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_cqe (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)
92 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define IO_RSRC_TAG_TABLE_SHIFT 9
105 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
106 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
108 #define IORING_MAX_REG_BUFFERS (1U << 14)
110 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
111 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
119 u32 head ____cacheline_aligned_in_smp
;
120 u32 tail ____cacheline_aligned_in_smp
;
124 * This data is shared with the application through the mmap at offsets
125 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
127 * The offsets to the member fields are published through struct
128 * io_sqring_offsets when calling io_uring_setup.
132 * Head and tail offsets into the ring; the offsets need to be
133 * masked to get valid indices.
135 * The kernel controls head of the sq ring and the tail of the cq ring,
136 * and the application controls tail of the sq ring and the head of the
139 struct io_uring sq
, cq
;
141 * Bitmasks to apply to head and tail offsets (constant, equals
144 u32 sq_ring_mask
, cq_ring_mask
;
145 /* Ring sizes (constant, power of 2) */
146 u32 sq_ring_entries
, cq_ring_entries
;
148 * Number of invalid entries dropped by the kernel due to
149 * invalid index stored in array
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * After a new SQ head value was read by the application this
156 * counter includes all submissions that were dropped reaching
157 * the new SQ head (and possibly more).
163 * Written by the kernel, shouldn't be modified by the
166 * The application needs a full memory barrier before checking
167 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
173 * Written by the application, shouldn't be modified by the
178 * Number of completion events lost because the queue was full;
179 * this should be avoided by the application by making sure
180 * there are not more requests pending than there is space in
181 * the completion queue.
183 * Written by the kernel, shouldn't be modified by the
184 * application (i.e. get number of "new events" by comparing to
187 * As completion events come in out of order this counter is not
188 * ordered with any other data.
192 * Ring buffer of completion events.
194 * The kernel writes completion events fresh every time they are
195 * produced, so the application is allowed to modify pending
198 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
201 enum io_uring_cmd_flags
{
202 IO_URING_F_NONBLOCK
= 1,
203 IO_URING_F_COMPLETE_DEFER
= 2,
206 struct io_mapped_ubuf
{
209 unsigned int nr_bvecs
;
210 unsigned long acct_pages
;
211 struct bio_vec bvec
[];
216 struct io_overflow_cqe
{
217 struct io_uring_cqe cqe
;
218 struct list_head list
;
221 struct io_fixed_file
{
222 /* file * with additional FFS_* flags */
223 unsigned long file_ptr
;
227 struct list_head list
;
232 struct io_mapped_ubuf
*buf
;
236 struct io_file_table
{
237 /* two level table */
238 struct io_fixed_file
**files
;
241 struct io_rsrc_node
{
242 struct percpu_ref refs
;
243 struct list_head node
;
244 struct list_head rsrc_list
;
245 struct io_rsrc_data
*rsrc_data
;
246 struct llist_node llist
;
250 typedef void (rsrc_put_fn
)(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
);
252 struct io_rsrc_data
{
253 struct io_ring_ctx
*ctx
;
259 struct completion done
;
264 struct list_head list
;
270 struct io_restriction
{
271 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
272 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
273 u8 sqe_flags_allowed
;
274 u8 sqe_flags_required
;
279 IO_SQ_THREAD_SHOULD_STOP
= 0,
280 IO_SQ_THREAD_SHOULD_PARK
,
285 atomic_t park_pending
;
288 /* ctx's that are using this sqd */
289 struct list_head ctx_list
;
291 struct task_struct
*thread
;
292 struct wait_queue_head wait
;
294 unsigned sq_thread_idle
;
300 struct completion exited
;
303 #define IO_IOPOLL_BATCH 8
304 #define IO_COMPL_BATCH 32
305 #define IO_REQ_CACHE_SIZE 32
306 #define IO_REQ_ALLOC_BATCH 8
308 struct io_comp_state
{
309 struct io_kiocb
*reqs
[IO_COMPL_BATCH
];
311 /* inline/task_work completion list, under ->uring_lock */
312 struct list_head free_list
;
315 struct io_submit_link
{
316 struct io_kiocb
*head
;
317 struct io_kiocb
*last
;
320 struct io_submit_state
{
321 struct blk_plug plug
;
322 struct io_submit_link link
;
325 * io_kiocb alloc cache
327 void *reqs
[IO_REQ_CACHE_SIZE
];
328 unsigned int free_reqs
;
333 * Batch completion logic
335 struct io_comp_state comp
;
338 * File reference cache
342 unsigned int file_refs
;
343 unsigned int ios_left
;
347 /* const or read-mostly hot data */
349 struct percpu_ref refs
;
351 struct io_rings
*rings
;
353 unsigned int compat
: 1;
354 unsigned int drain_next
: 1;
355 unsigned int eventfd_async
: 1;
356 unsigned int restricted
: 1;
357 unsigned int off_timeout_used
: 1;
358 unsigned int drain_active
: 1;
359 } ____cacheline_aligned_in_smp
;
361 /* submission data */
363 struct mutex uring_lock
;
366 * Ring buffer of indices into array of io_uring_sqe, which is
367 * mmapped by the application using the IORING_OFF_SQES offset.
369 * This indirection could e.g. be used to assign fixed
370 * io_uring_sqe entries to operations and only submit them to
371 * the queue when needed.
373 * The kernel modifies neither the indices array nor the entries
377 struct io_uring_sqe
*sq_sqes
;
378 unsigned cached_sq_head
;
380 struct list_head defer_list
;
383 * Fixed resources fast path, should be accessed only under
384 * uring_lock, and updated through io_uring_register(2)
386 struct io_rsrc_node
*rsrc_node
;
387 struct io_file_table file_table
;
388 unsigned nr_user_files
;
389 unsigned nr_user_bufs
;
390 struct io_mapped_ubuf
**user_bufs
;
392 struct io_submit_state submit_state
;
393 struct list_head timeout_list
;
394 struct list_head cq_overflow_list
;
395 struct xarray io_buffers
;
396 struct xarray personalities
;
398 unsigned sq_thread_idle
;
399 } ____cacheline_aligned_in_smp
;
401 /* IRQ completion list, under ->completion_lock */
402 struct list_head locked_free_list
;
403 unsigned int locked_free_nr
;
405 const struct cred
*sq_creds
; /* cred used for __io_sq_thread() */
406 struct io_sq_data
*sq_data
; /* if using sq thread polling */
408 struct wait_queue_head sqo_sq_wait
;
409 struct list_head sqd_list
;
411 unsigned long check_cq_overflow
;
414 unsigned cached_cq_tail
;
416 struct eventfd_ctx
*cq_ev_fd
;
417 struct wait_queue_head poll_wait
;
418 struct wait_queue_head cq_wait
;
420 atomic_t cq_timeouts
;
421 struct fasync_struct
*cq_fasync
;
422 unsigned cq_last_tm_flush
;
423 } ____cacheline_aligned_in_smp
;
426 spinlock_t completion_lock
;
429 * ->iopoll_list is protected by the ctx->uring_lock for
430 * io_uring instances that don't use IORING_SETUP_SQPOLL.
431 * For SQPOLL, only the single threaded io_sq_thread() will
432 * manipulate the list, hence no extra locking is needed there.
434 struct list_head iopoll_list
;
435 struct hlist_head
*cancel_hash
;
436 unsigned cancel_hash_bits
;
437 bool poll_multi_queue
;
438 } ____cacheline_aligned_in_smp
;
440 struct io_restriction restrictions
;
442 /* slow path rsrc auxilary data, used by update/register */
444 struct io_rsrc_node
*rsrc_backup_node
;
445 struct io_mapped_ubuf
*dummy_ubuf
;
446 struct io_rsrc_data
*file_data
;
447 struct io_rsrc_data
*buf_data
;
449 struct delayed_work rsrc_put_work
;
450 struct llist_head rsrc_put_llist
;
451 struct list_head rsrc_ref_list
;
452 spinlock_t rsrc_ref_lock
;
455 /* Keep this last, we don't need it for the fast path */
457 #if defined(CONFIG_UNIX)
458 struct socket
*ring_sock
;
460 /* hashed buffered write serialization */
461 struct io_wq_hash
*hash_map
;
463 /* Only used for accounting purposes */
464 struct user_struct
*user
;
465 struct mm_struct
*mm_account
;
467 /* ctx exit and cancelation */
468 struct llist_head fallback_llist
;
469 struct delayed_work fallback_work
;
470 struct work_struct exit_work
;
471 struct list_head tctx_list
;
472 struct completion ref_comp
;
476 struct io_uring_task
{
477 /* submission side */
480 struct wait_queue_head wait
;
481 const struct io_ring_ctx
*last
;
483 struct percpu_counter inflight
;
484 atomic_t inflight_tracked
;
487 spinlock_t task_lock
;
488 struct io_wq_work_list task_list
;
489 unsigned long task_state
;
490 struct callback_head task_work
;
494 * First field must be the file pointer in all the
495 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
497 struct io_poll_iocb
{
499 struct wait_queue_head
*head
;
503 struct wait_queue_entry wait
;
506 struct io_poll_update
{
512 bool update_user_data
;
520 struct io_timeout_data
{
521 struct io_kiocb
*req
;
522 struct hrtimer timer
;
523 struct timespec64 ts
;
524 enum hrtimer_mode mode
;
529 struct sockaddr __user
*addr
;
530 int __user
*addr_len
;
532 unsigned long nofile
;
552 struct list_head list
;
553 /* head of the link, used by linked timeouts only */
554 struct io_kiocb
*head
;
557 struct io_timeout_rem
{
562 struct timespec64 ts
;
567 /* NOTE: kiocb has the file as the first member, so don't do it here */
575 struct sockaddr __user
*addr
;
582 struct compat_msghdr __user
*umsg_compat
;
583 struct user_msghdr __user
*umsg
;
589 struct io_buffer
*kbuf
;
595 struct filename
*filename
;
597 unsigned long nofile
;
600 struct io_rsrc_update
{
626 struct epoll_event event
;
630 struct file
*file_out
;
631 struct file
*file_in
;
638 struct io_provide_buf
{
652 const char __user
*filename
;
653 struct statx __user
*buffer
;
665 struct filename
*oldpath
;
666 struct filename
*newpath
;
674 struct filename
*filename
;
677 struct io_completion
{
679 struct list_head list
;
683 struct io_async_connect
{
684 struct sockaddr_storage address
;
687 struct io_async_msghdr
{
688 struct iovec fast_iov
[UIO_FASTIOV
];
689 /* points to an allocated iov, if NULL we use fast_iov instead */
690 struct iovec
*free_iov
;
691 struct sockaddr __user
*uaddr
;
693 struct sockaddr_storage addr
;
697 struct iovec fast_iov
[UIO_FASTIOV
];
698 const struct iovec
*free_iovec
;
699 struct iov_iter iter
;
701 struct wait_page_queue wpq
;
705 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
706 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
707 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
708 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
709 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
710 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
712 /* first byte is taken by user flags, shift it to not overlap */
717 REQ_F_LINK_TIMEOUT_BIT
,
718 REQ_F_NEED_CLEANUP_BIT
,
720 REQ_F_BUFFER_SELECTED_BIT
,
721 REQ_F_LTIMEOUT_ACTIVE_BIT
,
722 REQ_F_COMPLETE_INLINE_BIT
,
724 REQ_F_DONT_REISSUE_BIT
,
726 /* keep async read/write and isreg together and in order */
727 REQ_F_ASYNC_READ_BIT
,
728 REQ_F_ASYNC_WRITE_BIT
,
731 /* not a real bit, just to check we're not overflowing the space */
737 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
738 /* drain existing IO first */
739 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
741 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
742 /* doesn't sever on completion < 0 */
743 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
745 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
746 /* IOSQE_BUFFER_SELECT */
747 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
749 /* fail rest of links */
750 REQ_F_FAIL
= BIT(REQ_F_FAIL_BIT
),
751 /* on inflight list, should be cancelled and waited on exit reliably */
752 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
753 /* read/write uses file position */
754 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
755 /* must not punt to workers */
756 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
757 /* has or had linked timeout */
758 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
760 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
761 /* already went through poll handler */
762 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
763 /* buffer already selected */
764 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
765 /* linked timeout is active, i.e. prepared by link's head */
766 REQ_F_LTIMEOUT_ACTIVE
= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT
),
767 /* completion is deferred through io_comp_state */
768 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
769 /* caller should reissue async */
770 REQ_F_REISSUE
= BIT(REQ_F_REISSUE_BIT
),
771 /* don't attempt request reissue, see io_rw_reissue() */
772 REQ_F_DONT_REISSUE
= BIT(REQ_F_DONT_REISSUE_BIT
),
773 /* supports async reads */
774 REQ_F_ASYNC_READ
= BIT(REQ_F_ASYNC_READ_BIT
),
775 /* supports async writes */
776 REQ_F_ASYNC_WRITE
= BIT(REQ_F_ASYNC_WRITE_BIT
),
778 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
779 /* has creds assigned */
780 REQ_F_CREDS
= BIT(REQ_F_CREDS_BIT
),
784 struct io_poll_iocb poll
;
785 struct io_poll_iocb
*double_poll
;
788 typedef void (*io_req_tw_func_t
)(struct io_kiocb
*req
);
790 struct io_task_work
{
792 struct io_wq_work_node node
;
793 struct llist_node fallback_node
;
795 io_req_tw_func_t func
;
799 IORING_RSRC_FILE
= 0,
800 IORING_RSRC_BUFFER
= 1,
804 * NOTE! Each of the iocb union members has the file pointer
805 * as the first entry in their struct definition. So you can
806 * access the file pointer through any of the sub-structs,
807 * or directly as just 'ki_filp' in this struct.
813 struct io_poll_iocb poll
;
814 struct io_poll_update poll_update
;
815 struct io_accept accept
;
817 struct io_cancel cancel
;
818 struct io_timeout timeout
;
819 struct io_timeout_rem timeout_rem
;
820 struct io_connect connect
;
821 struct io_sr_msg sr_msg
;
823 struct io_close close
;
824 struct io_rsrc_update rsrc_update
;
825 struct io_fadvise fadvise
;
826 struct io_madvise madvise
;
827 struct io_epoll epoll
;
828 struct io_splice splice
;
829 struct io_provide_buf pbuf
;
830 struct io_statx statx
;
831 struct io_shutdown shutdown
;
832 struct io_rename rename
;
833 struct io_unlink unlink
;
834 /* use only after cleaning per-op data, see io_clean_op() */
835 struct io_completion
compl;
838 /* opcode allocated if it needs to store data for async defer */
841 /* polled IO has completed */
847 struct io_ring_ctx
*ctx
;
850 struct task_struct
*task
;
853 struct io_kiocb
*link
;
854 struct percpu_ref
*fixed_rsrc_refs
;
856 /* used with ctx->iopoll_list with reads/writes */
857 struct list_head inflight_entry
;
858 struct io_task_work io_task_work
;
859 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
860 struct hlist_node hash_node
;
861 struct async_poll
*apoll
;
862 struct io_wq_work work
;
863 const struct cred
*creds
;
865 /* store used ubuf, so we can prevent reloading */
866 struct io_mapped_ubuf
*imu
;
869 struct io_tctx_node
{
870 struct list_head ctx_node
;
871 struct task_struct
*task
;
872 struct io_ring_ctx
*ctx
;
875 struct io_defer_entry
{
876 struct list_head list
;
877 struct io_kiocb
*req
;
882 /* needs req->file assigned */
883 unsigned needs_file
: 1;
884 /* hash wq insertion if file is a regular file */
885 unsigned hash_reg_file
: 1;
886 /* unbound wq insertion if file is a non-regular file */
887 unsigned unbound_nonreg_file
: 1;
888 /* opcode is not supported by this kernel */
889 unsigned not_supported
: 1;
890 /* set if opcode supports polled "wait" */
892 unsigned pollout
: 1;
893 /* op supports buffer selection */
894 unsigned buffer_select
: 1;
895 /* do prep async if is going to be punted */
896 unsigned needs_async_setup
: 1;
897 /* should block plug */
899 /* size of async data needed, if any */
900 unsigned short async_size
;
903 static const struct io_op_def io_op_defs
[] = {
904 [IORING_OP_NOP
] = {},
905 [IORING_OP_READV
] = {
907 .unbound_nonreg_file
= 1,
910 .needs_async_setup
= 1,
912 .async_size
= sizeof(struct io_async_rw
),
914 [IORING_OP_WRITEV
] = {
917 .unbound_nonreg_file
= 1,
919 .needs_async_setup
= 1,
921 .async_size
= sizeof(struct io_async_rw
),
923 [IORING_OP_FSYNC
] = {
926 [IORING_OP_READ_FIXED
] = {
928 .unbound_nonreg_file
= 1,
931 .async_size
= sizeof(struct io_async_rw
),
933 [IORING_OP_WRITE_FIXED
] = {
936 .unbound_nonreg_file
= 1,
939 .async_size
= sizeof(struct io_async_rw
),
941 [IORING_OP_POLL_ADD
] = {
943 .unbound_nonreg_file
= 1,
945 [IORING_OP_POLL_REMOVE
] = {},
946 [IORING_OP_SYNC_FILE_RANGE
] = {
949 [IORING_OP_SENDMSG
] = {
951 .unbound_nonreg_file
= 1,
953 .needs_async_setup
= 1,
954 .async_size
= sizeof(struct io_async_msghdr
),
956 [IORING_OP_RECVMSG
] = {
958 .unbound_nonreg_file
= 1,
961 .needs_async_setup
= 1,
962 .async_size
= sizeof(struct io_async_msghdr
),
964 [IORING_OP_TIMEOUT
] = {
965 .async_size
= sizeof(struct io_timeout_data
),
967 [IORING_OP_TIMEOUT_REMOVE
] = {
968 /* used by timeout updates' prep() */
970 [IORING_OP_ACCEPT
] = {
972 .unbound_nonreg_file
= 1,
975 [IORING_OP_ASYNC_CANCEL
] = {},
976 [IORING_OP_LINK_TIMEOUT
] = {
977 .async_size
= sizeof(struct io_timeout_data
),
979 [IORING_OP_CONNECT
] = {
981 .unbound_nonreg_file
= 1,
983 .needs_async_setup
= 1,
984 .async_size
= sizeof(struct io_async_connect
),
986 [IORING_OP_FALLOCATE
] = {
989 [IORING_OP_OPENAT
] = {},
990 [IORING_OP_CLOSE
] = {},
991 [IORING_OP_FILES_UPDATE
] = {},
992 [IORING_OP_STATX
] = {},
995 .unbound_nonreg_file
= 1,
999 .async_size
= sizeof(struct io_async_rw
),
1001 [IORING_OP_WRITE
] = {
1003 .unbound_nonreg_file
= 1,
1006 .async_size
= sizeof(struct io_async_rw
),
1008 [IORING_OP_FADVISE
] = {
1011 [IORING_OP_MADVISE
] = {},
1012 [IORING_OP_SEND
] = {
1014 .unbound_nonreg_file
= 1,
1017 [IORING_OP_RECV
] = {
1019 .unbound_nonreg_file
= 1,
1023 [IORING_OP_OPENAT2
] = {
1025 [IORING_OP_EPOLL_CTL
] = {
1026 .unbound_nonreg_file
= 1,
1028 [IORING_OP_SPLICE
] = {
1031 .unbound_nonreg_file
= 1,
1033 [IORING_OP_PROVIDE_BUFFERS
] = {},
1034 [IORING_OP_REMOVE_BUFFERS
] = {},
1038 .unbound_nonreg_file
= 1,
1040 [IORING_OP_SHUTDOWN
] = {
1043 [IORING_OP_RENAMEAT
] = {},
1044 [IORING_OP_UNLINKAT
] = {},
1047 static bool io_disarm_next(struct io_kiocb
*req
);
1048 static void io_uring_del_tctx_node(unsigned long index
);
1049 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1050 struct task_struct
*task
,
1052 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
);
1053 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
);
1055 static bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1056 long res
, unsigned int cflags
);
1057 static void io_put_req(struct io_kiocb
*req
);
1058 static void io_put_req_deferred(struct io_kiocb
*req
, int nr
);
1059 static void io_dismantle_req(struct io_kiocb
*req
);
1060 static void io_put_task(struct task_struct
*task
, int nr
);
1061 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
1062 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1063 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
1064 struct io_uring_rsrc_update2
*up
,
1066 static void io_clean_op(struct io_kiocb
*req
);
1067 static struct file
*io_file_get(struct io_submit_state
*state
,
1068 struct io_kiocb
*req
, int fd
, bool fixed
);
1069 static void __io_queue_sqe(struct io_kiocb
*req
);
1070 static void io_rsrc_put_work(struct work_struct
*work
);
1072 static void io_req_task_queue(struct io_kiocb
*req
);
1073 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
);
1074 static bool io_poll_remove_waitqs(struct io_kiocb
*req
);
1075 static int io_req_prep_async(struct io_kiocb
*req
);
1077 static void io_fallback_req_func(struct work_struct
*unused
);
1079 static struct kmem_cache
*req_cachep
;
1081 static const struct file_operations io_uring_fops
;
1083 struct sock
*io_uring_get_socket(struct file
*file
)
1085 #if defined(CONFIG_UNIX)
1086 if (file
->f_op
== &io_uring_fops
) {
1087 struct io_ring_ctx
*ctx
= file
->private_data
;
1089 return ctx
->ring_sock
->sk
;
1094 EXPORT_SYMBOL(io_uring_get_socket
);
1096 #define io_for_each_link(pos, head) \
1097 for (pos = (head); pos; pos = pos->link)
1099 static inline void io_req_set_rsrc_node(struct io_kiocb
*req
)
1101 struct io_ring_ctx
*ctx
= req
->ctx
;
1103 if (!req
->fixed_rsrc_refs
) {
1104 req
->fixed_rsrc_refs
= &ctx
->rsrc_node
->refs
;
1105 percpu_ref_get(req
->fixed_rsrc_refs
);
1109 static void io_refs_resurrect(struct percpu_ref
*ref
, struct completion
*compl)
1111 bool got
= percpu_ref_tryget(ref
);
1113 /* already at zero, wait for ->release() */
1115 wait_for_completion(compl);
1116 percpu_ref_resurrect(ref
);
1118 percpu_ref_put(ref
);
1121 static bool io_match_task(struct io_kiocb
*head
, struct task_struct
*task
,
1124 struct io_kiocb
*req
;
1126 if (task
&& head
->task
!= task
)
1131 io_for_each_link(req
, head
) {
1132 if (req
->flags
& REQ_F_INFLIGHT
)
1138 static inline void req_set_fail(struct io_kiocb
*req
)
1140 req
->flags
|= REQ_F_FAIL
;
1143 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1145 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1147 complete(&ctx
->ref_comp
);
1150 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1152 return !req
->timeout
.off
;
1155 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1157 struct io_ring_ctx
*ctx
;
1160 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1165 * Use 5 bits less than the max cq entries, that should give us around
1166 * 32 entries per hash list if totally full and uniformly spread.
1168 hash_bits
= ilog2(p
->cq_entries
);
1172 ctx
->cancel_hash_bits
= hash_bits
;
1173 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1175 if (!ctx
->cancel_hash
)
1177 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1179 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
1180 if (!ctx
->dummy_ubuf
)
1182 /* set invalid range, so io_import_fixed() fails meeting it */
1183 ctx
->dummy_ubuf
->ubuf
= -1UL;
1185 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1186 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1189 ctx
->flags
= p
->flags
;
1190 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1191 INIT_LIST_HEAD(&ctx
->sqd_list
);
1192 init_waitqueue_head(&ctx
->poll_wait
);
1193 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1194 init_completion(&ctx
->ref_comp
);
1195 xa_init_flags(&ctx
->io_buffers
, XA_FLAGS_ALLOC1
);
1196 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1197 mutex_init(&ctx
->uring_lock
);
1198 init_waitqueue_head(&ctx
->cq_wait
);
1199 spin_lock_init(&ctx
->completion_lock
);
1200 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1201 INIT_LIST_HEAD(&ctx
->defer_list
);
1202 INIT_LIST_HEAD(&ctx
->timeout_list
);
1203 spin_lock_init(&ctx
->rsrc_ref_lock
);
1204 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1205 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1206 init_llist_head(&ctx
->rsrc_put_llist
);
1207 INIT_LIST_HEAD(&ctx
->tctx_list
);
1208 INIT_LIST_HEAD(&ctx
->submit_state
.comp
.free_list
);
1209 INIT_LIST_HEAD(&ctx
->locked_free_list
);
1210 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
1213 kfree(ctx
->dummy_ubuf
);
1214 kfree(ctx
->cancel_hash
);
1219 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
1221 struct io_rings
*r
= ctx
->rings
;
1223 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
1227 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1229 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1230 struct io_ring_ctx
*ctx
= req
->ctx
;
1232 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
1238 static void io_req_track_inflight(struct io_kiocb
*req
)
1240 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1241 req
->flags
|= REQ_F_INFLIGHT
;
1242 atomic_inc(¤t
->io_uring
->inflight_tracked
);
1246 static void io_prep_async_work(struct io_kiocb
*req
)
1248 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1249 struct io_ring_ctx
*ctx
= req
->ctx
;
1251 if (!(req
->flags
& REQ_F_CREDS
)) {
1252 req
->flags
|= REQ_F_CREDS
;
1253 req
->creds
= get_current_cred();
1256 req
->work
.list
.next
= NULL
;
1257 req
->work
.flags
= 0;
1258 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1259 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1261 if (req
->flags
& REQ_F_ISREG
) {
1262 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1263 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1264 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
1265 if (def
->unbound_nonreg_file
)
1266 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1269 switch (req
->opcode
) {
1270 case IORING_OP_SPLICE
:
1272 if (!S_ISREG(file_inode(req
->splice
.file_in
)->i_mode
))
1273 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1278 static void io_prep_async_link(struct io_kiocb
*req
)
1280 struct io_kiocb
*cur
;
1282 io_for_each_link(cur
, req
)
1283 io_prep_async_work(cur
);
1286 static void io_queue_async_work(struct io_kiocb
*req
)
1288 struct io_ring_ctx
*ctx
= req
->ctx
;
1289 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1290 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1293 BUG_ON(!tctx
->io_wq
);
1295 /* init ->work of the whole link before punting */
1296 io_prep_async_link(req
);
1297 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1298 &req
->work
, req
->flags
);
1299 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
1301 io_queue_linked_timeout(link
);
1304 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1305 __must_hold(&req
->ctx
->completion_lock
)
1307 struct io_timeout_data
*io
= req
->async_data
;
1309 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1310 atomic_set(&req
->ctx
->cq_timeouts
,
1311 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1312 list_del_init(&req
->timeout
.list
);
1313 io_cqring_fill_event(req
->ctx
, req
->user_data
, status
, 0);
1314 io_put_req_deferred(req
, 1);
1318 static void io_queue_deferred(struct io_ring_ctx
*ctx
)
1320 while (!list_empty(&ctx
->defer_list
)) {
1321 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1322 struct io_defer_entry
, list
);
1324 if (req_need_defer(de
->req
, de
->seq
))
1326 list_del_init(&de
->list
);
1327 io_req_task_queue(de
->req
);
1332 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1334 u32 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1336 while (!list_empty(&ctx
->timeout_list
)) {
1337 u32 events_needed
, events_got
;
1338 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1339 struct io_kiocb
, timeout
.list
);
1341 if (io_is_timeout_noseq(req
))
1345 * Since seq can easily wrap around over time, subtract
1346 * the last seq at which timeouts were flushed before comparing.
