);
8935 static void io_sqpoll_cancel_cb(struct callback_head
*cb
)
8937 struct io_tctx_exit
*work
= container_of(cb
, struct io_tctx_exit
, task_work
);
8938 struct io_ring_ctx
*ctx
= work
->ctx
;
8939 struct io_sq_data
*sqd
= ctx
->sq_data
;
8942 io_uring_cancel_sqpoll(ctx
);
8943 complete(&work
->completion
);
8946 static void io_sqpoll_cancel_sync(struct io_ring_ctx
*ctx
)
8948 struct io_sq_data
*sqd
= ctx
->sq_data
;
8949 struct io_tctx_exit work
= { .ctx
= ctx
, };
8950 struct task_struct
*task
;
8952 io_sq_thread_park(sqd
);
8953 list_del_init(&ctx
->sqd_list
);
8954 io_sqd_update_thread_idle(sqd
);
8957 init_completion(&work
.completion
);
8958 init_task_work(&work
.task_work
, io_sqpoll_cancel_cb
);
8959 io_task_work_add_head(&sqd
->park_task_work
, &work
.task_work
);
8960 wake_up_process(task
);
8962 io_sq_thread_unpark(sqd
);
8965 wait_for_completion(&work
.completion
);
8968 void __io_uring_files_cancel(struct files_struct
*files
)
8970 struct io_uring_task
*tctx
= current
->io_uring
;
8971 struct io_tctx_node
*node
;
8972 unsigned long index
;
8974 /* make sure overflow events are dropped */
8975 atomic_inc(&tctx
->in_idle
);
8976 xa_for_each(&tctx
->xa
, index
, node
) {
8977 struct io_ring_ctx
*ctx
= node
->ctx
;
8980 io_sqpoll_cancel_sync(ctx
);
8983 io_uring_cancel_files(ctx
, current
, files
);
8985 io_uring_try_cancel_requests(ctx
, current
, NULL
);
8987 atomic_dec(&tctx
->in_idle
);
8990 io_uring_clean_tctx(tctx
);
8993 /* should only be called by SQPOLL task */
8994 static void io_uring_cancel_sqpoll(struct io_ring_ctx
*ctx
)
8996 struct io_sq_data
*sqd
= ctx
->sq_data
;
8997 struct io_uring_task
*tctx
= current
->io_uring
;
9001 WARN_ON_ONCE(!sqd
|| ctx
->sq_data
->thread
!= current
);
9003 atomic_inc(&tctx
->in_idle
);
9005 /* read completions before cancelations */
9006 inflight
= tctx_inflight(tctx
);
9009 io_uring_try_cancel_requests(ctx
, current
, NULL
);
9011 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9013 * If we've seen completions, retry without waiting. This
9014 * avoids a race where a completion comes in before we did
9015 * prepare_to_wait().
9017 if (inflight
== tctx_inflight(tctx
))
9019 finish_wait(&tctx
->wait
, &wait
);
9021 atomic_dec(&tctx
->in_idle
);
9025 * Find any io_uring fd that this task has registered or done IO on, and cancel
9028 void __io_uring_task_cancel(void)
9030 struct io_uring_task
*tctx
= current
->io_uring
;
9034 /* make sure overflow events are dropped */
9035 atomic_inc(&tctx
->in_idle
);
9036 __io_uring_files_cancel(NULL
);
9039 /* read completions before cancelations */
9040 inflight
= tctx_inflight(tctx
);
9043 __io_uring_files_cancel(NULL
);
9045 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9048 * If we've seen completions, retry without waiting. This
9049 * avoids a race where a completion comes in before we did
9050 * prepare_to_wait().