1347 * Assuming not more than 2^31-1 events have happened since,
1348 * these subtractions won't have wrapped, so we can check if
1349 * target is in [last_seq, current_seq] by comparing the two.
1351 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1352 events_got
= seq
- ctx
->cq_last_tm_flush
;
1353 if (events_got
< events_needed
)
1356 list_del_init(&req
->timeout
.list
);
1357 io_kill_timeout(req
, 0);
1359 ctx
->cq_last_tm_flush
= seq
;
1362 static void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
1364 if (ctx
->off_timeout_used
)
1365 io_flush_timeouts(ctx
);
1366 if (ctx
->drain_active
)
1367 io_queue_deferred(ctx
);
1370 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
1372 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
))
1373 __io_commit_cqring_flush(ctx
);
1374 /* order cqe stores with ring update */
1375 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1378 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1380 struct io_rings
*r
= ctx
->rings
;
1382 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
1385 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1387 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1390 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
1392 struct io_rings
*rings
= ctx
->rings
;
1393 unsigned tail
, mask
= ctx
->cq_entries
- 1;
1396 * writes to the cq entry need to come after reading head; the
1397 * control dependency is enough as we're using WRITE_ONCE to
1400 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
1403 tail
= ctx
->cached_cq_tail
++;
1404 return &rings
->cqes
[tail
& mask
];
1407 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1409 if (likely(!ctx
->cq_ev_fd
))
1411 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1413 return !ctx
->eventfd_async
|| io_wq_current_is_worker();
1416 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1418 /* see waitqueue_active() comment */
1421 if (waitqueue_active(&ctx
->cq_wait
))
1422 wake_up(&ctx
->cq_wait
);
1423 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1424 wake_up(&ctx
->sq_data
->wait
);
1425 if (io_should_trigger_evfd(ctx
))
1426 eventfd_signal(ctx
->cq_ev_fd
, 1);
1427 if (waitqueue_active(&ctx
->poll_wait
)) {
1428 wake_up_interruptible(&ctx
->poll_wait
);
1429 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1433 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1435 /* see waitqueue_active() comment */
1438 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1439 if (waitqueue_active(&ctx
->cq_wait
))
1440 wake_up(&ctx
->cq_wait
);
1442 if (io_should_trigger_evfd(ctx
))
1443 eventfd_signal(ctx
->cq_ev_fd
, 1);
1444 if (waitqueue_active(&ctx
->poll_wait
)) {
1445 wake_up_interruptible(&ctx
->poll_wait
);
1446 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1450 /* Returns true if there are no backlogged entries after the flush */
1451 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1453 unsigned long flags
;
1454 bool all_flushed
, posted
;
1456 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
1460 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1461 while (!list_empty(&ctx
->cq_overflow_list
)) {
1462 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
1463 struct io_overflow_cqe
*ocqe
;
1467 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
1468 struct io_overflow_cqe
, list
);
1470 memcpy(cqe
, &ocqe
->cqe
, sizeof(*cqe
));
1472 io_account_cq_overflow(ctx
);
1475 list_del(&ocqe
->list
);
1479 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1481 clear_bit(0, &ctx
->check_cq_overflow
);
1482 ctx
->rings
->sq_flags
&= ~IORING_SQ_CQ_OVERFLOW
;
1486 io_commit_cqring(ctx
);
1487 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1489 io_cqring_ev_posted(ctx
);
1493 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1497 if (test_bit(0, &ctx
->check_cq_overflow
)) {
1498 /* iopoll syncs against uring_lock, not completion_lock */
1499 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1500 mutex_lock(&ctx
->uring_lock
);
1501 ret
= __io_cqring_overflow_flush(ctx
, force
);
1502 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1503 mutex_unlock(&ctx
->uring_lock
);
1510 * Shamelessly stolen from the mm implementation of page reference checking,
1511 * see commit f958d7b528b1 for details.
1513 #define req_ref_zero_or_close_to_overflow(req) \
1514 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1516 static inline bool req_ref_inc_not_zero(struct io_kiocb
*req
)
1518 return atomic_inc_not_zero(&req
->refs
);
1521 static inline bool req_ref_sub_and_test(struct io_kiocb
*req
, int refs
)
1523 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1524 return atomic_sub_and_test(refs
, &req
->refs
);
1527 static inline bool req_ref_put_and_test(struct io_kiocb
*req
)
1529 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1530 return atomic_dec_and_test(&req
->refs
);
1533 static inline void req_ref_put(struct io_kiocb
*req
)
1535 WARN_ON_ONCE(req_ref_put_and_test(req
));
1538 static inline void req_ref_get(struct io_kiocb
*req
)
1540 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1541 atomic_inc(&req
->refs
);
1544 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
1545 long res
, unsigned int cflags
)
1547 struct io_overflow_cqe
*ocqe
;
1549 ocqe
= kmalloc(sizeof(*ocqe
), GFP_ATOMIC
| __GFP_ACCOUNT
);
1552 * If we're in ring overflow flush mode, or in task cancel mode,
1553 * or cannot allocate an overflow entry, then we need to drop it
1556 io_account_cq_overflow(ctx
);
1559 if (list_empty(&ctx
->cq_overflow_list
)) {
1560 set_bit(0, &ctx
->check_cq_overflow
);
1561 ctx
->rings
->sq_flags
|= IORING_SQ_CQ_OVERFLOW
;
1563 ocqe
->cqe
.user_data
= user_data
;
1564 ocqe
->cqe
.res
= res
;
1565 ocqe
->cqe
.flags
= cflags
;
1566 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
1570 static inline bool __io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1571 long res
, unsigned int cflags
)
1573 struct io_uring_cqe
*cqe
;
1575 trace_io_uring_complete(ctx
, user_data
, res
, cflags
);
1578 * If we can't get a cq entry, userspace overflowed the
1579 * submission (by quite a lot). Increment the overflow count in
1582 cqe
= io_get_cqe(ctx
);
1584 WRITE_ONCE(cqe
->user_data
, user_data
);
1585 WRITE_ONCE(cqe
->res
, res
);
1586 WRITE_ONCE(cqe
->flags
, cflags
);
1589 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
);
1592 /* not as hot to bloat with inlining */
1593 static noinline
bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1594 long res
, unsigned int cflags
)
1596 return __io_cqring_fill_event(ctx
, user_data
, res
, cflags
);
1599 static void io_req_complete_post(struct io_kiocb
*req
, long res
,
1600 unsigned int cflags
)
1602 struct io_ring_ctx
*ctx
= req
->ctx
;
1603 unsigned long flags
;
1605 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1606 __io_cqring_fill_event(ctx
, req
->user_data
, res
, cflags
);
1608 * If we're the last reference to this request, add to our locked
1611 if (req_ref_put_and_test(req
)) {
1612 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1613 if (req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_FAIL
))
1614 io_disarm_next(req
);
1616 io_req_task_queue(req
->link
);
1620 io_dismantle_req(req
);
1621 io_put_task(req
->task
, 1);
1622 list_add(&req
->compl.list
, &ctx
->locked_free_list
);
1623 ctx
->locked_free_nr
++;
1625 if (!percpu_ref_tryget(&ctx
->refs
))
1628 io_commit_cqring(ctx
);
1629 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1632 io_cqring_ev_posted(ctx
);
1633 percpu_ref_put(&ctx
->refs
);
1637 static inline bool io_req_needs_clean(struct io_kiocb
*req
)
1639 return req
->flags
& IO_REQ_CLEAN_FLAGS
;
1642 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1643 unsigned int cflags
)
1645 if (io_req_needs_clean(req
))
1648 req
->compl.cflags
= cflags
;
1649 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1652 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1653 long res
, unsigned cflags
)
1655 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1656 io_req_complete_state(req
, res
, cflags
);
1658 io_req_complete_post(req
, res
, cflags
);
1661 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1663 __io_req_complete(req
, 0, res
, 0);
1666 static void io_req_complete_failed(struct io_kiocb
*req
, long res
)
1670 io_req_complete_post(req
, res
, 0);
1673 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1674 struct io_comp_state
*cs
)
1676 spin_lock_irq(&ctx
->completion_lock
);
1677 list_splice_init(&ctx
->locked_free_list
, &cs
->free_list
);
1678 ctx
->locked_free_nr
= 0;
1679 spin_unlock_irq(&ctx
->completion_lock
);
1682 /* Returns true IFF there are requests in the cache */
1683 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1685 struct io_submit_state
*state
= &ctx
->submit_state
;
1686 struct io_comp_state
*cs
= &state
->comp
;
1690 * If we have more than a batch's worth of requests in our IRQ side
1691 * locked cache, grab the lock and move them over to our submission
1694 if (READ_ONCE(ctx
->locked_free_nr
) > IO_COMPL_BATCH
)
1695 io_flush_cached_locked_reqs(ctx
, cs
);
1697 nr
= state
->free_reqs
;
1698 while (!list_empty(&cs
->free_list
)) {
1699 struct io_kiocb
*req
= list_first_entry(&cs
->free_list
,
1700 struct io_kiocb
, compl.list
);
1702 list_del(&req
->compl.list
);
1703 state
->reqs
[nr
++] = req
;
1704 if (nr
== ARRAY_SIZE(state
->reqs
))
1708 state
->free_reqs
= nr
;
1712 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1714 struct io_submit_state
*state
= &ctx
->submit_state
;
1716 BUILD_BUG_ON(ARRAY_SIZE(state
->reqs
) < IO_REQ_ALLOC_BATCH
);
1718 if (!state
->free_reqs
) {
1719 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1722 if (io_flush_cached_reqs(ctx
))
1725 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1729 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1730 * retry single alloc to be on the safe side.
1732 if (unlikely(ret
<= 0)) {
1733 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1734 if (!state
->reqs
[0])
1740 * Don't initialise the fields below on every allocation, but
1741 * do that in advance and keep valid on free.
1743 for (i
= 0; i
< ret
; i
++) {
1744 struct io_kiocb
*req
= state
->reqs
[i
];
1748 req
->async_data
= NULL
;
1749 /* not necessary, but safer to zero */
1752 state
->free_reqs
= ret
;
1756 return state
->reqs
[state
->free_reqs
];
1759 static inline void io_put_file(struct file
*file
)
1765 static void io_dismantle_req(struct io_kiocb
*req
)
1767 unsigned int flags
= req
->flags
;
1769 if (io_req_needs_clean(req
))
1771 if (!(flags
& REQ_F_FIXED_FILE
))
1772 io_put_file(req
->file
);
1773 if (req
->fixed_rsrc_refs
)
1774 percpu_ref_put(req
->fixed_rsrc_refs
);
1775 if (req
->async_data
) {
1776 kfree(req
->async_data
);
1777 req
->async_data
= NULL
;
1781 /* must to be called somewhat shortly after putting a request */
1782 static inline void io_put_task(struct task_struct
*task
, int nr
)
1784 struct io_uring_task
*tctx
= task
->io_uring
;
1786 percpu_counter_sub(&tctx
->inflight
, nr
);
1787 if (unlikely(atomic_read(&tctx
->in_idle
)))
1788 wake_up(&tctx
->wait
);
1789 put_task_struct_many(task
, nr
);
1792 static void __io_free_req(struct io_kiocb
*req
)
1794 struct io_ring_ctx
*ctx
= req
->ctx
;
1796 io_dismantle_req(req
);
1797 io_put_task(req
->task
, 1);
1799 kmem_cache_free(req_cachep
, req
);
1800 percpu_ref_put(&ctx
->refs
);
1803 static inline void io_remove_next_linked(struct io_kiocb
*req
)
1805 struct io_kiocb
*nxt
= req
->link
;
1807 req
->link
= nxt
->link
;
1811 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
1812 __must_hold(&req
->ctx
->completion_lock
)
1814 struct io_kiocb
*link
= req
->link
;
1817 * Can happen if a linked timeout fired and link had been like
1818 * req -> link t-out -> link t-out [-> ...]
1820 if (link
&& (link
->flags
& REQ_F_LTIMEOUT_ACTIVE
)) {
1821 struct io_timeout_data
*io
= link
->async_data
;
1823 io_remove_next_linked(req
);
1824 link
->timeout
.head
= NULL
;
1825 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1826 io_cqring_fill_event(link
->ctx
, link
->user_data
,
1828 io_put_req_deferred(link
, 1);
1835 static void io_fail_links(struct io_kiocb
*req
)
1836 __must_hold(&req
->ctx
->completion_lock
)
1838 struct io_kiocb
*nxt
, *link
= req
->link
;
1845 trace_io_uring_fail_link(req
, link
);
1846 io_cqring_fill_event(link
->ctx
, link
->user_data
, -ECANCELED
, 0);
1847 io_put_req_deferred(link
, 2);
1852 static bool io_disarm_next(struct io_kiocb
*req
)
1853 __must_hold(&req
->ctx
->completion_lock
)
1855 bool posted
= false;
1857 if (likely(req
->flags
& REQ_F_LINK_TIMEOUT
))
1858 posted
= io_kill_linked_timeout(req
);
1859 if (unlikely((req
->flags
& REQ_F_FAIL
) &&
1860 !(req
->flags
& REQ_F_HARDLINK
))) {
1861 posted
|= (req
->link
!= NULL
);
1867 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
1869 struct io_kiocb
*nxt
;
1872 * If LINK is set, we have dependent requests in this chain. If we
1873 * didn't fail this request, queue the first one up, moving any other
1874 * dependencies to the next request. In case of failure, fail the rest
1877 if (req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_FAIL
)) {
1878 struct io_ring_ctx
*ctx
= req
->ctx
;
1879 unsigned long flags
;
1882 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1883 posted
= io_disarm_next(req
);
1885 io_commit_cqring(req
->ctx
);
1886 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1888 io_cqring_ev_posted(ctx
);
1895 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1897 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
1899 return __io_req_find_next(req
);
1902 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
)
1906 if (ctx
->submit_state
.comp
.nr
) {
1907 mutex_lock(&ctx
->uring_lock
);
1908 io_submit_flush_completions(ctx
);
1909 mutex_unlock(&ctx
->uring_lock
);
1911 percpu_ref_put(&ctx
->refs
);
1914 static void tctx_task_work(struct callback_head
*cb
)
1916 struct io_ring_ctx
*ctx
= NULL
;
1917 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
1921 struct io_wq_work_node
*node
;
1923 spin_lock_irq(&tctx
->task_lock
);
1924 node
= tctx
->task_list
.first
;
1925 INIT_WQ_LIST(&tctx
->task_list
);
1926 spin_unlock_irq(&tctx
->task_lock
);
1929 struct io_wq_work_node
*next
= node
->next
;
1930 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1933 if (req
->ctx
!= ctx
) {
1934 ctx_flush_and_put(ctx
);
1936 percpu_ref_get(&ctx
->refs
);
1938 req
->io_task_work
.func(req
);
1941 if (wq_list_empty(&tctx
->task_list
)) {
1942 clear_bit(0, &tctx
->task_state
);
1943 if (wq_list_empty(&tctx
->task_list
))
1945 /* another tctx_task_work() is enqueued, yield */
1946 if (test_and_set_bit(0, &tctx
->task_state
))
1952 ctx_flush_and_put(ctx
);
1955 static void io_req_task_work_add(struct io_kiocb
*req
)
1957 struct task_struct
*tsk
= req
->task
;
1958 struct io_uring_task
*tctx
= tsk
->io_uring
;
1959 enum task_work_notify_mode notify
;
1960 struct io_wq_work_node
*node
;
1961 unsigned long flags
;
1963 WARN_ON_ONCE(!tctx
);
1965 spin_lock_irqsave(&tctx
->task_lock
, flags
);
1966 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
1967 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
1969 /* task_work already pending, we're done */
1970 if (test_bit(0, &tctx
->task_state
) ||
1971 test_and_set_bit(0, &tctx
->task_state
))
1973 if (unlikely(tsk
->flags
& PF_EXITING
))
1977 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1978 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1979 * processing task_work. There's no reliable way to tell if TWA_RESUME
1982 notify
= (req
->ctx
->flags
& IORING_SETUP_SQPOLL
) ? TWA_NONE
: TWA_SIGNAL
;
1983 if (!task_work_add(tsk
, &tctx
->task_work
, notify
)) {
1984 wake_up_process(tsk
);
1988 clear_bit(0, &tctx
->task_state
);
1989 spin_lock_irqsave(&tctx
->task_lock
, flags
);
1990 node
= tctx
->task_list
.first
;
1991 INIT_WQ_LIST(&tctx
->task_list
);
1992 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
1995 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1997 if (llist_add(&req
->io_task_work
.fallback_node
,
1998 &req
->ctx
->fallback_llist
))
1999 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
2003 static void io_req_task_cancel(struct io_kiocb
*req
)
2005 struct io_ring_ctx
*ctx
= req
->ctx
;
2007 /* ctx is guaranteed to stay alive while we hold uring_lock */
2008 mutex_lock(&ctx
->uring_lock
);
2009 io_req_complete_failed(req
, req
->result
);
2010 mutex_unlock(&ctx
->uring_lock
);
2013 static void io_req_task_submit(struct io_kiocb
*req
)
2015 struct io_ring_ctx
*ctx
= req
->ctx
;
2017 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2018 mutex_lock(&ctx
->uring_lock
);
2019 if (!(req
->task
->flags
& PF_EXITING
) && !req
->task
->in_execve
)
2020 __io_queue_sqe(req
);
2022 io_req_complete_failed(req
, -EFAULT
);
2023 mutex_unlock(&ctx
->uring_lock
);
2026 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2029 req
->io_task_work
.func
= io_req_task_cancel
;
2030 io_req_task_work_add(req
);
2033 static void io_req_task_queue(struct io_kiocb
*req
)
2035 req
->io_task_work
.func
= io_req_task_submit
;
2036 io_req_task_work_add(req
);
2039 static inline void io_queue_next(struct io_kiocb
*req
)
2041 struct io_kiocb
*nxt
= io_req_find_next(req
);
2044 io_req_task_queue(nxt
);
2047 static void io_free_req(struct io_kiocb
*req
)
2054 struct task_struct
*task
;
2059 static inline void io_init_req_batch(struct req_batch
*rb
)
2066 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2067 struct req_batch
*rb
)
2070 io_put_task(rb
->task
, rb
->task_refs
);
2072 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2075 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2076 struct io_submit_state
*state
)
2079 io_dismantle_req(req
);
2081 if (req
->task
!= rb
->task
) {
2083 io_put_task(rb
->task
, rb
->task_refs
);
2084 rb
->task
= req
->task
;
2090 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2091 state
->reqs
[state
->free_reqs
++] = req
;
2093 list_add(&req
->compl.list
, &state
->comp
.free_list
);
2096 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
2098 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
2100 struct req_batch rb
;
2102 spin_lock_irq(&ctx
->completion_lock
);
2103 for (i
= 0; i
< nr
; i
++) {
2104 struct io_kiocb
*req
= cs
->reqs
[i
];
2106 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2109 io_commit_cqring(ctx
);
2110 spin_unlock_irq(&ctx
->completion_lock
);
2111 io_cqring_ev_posted(ctx
);
2113 io_init_req_batch(&rb
);
2114 for (i
= 0; i
< nr
; i
++) {
2115 struct io_kiocb
*req
= cs
->reqs
[i
];
2117 /* submission and completion refs */
2118 if (req_ref_sub_and_test(req
, 2))
2119 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2122 io_req_free_batch_finish(ctx
, &rb
);
2127 * Drop reference to request, return next in chain (if there is one) if this
2128 * was the last reference to this request.
2130 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2132 struct io_kiocb
*nxt
= NULL
;
2134 if (req_ref_put_and_test(req
)) {
2135 nxt
= io_req_find_next(req
);
2141 static inline void io_put_req(struct io_kiocb
*req
)
2143 if (req_ref_put_and_test(req
))
2147 static void io_free_req_deferred(struct io_kiocb
*req
)
2149 req
->io_task_work
.func
= io_free_req
;
2150 io_req_task_work_add(req
);
2153 static inline void io_put_req_deferred(struct io_kiocb
*req
, int refs
)
2155 if (req_ref_sub_and_test(req
, refs
))
2156 io_free_req_deferred(req
);
2159 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2161 /* See comment at the top of this file */
2163 return __io_cqring_events(ctx
);
2166 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2168 struct io_rings
*rings
= ctx
->rings
;
2170 /* make sure SQ entry isn't read before tail */
2171 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2174 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2176 unsigned int cflags
;
2178 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2179 cflags
|= IORING_CQE_F_BUFFER
;
2180 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2185 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2187 struct io_buffer
*kbuf
;
2189 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2190 return io_put_kbuf(req
, kbuf
);
2193 static inline bool io_run_task_work(void)
2195 if (current
->task_works
) {
2196 __set_current_state(TASK_RUNNING
);
2205 * Find and free completed poll iocbs
2207 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2208 struct list_head
*done
)
2210 struct req_batch rb
;
2211 struct io_kiocb
*req
;
2213 /* order with ->result store in io_complete_rw_iopoll() */
2216 io_init_req_batch(&rb
);
2217 while (!list_empty(done
)) {
2220 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2221 list_del(&req
->inflight_entry
);
2223 if (READ_ONCE(req
->result
) == -EAGAIN
&&
2224 !(req
->flags
& REQ_F_DONT_REISSUE
)) {
2225 req
->iopoll_completed
= 0;
2227 io_queue_async_work(req
);
2231 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2232 cflags
= io_put_rw_kbuf(req
);
2234 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
, cflags
);
2237 if (req_ref_put_and_test(req
))
2238 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2241 io_commit_cqring(ctx
);
2242 io_cqring_ev_posted_iopoll(ctx
);
2243 io_req_free_batch_finish(ctx
, &rb
);
2246 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2249 struct io_kiocb
*req
, *tmp
;
2255 * Only spin for completions if we don't have multiple devices hanging
2256 * off our complete list, and we're under the requested amount.
2258 spin
= !ctx
->poll_multi_queue
&& *nr_events
< min
;
2261 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2262 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2265 * Move completed and retryable entries to our local lists.
2266 * If we find a request that requires polling, break out
2267 * and complete those lists first, if we have entries there.
2269 if (READ_ONCE(req
->iopoll_completed
)) {
2270 list_move_tail(&req
->inflight_entry
, &done
);
2273 if (!list_empty(&done
))
2276 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2280 /* iopoll may have completed current req */
2281 if (READ_ONCE(req
->iopoll_completed
))
2282 list_move_tail(&req
->inflight_entry
, &done
);
2289 if (!list_empty(&done
))
2290 io_iopoll_complete(ctx
, nr_events
, &done
);
2296 * We can't just wait for polled events to come to us, we have to actively
2297 * find and complete them.
2299 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2301 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2304 mutex_lock(&ctx
->uring_lock
);
2305 while (!list_empty(&ctx
->iopoll_list
)) {
2306 unsigned int nr_events
= 0;
2308 io_do_iopoll(ctx
, &nr_events
, 0);
2310 /* let it sleep and repeat later if can't complete a request */
2314 * Ensure we allow local-to-the-cpu processing to take place,
2315 * in this case we need to ensure that we reap all events.
2316 * Also let task_work, etc. to progress by releasing the mutex
2318 if (need_resched()) {
2319 mutex_unlock(&ctx
->uring_lock
);
2321 mutex_lock(&ctx
->uring_lock
);
2324 mutex_unlock(&ctx
->uring_lock
);
2327 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2329 unsigned int nr_events
= 0;
2333 * We disallow the app entering submit/complete with polling, but we
2334 * still need to lock the ring to prevent racing with polled issue
2335 * that got punted to a workqueue.
2337 mutex_lock(&ctx
->uring_lock
);
2339 * Don't enter poll loop if we already have events pending.
2340 * If we do, we can potentially be spinning for commands that
2341 * already triggered a CQE (eg in error).
2343 if (test_bit(0, &ctx
->check_cq_overflow
))
2344 __io_cqring_overflow_flush(ctx
, false);
2345 if (io_cqring_events(ctx
))
2349 * If a submit got punted to a workqueue, we can have the
2350 * application entering polling for a command before it gets
2351 * issued. That app will hold the uring_lock for the duration
2352 * of the poll right here, so we need to take a breather every
2353 * now and then to ensure that the issue has a chance to add
2354 * the poll to the issued list. Otherwise we can spin here
2355 * forever, while the workqueue is stuck trying to acquire the
2358 if (list_empty(&ctx
->iopoll_list
)) {
2359 u32 tail
= ctx
->cached_cq_tail
;
2361 mutex_unlock(&ctx
->uring_lock
);
2363 mutex_lock(&ctx
->uring_lock
);
2365 /* some requests don't go through iopoll_list */
2366 if (tail
!= ctx
->cached_cq_tail
||
2367 list_empty(&ctx
->iopoll_list
))
2370 ret
= io_do_iopoll(ctx
, &nr_events
, min
);
2371 } while (!ret
&& nr_events
< min
&& !need_resched());
2373 mutex_unlock(&ctx
->uring_lock
);
2377 static void kiocb_end_write(struct io_kiocb
*req
)
2380 * Tell lockdep we inherited freeze protection from submission
2383 if (req
->flags
& REQ_F_ISREG
) {
2384 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
2386 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
2392 static bool io_resubmit_prep(struct io_kiocb
*req
)
2394 struct io_async_rw
*rw
= req
->async_data
;
2397 return !io_req_prep_async(req
);
2398 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2399 iov_iter_revert(&rw
->iter
, req
->result
- iov_iter_count(&rw
->iter
));
2403 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2405 umode_t mode
= file_inode(req
->file
)->i_mode
;
2406 struct io_ring_ctx
*ctx
= req
->ctx
;
2408 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2410 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2411 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2414 * If ref is dying, we might be running poll reap from the exit work.
2415 * Don't attempt to reissue from that path, just let it fail with
2418 if (percpu_ref_is_dying(&ctx
->refs
))
2423 static bool io_resubmit_prep(struct io_kiocb
*req
)
2427 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2433 static void io_fallback_req_func(struct work_struct
*work
)
2435 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
2436 fallback_work
.work
);
2437 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
2438 struct io_kiocb
*req
, *tmp
;
2440 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.fallback_node
)
2441 req
->io_task_work
.func(req
);
2444 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2445 unsigned int issue_flags
)
2449 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
)
2450 kiocb_end_write(req
);
2451 if (res
!= req
->result
) {
2452 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
2453 io_rw_should_reissue(req
)) {
2454 req
->flags
|= REQ_F_REISSUE
;
2459 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2460 cflags
= io_put_rw_kbuf(req
);
2461 __io_req_complete(req
, issue_flags
, res
, cflags
);
2464 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2466 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2468 __io_complete_rw(req
, res
, res2
, 0);
2471 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2473 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2475 if (kiocb
->ki_flags
& IOCB_WRITE
)
2476 kiocb_end_write(req
);
2477 if (unlikely(res
!= req
->result
)) {
2478 if (!(res
== -EAGAIN
&& io_rw_should_reissue(req
) &&
2479 io_resubmit_prep(req
))) {
2481 req
->flags
|= REQ_F_DONT_REISSUE
;
2485 WRITE_ONCE(req
->result
, res
);
2486 /* order with io_iopoll_complete() checking ->result */
2488 WRITE_ONCE(req
->iopoll_completed
, 1);
2492 * After the iocb has been issued, it's safe to be found on the poll list.
2493 * Adding the kiocb to the list AFTER submission ensures that we don't
2494 * find it from a io_do_iopoll() thread before the issuer is done
2495 * accessing the kiocb cookie.
2497 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2499 struct io_ring_ctx
*ctx
= req
->ctx
;
2500 const bool in_async
= io_wq_current_is_worker();
2502 /* workqueue context doesn't hold uring_lock, grab it now */
2503 if (unlikely(in_async
))
2504 mutex_lock(&ctx
->uring_lock
);
2507 * Track whether we have multiple files in our lists. This will impact
2508 * how we do polling eventually, not spinning if we're on potentially
2509 * different devices.
2511 if (list_empty(&ctx
->iopoll_list
)) {
2512 ctx
->poll_multi_queue
= false;
2513 } else if (!ctx
->poll_multi_queue
) {
2514 struct io_kiocb
*list_req
;
2515 unsigned int queue_num0
, queue_num1
;
2517 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2520 if (list_req
->file
!= req
->file
) {
2521 ctx
->poll_multi_queue
= true;
2523 queue_num0
= blk_qc_t_to_queue_num(list_req
->rw
.kiocb
.ki_cookie
);
2524 queue_num1
= blk_qc_t_to_queue_num(req
->rw
.kiocb
.ki_cookie
);
2525 if (queue_num0
!= queue_num1
)
2526 ctx
->poll_multi_queue
= true;
2531 * For fast devices, IO may have already completed. If it has, add
2532 * it to the front so we find it first.
2534 if (READ_ONCE(req
->iopoll_completed
))
2535 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2537 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2539 if (unlikely(in_async
)) {
2541 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2542 * in sq thread task context or in io worker task context. If
2543 * current task context is sq thread, we don't need to check
2544 * whether should wake up sq thread.
2546 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2547 wq_has_sleeper(&ctx
->sq_data
->wait
))
2548 wake_up(&ctx
->sq_data
->wait
);
2550 mutex_unlock(&ctx
->uring_lock
);
2554 static inline void io_state_file_put(struct io_submit_state
*state
)
2556 if (state
->file_refs
) {
2557 fput_many(state
->file
, state
->file_refs
);
2558 state
->file_refs
= 0;
2563 * Get as many references to a file as we have IOs left in this submission,
2564 * assuming most submissions are for one file, or at least that each file
2565 * has more than one submission.
2567 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2572 if (state
->file_refs
) {
2573 if (state
->fd
== fd
) {
2577 io_state_file_put(state
);
2579 state
->file
= fget_many(fd
, state
->ios_left
);
2580 if (unlikely(!state
->file
))
2584 state
->file_refs
= state
->ios_left
- 1;
2588 static bool io_bdev_nowait(struct block_device
*bdev
)
2590 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2594 * If we tracked the file through the SCM inflight mechanism, we could support
2595 * any file. For now, just ensure that anything potentially problematic is done
2598 static bool __io_file_supports_async(struct file
*file
, int rw
)
2600 umode_t mode
= file_inode(file
)->i_mode
;
2602 if (S_ISBLK(mode
)) {
2603 if (IS_ENABLED(CONFIG_BLOCK
) &&
2604 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2610 if (S_ISREG(mode
)) {
2611 if (IS_ENABLED(CONFIG_BLOCK
) &&
2612 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2613 file
->f_op
!= &io_uring_fops
)
2618 /* any ->read/write should understand O_NONBLOCK */
2619 if (file
->f_flags
& O_NONBLOCK
)
2622 if (!(file
->f_mode
& FMODE_NOWAIT
))
2626 return file
->f_op
->read_iter
!= NULL
;
2628 return file
->f_op
->write_iter
!= NULL
;
2631 static bool io_file_supports_async(struct io_kiocb
*req
, int rw
)
2633 if (rw
== READ
&& (req
->flags
& REQ_F_ASYNC_READ
))
2635 else if (rw
== WRITE
&& (req
->flags
& REQ_F_ASYNC_WRITE
))
2638 return __io_file_supports_async(req
->file
, rw
);
2641 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2643 struct io_ring_ctx
*ctx
= req
->ctx
;
2644 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2645 struct file
*file
= req
->file
;
2649 if (!(req
->flags
& REQ_F_ISREG
) && S_ISREG(file_inode(file
)->i_mode
))
2650 req
->flags
|= REQ_F_ISREG
;
2652 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2653 if (kiocb
->ki_pos
== -1 && !(file
->f_mode
& FMODE_STREAM
)) {
2654 req
->flags
|= REQ_F_CUR_POS
;
2655 kiocb
->ki_pos
= file
->f_pos
;
2657 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2658 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2659 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2663 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2664 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) || (file
->f_flags
& O_NONBLOCK
))
2665 req
->flags
|= REQ_F_NOWAIT
;
2667 ioprio
= READ_ONCE(sqe
->ioprio
);
2669 ret
= ioprio_check_cap(ioprio
);
2673 kiocb
->ki_ioprio
= ioprio
;
2675 kiocb
->ki_ioprio
= get_current_ioprio();
2677 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2678 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2679 !kiocb
->ki_filp
->f_op
->iopoll
)
2682 kiocb
->ki_flags
|= IOCB_HIPRI
;
2683 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2684 req
->iopoll_completed
= 0;
2686 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2688 kiocb
->ki_complete
= io_complete_rw
;
2691 if (req
->opcode
== IORING_OP_READ_FIXED
||
2692 req
->opcode
== IORING_OP_WRITE_FIXED
) {
2694 io_req_set_rsrc_node(req
);
2697 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2698 req
->rw
.len
= READ_ONCE(sqe
->len
);
2699 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2703 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2709 case -ERESTARTNOINTR
:
2710 case -ERESTARTNOHAND
:
2711 case -ERESTART_RESTARTBLOCK
:
2713 * We can't just restart the syscall, since previously
2714 * submitted sqes may already be in progress. Just fail this
2720 kiocb
->ki_complete(kiocb
, ret
, 0);
2724 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2725 unsigned int issue_flags
)
2727 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2728 struct io_async_rw
*io
= req
->async_data
;
2729 bool check_reissue
= kiocb
->ki_complete
== io_complete_rw
;
2731 /* add previously done IO, if any */
2732 if (io
&& io
->bytes_done
> 0) {
2734 ret
= io
->bytes_done
;
2736 ret
+= io
->bytes_done
;
2739 if (req
->flags
& REQ_F_CUR_POS
)
2740 req
->file
->f_pos
= kiocb
->ki_pos
;
2741 if (ret
>= 0 && check_reissue
)
2742 __io_complete_rw(req
, ret
, 0, issue_flags
);
2744 io_rw_done(kiocb
, ret
);
2746 if (check_reissue
&& (req
->flags
& REQ_F_REISSUE
)) {
2747 req
->flags
&= ~REQ_F_REISSUE
;
2748 if (io_resubmit_prep(req
)) {
2750 io_queue_async_work(req
);
2755 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2756 cflags
= io_put_rw_kbuf(req
);
2757 __io_req_complete(req
, issue_flags
, ret
, cflags
);
2762 static int __io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
2763 struct io_mapped_ubuf
*imu
)
2765 size_t len
= req
->rw
.len
;
2766 u64 buf_end
, buf_addr
= req
->rw
.addr
;
2769 if (unlikely(check_add_overflow(buf_addr
, (u64
)len
, &buf_end
)))
2771 /* not inside the mapped region */
2772 if (unlikely(buf_addr
< imu
->ubuf
|| buf_end
> imu
->ubuf_end
))
2776 * May not be a start of buffer, set size appropriately
2777 * and advance us to the beginning.
2779 offset
= buf_addr
- imu
->ubuf
;
2780 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
2784 * Don't use iov_iter_advance() here, as it's really slow for
2785 * using the latter parts of a big fixed buffer - it iterates
2786 * over each segment manually. We can cheat a bit here, because
2789 * 1) it's a BVEC iter, we set it up
2790 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2791 * first and last bvec
2793 * So just find our index, and adjust the iterator afterwards.
2794 * If the offset is within the first bvec (or the whole first
2795 * bvec, just use iov_iter_advance(). This makes it easier
2796 * since we can just skip the first segment, which may not
2797 * be PAGE_SIZE aligned.
2799 const struct bio_vec
*bvec
= imu
->bvec
;
2801 if (offset
<= bvec
->bv_len
) {
2802 iov_iter_advance(iter
, offset
);
2804 unsigned long seg_skip
;
2806 /* skip first vec */
2807 offset
-= bvec
->bv_len
;
2808 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
2810 iter
->bvec
= bvec
+ seg_skip
;
2811 iter
->nr_segs
-= seg_skip
;
2812 iter
->count
-= bvec
->bv_len
+ offset
;
2813 iter
->iov_offset
= offset
& ~PAGE_MASK
;
2820 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
2822 struct io_ring_ctx
*ctx
= req
->ctx
;
2823 struct io_mapped_ubuf
*imu
= req
->imu
;
2824 u16 index
, buf_index
= req
->buf_index
;
2827 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
2829 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
2830 imu
= READ_ONCE(ctx
->user_bufs
[index
]);
2833 return __io_import_fixed(req
, rw
, iter
, imu
);
2836 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2839 mutex_unlock(&ctx
->uring_lock
);
2842 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2845 * "Normal" inline submissions always hold the uring_lock, since we
2846 * grab it from the system call. Same is true for the SQPOLL offload.
2847 * The only exception is when we've detached the request and issue it
2848 * from an async worker thread, grab the lock for that case.
2851 mutex_lock(&ctx
->uring_lock
);
2854 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
2855 int bgid
, struct io_buffer
*kbuf
,
2858 struct io_buffer
*head
;
2860 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2863 io_ring_submit_lock(req
->ctx
, needs_lock
);
2865 lockdep_assert_held(&req
->ctx
->uring_lock
);
2867 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
2869 if (!list_empty(&head
->list
)) {
2870 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
2872 list_del(&kbuf
->list
);
2875 xa_erase(&req
->ctx
->io_buffers
, bgid
);
2877 if (*len
> kbuf
->len
)
2880 kbuf
= ERR_PTR(-ENOBUFS
);
2883 io_ring_submit_unlock(req
->ctx
, needs_lock
);
2888 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
2891 struct io_buffer
*kbuf
;
2894 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2895 bgid
= req
->buf_index
;
2896 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
2899 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
2900 req
->flags
|= REQ_F_BUFFER_SELECTED
;
2901 return u64_to_user_ptr(kbuf
->addr
);
2904 #ifdef CONFIG_COMPAT
2905 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
2908 struct compat_iovec __user
*uiov
;
2909 compat_ssize_t clen
;
2913 uiov
= u64_to_user_ptr(req
->rw
.addr
);
2914 if (!access_ok(uiov
, sizeof(*uiov
)))
2916 if (__get_user(clen
, &uiov
->iov_len
))
2922 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2924 return PTR_ERR(buf
);
2925 iov
[0].iov_base
= buf
;
2926 iov
[0].iov_len
= (compat_size_t
) len
;
2931 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2934 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
2938 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
2941 len
= iov
[0].iov_len
;
2944 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2946 return PTR_ERR(buf
);
2947 iov
[0].iov_base
= buf
;
2948 iov
[0].iov_len
= len
;
2952 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2955 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
2956 struct io_buffer
*kbuf
;
2958 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2959 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
2960 iov
[0].iov_len
= kbuf
->len
;
2963 if (req
->rw
.len
!= 1)
2966 #ifdef CONFIG_COMPAT
2967 if (req
->ctx
->compat
)
2968 return io_compat_import(req
, iov
, needs_lock
);
2971 return __io_iov_buffer_select(req
, iov
, needs_lock
);
2974 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
2975 struct iov_iter
*iter
, bool needs_lock
)
2977 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
2978 size_t sqe_len
= req
->rw
.len
;
2979 u8 opcode
= req
->opcode
;
2982 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
2984 return io_import_fixed(req
, rw
, iter
);
2987 /* buffer index only valid with fixed read/write, or buffer select */
2988 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
2991 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
2992 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
2993 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
2995 return PTR_ERR(buf
);
2996 req
->rw
.len
= sqe_len
;
2999 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3004 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3005 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3007 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3012 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3016 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3018 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3022 * For files that don't have ->read_iter() and ->write_iter(), handle them
3023 * by looping over ->read() or ->write() manually.
3025 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3027 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3028 struct file
*file
= req
->file
;
3032 * Don't support polled IO through this interface, and we can't
3033 * support non-blocking either. For the latter, this just causes
3034 * the kiocb to be handled from an async context.
3036 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3038 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3041 while (iov_iter_count(iter
)) {
3045 if (!iov_iter_is_bvec(iter
)) {
3046 iovec
= iov_iter_iovec(iter
);
3048 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3049 iovec
.iov_len
= req
->rw
.len
;
3053 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3054 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3056 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3057 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3066 if (nr
!= iovec
.iov_len
)
3070 iov_iter_advance(iter
, nr
);
3076 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3077 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3079 struct io_async_rw
*rw
= req
->async_data
;
3081 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3082 rw
->free_iovec
= iovec
;
3084 /* can only be fixed buffers, no need to do anything */
3085 if (iov_iter_is_bvec(iter
))
3088 unsigned iov_off
= 0;
3090 rw
->iter
.iov
= rw
->fast_iov
;
3091 if (iter
->iov
!= fast_iov
) {
3092 iov_off
= iter
->iov
- fast_iov
;
3093 rw
->iter
.iov
+= iov_off
;
3095 if (rw
->fast_iov
!= fast_iov
)
3096 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3097 sizeof(struct iovec
) * iter
->nr_segs
);
3099 req
->flags
|= REQ_F_NEED_CLEANUP
;
3103 static inline int io_alloc_async_data(struct io_kiocb
*req
)
3105 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3106 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3107 return req
->async_data
== NULL
;
3110 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3111 const struct iovec
*fast_iov
,
3112 struct iov_iter
*iter
, bool force
)
3114 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_setup
)
3116 if (!req
->async_data
) {
3117 if (io_alloc_async_data(req
)) {
3122 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3127 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3129 struct io_async_rw
*iorw
= req
->async_data
;
3130 struct iovec
*iov
= iorw
->fast_iov
;
3133 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3134 if (unlikely(ret
< 0))
3137 iorw
->bytes_done
= 0;
3138 iorw
->free_iovec
= iov
;
3140 req
->flags
|= REQ_F_NEED_CLEANUP
;
3144 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3146 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3148 return io_prep_rw(req
, sqe
);
3152 * This is our waitqueue callback handler, registered through lock_page_async()
3153 * when we initially tried to do the IO with the iocb armed our waitqueue.
3154 * This gets called when the page is unlocked, and we generally expect that to
3155 * happen when the page IO is completed and the page is now uptodate. This will
3156 * queue a task_work based retry of the operation, attempting to copy the data
3157 * again. If the latter fails because the page was NOT uptodate, then we will
3158 * do a thread based blocking retry of the operation. That's the unexpected
3161 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3162 int sync
, void *arg
)
3164 struct wait_page_queue
*wpq
;
3165 struct io_kiocb
*req
= wait
->private;
3166 struct wait_page_key
*key
= arg
;
3168 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3170 if (!wake_page_match(wpq
, key
))
3173 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3174 list_del_init(&wait
->entry
);
3176 /* submit ref gets dropped, acquire a new one */
3178 io_req_task_queue(req
);
3183 * This controls whether a given IO request should be armed for async page
3184 * based retry. If we return false here, the request is handed to the async
3185 * worker threads for retry. If we're doing buffered reads on a regular file,
3186 * we prepare a private wait_page_queue entry and retry the operation. This
3187 * will either succeed because the page is now uptodate and unlocked, or it
3188 * will register a callback when the page is unlocked at IO completion. Through
3189 * that callback, io_uring uses task_work to setup a retry of the operation.
3190 * That retry will attempt the buffered read again. The retry will generally
3191 * succeed, or in rare cases where it fails, we then fall back to using the
3192 * async worker threads for a blocking retry.
3194 static bool io_rw_should_retry(struct io_kiocb
*req
)
3196 struct io_async_rw
*rw
= req
->async_data
;
3197 struct wait_page_queue
*wait
= &rw
->wpq
;
3198 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3200 /* never retry for NOWAIT, we just complete with -EAGAIN */
3201 if (req
->flags
& REQ_F_NOWAIT
)
3204 /* Only for buffered IO */
3205 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3209 * just use poll if we can, and don't attempt if the fs doesn't
3210 * support callback based unlocks
3212 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3215 wait
->wait
.func
= io_async_buf_func
;
3216 wait
->wait
.private = req
;
3217 wait
->wait
.flags
= 0;
3218 INIT_LIST_HEAD(&wait
->wait
.entry
);
3219 kiocb
->ki_flags
|= IOCB_WAITQ
;
3220 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3221 kiocb
->ki_waitq
= wait
;
3225 static inline int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3227 if (req
->file
->f_op
->read_iter
)
3228 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3229 else if (req
->file
->f_op
->read
)
3230 return loop_rw_iter(READ
, req
, iter
);
3235 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3237 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3238 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3239 struct iov_iter __iter
, *iter
= &__iter
;
3240 struct io_async_rw
*rw
= req
->async_data
;
3241 ssize_t io_size
, ret
, ret2
;
3242 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3248 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3252 io_size
= iov_iter_count(iter
);
3253 req
->result
= io_size
;
3255 /* Ensure we clear previously set non-block flag */
3256 if (!force_nonblock
)
3257 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3259 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3261 /* If the file doesn't support async, just async punt */
3262 if (force_nonblock
&& !io_file_supports_async(req
, READ
)) {
3263 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3264 return ret
?: -EAGAIN
;
3267 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3268 if (unlikely(ret
)) {
3273 ret
= io_iter_do_read(req
, iter
);
3275 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
3276 req
->flags
&= ~REQ_F_REISSUE
;
3277 /* IOPOLL retry should happen for io-wq threads */
3278 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3280 /* no retry on NONBLOCK nor RWF_NOWAIT */
3281 if (req
->flags
& REQ_F_NOWAIT
)
3283 /* some cases will consume bytes even on error returns */
3284 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3286 } else if (ret
== -EIOCBQUEUED
) {
3288 } else if (ret
<= 0 || ret
== io_size
|| !force_nonblock
||
3289 (req
->flags
& REQ_F_NOWAIT
) || !(req
->flags
& REQ_F_ISREG
)) {
3290 /* read all, failed, already did sync or don't want to retry */
3294 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3299 rw
= req
->async_data
;
3300 /* now use our persistent iterator, if we aren't already */
3305 rw
->bytes_done
+= ret
;
3306 /* if we can retry, do so with the callbacks armed */
3307 if (!io_rw_should_retry(req
)) {
3308 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3313 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3314 * we get -EIOCBQUEUED, then we'll get a notification when the
3315 * desired page gets unlocked. We can also get a partial read
3316 * here, and if we do, then just retry at the new offset.
3318 ret
= io_iter_do_read(req
, iter
);
3319 if (ret
== -EIOCBQUEUED
)
3321 /* we got some bytes, but not all. retry. */
3322 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3323 } while (ret
> 0 && ret
< io_size
);
3325 kiocb_done(kiocb
, ret
, issue_flags
);
3327 /* it's faster to check here then delegate to kfree */
3333 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3335 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3337 return io_prep_rw(req
, sqe
);
3340 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3342 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3343 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3344 struct iov_iter __iter
, *iter
= &__iter
;
3345 struct io_async_rw
*rw
= req
->async_data
;
3346 ssize_t ret
, ret2
, io_size
;
3347 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3353 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3357 io_size
= iov_iter_count(iter
);
3358 req
->result
= io_size
;
3360 /* Ensure we clear previously set non-block flag */
3361 if (!force_nonblock
)
3362 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3364 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3366 /* If the file doesn't support async, just async punt */
3367 if (force_nonblock
&& !io_file_supports_async(req
, WRITE
))
3370 /* file path doesn't support NOWAIT for non-direct_IO */
3371 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3372 (req
->flags
& REQ_F_ISREG
))
3375 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), io_size
);
3380 * Open-code file_start_write here to grab freeze protection,
3381 * which will be released by another thread in
3382 * io_complete_rw(). Fool lockdep by telling it the lock got
3383 * released so that it doesn't complain about the held lock when
3384 * we return to userspace.
3386 if (req
->flags
& REQ_F_ISREG
) {
3387 sb_start_write(file_inode(req
->file
)->i_sb
);
3388 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3391 kiocb
->ki_flags
|= IOCB_WRITE
;
3393 if (req
->file
->f_op
->write_iter
)
3394 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3395 else if (req
->file
->f_op
->write
)
3396 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3400 if (req
->flags
& REQ_F_REISSUE
) {
3401 req
->flags
&= ~REQ_F_REISSUE
;
3406 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3407 * retry them without IOCB_NOWAIT.
3409 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3411 /* no retry on NONBLOCK nor RWF_NOWAIT */
3412 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3414 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3415 /* IOPOLL retry should happen for io-wq threads */
3416 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3419 kiocb_done(kiocb
, ret2
, issue_flags
);
3422 /* some cases will consume bytes even on error returns */
3423 iov_iter_revert(iter
, io_size
- iov_iter_count(iter
));
3424 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3425 return ret
?: -EAGAIN
;
3428 /* it's reportedly faster than delegating the null check to kfree() */
3434 static int io_renameat_prep(struct io_kiocb
*req
,
3435 const struct io_uring_sqe
*sqe
)
3437 struct io_rename
*ren
= &req
->rename
;
3438 const char __user
*oldf
, *newf
;
3440 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3442 if (sqe
->ioprio
|| sqe
->buf_index
)
3444 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3447 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3448 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3449 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3450 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3451 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3453 ren
->oldpath
= getname(oldf
);
3454 if (IS_ERR(ren
->oldpath
))
3455 return PTR_ERR(ren
->oldpath
);
3457 ren
->newpath
= getname(newf
);
3458 if (IS_ERR(ren
->newpath
)) {
3459 putname(ren
->oldpath
);
3460 return PTR_ERR(ren
->newpath
);
3463 req
->flags
|= REQ_F_NEED_CLEANUP
;
3467 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3469 struct io_rename
*ren
= &req
->rename
;
3472 if (issue_flags
& IO_URING_F_NONBLOCK
)
3475 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3476 ren
->newpath
, ren
->flags
);
3478 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3481 io_req_complete(req
, ret
);
3485 static int io_unlinkat_prep(struct io_kiocb
*req
,
3486 const struct io_uring_sqe
*sqe
)
3488 struct io_unlink
*un
= &req
->unlink
;
3489 const char __user
*fname
;
3491 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3493 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
3495 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3498 un
->dfd
= READ_ONCE(sqe
->fd
);
3500 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3501 if (un
->flags
& ~AT_REMOVEDIR
)
3504 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3505 un
->filename
= getname(fname
);
3506 if (IS_ERR(un
->filename
))
3507 return PTR_ERR(un
->filename
);
3509 req
->flags
|= REQ_F_NEED_CLEANUP
;
3513 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3515 struct io_unlink
*un
= &req
->unlink
;
3518 if (issue_flags
& IO_URING_F_NONBLOCK
)
3521 if (un
->flags
& AT_REMOVEDIR
)
3522 ret
= do_rmdir(un
->dfd
, un
->filename
);
3524 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3526 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3529 io_req_complete(req
, ret
);
3533 static int io_shutdown_prep(struct io_kiocb
*req
,
3534 const struct io_uring_sqe
*sqe
)
3536 #if defined(CONFIG_NET)
3537 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3539 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3543 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3550 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3552 #if defined(CONFIG_NET)
3553 struct socket
*sock
;
3556 if (issue_flags
& IO_URING_F_NONBLOCK
)
3559 sock
= sock_from_file(req
->file
);
3560 if (unlikely(!sock
))
3563 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3566 io_req_complete(req
, ret
);
3573 static int __io_splice_prep(struct io_kiocb
*req
,
3574 const struct io_uring_sqe
*sqe
)
3576 struct io_splice
*sp
= &req
->splice
;
3577 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3579 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3583 sp
->len
= READ_ONCE(sqe
->len
);
3584 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3586 if (unlikely(sp
->flags
& ~valid_flags
))
3589 sp
->file_in
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
),
3590 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3593 req
->flags
|= REQ_F_NEED_CLEANUP
;
3597 static int io_tee_prep(struct io_kiocb
*req
,
3598 const struct io_uring_sqe
*sqe
)
3600 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3602 return __io_splice_prep(req
, sqe
);
3605 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
3607 struct io_splice
*sp
= &req
->splice
;
3608 struct file
*in
= sp
->file_in
;
3609 struct file
*out
= sp
->file_out
;
3610 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3613 if (issue_flags
& IO_URING_F_NONBLOCK
)
3616 ret
= do_tee(in
, out
, sp
->len
, flags
);
3618 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
3620 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3624 io_req_complete(req
, ret
);
3628 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3630 struct io_splice
*sp
= &req
->splice
;
3632 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3633 sp
->off_out
= READ_ONCE(sqe
->off
);
3634 return __io_splice_prep(req
, sqe
);
3637 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
3639 struct io_splice
*sp
= &req
->splice
;
3640 struct file
*in
= sp
->file_in
;
3641 struct file
*out
= sp
->file_out
;
3642 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3643 loff_t
*poff_in
, *poff_out
;
3646 if (issue_flags
& IO_URING_F_NONBLOCK
)
3649 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3650 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3653 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3655 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
3657 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3661 io_req_complete(req
, ret
);
3666 * IORING_OP_NOP just posts a completion event, nothing else.