9052 if (inflight
== tctx_inflight(tctx
))
9054 finish_wait(&tctx
->wait
, &wait
);
9057 atomic_dec(&tctx
->in_idle
);
9059 io_uring_clean_tctx(tctx
);
9060 /* all current's requests should be gone, we can kill tctx */
9061 __io_uring_free(current
);
9064 static void *io_uring_validate_mmap_request(struct file
*file
,
9065 loff_t pgoff
, size_t sz
)
9067 struct io_ring_ctx
*ctx
= file
->private_data
;
9068 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9073 case IORING_OFF_SQ_RING
:
9074 case IORING_OFF_CQ_RING
:
9077 case IORING_OFF_SQES
:
9081 return ERR_PTR(-EINVAL
);
9084 page
= virt_to_head_page(ptr
);
9085 if (sz
> page_size(page
))
9086 return ERR_PTR(-EINVAL
);
9093 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9095 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9099 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9101 return PTR_ERR(ptr
);
9103 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9104 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9107 #else /* !CONFIG_MMU */
9109 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9111 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9114 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9116 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9119 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9120 unsigned long addr
, unsigned long len
,
9121 unsigned long pgoff
, unsigned long flags
)
9125 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9127 return PTR_ERR(ptr
);
9129 return (unsigned long) ptr
;
9132 #endif /* !CONFIG_MMU */
9134 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9139 if (!io_sqring_full(ctx
))
9141 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9143 if (!io_sqring_full(ctx
))
9146 } while (!signal_pending(current
));
9148 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9152 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9153 struct __kernel_timespec __user
**ts
,
9154 const sigset_t __user
**sig
)
9156 struct io_uring_getevents_arg arg
;
9159 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9160 * is just a pointer to the sigset_t.
9162 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9163 *sig
= (const sigset_t __user
*) argp
;
9169 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9170 * timespec and sigset_t pointers if good.
9172 if (*argsz
!= sizeof(arg
))
9174 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9176 *sig
= u64_to_user_ptr(arg
.sigmask
);
9177 *argsz
= arg
.sigmask_sz
;
9178 *ts
= u64_to_user_ptr(arg
.ts
);
9182 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9183 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9186 struct io_ring_ctx
*ctx
;
9193 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9194 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
))
9202 if (f
.file
->f_op
!= &io_uring_fops
)
9206 ctx
= f
.file
->private_data
;
9207 if (!percpu_ref_tryget(&ctx
->refs
))
9211 if (ctx
->flags
& IORING_SETUP_R_DISABLED
)
9215 * For SQ polling, the thread will do all submissions and completions.
9216 * Just return the requested submit count, and wake the thread if
9220 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9221 io_cqring_overflow_flush(ctx
, false, NULL
, NULL
);
9224 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
9227 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9228 wake_up(&ctx
->sq_data
->wait
);
9229 if (flags
& IORING_ENTER_SQ_WAIT
) {
9230 ret
= io_sqpoll_wait_sq(ctx
);
9234 submitted
= to_submit
;
9235 } else if (to_submit
) {
9236 ret
= io_uring_add_task_file(ctx
);
9239 mutex_lock(&ctx
->uring_lock
);
9240 submitted
= io_submit_sqes(ctx
, to_submit
);
9241 mutex_unlock(&ctx
->uring_lock
);
9243 if (submitted
!= to_submit
)
9246 if (flags
& IORING_ENTER_GETEVENTS
) {
9247 const sigset_t __user
*sig
;
9248 struct __kernel_timespec __user
*ts
;
9250 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9254 min_complete
= min(min_complete
, ctx
->cq_entries
);
9257 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9258 * space applications don't need to do io completion events
9259 * polling again, they can rely on io_sq_thread to do polling
9260 * work, which can reduce cpu usage and uring_lock contention.
9262 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9263 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9264 ret
= io_iopoll_check(ctx
, min_complete
);
9266 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9271 percpu_ref_put(&ctx
->refs
);
9274 return submitted
? submitted
: ret
;
9277 #ifdef CONFIG_PROC_FS
9278 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
9279 const struct cred
*cred
)
9281 struct user_namespace
*uns
= seq_user_ns(m
);
9282 struct group_info
*gi
;
9287 seq_printf(m
, "%5d\n", id
);
9288 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9289 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9290 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9291 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9292 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9293 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9294 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9295 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9296 seq_puts(m
, "\n\tGroups:\t");
9297 gi
= cred
->group_info
;
9298 for (g
= 0; g
< gi
->ngroups
; g
++) {
9299 seq_put_decimal_ull(m
, g
? " " : "",
9300 from_kgid_munged(uns
, gi
->gid
[g
]));
9302 seq_puts(m
, "\n\tCapEff:\t");
9303 cap
= cred
->cap_effective
;
9304 CAP_FOR_EACH_U32(__capi
)
9305 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9310 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9312 struct io_sq_data
*sq
= NULL
;
9317 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9318 * since fdinfo case grabs it in the opposite direction of normal use
9319 * cases. If we fail to get the lock, we just don't iterate any
9320 * structures that could be going away outside the io_uring mutex.
9322 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9324 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9330 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9331 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9332 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9333 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9334 struct file
*f
= *io_fixed_file_slot(ctx
->file_data
, i
);
9337 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9339 seq_printf(m
, "%5u: <none>\n", i
);
9341 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9342 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9343 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
9345 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
9346 (unsigned int) buf
->len
);
9348 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
9349 unsigned long index
;
9350 const struct cred
*cred
;
9352 seq_printf(m
, "Personalities:\n");
9353 xa_for_each(&ctx
->personalities
, index
, cred
)
9354 io_uring_show_cred(m
, index
, cred
);
9356 seq_printf(m
, "PollList:\n");
9357 spin_lock_irq(&ctx
->completion_lock
);
9358 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9359 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9360 struct io_kiocb
*req
;
9362 hlist_for_each_entry(req
, list
, hash_node
)
9363 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9364 req
->task
->task_works
!= NULL
);
9366 spin_unlock_irq(&ctx
->completion_lock
);
9368 mutex_unlock(&ctx
->uring_lock
);
9371 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9373 struct io_ring_ctx
*ctx
= f
->private_data
;
9375 if (percpu_ref_tryget(&ctx
->refs
)) {
9376 __io_uring_show_fdinfo(ctx
, m
);
9377 percpu_ref_put(&ctx
->refs
);
9382 static const struct file_operations io_uring_fops
= {
9383 .release
= io_uring_release
,
9384 .mmap
= io_uring_mmap
,
9386 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9387 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9389 .poll
= io_uring_poll
,
9390 .fasync
= io_uring_fasync
,
9391 #ifdef CONFIG_PROC_FS
9392 .show_fdinfo
= io_uring_show_fdinfo
,
9396 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9397 struct io_uring_params
*p
)
9399 struct io_rings
*rings
;
9400 size_t size
, sq_array_offset
;
9402 /* make sure these are sane, as we already accounted them */
9403 ctx
->sq_entries
= p
->sq_entries
;
9404 ctx
->cq_entries
= p
->cq_entries
;
9406 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9407 if (size
== SIZE_MAX
)
9410 rings
= io_mem_alloc(size
);
9415 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9416 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9417 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9418 rings
->sq_ring_entries
= p
->sq_entries
;
9419 rings
->cq_ring_entries
= p
->cq_entries
;
9420 ctx
->sq_mask
= rings
->sq_ring_mask
;
9421 ctx
->cq_mask
= rings
->cq_ring_mask
;
9423 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9424 if (size
== SIZE_MAX
) {
9425 io_mem_free(ctx
->rings
);
9430 ctx
->sq_sqes
= io_mem_alloc(size
);
9431 if (!ctx
->sq_sqes
) {
9432 io_mem_free(ctx
->rings
);
9440 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9444 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9448 ret
= io_uring_add_task_file(ctx
);
9453 fd_install(fd
, file
);
9458 * Allocate an anonymous fd, this is what constitutes the application
9459 * visible backing of an io_uring instance. The application mmaps this
9460 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9461 * we have to tie this fd to a socket for file garbage collection purposes.
9463 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9466 #if defined(CONFIG_UNIX)
9469 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9472 return ERR_PTR(ret
);
9475 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9476 O_RDWR
| O_CLOEXEC
);
9477 #if defined(CONFIG_UNIX)
9479 sock_release(ctx
->ring_sock
);
9480 ctx
->ring_sock
= NULL
;
9482 ctx
->ring_sock
->file
= file
;
9488 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9489 struct io_uring_params __user
*params
)
9491 struct io_ring_ctx
*ctx
;
9497 if (entries
> IORING_MAX_ENTRIES
) {
9498 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9500 entries
= IORING_MAX_ENTRIES
;
9504 * Use twice as many entries for the CQ ring. It's possible for the
9505 * application to drive a higher depth than the size of the SQ ring,
9506 * since the sqes are only used at submission time. This allows for
9507 * some flexibility in overcommitting a bit. If the application has
9508 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9509 * of CQ ring entries manually.
9511 p
->sq_entries
= roundup_pow_of_two(entries
);
9512 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9514 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9515 * to a power-of-two, if it isn't already. We do NOT impose
9516 * any cq vs sq ring sizing.