3668 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
3670 struct io_ring_ctx
*ctx
= req
->ctx
;
3672 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3675 __io_req_complete(req
, issue_flags
, 0, 0);
3679 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3681 struct io_ring_ctx
*ctx
= req
->ctx
;
3686 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3688 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3691 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3692 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3695 req
->sync
.off
= READ_ONCE(sqe
->off
);
3696 req
->sync
.len
= READ_ONCE(sqe
->len
);
3700 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
3702 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3705 /* fsync always requires a blocking context */
3706 if (issue_flags
& IO_URING_F_NONBLOCK
)
3709 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3710 end
> 0 ? end
: LLONG_MAX
,
3711 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3714 io_req_complete(req
, ret
);
3718 static int io_fallocate_prep(struct io_kiocb
*req
,
3719 const struct io_uring_sqe
*sqe
)
3721 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3723 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3726 req
->sync
.off
= READ_ONCE(sqe
->off
);
3727 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3728 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3732 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
3736 /* fallocate always requiring blocking context */
3737 if (issue_flags
& IO_URING_F_NONBLOCK
)
3739 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3743 io_req_complete(req
, ret
);
3747 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3749 const char __user
*fname
;
3752 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3754 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3757 /* open.how should be already initialised */
3758 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
3759 req
->open
.how
.flags
|= O_LARGEFILE
;
3761 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3762 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3763 req
->open
.filename
= getname(fname
);
3764 if (IS_ERR(req
->open
.filename
)) {
3765 ret
= PTR_ERR(req
->open
.filename
);
3766 req
->open
.filename
= NULL
;
3769 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
3770 req
->flags
|= REQ_F_NEED_CLEANUP
;
3774 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3778 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3780 mode
= READ_ONCE(sqe
->len
);
3781 flags
= READ_ONCE(sqe
->open_flags
);
3782 req
->open
.how
= build_open_how(flags
, mode
);
3783 return __io_openat_prep(req
, sqe
);
3786 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3788 struct open_how __user
*how
;
3792 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3794 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3795 len
= READ_ONCE(sqe
->len
);
3796 if (len
< OPEN_HOW_SIZE_VER0
)
3799 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
3804 return __io_openat_prep(req
, sqe
);
3807 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
3809 struct open_flags op
;
3812 bool resolve_nonblock
;
3815 ret
= build_open_flags(&req
->open
.how
, &op
);
3818 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
3819 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
3820 if (issue_flags
& IO_URING_F_NONBLOCK
) {
3822 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3823 * it'll always -EAGAIN
3825 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
3827 op
.lookup_flags
|= LOOKUP_CACHED
;
3828 op
.open_flag
|= O_NONBLOCK
;
3831 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
3835 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
3838 * We could hang on to this 'fd' on retrying, but seems like
3839 * marginal gain for something that is now known to be a slower
3840 * path. So just put it, and we'll get a new one when we retry.
3844 ret
= PTR_ERR(file
);
3845 /* only retry if RESOLVE_CACHED wasn't already set by application */
3846 if (ret
== -EAGAIN
&&
3847 (!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)))
3852 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
3853 file
->f_flags
&= ~O_NONBLOCK
;
3854 fsnotify_open(file
);
3855 fd_install(ret
, file
);
3857 putname(req
->open
.filename
);
3858 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3861 __io_req_complete(req
, issue_flags
, ret
, 0);
3865 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
3867 return io_openat2(req
, issue_flags
);
3870 static int io_remove_buffers_prep(struct io_kiocb
*req
,
3871 const struct io_uring_sqe
*sqe
)
3873 struct io_provide_buf
*p
= &req
->pbuf
;
3876 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
3879 tmp
= READ_ONCE(sqe
->fd
);
3880 if (!tmp
|| tmp
> USHRT_MAX
)
3883 memset(p
, 0, sizeof(*p
));
3885 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3889 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
3890 int bgid
, unsigned nbufs
)
3894 /* shouldn't happen */
3898 /* the head kbuf is the list itself */
3899 while (!list_empty(&buf
->list
)) {
3900 struct io_buffer
*nxt
;
3902 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
3903 list_del(&nxt
->list
);
3910 xa_erase(&ctx
->io_buffers
, bgid
);
3915 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
3917 struct io_provide_buf
*p
= &req
->pbuf
;
3918 struct io_ring_ctx
*ctx
= req
->ctx
;
3919 struct io_buffer
*head
;
3921 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3923 io_ring_submit_lock(ctx
, !force_nonblock
);
3925 lockdep_assert_held(&ctx
->uring_lock
);
3928 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
3930 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
3934 /* complete before unlock, IOPOLL may need the lock */
3935 __io_req_complete(req
, issue_flags
, ret
, 0);
3936 io_ring_submit_unlock(ctx
, !force_nonblock
);
3940 static int io_provide_buffers_prep(struct io_kiocb
*req
,
3941 const struct io_uring_sqe
*sqe
)
3943 unsigned long size
, tmp_check
;
3944 struct io_provide_buf
*p
= &req
->pbuf
;
3947 if (sqe
->ioprio
|| sqe
->rw_flags
)
3950 tmp
= READ_ONCE(sqe
->fd
);
3951 if (!tmp
|| tmp
> USHRT_MAX
)
3954 p
->addr
= READ_ONCE(sqe
->addr
);
3955 p
->len
= READ_ONCE(sqe
->len
);
3957 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
3960 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
3963 size
= (unsigned long)p
->len
* p
->nbufs
;
3964 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
3967 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3968 tmp
= READ_ONCE(sqe
->off
);
3969 if (tmp
> USHRT_MAX
)
3975 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
3977 struct io_buffer
*buf
;
3978 u64 addr
= pbuf
->addr
;
3979 int i
, bid
= pbuf
->bid
;
3981 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
3982 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
3987 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
3992 INIT_LIST_HEAD(&buf
->list
);
3995 list_add_tail(&buf
->list
, &(*head
)->list
);
3999 return i
? i
: -ENOMEM
;
4002 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4004 struct io_provide_buf
*p
= &req
->pbuf
;
4005 struct io_ring_ctx
*ctx
= req
->ctx
;
4006 struct io_buffer
*head
, *list
;
4008 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4010 io_ring_submit_lock(ctx
, !force_nonblock
);
4012 lockdep_assert_held(&ctx
->uring_lock
);
4014 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4016 ret
= io_add_buffers(p
, &head
);
4017 if (ret
>= 0 && !list
) {
4018 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4020 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4024 /* complete before unlock, IOPOLL may need the lock */
4025 __io_req_complete(req
, issue_flags
, ret
, 0);
4026 io_ring_submit_unlock(ctx
, !force_nonblock
);
4030 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4031 const struct io_uring_sqe
*sqe
)
4033 #if defined(CONFIG_EPOLL)
4034 if (sqe
->ioprio
|| sqe
->buf_index
)
4036 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4039 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4040 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4041 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4043 if (ep_op_has_event(req
->epoll
.op
)) {
4044 struct epoll_event __user
*ev
;
4046 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4047 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4057 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4059 #if defined(CONFIG_EPOLL)
4060 struct io_epoll
*ie
= &req
->epoll
;
4062 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4064 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4065 if (force_nonblock
&& ret
== -EAGAIN
)
4070 __io_req_complete(req
, issue_flags
, ret
, 0);
4077 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4079 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4080 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
4082 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4085 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4086 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4087 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4094 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4096 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4097 struct io_madvise
*ma
= &req
->madvise
;
4100 if (issue_flags
& IO_URING_F_NONBLOCK
)
4103 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4106 io_req_complete(req
, ret
);
4113 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4115 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
4117 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4120 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4121 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4122 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4126 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4128 struct io_fadvise
*fa
= &req
->fadvise
;
4131 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4132 switch (fa
->advice
) {
4133 case POSIX_FADV_NORMAL
:
4134 case POSIX_FADV_RANDOM
:
4135 case POSIX_FADV_SEQUENTIAL
:
4142 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4145 __io_req_complete(req
, issue_flags
, ret
, 0);
4149 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4151 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4153 if (sqe
->ioprio
|| sqe
->buf_index
)
4155 if (req
->flags
& REQ_F_FIXED_FILE
)
4158 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4159 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4160 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4161 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4162 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4167 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4169 struct io_statx
*ctx
= &req
->statx
;
4172 if (issue_flags
& IO_URING_F_NONBLOCK
)
4175 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4180 io_req_complete(req
, ret
);
4184 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4186 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4188 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4189 sqe
->rw_flags
|| sqe
->buf_index
)
4191 if (req
->flags
& REQ_F_FIXED_FILE
)
4194 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4198 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4200 struct files_struct
*files
= current
->files
;
4201 struct io_close
*close
= &req
->close
;
4202 struct fdtable
*fdt
;
4203 struct file
*file
= NULL
;
4206 spin_lock(&files
->file_lock
);
4207 fdt
= files_fdtable(files
);
4208 if (close
->fd
>= fdt
->max_fds
) {
4209 spin_unlock(&files
->file_lock
);
4212 file
= fdt
->fd
[close
->fd
];
4213 if (!file
|| file
->f_op
== &io_uring_fops
) {
4214 spin_unlock(&files
->file_lock
);
4219 /* if the file has a flush method, be safe and punt to async */
4220 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4221 spin_unlock(&files
->file_lock
);
4225 ret
= __close_fd_get_file(close
->fd
, &file
);
4226 spin_unlock(&files
->file_lock
);
4233 /* No ->flush() or already async, safely close from here */
4234 ret
= filp_close(file
, current
->files
);
4240 __io_req_complete(req
, issue_flags
, ret
, 0);
4244 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4246 struct io_ring_ctx
*ctx
= req
->ctx
;
4248 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4250 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
4253 req
->sync
.off
= READ_ONCE(sqe
->off
);
4254 req
->sync
.len
= READ_ONCE(sqe
->len
);
4255 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4259 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4263 /* sync_file_range always requires a blocking context */
4264 if (issue_flags
& IO_URING_F_NONBLOCK
)
4267 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4271 io_req_complete(req
, ret
);
4275 #if defined(CONFIG_NET)
4276 static int io_setup_async_msg(struct io_kiocb
*req
,
4277 struct io_async_msghdr
*kmsg
)
4279 struct io_async_msghdr
*async_msg
= req
->async_data
;
4283 if (io_alloc_async_data(req
)) {
4284 kfree(kmsg
->free_iov
);
4287 async_msg
= req
->async_data
;
4288 req
->flags
|= REQ_F_NEED_CLEANUP
;
4289 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4290 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4291 /* if were using fast_iov, set it to the new one */
4292 if (!async_msg
->free_iov
)
4293 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4298 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4299 struct io_async_msghdr
*iomsg
)
4301 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4302 iomsg
->free_iov
= iomsg
->fast_iov
;
4303 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4304 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4307 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4311 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4313 req
->flags
|= REQ_F_NEED_CLEANUP
;
4317 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4319 struct io_sr_msg
*sr
= &req
->sr_msg
;
4321 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4324 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4325 sr
->len
= READ_ONCE(sqe
->len
);
4326 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4327 if (sr
->msg_flags
& MSG_DONTWAIT
)
4328 req
->flags
|= REQ_F_NOWAIT
;
4330 #ifdef CONFIG_COMPAT
4331 if (req
->ctx
->compat
)
4332 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4337 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4339 struct io_async_msghdr iomsg
, *kmsg
;
4340 struct socket
*sock
;
4345 sock
= sock_from_file(req
->file
);
4346 if (unlikely(!sock
))
4349 kmsg
= req
->async_data
;
4351 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4357 flags
= req
->sr_msg
.msg_flags
;
4358 if (issue_flags
& IO_URING_F_NONBLOCK
)
4359 flags
|= MSG_DONTWAIT
;
4360 if (flags
& MSG_WAITALL
)
4361 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4363 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4364 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4365 return io_setup_async_msg(req
, kmsg
);
4366 if (ret
== -ERESTARTSYS
)
4369 /* fast path, check for non-NULL to avoid function call */
4371 kfree(kmsg
->free_iov
);
4372 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4375 __io_req_complete(req
, issue_flags
, ret
, 0);
4379 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4381 struct io_sr_msg
*sr
= &req
->sr_msg
;
4384 struct socket
*sock
;
4389 sock
= sock_from_file(req
->file
);
4390 if (unlikely(!sock
))
4393 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4397 msg
.msg_name
= NULL
;
4398 msg
.msg_control
= NULL
;
4399 msg
.msg_controllen
= 0;
4400 msg
.msg_namelen
= 0;
4402 flags
= req
->sr_msg
.msg_flags
;
4403 if (issue_flags
& IO_URING_F_NONBLOCK
)
4404 flags
|= MSG_DONTWAIT
;
4405 if (flags
& MSG_WAITALL
)
4406 min_ret
= iov_iter_count(&msg
.msg_iter
);
4408 msg
.msg_flags
= flags
;
4409 ret
= sock_sendmsg(sock
, &msg
);
4410 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4412 if (ret
== -ERESTARTSYS
)
4417 __io_req_complete(req
, issue_flags
, ret
, 0);
4421 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4422 struct io_async_msghdr
*iomsg
)
4424 struct io_sr_msg
*sr
= &req
->sr_msg
;
4425 struct iovec __user
*uiov
;
4429 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4430 &iomsg
->uaddr
, &uiov
, &iov_len
);
4434 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4437 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4439 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4440 iomsg
->free_iov
= NULL
;
4442 iomsg
->free_iov
= iomsg
->fast_iov
;
4443 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4444 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4453 #ifdef CONFIG_COMPAT
4454 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4455 struct io_async_msghdr
*iomsg
)
4457 struct io_sr_msg
*sr
= &req
->sr_msg
;
4458 struct compat_iovec __user
*uiov
;
4463 ret
= __get_compat_msghdr(&iomsg
->msg
, sr
->umsg_compat
, &iomsg
->uaddr
,
4468 uiov
= compat_ptr(ptr
);
4469 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4470 compat_ssize_t clen
;
4474 if (!access_ok(uiov
, sizeof(*uiov
)))
4476 if (__get_user(clen
, &uiov
->iov_len
))
4481 iomsg
->free_iov
= NULL
;
4483 iomsg
->free_iov
= iomsg
->fast_iov
;
4484 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4485 UIO_FASTIOV
, &iomsg
->free_iov
,
4486 &iomsg
->msg
.msg_iter
, true);
4495 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4496 struct io_async_msghdr
*iomsg
)
4498 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4500 #ifdef CONFIG_COMPAT
4501 if (req
->ctx
->compat
)
4502 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4505 return __io_recvmsg_copy_hdr(req
, iomsg
);
4508 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4511 struct io_sr_msg
*sr
= &req
->sr_msg
;
4512 struct io_buffer
*kbuf
;
4514 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4519 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4523 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4525 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4528 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
4532 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
4534 req
->flags
|= REQ_F_NEED_CLEANUP
;
4538 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4540 struct io_sr_msg
*sr
= &req
->sr_msg
;
4542 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4545 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4546 sr
->len
= READ_ONCE(sqe
->len
);
4547 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4548 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4549 if (sr
->msg_flags
& MSG_DONTWAIT
)
4550 req
->flags
|= REQ_F_NOWAIT
;
4552 #ifdef CONFIG_COMPAT
4553 if (req
->ctx
->compat
)
4554 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4559 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4561 struct io_async_msghdr iomsg
, *kmsg
;
4562 struct socket
*sock
;
4563 struct io_buffer
*kbuf
;
4566 int ret
, cflags
= 0;
4567 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4569 sock
= sock_from_file(req
->file
);
4570 if (unlikely(!sock
))
4573 kmsg
= req
->async_data
;
4575 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
4581 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4582 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4584 return PTR_ERR(kbuf
);
4585 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4586 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
4587 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
4588 1, req
->sr_msg
.len
);
4591 flags
= req
->sr_msg
.msg_flags
;
4593 flags
|= MSG_DONTWAIT
;
4594 if (flags
& MSG_WAITALL
)
4595 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4597 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
4598 kmsg
->uaddr
, flags
);
4599 if (force_nonblock
&& ret
== -EAGAIN
)
4600 return io_setup_async_msg(req
, kmsg
);
4601 if (ret
== -ERESTARTSYS
)
4604 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4605 cflags
= io_put_recv_kbuf(req
);
4606 /* fast path, check for non-NULL to avoid function call */
4608 kfree(kmsg
->free_iov
);
4609 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4610 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4612 __io_req_complete(req
, issue_flags
, ret
, cflags
);
4616 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
4618 struct io_buffer
*kbuf
;
4619 struct io_sr_msg
*sr
= &req
->sr_msg
;
4621 void __user
*buf
= sr
->buf
;
4622 struct socket
*sock
;
4626 int ret
, cflags
= 0;
4627 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4629 sock
= sock_from_file(req
->file
);
4630 if (unlikely(!sock
))
4633 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4634 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
4636 return PTR_ERR(kbuf
);
4637 buf
= u64_to_user_ptr(kbuf
->addr
);
4640 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4644 msg
.msg_name
= NULL
;
4645 msg
.msg_control
= NULL
;
4646 msg
.msg_controllen
= 0;
4647 msg
.msg_namelen
= 0;
4648 msg
.msg_iocb
= NULL
;
4651 flags
= req
->sr_msg
.msg_flags
;
4653 flags
|= MSG_DONTWAIT
;
4654 if (flags
& MSG_WAITALL
)
4655 min_ret
= iov_iter_count(&msg
.msg_iter
);
4657 ret
= sock_recvmsg(sock
, &msg
, flags
);
4658 if (force_nonblock
&& ret
== -EAGAIN
)
4660 if (ret
== -ERESTARTSYS
)
4663 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
4664 cflags
= io_put_recv_kbuf(req
);
4665 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
4667 __io_req_complete(req
, issue_flags
, ret
, cflags
);
4671 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4673 struct io_accept
*accept
= &req
->accept
;
4675 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4677 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4680 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4681 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4682 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4683 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4687 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
4689 struct io_accept
*accept
= &req
->accept
;
4690 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4691 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4694 if (req
->file
->f_flags
& O_NONBLOCK
)
4695 req
->flags
|= REQ_F_NOWAIT
;
4697 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4698 accept
->addr_len
, accept
->flags
,
4700 if (ret
== -EAGAIN
&& force_nonblock
)
4703 if (ret
== -ERESTARTSYS
)
4707 __io_req_complete(req
, issue_flags
, ret
, 0);
4711 static int io_connect_prep_async(struct io_kiocb
*req
)
4713 struct io_async_connect
*io
= req
->async_data
;
4714 struct io_connect
*conn
= &req
->connect
;
4716 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
4719 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4721 struct io_connect
*conn
= &req
->connect
;
4723 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4725 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4728 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4729 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4733 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
4735 struct io_async_connect __io
, *io
;
4736 unsigned file_flags
;
4738 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4740 if (req
->async_data
) {
4741 io
= req
->async_data
;
4743 ret
= move_addr_to_kernel(req
->connect
.addr
,
4744 req
->connect
.addr_len
,
4751 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4753 ret
= __sys_connect_file(req
->file
, &io
->address
,
4754 req
->connect
.addr_len
, file_flags
);
4755 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4756 if (req
->async_data
)
4758 if (io_alloc_async_data(req
)) {
4762 memcpy(req
->async_data
, &__io
, sizeof(__io
));
4765 if (ret
== -ERESTARTSYS
)
4770 __io_req_complete(req
, issue_flags
, ret
, 0);
4773 #else /* !CONFIG_NET */
4774 #define IO_NETOP_FN(op) \
4775 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4777 return -EOPNOTSUPP; \
4780 #define IO_NETOP_PREP(op) \
4782 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4784 return -EOPNOTSUPP; \
4787 #define IO_NETOP_PREP_ASYNC(op) \
4789 static int io_##op##_prep_async(struct io_kiocb *req) \
4791 return -EOPNOTSUPP; \
4794 IO_NETOP_PREP_ASYNC(sendmsg
);
4795 IO_NETOP_PREP_ASYNC(recvmsg
);
4796 IO_NETOP_PREP_ASYNC(connect
);
4797 IO_NETOP_PREP(accept
);
4800 #endif /* CONFIG_NET */
4802 struct io_poll_table
{
4803 struct poll_table_struct pt
;
4804 struct io_kiocb
*req
;
4808 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
4809 __poll_t mask
, io_req_tw_func_t func
)
4811 /* for instances that support it check for an event match first: */
4812 if (mask
&& !(mask
& poll
->events
))
4815 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
4817 list_del_init(&poll
->wait
.entry
);
4820 req
->io_task_work
.func
= func
;
4823 * If this fails, then the task is exiting. When a task exits, the
4824 * work gets canceled, so just cancel this request as well instead
4825 * of executing it. We can't safely execute it anyway, as we may not
4826 * have the needed state needed for it anyway.
4828 io_req_task_work_add(req
);
4832 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
4833 __acquires(&req
->ctx
->completion_lock
)
4835 struct io_ring_ctx
*ctx
= req
->ctx
;
4837 if (unlikely(req
->task
->flags
& PF_EXITING
))
4838 WRITE_ONCE(poll
->canceled
, true);
4840 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4841 struct poll_table_struct pt
= { ._key
= poll
->events
};
4843 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
4846 spin_lock_irq(&ctx
->completion_lock
);
4847 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4848 add_wait_queue(poll
->head
, &poll
->wait
);
4855 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
4857 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4858 if (req
->opcode
== IORING_OP_POLL_ADD
)
4859 return req
->async_data
;
4860 return req
->apoll
->double_poll
;
4863 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
4865 if (req
->opcode
== IORING_OP_POLL_ADD
)
4867 return &req
->apoll
->poll
;
4870 static void io_poll_remove_double(struct io_kiocb
*req
)
4871 __must_hold(&req
->ctx
->completion_lock
)
4873 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
4875 lockdep_assert_held(&req
->ctx
->completion_lock
);
4877 if (poll
&& poll
->head
) {
4878 struct wait_queue_head
*head
= poll
->head
;
4880 spin_lock(&head
->lock
);
4881 list_del_init(&poll
->wait
.entry
);
4882 if (poll
->wait
.private)
4885 spin_unlock(&head
->lock
);
4889 static bool io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
4890 __must_hold(&req
->ctx
->completion_lock
)
4892 struct io_ring_ctx
*ctx
= req
->ctx
;
4893 unsigned flags
= IORING_CQE_F_MORE
;
4896 if (READ_ONCE(req
->poll
.canceled
)) {
4898 req
->poll
.events
|= EPOLLONESHOT
;
4900 error
= mangle_poll(mask
);
4902 if (req
->poll
.events
& EPOLLONESHOT
)
4904 if (!io_cqring_fill_event(ctx
, req
->user_data
, error
, flags
)) {
4905 io_poll_remove_waitqs(req
);
4906 req
->poll
.done
= true;
4909 if (flags
& IORING_CQE_F_MORE
)
4912 io_commit_cqring(ctx
);
4913 return !(flags
& IORING_CQE_F_MORE
);
4916 static void io_poll_task_func(struct io_kiocb
*req
)
4918 struct io_ring_ctx
*ctx
= req
->ctx
;
4919 struct io_kiocb
*nxt
;
4921 if (io_poll_rewait(req
, &req
->poll
)) {
4922 spin_unlock_irq(&ctx
->completion_lock
);
4926 done
= io_poll_complete(req
, req
->result
);
4928 hash_del(&req
->hash_node
);
4931 add_wait_queue(req
->poll
.head
, &req
->poll
.wait
);
4933 spin_unlock_irq(&ctx
->completion_lock
);
4934 io_cqring_ev_posted(ctx
);
4937 nxt
= io_put_req_find_next(req
);
4939 io_req_task_submit(nxt
);
4944 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
4945 int sync
, void *key
)
4947 struct io_kiocb
*req
= wait
->private;
4948 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
4949 __poll_t mask
= key_to_poll(key
);
4951 /* for instances that support it check for an event match first: */
4952 if (mask
&& !(mask
& poll
->events
))
4954 if (!(poll
->events
& EPOLLONESHOT
))
4955 return poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
4957 list_del_init(&wait
->entry
);
4962 spin_lock(&poll
->head
->lock
);
4963 done
= list_empty(&poll
->wait
.entry
);
4965 list_del_init(&poll
->wait
.entry
);
4966 /* make sure double remove sees this as being gone */
4967 wait
->private = NULL
;
4968 spin_unlock(&poll
->head
->lock
);
4970 /* use wait func handler, so it matches the rq type */
4971 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
4978 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
4979 wait_queue_func_t wake_func
)
4983 poll
->canceled
= false;
4984 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4985 /* mask in events that we always want/need */
4986 poll
->events
= events
| IO_POLL_UNMASK
;
4987 INIT_LIST_HEAD(&poll
->wait
.entry
);
4988 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
4991 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
4992 struct wait_queue_head
*head
,
4993 struct io_poll_iocb
**poll_ptr
)
4995 struct io_kiocb
*req
= pt
->req
;
4998 * If poll->head is already set, it's because the file being polled
4999 * uses multiple waitqueues for poll handling (eg one for read, one
5000 * for write). Setup a separate io_poll_iocb if this happens.
5002 if (unlikely(poll
->head
)) {
5003 struct io_poll_iocb
*poll_one
= poll
;
5005 /* already have a 2nd entry, fail a third attempt */
5007 pt
->error
= -EINVAL
;
5011 * Can't handle multishot for double wait for now, turn it
5012 * into one-shot mode.