9520 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9521 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9523 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9525 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9526 if (p
->cq_entries
< p
->sq_entries
)
9529 p
->cq_entries
= 2 * p
->sq_entries
;
9532 ctx
= io_ring_ctx_alloc(p
);
9535 ctx
->compat
= in_compat_syscall();
9536 if (!capable(CAP_IPC_LOCK
))
9537 ctx
->user
= get_uid(current_user());
9540 * This is just grabbed for accounting purposes. When a process exits,
9541 * the mm is exited and dropped before the files, hence we need to hang
9542 * on to this mm purely for the purposes of being able to unaccount
9543 * memory (locked/pinned vm). It's not used for anything else.
9545 mmgrab(current
->mm
);
9546 ctx
->mm_account
= current
->mm
;
9548 ret
= io_allocate_scq_urings(ctx
, p
);
9552 ret
= io_sq_offload_create(ctx
, p
);
9556 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9557 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9558 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9559 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9560 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9561 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9562 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9563 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9565 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9566 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9567 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9568 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9569 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9570 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9571 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9572 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9574 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9575 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9576 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9577 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9578 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
;
9580 if (copy_to_user(params
, p
, sizeof(*p
))) {
9585 file
= io_uring_get_file(ctx
);
9587 ret
= PTR_ERR(file
);
9592 * Install ring fd as the very last thing, so we don't risk someone
9593 * having closed it before we finish setup
9595 ret
= io_uring_install_fd(ctx
, file
);
9597 /* fput will clean it up */
9602 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9605 io_ring_ctx_wait_and_kill(ctx
);
9610 * Sets up an aio uring context, and returns the fd. Applications asks for a
9611 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9612 * params structure passed in.
9614 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9616 struct io_uring_params p
;
9619 if (copy_from_user(&p
, params
, sizeof(p
)))
9621 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9626 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9627 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9628 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9629 IORING_SETUP_R_DISABLED
))
9632 return io_uring_create(entries
, &p
, params
);
9635 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9636 struct io_uring_params __user
*, params
)
9638 return io_uring_setup(entries
, params
);
9641 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9643 struct io_uring_probe
*p
;
9647 size
= struct_size(p
, ops
, nr_args
);
9648 if (size
== SIZE_MAX
)
9650 p
= kzalloc(size
, GFP_KERNEL
);
9655 if (copy_from_user(p
, arg
, size
))
9658 if (memchr_inv(p
, 0, size
))
9661 p
->last_op
= IORING_OP_LAST
- 1;
9662 if (nr_args
> IORING_OP_LAST
)
9663 nr_args
= IORING_OP_LAST
;
9665 for (i
= 0; i
< nr_args
; i
++) {
9667 if (!io_op_defs
[i
].not_supported
)
9668 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9673 if (copy_to_user(arg
, p
, size
))
9680 static int io_register_personality(struct io_ring_ctx
*ctx
)
9682 const struct cred
*creds
;
9686 creds
= get_current_cred();
9688 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
9689 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
9696 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9697 unsigned int nr_args
)
9699 struct io_uring_restriction
*res
;
9703 /* Restrictions allowed only if rings started disabled */
9704 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9707 /* We allow only a single restrictions registration */
9708 if (ctx
->restrictions
.registered
)
9711 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9714 size
= array_size(nr_args
, sizeof(*res
));
9715 if (size
== SIZE_MAX
)
9718 res
= memdup_user(arg
, size
);
9720 return PTR_ERR(res
);
9724 for (i
= 0; i
< nr_args
; i
++) {
9725 switch (res
[i
].opcode
) {
9726 case IORING_RESTRICTION_REGISTER_OP
:
9727 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9732 __set_bit(res
[i
].register_op
,
9733 ctx
->restrictions
.register_op
);
9735 case IORING_RESTRICTION_SQE_OP
:
9736 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9741 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9743 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9744 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9746 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9747 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9756 /* Reset all restrictions if an error happened */
9758 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9760 ctx
->restrictions
.registered
= true;
9766 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
9768 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9771 if (ctx
->restrictions
.registered
)
9772 ctx
->restricted
= 1;
9774 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
9775 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
9776 wake_up(&ctx
->sq_data
->wait
);
9780 static bool io_register_op_must_quiesce(int op
)
9783 case IORING_UNREGISTER_FILES
:
9784 case IORING_REGISTER_FILES_UPDATE
:
9785 case IORING_REGISTER_PROBE
:
9786 case IORING_REGISTER_PERSONALITY
:
9787 case IORING_UNREGISTER_PERSONALITY
:
9794 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
9795 void __user
*arg
, unsigned nr_args
)
9796 __releases(ctx
->uring_lock
)
9797 __acquires(ctx
->uring_lock
)
9802 * We're inside the ring mutex, if the ref is already dying, then
9803 * someone else killed the ctx or is already going through
9804 * io_uring_register().
9806 if (percpu_ref_is_dying(&ctx
->refs
))
9809 if (io_register_op_must_quiesce(opcode
)) {
9810 percpu_ref_kill(&ctx
->refs
);
9813 * Drop uring mutex before waiting for references to exit. If
9814 * another thread is currently inside io_uring_enter() it might
9815 * need to grab the uring_lock to make progress. If we hold it
9816 * here across the drain wait, then we can deadlock. It's safe
9817 * to drop the mutex here, since no new references will come in
9818 * after we've killed the percpu ref.
9820 mutex_unlock(&ctx
->uring_lock
);
9822 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
9825 ret
= io_run_task_work_sig();
9830 mutex_lock(&ctx
->uring_lock
);
9833 percpu_ref_resurrect(&ctx
->refs
);
9838 if (ctx
->restricted
) {
9839 if (opcode
>= IORING_REGISTER_LAST
) {
9844 if (!test_bit(opcode
, ctx
->restrictions
.register_op
)) {
9851 case IORING_REGISTER_BUFFERS
:
9852 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
);
9854 case IORING_UNREGISTER_BUFFERS
:
9858 ret
= io_sqe_buffers_unregister(ctx
);
9860 case IORING_REGISTER_FILES
:
9861 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
9863 case IORING_UNREGISTER_FILES
:
9867 ret
= io_sqe_files_unregister(ctx
);
9869 case IORING_REGISTER_FILES_UPDATE
:
9870 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
9872 case IORING_REGISTER_EVENTFD
:
9873 case IORING_REGISTER_EVENTFD_ASYNC
:
9877 ret
= io_eventfd_register(ctx
, arg
);
9880 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
9881 ctx
->eventfd_async
= 1;
9883 ctx
->eventfd_async
= 0;
9885 case IORING_UNREGISTER_EVENTFD
:
9889 ret
= io_eventfd_unregister(ctx
);
9891 case IORING_REGISTER_PROBE
:
9893 if (!arg
|| nr_args
> 256)
9895 ret
= io_probe(ctx
, arg
, nr_args
);
9897 case IORING_REGISTER_PERSONALITY
:
9901 ret
= io_register_personality(ctx
);
9903 case IORING_UNREGISTER_PERSONALITY
:
9907 ret
= io_unregister_personality(ctx
, nr_args
);
9909 case IORING_REGISTER_ENABLE_RINGS
:
9913 ret
= io_register_enable_rings(ctx
);
9915 case IORING_REGISTER_RESTRICTIONS
:
9916 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
9924 if (io_register_op_must_quiesce(opcode
)) {
9925 /* bring the ctx back to life */
9926 percpu_ref_reinit(&ctx
->refs
);
9928 reinit_completion(&ctx
->ref_comp
);
9933 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
9934 void __user
*, arg
, unsigned int, nr_args
)
9936 struct io_ring_ctx
*ctx
;
9945 if (f
.file
->f_op
!= &io_uring_fops
)
9948 ctx
= f
.file
->private_data
;
9952 mutex_lock(&ctx
->uring_lock
);
9953 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
9954 mutex_unlock(&ctx
->uring_lock
);
9955 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
9956 ctx
->cq_ev_fd
!= NULL
, ret
);
9962 static int __init
io_uring_init(void)
9964 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9965 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9966 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9969 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9970 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9971 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
9972 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
9973 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
9974 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
9975 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
9976 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
9977 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
9978 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
9979 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
9980 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
9981 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
9982 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
9983 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
9984 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
9985 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
9986 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
9987 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
9988 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
9989 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
9990 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
9991 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
9992 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
9993 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
9994 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
9995 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
9996 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
9997 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
9998 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
9999 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10001 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10002 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
10003 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
10007 __initcall(io_uring_init
);