5014 if (!(poll_one
->events
& EPOLLONESHOT
))
5015 poll_one
->events
|= EPOLLONESHOT
;
5016 /* double add on the same waitqueue head, ignore */
5017 if (poll_one
->head
== head
)
5019 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5021 pt
->error
= -ENOMEM
;
5024 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5026 poll
->wait
.private = req
;
5033 if (poll
->events
& EPOLLEXCLUSIVE
)
5034 add_wait_queue_exclusive(head
, &poll
->wait
);
5036 add_wait_queue(head
, &poll
->wait
);
5039 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5040 struct poll_table_struct
*p
)
5042 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5043 struct async_poll
*apoll
= pt
->req
->apoll
;
5045 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5048 static void io_async_task_func(struct io_kiocb
*req
)
5050 struct async_poll
*apoll
= req
->apoll
;
5051 struct io_ring_ctx
*ctx
= req
->ctx
;
5053 trace_io_uring_task_run(req
->ctx
, req
, req
->opcode
, req
->user_data
);
5055 if (io_poll_rewait(req
, &apoll
->poll
)) {
5056 spin_unlock_irq(&ctx
->completion_lock
);
5060 hash_del(&req
->hash_node
);
5061 io_poll_remove_double(req
);
5062 spin_unlock_irq(&ctx
->completion_lock
);
5064 if (!READ_ONCE(apoll
->poll
.canceled
))
5065 io_req_task_submit(req
);
5067 io_req_complete_failed(req
, -ECANCELED
);
5070 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5073 struct io_kiocb
*req
= wait
->private;
5074 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5076 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5079 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5082 static void io_poll_req_insert(struct io_kiocb
*req
)
5084 struct io_ring_ctx
*ctx
= req
->ctx
;
5085 struct hlist_head
*list
;
5087 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5088 hlist_add_head(&req
->hash_node
, list
);
5091 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5092 struct io_poll_iocb
*poll
,
5093 struct io_poll_table
*ipt
, __poll_t mask
,
5094 wait_queue_func_t wake_func
)
5095 __acquires(&ctx
->completion_lock
)
5097 struct io_ring_ctx
*ctx
= req
->ctx
;
5098 bool cancel
= false;
5100 INIT_HLIST_NODE(&req
->hash_node
);
5101 io_init_poll_iocb(poll
, mask
, wake_func
);
5102 poll
->file
= req
->file
;
5103 poll
->wait
.private = req
;
5105 ipt
->pt
._key
= mask
;
5107 ipt
->error
= -EINVAL
;
5109 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5111 spin_lock_irq(&ctx
->completion_lock
);
5112 if (likely(poll
->head
)) {
5113 spin_lock(&poll
->head
->lock
);
5114 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5120 if ((mask
&& (poll
->events
& EPOLLONESHOT
)) || ipt
->error
)
5121 list_del_init(&poll
->wait
.entry
);
5123 WRITE_ONCE(poll
->canceled
, true);
5124 else if (!poll
->done
) /* actually waiting for an event */
5125 io_poll_req_insert(req
);
5126 spin_unlock(&poll
->head
->lock
);
5138 static int io_arm_poll_handler(struct io_kiocb
*req
)
5140 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5141 struct io_ring_ctx
*ctx
= req
->ctx
;
5142 struct async_poll
*apoll
;
5143 struct io_poll_table ipt
;
5144 __poll_t ret
, mask
= EPOLLONESHOT
| POLLERR
| POLLPRI
;
5147 if (!req
->file
|| !file_can_poll(req
->file
))
5148 return IO_APOLL_ABORTED
;
5149 if (req
->flags
& REQ_F_POLLED
)
5150 return IO_APOLL_ABORTED
;
5151 if (!def
->pollin
&& !def
->pollout
)
5152 return IO_APOLL_ABORTED
;
5156 mask
|= POLLIN
| POLLRDNORM
;
5158 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5159 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5160 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5164 mask
|= POLLOUT
| POLLWRNORM
;
5167 /* if we can't nonblock try, then no point in arming a poll handler */
5168 if (!io_file_supports_async(req
, rw
))
5169 return IO_APOLL_ABORTED
;
5171 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5172 if (unlikely(!apoll
))
5173 return IO_APOLL_ABORTED
;
5174 apoll
->double_poll
= NULL
;
5176 req
->flags
|= REQ_F_POLLED
;
5177 ipt
.pt
._qproc
= io_async_queue_proc
;
5179 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5181 if (ret
|| ipt
.error
) {
5182 io_poll_remove_double(req
);
5183 spin_unlock_irq(&ctx
->completion_lock
);
5185 return IO_APOLL_READY
;
5186 return IO_APOLL_ABORTED
;
5188 spin_unlock_irq(&ctx
->completion_lock
);
5189 trace_io_uring_poll_arm(ctx
, req
, req
->opcode
, req
->user_data
,
5190 mask
, apoll
->poll
.events
);
5194 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5195 struct io_poll_iocb
*poll
, bool do_cancel
)
5196 __must_hold(&req
->ctx
->completion_lock
)
5198 bool do_complete
= false;
5202 spin_lock(&poll
->head
->lock
);
5204 WRITE_ONCE(poll
->canceled
, true);
5205 if (!list_empty(&poll
->wait
.entry
)) {
5206 list_del_init(&poll
->wait
.entry
);
5209 spin_unlock(&poll
->head
->lock
);
5210 hash_del(&req
->hash_node
);
5214 static bool io_poll_remove_waitqs(struct io_kiocb
*req
)
5215 __must_hold(&req
->ctx
->completion_lock
)
5219 io_poll_remove_double(req
);
5220 do_complete
= __io_poll_remove_one(req
, io_poll_get_single(req
), true);
5222 if (req
->opcode
!= IORING_OP_POLL_ADD
&& do_complete
) {
5223 /* non-poll requests have submit ref still */
5229 static bool io_poll_remove_one(struct io_kiocb
*req
)
5230 __must_hold(&req
->ctx
->completion_lock
)
5234 do_complete
= io_poll_remove_waitqs(req
);
5236 io_cqring_fill_event(req
->ctx
, req
->user_data
, -ECANCELED
, 0);
5237 io_commit_cqring(req
->ctx
);
5239 io_put_req_deferred(req
, 1);
5246 * Returns true if we found and killed one or more poll requests
5248 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5251 struct hlist_node
*tmp
;
5252 struct io_kiocb
*req
;
5255 spin_lock_irq(&ctx
->completion_lock
);
5256 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5257 struct hlist_head
*list
;
5259 list
= &ctx
->cancel_hash
[i
];
5260 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5261 if (io_match_task(req
, tsk
, cancel_all
))
5262 posted
+= io_poll_remove_one(req
);
5265 spin_unlock_irq(&ctx
->completion_lock
);
5268 io_cqring_ev_posted(ctx
);
5273 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5275 __must_hold(&ctx
->completion_lock
)
5277 struct hlist_head
*list
;
5278 struct io_kiocb
*req
;
5280 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5281 hlist_for_each_entry(req
, list
, hash_node
) {
5282 if (sqe_addr
!= req
->user_data
)
5284 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
5291 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5293 __must_hold(&ctx
->completion_lock
)
5295 struct io_kiocb
*req
;
5297 req
= io_poll_find(ctx
, sqe_addr
, poll_only
);
5300 if (io_poll_remove_one(req
))
5306 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
5311 events
= READ_ONCE(sqe
->poll32_events
);
5313 events
= swahw32(events
);
5315 if (!(flags
& IORING_POLL_ADD_MULTI
))
5316 events
|= EPOLLONESHOT
;
5317 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
5320 static int io_poll_update_prep(struct io_kiocb
*req
,
5321 const struct io_uring_sqe
*sqe
)
5323 struct io_poll_update
*upd
= &req
->poll_update
;
5326 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5328 if (sqe
->ioprio
|| sqe
->buf_index
)
5330 flags
= READ_ONCE(sqe
->len
);
5331 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
5332 IORING_POLL_ADD_MULTI
))
5334 /* meaningless without update */
5335 if (flags
== IORING_POLL_ADD_MULTI
)
5338 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
5339 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
5340 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
5342 upd
->new_user_data
= READ_ONCE(sqe
->off
);
5343 if (!upd
->update_user_data
&& upd
->new_user_data
)
5345 if (upd
->update_events
)
5346 upd
->events
= io_poll_parse_events(sqe
, flags
);
5347 else if (sqe
->poll32_events
)
5353 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5356 struct io_kiocb
*req
= wait
->private;
5357 struct io_poll_iocb
*poll
= &req
->poll
;
5359 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5362 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5363 struct poll_table_struct
*p
)
5365 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5367 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5370 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5372 struct io_poll_iocb
*poll
= &req
->poll
;
5375 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5377 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
5379 flags
= READ_ONCE(sqe
->len
);
5380 if (flags
& ~IORING_POLL_ADD_MULTI
)
5383 poll
->events
= io_poll_parse_events(sqe
, flags
);
5387 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5389 struct io_poll_iocb
*poll
= &req
->poll
;
5390 struct io_ring_ctx
*ctx
= req
->ctx
;
5391 struct io_poll_table ipt
;
5394 ipt
.pt
._qproc
= io_poll_queue_proc
;
5396 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5399 if (mask
) { /* no async, we'd stolen it */
5401 io_poll_complete(req
, mask
);
5403 spin_unlock_irq(&ctx
->completion_lock
);
5406 io_cqring_ev_posted(ctx
);
5407 if (poll
->events
& EPOLLONESHOT
)
5413 static int io_poll_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5415 struct io_ring_ctx
*ctx
= req
->ctx
;
5416 struct io_kiocb
*preq
;
5420 spin_lock_irq(&ctx
->completion_lock
);
5421 preq
= io_poll_find(ctx
, req
->poll_update
.old_user_data
, true);
5427 if (!req
->poll_update
.update_events
&& !req
->poll_update
.update_user_data
) {
5429 ret
= io_poll_remove_one(preq
) ? 0 : -EALREADY
;
5434 * Don't allow racy completion with singleshot, as we cannot safely
5435 * update those. For multishot, if we're racing with completion, just
5436 * let completion re-add it.
5438 completing
= !__io_poll_remove_one(preq
, &preq
->poll
, false);
5439 if (completing
&& (preq
->poll
.events
& EPOLLONESHOT
)) {
5443 /* we now have a detached poll request. reissue. */
5447 spin_unlock_irq(&ctx
->completion_lock
);
5449 io_req_complete(req
, ret
);
5452 /* only mask one event flags, keep behavior flags */
5453 if (req
->poll_update
.update_events
) {
5454 preq
->poll
.events
&= ~0xffff;
5455 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
5456 preq
->poll
.events
|= IO_POLL_UNMASK
;
5458 if (req
->poll_update
.update_user_data
)
5459 preq
->user_data
= req
->poll_update
.new_user_data
;
5460 spin_unlock_irq(&ctx
->completion_lock
);
5462 /* complete update request, we're done with it */
5463 io_req_complete(req
, ret
);
5466 ret
= io_poll_add(preq
, issue_flags
);
5469 io_req_complete(preq
, ret
);
5475 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5477 struct io_timeout_data
*data
= container_of(timer
,
5478 struct io_timeout_data
, timer
);
5479 struct io_kiocb
*req
= data
->req
;
5480 struct io_ring_ctx
*ctx
= req
->ctx
;
5481 unsigned long flags
;
5483 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5484 list_del_init(&req
->timeout
.list
);
5485 atomic_set(&req
->ctx
->cq_timeouts
,
5486 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5488 io_cqring_fill_event(ctx
, req
->user_data
, -ETIME
, 0);
5489 io_commit_cqring(ctx
);
5490 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5492 io_cqring_ev_posted(ctx
);
5495 return HRTIMER_NORESTART
;
5498 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5500 __must_hold(&ctx
->completion_lock
)
5502 struct io_timeout_data
*io
;
5503 struct io_kiocb
*req
;
5506 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5507 found
= user_data
== req
->user_data
;
5512 return ERR_PTR(-ENOENT
);
5514 io
= req
->async_data
;
5515 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
5516 return ERR_PTR(-EALREADY
);
5517 list_del_init(&req
->timeout
.list
);
5521 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5522 __must_hold(&ctx
->completion_lock
)
5524 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5527 return PTR_ERR(req
);
5530 io_cqring_fill_event(ctx
, req
->user_data
, -ECANCELED
, 0);
5531 io_put_req_deferred(req
, 1);
5535 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5536 struct timespec64
*ts
, enum hrtimer_mode mode
)
5537 __must_hold(&ctx
->completion_lock
)
5539 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5540 struct io_timeout_data
*data
;
5543 return PTR_ERR(req
);
5545 req
->timeout
.off
= 0; /* noseq */
5546 data
= req
->async_data
;
5547 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
5548 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, mode
);
5549 data
->timer
.function
= io_timeout_fn
;
5550 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
5554 static int io_timeout_remove_prep(struct io_kiocb
*req
,
5555 const struct io_uring_sqe
*sqe
)
5557 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5559 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5561 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5563 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
5566 tr
->addr
= READ_ONCE(sqe
->addr
);
5567 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
5568 if (tr
->flags
& IORING_TIMEOUT_UPDATE
) {
5569 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE
|IORING_TIMEOUT_ABS
))
5571 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
5573 } else if (tr
->flags
) {
5574 /* timeout removal doesn't support flags */
5581 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
5583 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
5588 * Remove or update an existing timeout command
5590 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
5592 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
5593 struct io_ring_ctx
*ctx
= req
->ctx
;
5596 spin_lock_irq(&ctx
->completion_lock
);
5597 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
))
5598 ret
= io_timeout_cancel(ctx
, tr
->addr
);
5600 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
,
5601 io_translate_timeout_mode(tr
->flags
));
5603 io_cqring_fill_event(ctx
, req
->user_data
, ret
, 0);
5604 io_commit_cqring(ctx
);
5605 spin_unlock_irq(&ctx
->completion_lock
);
5606 io_cqring_ev_posted(ctx
);
5613 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5614 bool is_timeout_link
)
5616 struct io_timeout_data
*data
;
5618 u32 off
= READ_ONCE(sqe
->off
);
5620 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5622 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
5624 if (off
&& is_timeout_link
)
5626 flags
= READ_ONCE(sqe
->timeout_flags
);
5627 if (flags
& ~IORING_TIMEOUT_ABS
)
5630 req
->timeout
.off
= off
;
5631 if (unlikely(off
&& !req
->ctx
->off_timeout_used
))
5632 req
->ctx
->off_timeout_used
= true;
5634 if (!req
->async_data
&& io_alloc_async_data(req
))
5637 data
= req
->async_data
;
5640 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5643 data
->mode
= io_translate_timeout_mode(flags
);
5644 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5645 if (is_timeout_link
)
5646 io_req_track_inflight(req
);
5650 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
5652 struct io_ring_ctx
*ctx
= req
->ctx
;
5653 struct io_timeout_data
*data
= req
->async_data
;
5654 struct list_head
*entry
;
5655 u32 tail
, off
= req
->timeout
.off
;
5657 spin_lock_irq(&ctx
->completion_lock
);
5660 * sqe->off holds how many events that need to occur for this
5661 * timeout event to be satisfied. If it isn't set, then this is
5662 * a pure timeout request, sequence isn't used.
5664 if (io_is_timeout_noseq(req
)) {
5665 entry
= ctx
->timeout_list
.prev
;
5669 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5670 req
->timeout
.target_seq
= tail
+ off
;
5672 /* Update the last seq here in case io_flush_timeouts() hasn't.
5673 * This is safe because ->completion_lock is held, and submissions
5674 * and completions are never mixed in the same ->completion_lock section.
5676 ctx
->cq_last_tm_flush
= tail
;
5679 * Insertion sort, ensuring the first entry in the list is always
5680 * the one we need first.
5682 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5683 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
5686 if (io_is_timeout_noseq(nxt
))
5688 /* nxt.seq is behind @tail, otherwise would've been completed */
5689 if (off
>= nxt
->timeout
.target_seq
- tail
)
5693 list_add(&req
->timeout
.list
, entry
);
5694 data
->timer
.function
= io_timeout_fn
;
5695 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5696 spin_unlock_irq(&ctx
->completion_lock
);
5700 struct io_cancel_data
{
5701 struct io_ring_ctx
*ctx
;
5705 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5707 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5708 struct io_cancel_data
*cd
= data
;
5710 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
5713 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
5714 struct io_ring_ctx
*ctx
)
5716 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
5717 enum io_wq_cancel cancel_ret
;
5720 if (!tctx
|| !tctx
->io_wq
)
5723 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
5724 switch (cancel_ret
) {
5725 case IO_WQ_CANCEL_OK
:
5728 case IO_WQ_CANCEL_RUNNING
:
5731 case IO_WQ_CANCEL_NOTFOUND
:
5739 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5740 struct io_kiocb
*req
, __u64 sqe_addr
,
5743 unsigned long flags
;
5746 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
5747 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5750 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5753 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
5757 io_cqring_fill_event(ctx
, req
->user_data
, ret
, 0);
5758 io_commit_cqring(ctx
);
5759 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5760 io_cqring_ev_posted(ctx
);
5766 static int io_async_cancel_prep(struct io_kiocb
*req
,
5767 const struct io_uring_sqe
*sqe
)
5769 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5771 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5773 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
)
5776 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5780 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
5782 struct io_ring_ctx
*ctx
= req
->ctx
;
5783 u64 sqe_addr
= req
->cancel
.addr
;
5784 struct io_tctx_node
*node
;
5787 /* tasks should wait for their io-wq threads, so safe w/o sync */
5788 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
5789 spin_lock_irq(&ctx
->completion_lock
);
5792 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5795 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
5798 spin_unlock_irq(&ctx
->completion_lock
);
5800 /* slow path, try all io-wq's */
5801 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
5803 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
5804 struct io_uring_task
*tctx
= node
->task
->io_uring
;
5806 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
5810 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
5812 spin_lock_irq(&ctx
->completion_lock
);
5814 io_cqring_fill_event(ctx
, req
->user_data
, ret
, 0);
5815 io_commit_cqring(ctx
);
5816 spin_unlock_irq(&ctx
->completion_lock
);
5817 io_cqring_ev_posted(ctx
);
5825 static int io_rsrc_update_prep(struct io_kiocb
*req
,
5826 const struct io_uring_sqe
*sqe
)
5828 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
5830 if (sqe
->ioprio
|| sqe
->rw_flags
)
5833 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
5834 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
5835 if (!req
->rsrc_update
.nr_args
)
5837 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
5841 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5843 struct io_ring_ctx
*ctx
= req
->ctx
;
5844 struct io_uring_rsrc_update2 up
;
5847 if (issue_flags
& IO_URING_F_NONBLOCK
)
5850 up
.offset
= req
->rsrc_update
.offset
;
5851 up
.data
= req
->rsrc_update
.arg
;
5856 mutex_lock(&ctx
->uring_lock
);
5857 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
5858 &up
, req
->rsrc_update
.nr_args
);
5859 mutex_unlock(&ctx
->uring_lock
);
5863 __io_req_complete(req
, issue_flags
, ret
, 0);
5867 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5869 switch (req
->opcode
) {
5872 case IORING_OP_READV
:
5873 case IORING_OP_READ_FIXED
:
5874 case IORING_OP_READ
:
5875 return io_read_prep(req
, sqe
);
5876 case IORING_OP_WRITEV
:
5877 case IORING_OP_WRITE_FIXED
:
5878 case IORING_OP_WRITE
:
5879 return io_write_prep(req
, sqe
);
5880 case IORING_OP_POLL_ADD
:
5881 return io_poll_add_prep(req
, sqe
);
5882 case IORING_OP_POLL_REMOVE
:
5883 return io_poll_update_prep(req
, sqe
);
5884 case IORING_OP_FSYNC
:
5885 return io_fsync_prep(req
, sqe
);
5886 case IORING_OP_SYNC_FILE_RANGE
:
5887 return io_sfr_prep(req
, sqe
);
5888 case IORING_OP_SENDMSG
:
5889 case IORING_OP_SEND
:
5890 return io_sendmsg_prep(req
, sqe
);
5891 case IORING_OP_RECVMSG
:
5892 case IORING_OP_RECV
:
5893 return io_recvmsg_prep(req
, sqe
);
5894 case IORING_OP_CONNECT
:
5895 return io_connect_prep(req
, sqe
);
5896 case IORING_OP_TIMEOUT
:
5897 return io_timeout_prep(req
, sqe
, false);
5898 case IORING_OP_TIMEOUT_REMOVE
:
5899 return io_timeout_remove_prep(req
, sqe
);
5900 case IORING_OP_ASYNC_CANCEL
:
5901 return io_async_cancel_prep(req
, sqe
);
5902 case IORING_OP_LINK_TIMEOUT
:
5903 return io_timeout_prep(req
, sqe
, true);
5904 case IORING_OP_ACCEPT
:
5905 return io_accept_prep(req
, sqe
);
5906 case IORING_OP_FALLOCATE
:
5907 return io_fallocate_prep(req
, sqe
);
5908 case IORING_OP_OPENAT
:
5909 return io_openat_prep(req
, sqe
);
5910 case IORING_OP_CLOSE
:
5911 return io_close_prep(req
, sqe
);
5912 case IORING_OP_FILES_UPDATE
:
5913 return io_rsrc_update_prep(req
, sqe
);
5914 case IORING_OP_STATX
:
5915 return io_statx_prep(req
, sqe
);
5916 case IORING_OP_FADVISE
:
5917 return io_fadvise_prep(req
, sqe
);
5918 case IORING_OP_MADVISE
:
5919 return io_madvise_prep(req
, sqe
);
5920 case IORING_OP_OPENAT2
:
5921 return io_openat2_prep(req
, sqe
);
5922 case IORING_OP_EPOLL_CTL
:
5923 return io_epoll_ctl_prep(req
, sqe
);
5924 case IORING_OP_SPLICE
:
5925 return io_splice_prep(req
, sqe
);
5926 case IORING_OP_PROVIDE_BUFFERS
:
5927 return io_provide_buffers_prep(req
, sqe
);
5928 case IORING_OP_REMOVE_BUFFERS
:
5929 return io_remove_buffers_prep(req
, sqe
);
5931 return io_tee_prep(req
, sqe
);
5932 case IORING_OP_SHUTDOWN
:
5933 return io_shutdown_prep(req
, sqe
);
5934 case IORING_OP_RENAMEAT
:
5935 return io_renameat_prep(req
, sqe
);
5936 case IORING_OP_UNLINKAT
:
5937 return io_unlinkat_prep(req
, sqe
);
5940 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5945 static int io_req_prep_async(struct io_kiocb
*req
)
5947 if (!io_op_defs
[req
->opcode
].needs_async_setup
)
5949 if (WARN_ON_ONCE(req
->async_data
))
5951 if (io_alloc_async_data(req
))
5954 switch (req
->opcode
) {
5955 case IORING_OP_READV
:
5956 return io_rw_prep_async(req
, READ
);
5957 case IORING_OP_WRITEV
:
5958 return io_rw_prep_async(req
, WRITE
);
5959 case IORING_OP_SENDMSG
:
5960 return io_sendmsg_prep_async(req
);
5961 case IORING_OP_RECVMSG
:
5962 return io_recvmsg_prep_async(req
);
5963 case IORING_OP_CONNECT
:
5964 return io_connect_prep_async(req
);
5966 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
5971 static u32
io_get_sequence(struct io_kiocb
*req
)
5973 u32 seq
= req
->ctx
->cached_sq_head
;
5975 /* need original cached_sq_head, but it was increased for each req */
5976 io_for_each_link(req
, req
)
5981 static bool io_drain_req(struct io_kiocb
*req
)
5983 struct io_kiocb
*pos
;
5984 struct io_ring_ctx
*ctx
= req
->ctx
;
5985 struct io_defer_entry
*de
;
5990 * If we need to drain a request in the middle of a link, drain the
5991 * head request and the next request/link after the current link.
5992 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
5993 * maintained for every request of our link.
5995 if (ctx
->drain_next
) {
5996 req
->flags
|= REQ_F_IO_DRAIN
;
5997 ctx
->drain_next
= false;
5999 /* not interested in head, start from the first linked */
6000 io_for_each_link(pos
, req
->link
) {
6001 if (pos
->flags
& REQ_F_IO_DRAIN
) {
6002 ctx
->drain_next
= true;
6003 req
->flags
|= REQ_F_IO_DRAIN
;
6008 /* Still need defer if there is pending req in defer list. */
6009 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6010 !(req
->flags
& REQ_F_IO_DRAIN
))) {
6011 ctx
->drain_active
= false;
6015 seq
= io_get_sequence(req
);
6016 /* Still a chance to pass the sequence check */
6017 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6020 ret
= io_req_prep_async(req
);
6023 io_prep_async_link(req
);
6024 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6028 io_req_complete_failed(req
, ret
);
6032 spin_lock_irq(&ctx
->completion_lock
);
6033 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6034 spin_unlock_irq(&ctx
->completion_lock
);
6036 io_queue_async_work(req
);
6040 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6043 list_add_tail(&de
->list
, &ctx
->defer_list
);
6044 spin_unlock_irq(&ctx
->completion_lock
);
6048 static void io_clean_op(struct io_kiocb
*req
)
6050 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6051 switch (req
->opcode
) {
6052 case IORING_OP_READV
:
6053 case IORING_OP_READ_FIXED
:
6054 case IORING_OP_READ
:
6055 kfree((void *)(unsigned long)req
->rw
.addr
);
6057 case IORING_OP_RECVMSG
:
6058 case IORING_OP_RECV
:
6059 kfree(req
->sr_msg
.kbuf
);
6064 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6065 switch (req
->opcode
) {
6066 case IORING_OP_READV
:
6067 case IORING_OP_READ_FIXED
:
6068 case IORING_OP_READ
:
6069 case IORING_OP_WRITEV
:
6070 case IORING_OP_WRITE_FIXED
:
6071 case IORING_OP_WRITE
: {
6072 struct io_async_rw
*io
= req
->async_data
;
6074 kfree(io
->free_iovec
);
6077 case IORING_OP_RECVMSG
:
6078 case IORING_OP_SENDMSG
: {
6079 struct io_async_msghdr
*io
= req
->async_data
;
6081 kfree(io
->free_iov
);
6084 case IORING_OP_SPLICE
:
6086 if (!(req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
))
6087 io_put_file(req
->splice
.file_in
);
6089 case IORING_OP_OPENAT
:
6090 case IORING_OP_OPENAT2
:
6091 if (req
->open
.filename
)
6092 putname(req
->open
.filename
);
6094 case IORING_OP_RENAMEAT
:
6095 putname(req
->rename
.oldpath
);
6096 putname(req
->rename
.newpath
);
6098 case IORING_OP_UNLINKAT
:
6099 putname(req
->unlink
.filename
);
6103 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
6104 kfree(req
->apoll
->double_poll
);
6108 if (req
->flags
& REQ_F_INFLIGHT
) {
6109 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6111 atomic_dec(&tctx
->inflight_tracked
);
6113 if (req
->flags
& REQ_F_CREDS
)
6114 put_cred(req
->creds
);
6116 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
6119 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6121 struct io_ring_ctx
*ctx
= req
->ctx
;
6122 const struct cred
*creds
= NULL
;
6125 if ((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred())
6126 creds
= override_creds(req
->creds
);
6128 switch (req
->opcode
) {
6130 ret
= io_nop(req
, issue_flags
);
6132 case IORING_OP_READV
:
6133 case IORING_OP_READ_FIXED
:
6134 case IORING_OP_READ
:
6135 ret
= io_read(req
, issue_flags
);
6137 case IORING_OP_WRITEV
:
6138 case IORING_OP_WRITE_FIXED
:
6139 case IORING_OP_WRITE
:
6140 ret
= io_write(req
, issue_flags
);
6142 case IORING_OP_FSYNC
:
6143 ret
= io_fsync(req
, issue_flags
);
6145 case IORING_OP_POLL_ADD
:
6146 ret
= io_poll_add(req
, issue_flags
);
6148 case IORING_OP_POLL_REMOVE
:
6149 ret
= io_poll_update(req
, issue_flags
);
6151 case IORING_OP_SYNC_FILE_RANGE
:
6152 ret
= io_sync_file_range(req
, issue_flags
);
6154 case IORING_OP_SENDMSG
:
6155 ret
= io_sendmsg(req
, issue_flags
);
6157 case IORING_OP_SEND
:
6158 ret
= io_send(req
, issue_flags
);
6160 case IORING_OP_RECVMSG
:
6161 ret
= io_recvmsg(req
, issue_flags
);
6163 case IORING_OP_RECV
:
6164 ret
= io_recv(req
, issue_flags
);
6166 case IORING_OP_TIMEOUT
:
6167 ret
= io_timeout(req
, issue_flags
);
6169 case IORING_OP_TIMEOUT_REMOVE
:
6170 ret
= io_timeout_remove(req
, issue_flags
);
6172 case IORING_OP_ACCEPT
:
6173 ret
= io_accept(req
, issue_flags
);
6175 case IORING_OP_CONNECT
:
6176 ret
= io_connect(req
, issue_flags
);
6178 case IORING_OP_ASYNC_CANCEL
:
6179 ret
= io_async_cancel(req
, issue_flags
);
6181 case IORING_OP_FALLOCATE
:
6182 ret
= io_fallocate(req
, issue_flags
);
6184 case IORING_OP_OPENAT
:
6185 ret
= io_openat(req
, issue_flags
);
6187 case IORING_OP_CLOSE
:
6188 ret
= io_close(req
, issue_flags
);
6190 case IORING_OP_FILES_UPDATE
:
6191 ret
= io_files_update(req
, issue_flags
);
6193 case IORING_OP_STATX
:
6194 ret
= io_statx(req
, issue_flags
);
6196 case IORING_OP_FADVISE
:
6197 ret
= io_fadvise(req
, issue_flags
);
6199 case IORING_OP_MADVISE
:
6200 ret
= io_madvise(req
, issue_flags
);
6202 case IORING_OP_OPENAT2
:
6203 ret
= io_openat2(req
, issue_flags
);
6205 case IORING_OP_EPOLL_CTL
:
6206 ret
= io_epoll_ctl(req
, issue_flags
);
6208 case IORING_OP_SPLICE
:
6209 ret
= io_splice(req
, issue_flags
);
6211 case IORING_OP_PROVIDE_BUFFERS
:
6212 ret
= io_provide_buffers(req
, issue_flags
);
6214 case IORING_OP_REMOVE_BUFFERS
:
6215 ret
= io_remove_buffers(req
, issue_flags
);
6218 ret
= io_tee(req
, issue_flags
);
6220 case IORING_OP_SHUTDOWN
:
6221 ret
= io_shutdown(req
, issue_flags
);
6223 case IORING_OP_RENAMEAT
:
6224 ret
= io_renameat(req
, issue_flags
);
6226 case IORING_OP_UNLINKAT
:
6227 ret
= io_unlinkat(req
, issue_flags
);
6235 revert_creds(creds
);
6238 /* If the op doesn't have a file, we're not polling for it */
6239 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
6240 io_iopoll_req_issued(req
);
6245 static void io_wq_submit_work(struct io_wq_work
*work
)
6247 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6248 struct io_kiocb
*timeout
;
6251 timeout
= io_prep_linked_timeout(req
);
6253 io_queue_linked_timeout(timeout
);
6255 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6260 ret
= io_issue_sqe(req
, 0);
6262 * We can get EAGAIN for polled IO even though we're
6263 * forcing a sync submission from here, since we can't
6264 * wait for request slots on the block side.
6272 /* avoid locking problems by failing it from a clean context */
6274 /* io-wq is going to take one down */
6276 io_req_task_queue_fail(req
, ret
);
6280 #define FFS_ASYNC_READ 0x1UL
6281 #define FFS_ASYNC_WRITE 0x2UL
6283 #define FFS_ISREG 0x4UL
6285 #define FFS_ISREG 0x0UL
6287 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6289 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
6292 struct io_fixed_file
*table_l2
;
6294 table_l2
= table
->files
[i
>> IORING_FILE_TABLE_SHIFT
];
6295 return &table_l2
[i
& IORING_FILE_TABLE_MASK
];
6298 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6301 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
6303 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
6306 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
6308 unsigned long file_ptr
= (unsigned long) file
;
6310 if (__io_file_supports_async(file
, READ
))
6311 file_ptr
|= FFS_ASYNC_READ
;
6312 if (__io_file_supports_async(file
, WRITE
))
6313 file_ptr
|= FFS_ASYNC_WRITE
;
6314 if (S_ISREG(file_inode(file
)->i_mode
))
6315 file_ptr
|= FFS_ISREG
;
6316 file_slot
->file_ptr
= file_ptr
;
6319 static struct file
*io_file_get(struct io_submit_state
*state
,
6320 struct io_kiocb
*req
, int fd
, bool fixed
)
6322 struct io_ring_ctx
*ctx
= req
->ctx
;
6326 unsigned long file_ptr
;
6328 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6330 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6331 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
6332 file
= (struct file
*) (file_ptr
& FFS_MASK
);
6333 file_ptr
&= ~FFS_MASK
;
6334 /* mask in overlapping REQ_F and FFS bits */
6335 req
->flags
|= (file_ptr
<< REQ_F_ASYNC_READ_BIT
);
6336 io_req_set_rsrc_node(req
);
6338 trace_io_uring_file_get(ctx
, fd
);
6339 file
= __io_file_get(state
, fd
);
6341 /* we don't allow fixed io_uring files */
6342 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6343 io_req_track_inflight(req
);
6349 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6351 struct io_timeout_data
*data
= container_of(timer
,
6352 struct io_timeout_data
, timer
);
6353 struct io_kiocb
*prev
, *req
= data
->req
;
6354 struct io_ring_ctx
*ctx
= req
->ctx
;
6355 unsigned long flags
;
6357 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
6358 prev
= req
->timeout
.head
;
6359 req
->timeout
.head
= NULL
;
6362 * We don't expect the list to be empty, that will only happen if we
6363 * race with the completion of the linked work.
6366 io_remove_next_linked(prev
);
6367 if (!req_ref_inc_not_zero(prev
))
6370 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
6373 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
6374 io_put_req_deferred(prev
, 1);
6375 io_put_req_deferred(req
, 1);
6377 io_req_complete_post(req
, -ETIME
, 0);
6379 return HRTIMER_NORESTART
;
6382 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6384 struct io_ring_ctx
*ctx
= req
->ctx
;
6386 spin_lock_irq(&ctx
->completion_lock
);
6388 * If the back reference is NULL, then our linked request finished
6389 * before we got a chance to setup the timer
6391 if (req
->timeout
.head
) {
6392 struct io_timeout_data
*data
= req
->async_data
;
6394 data
->timer
.function
= io_link_timeout_fn
;
6395 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6398 spin_unlock_irq(&ctx
->completion_lock
);
6399 /* drop submission reference */
6403 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
6405 struct io_kiocb
*nxt
= req
->link
;
6407 if (!nxt
|| (req
->flags
& REQ_F_LINK_TIMEOUT
) ||
6408 nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
6411 nxt
->timeout
.head
= req
;
6412 nxt
->flags
|= REQ_F_LTIMEOUT_ACTIVE
;
6413 req
->flags
|= REQ_F_LINK_TIMEOUT
;
6417 static void __io_queue_sqe(struct io_kiocb
*req
)
6419 struct io_kiocb
*linked_timeout
= io_prep_linked_timeout(req
);
6423 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
6426 * We async punt it if the file wasn't marked NOWAIT, or if the file
6427 * doesn't support non-blocking read/write attempts
6430 /* drop submission reference */
6431 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
6432 struct io_ring_ctx
*ctx
= req
->ctx
;
6433 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
6435 cs
->reqs
[cs
->nr
++] = req
;
6436 if (cs
->nr
== ARRAY_SIZE(cs
->reqs
))
6437 io_submit_flush_completions(ctx
);
6441 } else if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6442 switch (io_arm_poll_handler(req
)) {
6443 case IO_APOLL_READY
:
6445 case IO_APOLL_ABORTED
:
6447 * Queued up for async execution, worker will release
6448 * submit reference when the iocb is actually submitted.
6450 io_queue_async_work(req
);
6454 io_req_complete_failed(req
, ret
);
6457 io_queue_linked_timeout(linked_timeout
);
6460 static inline void io_queue_sqe(struct io_kiocb
*req
)
6462 if (unlikely(req
->ctx
->drain_active
) && io_drain_req(req
))
6465 if (likely(!(req
->flags
& REQ_F_FORCE_ASYNC
))) {
6466 __io_queue_sqe(req
);
6468 int ret
= io_req_prep_async(req
);
6471 io_req_complete_failed(req
, ret
);
6473 io_queue_async_work(req
);
6478 * Check SQE restrictions (opcode and flags).
6480 * Returns 'true' if SQE is allowed, 'false' otherwise.
6482 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
6483 struct io_kiocb
*req
,
6484 unsigned int sqe_flags
)
6486 if (likely(!ctx
->restricted
))
6489 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
6492 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
6493 ctx
->restrictions
.sqe_flags_required
)
6496 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
6497 ctx
->restrictions
.sqe_flags_required
))
6503 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6504 const struct io_uring_sqe
*sqe
)
6506 struct io_submit_state
*state
;
6507 unsigned int sqe_flags
;
6508 int personality
, ret
= 0;
6510 req
->opcode
= READ_ONCE(sqe
->opcode
);
6511 /* same numerical values with corresponding REQ_F_*, safe to copy */
6512 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
6513 req
->user_data
= READ_ONCE(sqe
->user_data
);
6515 req
->fixed_rsrc_refs
= NULL
;
6516 /* one is dropped after submission, the other at completion */
6517 atomic_set(&req
->refs
, 2);
6518 req
->task
= current
;
6520 /* enforce forwards compatibility on users */
6521 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6523 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6525 if (!io_check_restriction(ctx
, req
, sqe_flags
))
6528 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6529 !io_op_defs
[req
->opcode
].buffer_select
)
6531 if (unlikely(sqe_flags
& IOSQE_IO_DRAIN
))
6532 ctx
->drain_active
= true;
6534 personality
= READ_ONCE(sqe
->personality
);
6536 req
->creds
= xa_load(&ctx
->personalities
, personality
);
6539 get_cred(req
->creds
);
6540 req
->flags
|= REQ_F_CREDS
;
6542 state
= &ctx
->submit_state
;
6545 * Plug now if we have more than 1 IO left after this, and the target
6546 * is potentially a read/write to block based storage.
6548 if (!state
->plug_started
&& state
->ios_left
> 1 &&
6549 io_op_defs
[req
->opcode
].plug
) {
6550 blk_start_plug(&state
->plug
);
6551 state
->plug_started
= true;
6554 if (io_op_defs
[req
->opcode
].needs_file
) {
6555 bool fixed
= req
->flags
& REQ_F_FIXED_FILE
;
6557 req
->file
= io_file_get(state
, req
, READ_ONCE(sqe
->fd
), fixed
);
6558 if (unlikely(!req
->file
))
6566 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6567 const struct io_uring_sqe
*sqe
)
6569 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
6572 ret
= io_init_req(ctx
, req
, sqe
);
6573 if (unlikely(ret
)) {
6576 /* fail even hard links since we don't submit */
6577 req_set_fail(link
->head
);
6578 io_req_complete_failed(link
->head
, -ECANCELED
);
6581 io_req_complete_failed(req
, ret
);
6585 ret
= io_req_prep(req
, sqe
);
6589 /* don't need @sqe from now on */
6590 trace_io_uring_submit_sqe(ctx
, req
, req
->opcode
, req
->user_data
,
6592 ctx
->flags
& IORING_SETUP_SQPOLL
);
6595 * If we already have a head request, queue this one for async
6596 * submittal once the head completes. If we don't have a head but
6597 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6598 * submitted sync once the chain is complete. If none of those
6599 * conditions are true (normal request), then just queue it.
6602 struct io_kiocb
*head
= link
->head
;
6604 ret
= io_req_prep_async(req
);
6607 trace_io_uring_link(ctx
, req
, head
);
6608 link
->last
->link
= req
;
6611 /* last request of a link, enqueue the link */
6612 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
6617 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6629 * Batched submission is done, ensure local IO is flushed out.
6631 static void io_submit_state_end(struct io_submit_state
*state
,
6632 struct io_ring_ctx
*ctx
)
6634 if (state
->link
.head
)
6635 io_queue_sqe(state
->link
.head
);
6637 io_submit_flush_completions(ctx
);
6638 if (state
->plug_started
)
6639 blk_finish_plug(&state
->plug
);
6640 io_state_file_put(state
);
6644 * Start submission side cache.
6646 static void io_submit_state_start(struct io_submit_state
*state
,
6647 unsigned int max_ios
)
6649 state
->plug_started
= false;
6650 state
->ios_left
= max_ios
;
6651 /* set only head, no need to init link_last in advance */
6652 state
->link
.head
= NULL
;
6655 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6657 struct io_rings
*rings
= ctx
->rings
;
6660 * Ensure any loads from the SQEs are done at this point,
6661 * since once we write the new head, the application could
6662 * write new data to them.
6664 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6668 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6669 * that is mapped by userspace. This means that care needs to be taken to
6670 * ensure that reads are stable, as we cannot rely on userspace always
6671 * being a good citizen. If members of the sqe are validated and then later
6672 * used, it's important that those reads are done through READ_ONCE() to
6673 * prevent a re-load down the line.
6675 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6677 unsigned head
, mask
= ctx
->sq_entries
- 1;
6678 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
6681 * The cached sq head (or cq tail) serves two purposes:
6683 * 1) allows us to batch the cost of updating the user visible
6685 * 2) allows the kernel side to track the head on its own, even
6686 * though the application is the one updating it.
6688 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
6689 if (likely(head
< ctx
->sq_entries
))
6690 return &ctx
->sq_sqes
[head
];
6692 /* drop invalid entries */
6694 WRITE_ONCE(ctx
->rings
->sq_dropped
,
6695 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
6699 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
6701 struct io_uring_task
*tctx
;
6704 /* make sure SQ entry isn't read before tail */
6705 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6706 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6709 tctx
= current
->io_uring
;
6710 tctx
->cached_refs
-= nr
;
6711 if (unlikely(tctx
->cached_refs
< 0)) {
6712 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
6714 percpu_counter_add(&tctx
->inflight
, refill
);
6715 refcount_add(refill
, ¤t
->usage
);
6716 tctx
->cached_refs
+= refill
;
6718 io_submit_state_start(&ctx
->submit_state
, nr
);
6720 while (submitted
< nr
) {
6721 const struct io_uring_sqe
*sqe
;
6722 struct io_kiocb
*req
;
6724 req
= io_alloc_req(ctx
);
6725 if (unlikely(!req
)) {
6727 submitted
= -EAGAIN
;
6730 sqe
= io_get_sqe(ctx
);
6731 if (unlikely(!sqe
)) {
6732 kmem_cache_free(req_cachep
, req
);
6735 /* will complete beyond this point, count as submitted */
6737 if (io_submit_sqe(ctx
, req
, sqe
))
6741 if (unlikely(submitted
!= nr
)) {
6742 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6743 int unused
= nr
- ref_used
;
6745 current
->io_uring
->cached_refs
+= unused
;
6746 percpu_ref_put_many(&ctx
->refs
, unused
);
6749 io_submit_state_end(&ctx
->submit_state
, ctx
);
6750 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6751 io_commit_sqring(ctx
);
6756 static inline bool io_sqd_events_pending(struct io_sq_data
*sqd
)
6758 return READ_ONCE(sqd
->state
);
6761 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
6763 /* Tell userspace we may need a wakeup call */
6764 spin_lock_irq(&ctx
->completion_lock
);
6765 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6766 spin_unlock_irq(&ctx
->completion_lock
);
6769 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
6771 spin_lock_irq(&ctx
->completion_lock
);
6772 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6773 spin_unlock_irq(&ctx
->completion_lock
);
6776 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
6778 unsigned int to_submit
;
6781 to_submit
= io_sqring_entries(ctx
);
6782 /* if we're handling multiple rings, cap submit size for fairness */
6783 if (cap_entries
&& to_submit
> IORING_SQPOLL_CAP_ENTRIES_VALUE
)
6784 to_submit
= IORING_SQPOLL_CAP_ENTRIES_VALUE
;
6786 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
6787 unsigned nr_events
= 0;
6788 const struct cred
*creds
= NULL
;
6790 if (ctx
->sq_creds
!= current_cred())
6791 creds
= override_creds(ctx
->sq_creds
);
6793 mutex_lock(&ctx
->uring_lock
);
6794 if (!list_empty(&ctx
->iopoll_list
))
6795 io_do_iopoll(ctx
, &nr_events
, 0);
6798 * Don't submit if refs are dying, good for io_uring_register(),
6799 * but also it is relied upon by io_ring_exit_work()
6801 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
6802 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
6803 ret
= io_submit_sqes(ctx
, to_submit
);
6804 mutex_unlock(&ctx
->uring_lock
);
6806 if (to_submit
&& wq_has_sleeper(&ctx
->sqo_sq_wait
))
6807 wake_up(&ctx
->sqo_sq_wait
);
6809 revert_creds(creds
);
6815 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
6817 struct io_ring_ctx
*ctx
;
6818 unsigned sq_thread_idle
= 0;
6820 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6821 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
6822 sqd
->sq_thread_idle
= sq_thread_idle
;
6825 static bool io_sqd_handle_event(struct io_sq_data
*sqd
)
6827 bool did_sig
= false;
6828 struct ksignal ksig
;
6830 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
6831 signal_pending(current
)) {
6832 mutex_unlock(&sqd
->lock
);
6833 if (signal_pending(current
))
6834 did_sig
= get_signal(&ksig
);
6836 mutex_lock(&sqd
->lock
);
6838 return did_sig
|| test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
6841 static int io_sq_thread(void *data
)
6843 struct io_sq_data
*sqd
= data
;
6844 struct io_ring_ctx
*ctx
;
6845 unsigned long timeout
= 0;
6846 char buf
[TASK_COMM_LEN
];
6849 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
6850 set_task_comm(current
, buf
);
6852 if (sqd
->sq_cpu
!= -1)
6853 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
6855 set_cpus_allowed_ptr(current
, cpu_online_mask
);
6856 current
->flags
|= PF_NO_SETAFFINITY
;
6858 mutex_lock(&sqd
->lock
);
6860 bool cap_entries
, sqt_spin
= false;
6862 if (io_sqd_events_pending(sqd
) || signal_pending(current
)) {
6863 if (io_sqd_handle_event(sqd
))
6865 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6868 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
6869 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6870 int ret
= __io_sq_thread(ctx
, cap_entries
);
6872 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
6875 if (io_run_task_work())
6878 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
6881 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6885 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
6886 if (!io_sqd_events_pending(sqd
) && !current
->task_works
) {
6887 bool needs_sched
= true;
6889 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
6890 io_ring_set_wakeup_flag(ctx
);
6892 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6893 !list_empty_careful(&ctx
->iopoll_list
)) {
6894 needs_sched
= false;
6897 if (io_sqring_entries(ctx
)) {
6898 needs_sched
= false;
6904 mutex_unlock(&sqd
->lock
);
6906 mutex_lock(&sqd
->lock
);
6908 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6909 io_ring_clear_wakeup_flag(ctx
);
6912 finish_wait(&sqd
->wait
, &wait
);
6913 timeout
= jiffies
+ sqd
->sq_thread_idle
;
6916 io_uring_cancel_generic(true, sqd
);
6918 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
6919 io_ring_set_wakeup_flag(ctx
);
6921 mutex_unlock(&sqd
->lock
);
6923 complete(&sqd
->exited
);
6927 struct io_wait_queue
{
6928 struct wait_queue_entry wq
;
6929 struct io_ring_ctx
*ctx
;
6931 unsigned nr_timeouts
;
6934 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
6936 struct io_ring_ctx
*ctx
= iowq
->ctx
;
6939 * Wake up if we have enough events, or if a timeout occurred since we
6940 * started waiting. For timeouts, we always want to return to userspace,
6941 * regardless of event count.
6943 return io_cqring_events(ctx
) >= iowq
->to_wait
||
6944 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
6947 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
6948 int wake_flags
, void *key
)
6950 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
6954 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6955 * the task, and the next invocation will do it.
6957 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->check_cq_overflow
))
6958 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
6962 static int io_run_task_work_sig(void)
6964 if (io_run_task_work())
6966 if (!signal_pending(current
))
6968 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
6969 return -ERESTARTSYS
;
6973 /* when returns >0, the caller should retry */
6974 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
6975 struct io_wait_queue
*iowq
,
6976 signed long *timeout
)
6980 /* make sure we run task_work before checking for signals */
6981 ret
= io_run_task_work_sig();
6982 if (ret
|| io_should_wake(iowq
))
6984 /* let the caller flush overflows, retry */
6985 if (test_bit(0, &ctx
->check_cq_overflow
))
6988 *timeout
= schedule_timeout(*timeout
);
6989 return !*timeout
? -ETIME
: 1;
6993 * Wait until events become available, if we don't already have some. The
6994 * application must reap them itself, as they reside on the shared cq ring.
6996 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
6997 const sigset_t __user
*sig
, size_t sigsz
,
6998 struct __kernel_timespec __user
*uts
)
7000 struct io_wait_queue iowq
= {
7003 .func
= io_wake_function
,
7004 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
7007 .to_wait
= min_events
,
7009 struct io_rings
*rings
= ctx
->rings
;
7010 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
7014 io_cqring_overflow_flush(ctx
, false);
7015 if (io_cqring_events(ctx
) >= min_events
)
7017 if (!io_run_task_work())
7022 #ifdef CONFIG_COMPAT
7023 if (in_compat_syscall())
7024 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7028 ret
= set_user_sigmask(sig
, sigsz
);
7035 struct timespec64 ts
;
7037 if (get_timespec64(&ts
, uts
))
7039 timeout
= timespec64_to_jiffies(&ts
);
7042 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7043 trace_io_uring_cqring_wait(ctx
, min_events
);
7045 /* if we can't even flush overflow, don't wait for more */
7046 if (!io_cqring_overflow_flush(ctx
, false)) {
7050 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
7051 TASK_INTERRUPTIBLE
);
7052 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
7053 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
7057 restore_saved_sigmask_unless(ret
== -EINTR
);
7059 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7062 static void io_free_page_table(void **table
, size_t size
)
7064 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7066 for (i
= 0; i
< nr_tables
; i
++)
7071 static void **io_alloc_page_table(size_t size
)
7073 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7074 size_t init_size
= size
;
7077 table
= kcalloc(nr_tables
, sizeof(*table
), GFP_KERNEL
);
7081 for (i
= 0; i
< nr_tables
; i
++) {
7082 unsigned int this_size
= min_t(size_t, size
, PAGE_SIZE
);
7084 table
[i
] = kzalloc(this_size
, GFP_KERNEL
);
7086 io_free_page_table(table
, init_size
);
7094 static inline void io_rsrc_ref_lock(struct io_ring_ctx
*ctx
)
7096 spin_lock_bh(&ctx
->rsrc_ref_lock
);
7099 static inline void io_rsrc_ref_unlock(struct io_ring_ctx
*ctx
)
7101 spin_unlock_bh(&ctx
->rsrc_ref_lock
);
7104 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
7106 percpu_ref_exit(&ref_node
->refs
);
7110 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
7111 struct io_rsrc_data
*data_to_kill
)
7113 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
7114 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
7117 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
7119 rsrc_node
->rsrc_data
= data_to_kill
;
7120 io_rsrc_ref_lock(ctx
);
7121 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
7122 io_rsrc_ref_unlock(ctx
);
7124 atomic_inc(&data_to_kill
->refs
);
7125 percpu_ref_kill(&rsrc_node
->refs
);
7126 ctx
->rsrc_node
= NULL
;
7129 if (!ctx
->rsrc_node
) {
7130 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
7131 ctx
->rsrc_backup_node
= NULL
;
7135 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
7137 if (ctx
->rsrc_backup_node
)
7139 ctx
->rsrc_backup_node
= io_rsrc_node_alloc(ctx
);
7140 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
7143 static int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
, struct io_ring_ctx
*ctx
)
7147 /* As we may drop ->uring_lock, other task may have started quiesce */
7151 data
->quiesce
= true;
7153 ret
= io_rsrc_node_switch_start(ctx
);
7156 io_rsrc_node_switch(ctx
, data
);
7158 /* kill initial ref, already quiesced if zero */
7159 if (atomic_dec_and_test(&data
->refs
))
7161 flush_delayed_work(&ctx
->rsrc_put_work
);
7162 ret
= wait_for_completion_interruptible(&data
->done
);
7166 atomic_inc(&data
->refs
);
7167 /* wait for all works potentially completing data->done */
7168 flush_delayed_work(&ctx
->rsrc_put_work
);
7169 reinit_completion(&data
->done
);
7171 mutex_unlock(&ctx
->uring_lock
);
7172 ret
= io_run_task_work_sig();
7173 mutex_lock(&ctx
->uring_lock
);
7175 data
->quiesce
= false;
7180 static u64
*io_get_tag_slot(struct io_rsrc_data
*data
, unsigned int idx
)
7182 unsigned int off
= idx
& IO_RSRC_TAG_TABLE_MASK
;
7183 unsigned int table_idx
= idx
>> IO_RSRC_TAG_TABLE_SHIFT
;
7185 return &data
->tags
[table_idx
][off
];
7188 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
7190 size_t size
= data
->nr
* sizeof(data
->tags
[0][0]);
7193 io_free_page_table((void **)data
->tags
, size
);
7197 static int io_rsrc_data_alloc(struct io_ring_ctx
*ctx
, rsrc_put_fn
*do_put
,
7198 u64 __user
*utags
, unsigned nr
,
7199 struct io_rsrc_data
**pdata
)
7201 struct io_rsrc_data
*data
;
7205 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7208 data
->tags
= (u64
**)io_alloc_page_table(nr
* sizeof(data
->tags
[0][0]));
7216 data
->do_put
= do_put
;
7219 for (i
= 0; i
< nr
; i
++) {
7220 u64
*tag_slot
= io_get_tag_slot(data
, i
);
7222 if (copy_from_user(tag_slot
, &utags
[i
],
7228 atomic_set(&data
->refs
, 1);
7229 init_completion(&data
->done
);
7233 io_rsrc_data_free(data
);
7237 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7239 size_t size
= nr_files
* sizeof(struct io_fixed_file
);
7241 table
->files
= (struct io_fixed_file
**)io_alloc_page_table(size
);
7242 return !!table
->files
;
7245 static void io_free_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7247 size_t size
= nr_files
* sizeof(struct io_fixed_file
);
7249 io_free_page_table((void **)table
->files
, size
);
7250 table
->files
= NULL
;
7253 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7255 #if defined(CONFIG_UNIX)
7256 if (ctx
->ring_sock
) {
7257 struct sock
*sock
= ctx
->ring_sock
->sk
;
7258 struct sk_buff
*skb
;
7260 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7266 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7269 file
= io_file_from_index(ctx
, i
);
7274 io_free_file_tables(&ctx
->file_table
, ctx
->nr_user_files
);
7275 io_rsrc_data_free(ctx
->file_data
);
7276 ctx
->file_data
= NULL
;
7277 ctx
->nr_user_files
= 0;
7280 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7284 if (!ctx
->file_data
)
7286 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
7288 __io_sqe_files_unregister(ctx
);
7292 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7293 __releases(&sqd
->lock
)
7295 WARN_ON_ONCE(sqd
->thread
== current
);
7298 * Do the dance but not conditional clear_bit() because it'd race with
7299 * other threads incrementing park_pending and setting the bit.
7301 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7302 if (atomic_dec_return(&sqd
->park_pending
))
7303 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7304 mutex_unlock(&sqd
->lock
);
7307 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7308 __acquires(&sqd
->lock
)
7310 WARN_ON_ONCE(sqd
->thread
== current
);
7312 atomic_inc(&sqd
->park_pending
);
7313 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7314 mutex_lock(&sqd
->lock
);
7316 wake_up_process(sqd
->thread
);
7319 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7321 WARN_ON_ONCE(sqd
->thread
== current
);
7322 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
7324 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7325 mutex_lock(&sqd
->lock
);
7327 wake_up_process(sqd
->thread
);
7328 mutex_unlock(&sqd
->lock
);
7329 wait_for_completion(&sqd
->exited
);
7332 static void io_put_sq_data(struct io_sq_data
*sqd
)
7334 if (refcount_dec_and_test(&sqd
->refs
)) {
7335 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
7337 io_sq_thread_stop(sqd
);
7342 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
7344 struct io_sq_data
*sqd
= ctx
->sq_data
;
7347 io_sq_thread_park(sqd
);
7348 list_del_init(&ctx
->sqd_list
);
7349 io_sqd_update_thread_idle(sqd
);
7350 io_sq_thread_unpark(sqd
);
7352 io_put_sq_data(sqd
);
7353 ctx
->sq_data
= NULL
;
7357 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7359 struct io_ring_ctx
*ctx_attach
;
7360 struct io_sq_data
*sqd
;
7363 f
= fdget(p
->wq_fd
);
7365 return ERR_PTR(-ENXIO
);
7366 if (f
.file
->f_op
!= &io_uring_fops
) {
7368 return ERR_PTR(-EINVAL
);
7371 ctx_attach
= f
.file
->private_data
;
7372 sqd
= ctx_attach
->sq_data
;
7375 return ERR_PTR(-EINVAL
);
7377 if (sqd
->task_tgid
!= current
->tgid
) {
7379 return ERR_PTR(-EPERM
);
7382 refcount_inc(&sqd
->refs
);
7387 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
7390 struct io_sq_data
*sqd
;
7393 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
7394 sqd
= io_attach_sq_data(p
);
7399 /* fall through for EPERM case, setup new sqd/task */
7400 if (PTR_ERR(sqd
) != -EPERM
)
7404 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7406 return ERR_PTR(-ENOMEM
);
7408 atomic_set(&sqd
->park_pending
, 0);
7409 refcount_set(&sqd
->refs
, 1);
7410 INIT_LIST_HEAD(&sqd
->ctx_list
);
7411 mutex_init(&sqd
->lock
);
7412 init_waitqueue_head(&sqd
->wait
);
7413 init_completion(&sqd
->exited
);
7417 #if defined(CONFIG_UNIX)
7419 * Ensure the UNIX gc is aware of our file set, so we are certain that
7420 * the io_uring can be safely unregistered on process exit, even if we have
7421 * loops in the file referencing.
7423 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
7425 struct sock
*sk
= ctx
->ring_sock
->sk
;
7426 struct scm_fp_list
*fpl
;
7427 struct sk_buff
*skb
;
7430 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
7434 skb
= alloc_skb(0, GFP_KERNEL
);
7443 fpl
->user
= get_uid(current_user());
7444 for (i
= 0; i
< nr
; i
++) {
7445 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
7449 fpl
->fp
[nr_files
] = get_file(file
);
7450 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
7455 fpl
->max
= SCM_MAX_FD
;
7456 fpl
->count
= nr_files
;
7457 UNIXCB(skb
).fp
= fpl
;
7458 skb
->destructor
= unix_destruct_scm
;
7459 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
7460 skb_queue_head(&sk
->sk_receive_queue
, skb
);
7462 for (i
= 0; i
< nr_files
; i
++)
7473 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7474 * causes regular reference counting to break down. We rely on the UNIX
7475 * garbage collection to take care of this problem for us.
7477 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7479 unsigned left
, total
;
7483 left
= ctx
->nr_user_files
;
7485 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
7487 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
7491 total
+= this_files
;
7497 while (total
< ctx
->nr_user_files
) {
7498 struct file
*file
= io_file_from_index(ctx
, total
);
7508 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
7514 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
7516 struct file
*file
= prsrc
->file
;
7517 #if defined(CONFIG_UNIX)
7518 struct sock
*sock
= ctx
->ring_sock
->sk
;
7519 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
7520 struct sk_buff
*skb
;
7523 __skb_queue_head_init(&list
);
7526 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7527 * remove this entry and rearrange the file array.
7529 skb
= skb_dequeue(head
);
7531 struct scm_fp_list
*fp
;
7533 fp
= UNIXCB(skb
).fp
;
7534 for (i
= 0; i
< fp
->count
; i
++) {
7537 if (fp
->fp
[i
] != file
)
7540 unix_notinflight(fp
->user
, fp
->fp
[i
]);
7541 left
= fp
->count
- 1 - i
;
7543 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
7544 left
* sizeof(struct file
*));
7551 __skb_queue_tail(&list
, skb
);
7561 __skb_queue_tail(&list
, skb
);
7563 skb
= skb_dequeue(head
);
7566 if (skb_peek(&list
)) {
7567 spin_lock_irq(&head
->lock
);
7568 while ((skb
= __skb_dequeue(&list
)) != NULL
)
7569 __skb_queue_tail(head
, skb
);
7570 spin_unlock_irq(&head
->lock
);
7577 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
7579 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
7580 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
7581 struct io_rsrc_put
*prsrc
, *tmp
;
7583 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
7584 list_del(&prsrc
->list
);
7587 bool lock_ring
= ctx
->flags
& IORING_SETUP_IOPOLL
;
7589 io_ring_submit_lock(ctx
, lock_ring
);
7590 spin_lock_irq(&ctx
->completion_lock
);
7591 io_cqring_fill_event(ctx
, prsrc
->tag
, 0, 0);
7593 io_commit_cqring(ctx
);
7594 spin_unlock_irq(&ctx
->completion_lock
);
7595 io_cqring_ev_posted(ctx
);
7596 io_ring_submit_unlock(ctx
, lock_ring
);
7599 rsrc_data
->do_put(ctx
, prsrc
);
7603 io_rsrc_node_destroy(ref_node
);
7604 if (atomic_dec_and_test(&rsrc_data
->refs
))
7605 complete(&rsrc_data
->done
);
7608 static void io_rsrc_put_work(struct work_struct
*work
)
7610 struct io_ring_ctx
*ctx
;
7611 struct llist_node
*node
;
7613 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
7614 node
= llist_del_all(&ctx
->rsrc_put_llist
);
7617 struct io_rsrc_node
*ref_node
;
7618 struct llist_node
*next
= node
->next
;
7620 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
7621 __io_rsrc_put_work(ref_node
);
7626 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7628 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
7629 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
7630 bool first_add
= false;
7632 io_rsrc_ref_lock(ctx
);
7635 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7636 node
= list_first_entry(&ctx
->rsrc_ref_list
,
7637 struct io_rsrc_node
, node
);
7638 /* recycle ref nodes in order */
7641 list_del(&node
->node
);
7642 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
7644 io_rsrc_ref_unlock(ctx
);
7647 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, HZ
);
7650 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
)
7652 struct io_rsrc_node
*ref_node
;
7654 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7658 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7663 INIT_LIST_HEAD(&ref_node
->node
);
7664 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7665 ref_node
->done
= false;
7669 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7670 unsigned nr_args
, u64 __user
*tags
)
7672 __s32 __user
*fds
= (__s32 __user
*) arg
;
7681 if (nr_args
> IORING_MAX_FIXED_FILES
)
7683 ret
= io_rsrc_node_switch_start(ctx
);
7686 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, tags
, nr_args
,
7692 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
))
7695 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
7696 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
7700 /* allow sparse sets */
7703 if (unlikely(*io_get_tag_slot(ctx
->file_data
, i
)))
7710 if (unlikely(!file
))
7714 * Don't allow io_uring instances to be registered. If UNIX
7715 * isn't enabled, then this causes a reference cycle and this
7716 * instance can never get freed. If UNIX is enabled we'll
7717 * handle it just fine, but there's still no point in allowing
7718 * a ring fd as it doesn't support regular read/write anyway.
7720 if (file
->f_op
== &io_uring_fops
) {
7724 io_fixed_file_set(io_fixed_file_slot(&ctx
->file_table
, i
), file
);
7727 ret
= io_sqe_files_scm(ctx
);
7729 __io_sqe_files_unregister(ctx
);
7733 io_rsrc_node_switch(ctx
, NULL
);
7736 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7737 file
= io_file_from_index(ctx
, i
);
7741 io_free_file_tables(&ctx
->file_table
, nr_args
);
7742 ctx
->nr_user_files
= 0;
7744 io_rsrc_data_free(ctx
->file_data
);
7745 ctx
->file_data
= NULL
;
7749 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
7752 #if defined(CONFIG_UNIX)
7753 struct sock
*sock
= ctx
->ring_sock
->sk
;
7754 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
7755 struct sk_buff
*skb
;
7758 * See if we can merge this file into an existing skb SCM_RIGHTS
7759 * file set. If there's no room, fall back to allocating a new skb
7760 * and filling it in.
7762 spin_lock_irq(&head
->lock
);
7763 skb
= skb_peek(head
);
7765 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
7767 if (fpl
->count
< SCM_MAX_FD
) {
7768 __skb_unlink(skb
, head
);
7769 spin_unlock_irq(&head
->lock
);
7770 fpl
->fp
[fpl
->count
] = get_file(file
);
7771 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
7773 spin_lock_irq(&head
->lock
);
7774 __skb_queue_head(head
, skb
);
7779 spin_unlock_irq(&head
->lock
);
7786 return __io_sqe_files_scm(ctx
, 1, index
);
7792 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
7793 struct io_rsrc_node
*node
, void *rsrc
)
7795 struct io_rsrc_put
*prsrc
;
7797 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
7801 prsrc
->tag
= *io_get_tag_slot(data
, idx
);
7803 list_add(&prsrc
->list
, &node
->rsrc_list
);
7807 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
7808 struct io_uring_rsrc_update2
*up
,
7811 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
7812 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
7813 struct io_rsrc_data
*data
= ctx
->file_data
;
7814 struct io_fixed_file
*file_slot
;
7818 bool needs_switch
= false;
7820 if (!ctx
->file_data
)
7822 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
7825 for (done
= 0; done
< nr_args
; done
++) {
7828 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
7829 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7833 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
7837 if (fd
== IORING_REGISTER_FILES_SKIP
)
7840 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
7841 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
7843 if (file_slot
->file_ptr
) {
7844 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
7845 err
= io_queue_rsrc_removal(data
, up
->offset
+ done
,
7846 ctx
->rsrc_node
, file
);
7849 file_slot
->file_ptr
= 0;
7850 needs_switch
= true;
7859 * Don't allow io_uring instances to be registered. If
7860 * UNIX isn't enabled, then this causes a reference
7861 * cycle and this instance can never get freed. If UNIX
7862 * is enabled we'll handle it just fine, but there's
7863 * still no point in allowing a ring fd as it doesn't
7864 * support regular read/write anyway.
7866 if (file
->f_op
== &io_uring_fops
) {
7871 *io_get_tag_slot(data
, up
->offset
+ done
) = tag
;
7872 io_fixed_file_set(file_slot
, file
);
7873 err
= io_sqe_file_register(ctx
, file
, i
);
7875 file_slot
->file_ptr
= 0;
7883 io_rsrc_node_switch(ctx
, data
);
7884 return done
? done
: err
;
7887 static struct io_wq_work
*io_free_work(struct io_wq_work
*work
)
7889 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7891 req
= io_put_req_find_next(req
);
7892 return req
? &req
->work
: NULL
;
7895 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
7896 struct task_struct
*task
)
7898 struct io_wq_hash
*hash
;
7899 struct io_wq_data data
;
7900 unsigned int concurrency
;
7902 hash
= ctx
->hash_map
;
7904 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
7906 return ERR_PTR(-ENOMEM
);
7907 refcount_set(&hash
->refs
, 1);
7908 init_waitqueue_head(&hash
->wait
);
7909 ctx
->hash_map
= hash
;
7914 data
.free_work
= io_free_work
;
7915 data
.do_work
= io_wq_submit_work
;
7917 /* Do QD, or 4 * CPUS, whatever is smallest */
7918 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
7920 return io_wq_create(concurrency
, &data
);
7923 static int io_uring_alloc_task_context(struct task_struct
*task
,
7924 struct io_ring_ctx
*ctx
)
7926 struct io_uring_task
*tctx
;
7929 tctx
= kzalloc(sizeof(*tctx
), GFP_KERNEL
);
7930 if (unlikely(!tctx
))
7933 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
7934 if (unlikely(ret
)) {
7939 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
7940 if (IS_ERR(tctx
->io_wq
)) {
7941 ret
= PTR_ERR(tctx
->io_wq
);
7942 percpu_counter_destroy(&tctx
->inflight
);
7948 init_waitqueue_head(&tctx
->wait
);
7949 atomic_set(&tctx
->in_idle
, 0);
7950 atomic_set(&tctx
->inflight_tracked
, 0);
7951 task
->io_uring
= tctx
;
7952 spin_lock_init(&tctx
->task_lock
);
7953 INIT_WQ_LIST(&tctx
->task_list
);
7954 init_task_work(&tctx
->task_work
, tctx_task_work
);
7958 void __io_uring_free(struct task_struct
*tsk
)
7960 struct io_uring_task
*tctx
= tsk
->io_uring
;
7962 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
7963 WARN_ON_ONCE(tctx
->io_wq
);
7964 WARN_ON_ONCE(tctx
->cached_refs
);
7966 percpu_counter_destroy(&tctx
->inflight
);
7968 tsk
->io_uring
= NULL
;
7971 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
7972 struct io_uring_params
*p
)
7976 /* Retain compatibility with failing for an invalid attach attempt */
7977 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
7978 IORING_SETUP_ATTACH_WQ
) {
7981 f
= fdget(p
->wq_fd
);
7985 if (f
.file
->f_op
!= &io_uring_fops
)
7988 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7989 struct task_struct
*tsk
;
7990 struct io_sq_data
*sqd
;
7993 sqd
= io_get_sq_data(p
, &attached
);
7999 ctx
->sq_creds
= get_current_cred();
8001 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8002 if (!ctx
->sq_thread_idle
)
8003 ctx
->sq_thread_idle
= HZ
;
8005 io_sq_thread_park(sqd
);
8006 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
8007 io_sqd_update_thread_idle(sqd
);
8008 /* don't attach to a dying SQPOLL thread, would be racy */
8009 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
8010 io_sq_thread_unpark(sqd
);
8017 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8018 int cpu
= p
->sq_thread_cpu
;
8021 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
8028 sqd
->task_pid
= current
->pid
;
8029 sqd
->task_tgid
= current
->tgid
;
8030 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
8037 ret
= io_uring_alloc_task_context(tsk
, ctx
);
8038 wake_up_new_task(tsk
);
8041 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8042 /* Can't have SQ_AFF without SQPOLL */
8049 complete(&ctx
->sq_data
->exited
);
8051 io_sq_thread_finish(ctx
);
8055 static inline void __io_unaccount_mem(struct user_struct
*user
,
8056 unsigned long nr_pages
)
8058 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8061 static inline int __io_account_mem(struct user_struct
*user
,
8062 unsigned long nr_pages
)
8064 unsigned long page_limit
, cur_pages
, new_pages
;
8066 /* Don't allow more pages than we can safely lock */
8067 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8070 cur_pages
= atomic_long_read(&user
->locked_vm
);
8071 new_pages
= cur_pages
+ nr_pages
;
8072 if (new_pages
> page_limit
)
8074 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8075 new_pages
) != cur_pages
);
8080 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8083 __io_unaccount_mem(ctx
->user
, nr_pages
);
8085 if (ctx
->mm_account
)
8086 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8089 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8094 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8099 if (ctx
->mm_account
)
8100 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8105 static void io_mem_free(void *ptr
)
8112 page
= virt_to_head_page(ptr
);
8113 if (put_page_testzero(page
))
8114 free_compound_page(page
);
8117 static void *io_mem_alloc(size_t size
)
8119 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8120 __GFP_NORETRY
| __GFP_ACCOUNT
;
8122 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8125 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8128 struct io_rings
*rings
;
8129 size_t off
, sq_array_size
;
8131 off
= struct_size(rings
, cqes
, cq_entries
);
8132 if (off
== SIZE_MAX
)
8136 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8144 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8145 if (sq_array_size
== SIZE_MAX
)
8148 if (check_add_overflow(off
, sq_array_size
, &off
))
8154 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
8156 struct io_mapped_ubuf
*imu
= *slot
;
8159 if (imu
!= ctx
->dummy_ubuf
) {
8160 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
8161 unpin_user_page(imu
->bvec
[i
].bv_page
);
8162 if (imu
->acct_pages
)
8163 io_unaccount_mem(ctx
, imu
->acct_pages
);
8169 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8171 io_buffer_unmap(ctx
, &prsrc
->buf
);
8175 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8179 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
8180 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
8181 kfree(ctx
->user_bufs
);
8182 io_rsrc_data_free(ctx
->buf_data
);
8183 ctx
->user_bufs
= NULL
;
8184 ctx
->buf_data
= NULL
;
8185 ctx
->nr_user_bufs
= 0;
8188 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8195 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
8197 __io_sqe_buffers_unregister(ctx
);
8201 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8202 void __user
*arg
, unsigned index
)
8204 struct iovec __user
*src
;
8206 #ifdef CONFIG_COMPAT
8208 struct compat_iovec __user
*ciovs
;
8209 struct compat_iovec ciov
;
8211 ciovs
= (struct compat_iovec __user
*) arg
;
8212 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8215 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8216 dst
->iov_len
= ciov
.iov_len
;
8220 src
= (struct iovec __user
*) arg
;
8221 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8227 * Not super efficient, but this is just a registration time. And we do cache
8228 * the last compound head, so generally we'll only do a full search if we don't
8231 * We check if the given compound head page has already been accounted, to
8232 * avoid double accounting it. This allows us to account the full size of the
8233 * page, not just the constituent pages of a huge page.
8235 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8236 int nr_pages
, struct page
*hpage
)
8240 /* check current page array */
8241 for (i
= 0; i
< nr_pages
; i
++) {
8242 if (!PageCompound(pages
[i
]))
8244 if (compound_head(pages
[i
]) == hpage
)
8248 /* check previously registered pages */
8249 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8250 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
8252 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8253 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8255 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8263 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8264 int nr_pages
, struct io_mapped_ubuf
*imu
,
8265 struct page
**last_hpage
)
8269 imu
->acct_pages
= 0;
8270 for (i
= 0; i
< nr_pages
; i
++) {
8271 if (!PageCompound(pages
[i
])) {
8276 hpage
= compound_head(pages
[i
]);
8277 if (hpage
== *last_hpage
)
8279 *last_hpage
= hpage
;
8280 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8282 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8286 if (!imu
->acct_pages
)
8289 ret
= io_account_mem(ctx
, imu
->acct_pages
);
8291 imu
->acct_pages
= 0;
8295 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
8296 struct io_mapped_ubuf
**pimu
,
8297 struct page
**last_hpage
)
8299 struct io_mapped_ubuf
*imu
= NULL
;
8300 struct vm_area_struct
**vmas
= NULL
;
8301 struct page
**pages
= NULL
;
8302 unsigned long off
, start
, end
, ubuf
;
8304 int ret
, pret
, nr_pages
, i
;
8306 if (!iov
->iov_base
) {
8307 *pimu
= ctx
->dummy_ubuf
;
8311 ubuf
= (unsigned long) iov
->iov_base
;
8312 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8313 start
= ubuf
>> PAGE_SHIFT
;
8314 nr_pages
= end
- start
;
8319 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
8323 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
8328 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
8333 mmap_read_lock(current
->mm
);
8334 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
8336 if (pret
== nr_pages
) {
8337 /* don't support file backed memory */
8338 for (i
= 0; i
< nr_pages
; i
++) {
8339 struct vm_area_struct
*vma
= vmas
[i
];
8341 if (vma_is_shmem(vma
))
8344 !is_file_hugepages(vma
->vm_file
)) {
8350 ret
= pret
< 0 ? pret
: -EFAULT
;
8352 mmap_read_unlock(current
->mm
);
8355 * if we did partial map, or found file backed vmas,
8356 * release any pages we did get
8359 unpin_user_pages(pages
, pret
);
8363 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
8365 unpin_user_pages(pages
, pret
);
8369 off
= ubuf
& ~PAGE_MASK
;
8370 size
= iov
->iov_len
;
8371 for (i
= 0; i
< nr_pages
; i
++) {
8374 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
8375 imu
->bvec
[i
].bv_page
= pages
[i
];
8376 imu
->bvec
[i
].bv_len
= vec_len
;
8377 imu
->bvec
[i
].bv_offset
= off
;
8381 /* store original address for later verification */
8383 imu
->ubuf_end
= ubuf
+ iov
->iov_len
;
8384 imu
->nr_bvecs
= nr_pages
;
8395 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
8397 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
8398 return ctx
->user_bufs
? 0 : -ENOMEM
;
8401 static int io_buffer_validate(struct iovec
*iov
)
8403 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
8406 * Don't impose further limits on the size and buffer
8407 * constraints here, we'll -EINVAL later when IO is
8408 * submitted if they are wrong.
8411 return iov
->iov_len
? -EFAULT
: 0;
8415 /* arbitrary limit, but we need something */
8416 if (iov
->iov_len
> SZ_1G
)
8419 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
8425 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8426 unsigned int nr_args
, u64 __user
*tags
)
8428 struct page
*last_hpage
= NULL
;
8429 struct io_rsrc_data
*data
;
8435 if (!nr_args
|| nr_args
> IORING_MAX_REG_BUFFERS
)
8437 ret
= io_rsrc_node_switch_start(ctx
);
8440 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, tags
, nr_args
, &data
);
8443 ret
= io_buffers_map_alloc(ctx
, nr_args
);
8445 io_rsrc_data_free(data
);
8449 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
8450 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
8453 ret
= io_buffer_validate(&iov
);
8456 if (!iov
.iov_base
&& *io_get_tag_slot(data
, i
)) {
8461 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
8467 WARN_ON_ONCE(ctx
->buf_data
);
8469 ctx
->buf_data
= data
;
8471 __io_sqe_buffers_unregister(ctx
);
8473 io_rsrc_node_switch(ctx
, NULL
);
8477 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
8478 struct io_uring_rsrc_update2
*up
,
8479 unsigned int nr_args
)
8481 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
8482 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
8483 struct page
*last_hpage
= NULL
;
8484 bool needs_switch
= false;
8490 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
8493 for (done
= 0; done
< nr_args
; done
++) {
8494 struct io_mapped_ubuf
*imu
;
8495 int offset
= up
->offset
+ done
;
8498 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
8501 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
8505 err
= io_buffer_validate(&iov
);
8508 if (!iov
.iov_base
&& tag
) {
8512 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
8516 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
8517 if (ctx
->user_bufs
[i
] != ctx
->dummy_ubuf
) {
8518 err
= io_queue_rsrc_removal(ctx
->buf_data
, offset
,
8519 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
8520 if (unlikely(err
)) {
8521 io_buffer_unmap(ctx
, &imu
);
8524 ctx
->user_bufs
[i
] = NULL
;
8525 needs_switch
= true;
8528 ctx
->user_bufs
[i
] = imu
;
8529 *io_get_tag_slot(ctx
->buf_data
, offset
) = tag
;
8533 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
8534 return done
? done
: err
;
8537 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
8539 __s32 __user
*fds
= arg
;
8545 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
8548 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
8549 if (IS_ERR(ctx
->cq_ev_fd
)) {
8550 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
8552 ctx
->cq_ev_fd
= NULL
;
8559 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
8561 if (ctx
->cq_ev_fd
) {
8562 eventfd_ctx_put(ctx
->cq_ev_fd
);
8563 ctx
->cq_ev_fd
= NULL
;
8570 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
8572 struct io_buffer
*buf
;
8573 unsigned long index
;
8575 xa_for_each(&ctx
->io_buffers
, index
, buf
)
8576 __io_remove_buffers(ctx
, buf
, index
, -1U);
8579 static void io_req_cache_free(struct list_head
*list
, struct task_struct
*tsk
)
8581 struct io_kiocb
*req
, *nxt
;
8583 list_for_each_entry_safe(req
, nxt
, list
, compl.list
) {
8584 if (tsk
&& req
->task
!= tsk
)
8586 list_del(&req
->compl.list
);
8587 kmem_cache_free(req_cachep
, req
);
8591 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
8593 struct io_submit_state
*submit_state
= &ctx
->submit_state
;
8594 struct io_comp_state
*cs
= &ctx
->submit_state
.comp
;
8596 mutex_lock(&ctx
->uring_lock
);
8598 if (submit_state
->free_reqs
) {
8599 kmem_cache_free_bulk(req_cachep
, submit_state
->free_reqs
,
8600 submit_state
->reqs
);
8601 submit_state
->free_reqs
= 0;
8604 io_flush_cached_locked_reqs(ctx
, cs
);
8605 io_req_cache_free(&cs
->free_list
, NULL
);
8606 mutex_unlock(&ctx
->uring_lock
);
8609 static bool io_wait_rsrc_data(struct io_rsrc_data
*data
)
8613 if (!atomic_dec_and_test(&data
->refs
))
8614 wait_for_completion(&data
->done
);
8618 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
8620 io_sq_thread_finish(ctx
);
8622 if (ctx
->mm_account
) {
8623 mmdrop(ctx
->mm_account
);
8624 ctx
->mm_account
= NULL
;
8627 mutex_lock(&ctx
->uring_lock
);
8628 if (io_wait_rsrc_data(ctx
->buf_data
))
8629 __io_sqe_buffers_unregister(ctx
);
8630 if (io_wait_rsrc_data(ctx
->file_data
))
8631 __io_sqe_files_unregister(ctx
);
8633 __io_cqring_overflow_flush(ctx
, true);
8634 mutex_unlock(&ctx
->uring_lock
);
8635 io_eventfd_unregister(ctx
);
8636 io_destroy_buffers(ctx
);
8638 put_cred(ctx
->sq_creds
);
8640 /* there are no registered resources left, nobody uses it */
8642 io_rsrc_node_destroy(ctx
->rsrc_node
);
8643 if (ctx
->rsrc_backup_node
)
8644 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
8645 flush_delayed_work(&ctx
->rsrc_put_work
);
8647 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
8648 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
8650 #if defined(CONFIG_UNIX)
8651 if (ctx
->ring_sock
) {
8652 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
8653 sock_release(ctx
->ring_sock
);
8657 io_mem_free(ctx
->rings
);
8658 io_mem_free(ctx
->sq_sqes
);
8660 percpu_ref_exit(&ctx
->refs
);
8661 free_uid(ctx
->user
);
8662 io_req_caches_free(ctx
);
8664 io_wq_put_hash(ctx
->hash_map
);
8665 kfree(ctx
->cancel_hash
);
8666 kfree(ctx
->dummy_ubuf
);
8670 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
8672 struct io_ring_ctx
*ctx
= file
->private_data
;
8675 poll_wait(file
, &ctx
->poll_wait
, wait
);
8677 * synchronizes with barrier from wq_has_sleeper call in
8681 if (!io_sqring_full(ctx
))
8682 mask
|= EPOLLOUT
| EPOLLWRNORM
;
8685 * Don't flush cqring overflow list here, just do a simple check.
8686 * Otherwise there could possible be ABBA deadlock:
8689 * lock(&ctx->uring_lock);
8691 * lock(&ctx->uring_lock);
8694 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8695 * pushs them to do the flush.
8697 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->check_cq_overflow
))
8698 mask
|= EPOLLIN
| EPOLLRDNORM
;
8703 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
8705 struct io_ring_ctx
*ctx
= file
->private_data
;
8707 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
8710 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8712 const struct cred
*creds
;
8714 creds
= xa_erase(&ctx
->personalities
, id
);
8723 struct io_tctx_exit
{
8724 struct callback_head task_work
;
8725 struct completion completion
;
8726 struct io_ring_ctx
*ctx
;
8729 static void io_tctx_exit_cb(struct callback_head
*cb
)
8731 struct io_uring_task
*tctx
= current
->io_uring
;
8732 struct io_tctx_exit
*work
;
8734 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
8736 * When @in_idle, we're in cancellation and it's racy to remove the
8737 * node. It'll be removed by the end of cancellation, just ignore it.
8739 if (!atomic_read(&tctx
->in_idle
))
8740 io_uring_del_tctx_node((unsigned long)work
->ctx
);
8741 complete(&work
->completion
);
8744 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
8746 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8748 return req
->ctx
== data
;
8751 static void io_ring_exit_work(struct work_struct
*work
)
8753 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
8754 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
8755 struct io_tctx_exit exit
;
8756 struct io_tctx_node
*node
;
8760 * If we're doing polled IO and end up having requests being
8761 * submitted async (out-of-line), then completions can come in while
8762 * we're waiting for refs to drop. We need to reap these manually,
8763 * as nobody else will be looking for them.
8766 io_uring_try_cancel_requests(ctx
, NULL
, true);
8768 struct io_sq_data
*sqd
= ctx
->sq_data
;
8769 struct task_struct
*tsk
;
8771 io_sq_thread_park(sqd
);
8773 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
8774 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
8775 io_cancel_ctx_cb
, ctx
, true);
8776 io_sq_thread_unpark(sqd
);
8779 WARN_ON_ONCE(time_after(jiffies
, timeout
));
8780 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20));
8782 init_completion(&exit
.completion
);
8783 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
8786 * Some may use context even when all refs and requests have been put,
8787 * and they are free to do so while still holding uring_lock or
8788 * completion_lock, see io_req_task_submit(). Apart from other work,
8789 * this lock/unlock section also waits them to finish.
8791 mutex_lock(&ctx
->uring_lock
);
8792 while (!list_empty(&ctx
->tctx_list
)) {
8793 WARN_ON_ONCE(time_after(jiffies
, timeout
));
8795 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
8797 /* don't spin on a single task if cancellation failed */
8798 list_rotate_left(&ctx
->tctx_list
);
8799 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
8800 if (WARN_ON_ONCE(ret
))
8802 wake_up_process(node
->task
);
8804 mutex_unlock(&ctx
->uring_lock
);
8805 wait_for_completion(&exit
.completion
);
8806 mutex_lock(&ctx
->uring_lock
);
8808 mutex_unlock(&ctx
->uring_lock
);
8809 spin_lock_irq(&ctx
->completion_lock
);
8810 spin_unlock_irq(&ctx
->completion_lock
);
8812 io_ring_ctx_free(ctx
);
8815 /* Returns true if we found and killed one or more timeouts */
8816 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
8819 struct io_kiocb
*req
, *tmp
;
8822 spin_lock_irq(&ctx
->completion_lock
);
8823 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
8824 if (io_match_task(req
, tsk
, cancel_all
)) {
8825 io_kill_timeout(req
, -ECANCELED
);
8830 io_commit_cqring(ctx
);
8831 spin_unlock_irq(&ctx
->completion_lock
);
8833 io_cqring_ev_posted(ctx
);
8834 return canceled
!= 0;
8837 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
8839 unsigned long index
;
8840 struct creds
*creds
;
8842 mutex_lock(&ctx
->uring_lock
);
8843 percpu_ref_kill(&ctx
->refs
);
8845 __io_cqring_overflow_flush(ctx
, true);
8846 xa_for_each(&ctx
->personalities
, index
, creds
)
8847 io_unregister_personality(ctx
, index
);
8848 mutex_unlock(&ctx
->uring_lock
);
8850 io_kill_timeouts(ctx
, NULL
, true);
8851 io_poll_remove_all(ctx
, NULL
, true);
8853 /* if we failed setting up the ctx, we might not have any rings */
8854 io_iopoll_try_reap_events(ctx
);
8856 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
8858 * Use system_unbound_wq to avoid spawning tons of event kworkers
8859 * if we're exiting a ton of rings at the same time. It just adds
8860 * noise and overhead, there's no discernable change in runtime
8861 * over using system_wq.
8863 queue_work(system_unbound_wq
, &ctx
->exit_work
);
8866 static int io_uring_release(struct inode
*inode
, struct file
*file
)
8868 struct io_ring_ctx
*ctx
= file
->private_data
;
8870 file
->private_data
= NULL
;
8871 io_ring_ctx_wait_and_kill(ctx
);
8875 struct io_task_cancel
{
8876 struct task_struct
*task
;
8880 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
8882 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
8883 struct io_task_cancel
*cancel
= data
;
8886 if (!cancel
->all
&& (req
->flags
& REQ_F_LINK_TIMEOUT
)) {
8887 unsigned long flags
;
8888 struct io_ring_ctx
*ctx
= req
->ctx
;
8890 /* protect against races with linked timeouts */
8891 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
8892 ret
= io_match_task(req
, cancel
->task
, cancel
->all
);
8893 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
8895 ret
= io_match_task(req
, cancel
->task
, cancel
->all
);
8900 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
8901 struct task_struct
*task
, bool cancel_all
)
8903 struct io_defer_entry
*de
;
8906 spin_lock_irq(&ctx
->completion_lock
);
8907 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
8908 if (io_match_task(de
->req
, task
, cancel_all
)) {
8909 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
8913 spin_unlock_irq(&ctx
->completion_lock
);
8914 if (list_empty(&list
))
8917 while (!list_empty(&list
)) {
8918 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
8919 list_del_init(&de
->list
);
8920 io_req_complete_failed(de
->req
, -ECANCELED
);
8926 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
8928 struct io_tctx_node
*node
;
8929 enum io_wq_cancel cret
;
8932 mutex_lock(&ctx
->uring_lock
);
8933 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
8934 struct io_uring_task
*tctx
= node
->task
->io_uring
;
8937 * io_wq will stay alive while we hold uring_lock, because it's
8938 * killed after ctx nodes, which requires to take the lock.
8940 if (!tctx
|| !tctx
->io_wq
)
8942 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
8943 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8945 mutex_unlock(&ctx
->uring_lock
);
8950 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
8951 struct task_struct
*task
,
8954 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
8955 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
8958 enum io_wq_cancel cret
;
8962 ret
|= io_uring_try_cancel_iowq(ctx
);
8963 } else if (tctx
&& tctx
->io_wq
) {
8965 * Cancels requests of all rings, not only @ctx, but
8966 * it's fine as the task is in exit/exec.
8968 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
8970 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
8973 /* SQPOLL thread does its own polling */
8974 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
8975 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
8976 while (!list_empty_careful(&ctx
->iopoll_list
)) {
8977 io_iopoll_try_reap_events(ctx
);
8982 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
8983 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
8984 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
8986 ret
|= io_run_task_work();
8993 static int __io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
8995 struct io_uring_task
*tctx
= current
->io_uring
;
8996 struct io_tctx_node
*node
;
8999 if (unlikely(!tctx
)) {
9000 ret
= io_uring_alloc_task_context(current
, ctx
);
9003 tctx
= current
->io_uring
;
9005 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
9006 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
9010 node
->task
= current
;
9012 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
9019 mutex_lock(&ctx
->uring_lock
);
9020 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
9021 mutex_unlock(&ctx
->uring_lock
);
9028 * Note that this task has used io_uring. We use it for cancelation purposes.
9030 static inline int io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9032 struct io_uring_task
*tctx
= current
->io_uring
;
9034 if (likely(tctx
&& tctx
->last
== ctx
))
9036 return __io_uring_add_tctx_node(ctx
);
9040 * Remove this io_uring_file -> task mapping.
9042 static void io_uring_del_tctx_node(unsigned long index
)
9044 struct io_uring_task
*tctx
= current
->io_uring
;
9045 struct io_tctx_node
*node
;
9049 node
= xa_erase(&tctx
->xa
, index
);
9053 WARN_ON_ONCE(current
!= node
->task
);
9054 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
9056 mutex_lock(&node
->ctx
->uring_lock
);
9057 list_del(&node
->ctx_node
);
9058 mutex_unlock(&node
->ctx
->uring_lock
);
9060 if (tctx
->last
== node
->ctx
)
9065 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
9067 struct io_wq
*wq
= tctx
->io_wq
;
9068 struct io_tctx_node
*node
;
9069 unsigned long index
;
9071 xa_for_each(&tctx
->xa
, index
, node
)
9072 io_uring_del_tctx_node(index
);
9075 * Must be after io_uring_del_task_file() (removes nodes under
9076 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9079 io_wq_put_and_exit(wq
);
9083 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
9086 return atomic_read(&tctx
->inflight_tracked
);
9087 return percpu_counter_sum(&tctx
->inflight
);
9090 static void io_uring_drop_tctx_refs(struct task_struct
*task
)
9092 struct io_uring_task
*tctx
= task
->io_uring
;
9093 unsigned int refs
= tctx
->cached_refs
;
9095 tctx
->cached_refs
= 0;
9096 percpu_counter_sub(&tctx
->inflight
, refs
);
9097 put_task_struct_many(task
, refs
);
9101 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9102 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9104 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
9106 struct io_uring_task
*tctx
= current
->io_uring
;
9107 struct io_ring_ctx
*ctx
;
9111 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
9113 if (!current
->io_uring
)
9116 io_wq_exit_start(tctx
->io_wq
);
9118 io_uring_drop_tctx_refs(current
);
9119 atomic_inc(&tctx
->in_idle
);
9121 /* read completions before cancelations */
9122 inflight
= tctx_inflight(tctx
, !cancel_all
);
9127 struct io_tctx_node
*node
;
9128 unsigned long index
;
9130 xa_for_each(&tctx
->xa
, index
, node
) {
9131 /* sqpoll task will cancel all its requests */
9132 if (node
->ctx
->sq_data
)
9134 io_uring_try_cancel_requests(node
->ctx
, current
,
9138 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9139 io_uring_try_cancel_requests(ctx
, current
,
9143 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9145 * If we've seen completions, retry without waiting. This
9146 * avoids a race where a completion comes in before we did
9147 * prepare_to_wait().
9149 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
9151 finish_wait(&tctx
->wait
, &wait
);
9153 atomic_dec(&tctx
->in_idle
);
9155 io_uring_clean_tctx(tctx
);
9157 /* for exec all current's requests should be gone, kill tctx */
9158 __io_uring_free(current
);
9162 void __io_uring_cancel(struct files_struct
*files
)
9164 io_uring_cancel_generic(!files
, NULL
);
9167 static void *io_uring_validate_mmap_request(struct file
*file
,
9168 loff_t pgoff
, size_t sz
)
9170 struct io_ring_ctx
*ctx
= file
->private_data
;
9171 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9176 case IORING_OFF_SQ_RING
:
9177 case IORING_OFF_CQ_RING
:
9180 case IORING_OFF_SQES
:
9184 return ERR_PTR(-EINVAL
);
9187 page
= virt_to_head_page(ptr
);
9188 if (sz
> page_size(page
))
9189 return ERR_PTR(-EINVAL
);
9196 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9198 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9202 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9204 return PTR_ERR(ptr
);
9206 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9207 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9210 #else /* !CONFIG_MMU */
9212 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9214 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9217 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9219 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9222 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9223 unsigned long addr
, unsigned long len
,
9224 unsigned long pgoff
, unsigned long flags
)
9228 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9230 return PTR_ERR(ptr
);
9232 return (unsigned long) ptr
;
9235 #endif /* !CONFIG_MMU */
9237 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9242 if (!io_sqring_full(ctx
))
9244 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9246 if (!io_sqring_full(ctx
))
9249 } while (!signal_pending(current
));
9251 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9255 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9256 struct __kernel_timespec __user
**ts
,
9257 const sigset_t __user
**sig
)
9259 struct io_uring_getevents_arg arg
;
9262 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9263 * is just a pointer to the sigset_t.
9265 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9266 *sig
= (const sigset_t __user
*) argp
;
9272 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9273 * timespec and sigset_t pointers if good.
9275 if (*argsz
!= sizeof(arg
))
9277 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9279 *sig
= u64_to_user_ptr(arg
.sigmask
);
9280 *argsz
= arg
.sigmask_sz
;
9281 *ts
= u64_to_user_ptr(arg
.ts
);
9285 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9286 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9289 struct io_ring_ctx
*ctx
;
9296 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9297 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
)))
9301 if (unlikely(!f
.file
))
9305 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
9309 ctx
= f
.file
->private_data
;
9310 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
9314 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9318 * For SQ polling, the thread will do all submissions and completions.
9319 * Just return the requested submit count, and wake the thread if
9323 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9324 io_cqring_overflow_flush(ctx
, false);
9327 if (unlikely(ctx
->sq_data
->thread
== NULL
))
9329 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9330 wake_up(&ctx
->sq_data
->wait
);
9331 if (flags
& IORING_ENTER_SQ_WAIT
) {
9332 ret
= io_sqpoll_wait_sq(ctx
);
9336 submitted
= to_submit
;
9337 } else if (to_submit
) {
9338 ret
= io_uring_add_tctx_node(ctx
);
9341 mutex_lock(&ctx
->uring_lock
);
9342 submitted
= io_submit_sqes(ctx
, to_submit
);
9343 mutex_unlock(&ctx
->uring_lock
);
9345 if (submitted
!= to_submit
)
9348 if (flags
& IORING_ENTER_GETEVENTS
) {
9349 const sigset_t __user
*sig
;
9350 struct __kernel_timespec __user
*ts
;
9352 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9356 min_complete
= min(min_complete
, ctx
->cq_entries
);
9359 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9360 * space applications don't need to do io completion events
9361 * polling again, they can rely on io_sq_thread to do polling
9362 * work, which can reduce cpu usage and uring_lock contention.
9364 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
9365 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9366 ret
= io_iopoll_check(ctx
, min_complete
);
9368 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
9373 percpu_ref_put(&ctx
->refs
);
9376 return submitted
? submitted
: ret
;
9379 #ifdef CONFIG_PROC_FS
9380 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
9381 const struct cred
*cred
)
9383 struct user_namespace
*uns
= seq_user_ns(m
);
9384 struct group_info
*gi
;
9389 seq_printf(m
, "%5d\n", id
);
9390 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
9391 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
9392 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
9393 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
9394 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
9395 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
9396 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
9397 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
9398 seq_puts(m
, "\n\tGroups:\t");
9399 gi
= cred
->group_info
;
9400 for (g
= 0; g
< gi
->ngroups
; g
++) {
9401 seq_put_decimal_ull(m
, g
? " " : "",
9402 from_kgid_munged(uns
, gi
->gid
[g
]));
9404 seq_puts(m
, "\n\tCapEff:\t");
9405 cap
= cred
->cap_effective
;
9406 CAP_FOR_EACH_U32(__capi
)
9407 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
9412 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
9414 struct io_sq_data
*sq
= NULL
;
9419 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9420 * since fdinfo case grabs it in the opposite direction of normal use
9421 * cases. If we fail to get the lock, we just don't iterate any
9422 * structures that could be going away outside the io_uring mutex.
9424 has_lock
= mutex_trylock(&ctx
->uring_lock
);
9426 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
9432 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
9433 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
9434 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
9435 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
9436 struct file
*f
= io_file_from_index(ctx
, i
);
9439 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
9441 seq_printf(m
, "%5u: <none>\n", i
);
9443 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
9444 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
9445 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
9446 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
9448 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
9450 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
9451 unsigned long index
;
9452 const struct cred
*cred
;
9454 seq_printf(m
, "Personalities:\n");
9455 xa_for_each(&ctx
->personalities
, index
, cred
)
9456 io_uring_show_cred(m
, index
, cred
);
9458 seq_printf(m
, "PollList:\n");
9459 spin_lock_irq(&ctx
->completion_lock
);
9460 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
9461 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
9462 struct io_kiocb
*req
;
9464 hlist_for_each_entry(req
, list
, hash_node
)
9465 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
9466 req
->task
->task_works
!= NULL
);
9468 spin_unlock_irq(&ctx
->completion_lock
);
9470 mutex_unlock(&ctx
->uring_lock
);
9473 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
9475 struct io_ring_ctx
*ctx
= f
->private_data
;
9477 if (percpu_ref_tryget(&ctx
->refs
)) {
9478 __io_uring_show_fdinfo(ctx
, m
);
9479 percpu_ref_put(&ctx
->refs
);
9484 static const struct file_operations io_uring_fops
= {
9485 .release
= io_uring_release
,
9486 .mmap
= io_uring_mmap
,
9488 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
9489 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
9491 .poll
= io_uring_poll
,
9492 .fasync
= io_uring_fasync
,
9493 #ifdef CONFIG_PROC_FS
9494 .show_fdinfo
= io_uring_show_fdinfo
,
9498 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
9499 struct io_uring_params
*p
)
9501 struct io_rings
*rings
;
9502 size_t size
, sq_array_offset
;
9504 /* make sure these are sane, as we already accounted them */
9505 ctx
->sq_entries
= p
->sq_entries
;
9506 ctx
->cq_entries
= p
->cq_entries
;
9508 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
9509 if (size
== SIZE_MAX
)
9512 rings
= io_mem_alloc(size
);
9517 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
9518 rings
->sq_ring_mask
= p
->sq_entries
- 1;
9519 rings
->cq_ring_mask
= p
->cq_entries
- 1;
9520 rings
->sq_ring_entries
= p
->sq_entries
;
9521 rings
->cq_ring_entries
= p
->cq_entries
;
9523 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
9524 if (size
== SIZE_MAX
) {
9525 io_mem_free(ctx
->rings
);
9530 ctx
->sq_sqes
= io_mem_alloc(size
);
9531 if (!ctx
->sq_sqes
) {
9532 io_mem_free(ctx
->rings
);
9540 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
9544 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
9548 ret
= io_uring_add_tctx_node(ctx
);
9553 fd_install(fd
, file
);
9558 * Allocate an anonymous fd, this is what constitutes the application
9559 * visible backing of an io_uring instance. The application mmaps this
9560 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9561 * we have to tie this fd to a socket for file garbage collection purposes.
9563 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
9566 #if defined(CONFIG_UNIX)
9569 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
9572 return ERR_PTR(ret
);
9575 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
9576 O_RDWR
| O_CLOEXEC
);
9577 #if defined(CONFIG_UNIX)
9579 sock_release(ctx
->ring_sock
);
9580 ctx
->ring_sock
= NULL
;
9582 ctx
->ring_sock
->file
= file
;
9588 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
9589 struct io_uring_params __user
*params
)
9591 struct io_ring_ctx
*ctx
;
9597 if (entries
> IORING_MAX_ENTRIES
) {
9598 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9600 entries
= IORING_MAX_ENTRIES
;
9604 * Use twice as many entries for the CQ ring. It's possible for the
9605 * application to drive a higher depth than the size of the SQ ring,
9606 * since the sqes are only used at submission time. This allows for
9607 * some flexibility in overcommitting a bit. If the application has
9608 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9609 * of CQ ring entries manually.
9611 p
->sq_entries
= roundup_pow_of_two(entries
);
9612 if (p
->flags
& IORING_SETUP_CQSIZE
) {
9614 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9615 * to a power-of-two, if it isn't already. We do NOT impose
9616 * any cq vs sq ring sizing.
9620 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
9621 if (!(p
->flags
& IORING_SETUP_CLAMP
))
9623 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
9625 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
9626 if (p
->cq_entries
< p
->sq_entries
)
9629 p
->cq_entries
= 2 * p
->sq_entries
;
9632 ctx
= io_ring_ctx_alloc(p
);
9635 ctx
->compat
= in_compat_syscall();
9636 if (!capable(CAP_IPC_LOCK
))
9637 ctx
->user
= get_uid(current_user());
9640 * This is just grabbed for accounting purposes. When a process exits,
9641 * the mm is exited and dropped before the files, hence we need to hang
9642 * on to this mm purely for the purposes of being able to unaccount
9643 * memory (locked/pinned vm). It's not used for anything else.
9645 mmgrab(current
->mm
);
9646 ctx
->mm_account
= current
->mm
;
9648 ret
= io_allocate_scq_urings(ctx
, p
);
9652 ret
= io_sq_offload_create(ctx
, p
);
9655 /* always set a rsrc node */
9656 ret
= io_rsrc_node_switch_start(ctx
);
9659 io_rsrc_node_switch(ctx
, NULL
);
9661 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
9662 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
9663 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
9664 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
9665 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
9666 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
9667 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
9668 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
9670 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
9671 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
9672 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
9673 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
9674 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
9675 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
9676 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
9677 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
9679 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
9680 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
9681 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
9682 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
9683 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
9684 IORING_FEAT_RSRC_TAGS
;
9686 if (copy_to_user(params
, p
, sizeof(*p
))) {
9691 file
= io_uring_get_file(ctx
);
9693 ret
= PTR_ERR(file
);
9698 * Install ring fd as the very last thing, so we don't risk someone
9699 * having closed it before we finish setup
9701 ret
= io_uring_install_fd(ctx
, file
);
9703 /* fput will clean it up */
9708 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
9711 io_ring_ctx_wait_and_kill(ctx
);
9716 * Sets up an aio uring context, and returns the fd. Applications asks for a
9717 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9718 * params structure passed in.
9720 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
9722 struct io_uring_params p
;
9725 if (copy_from_user(&p
, params
, sizeof(p
)))
9727 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
9732 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
9733 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
9734 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
9735 IORING_SETUP_R_DISABLED
))
9738 return io_uring_create(entries
, &p
, params
);
9741 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
9742 struct io_uring_params __user
*, params
)
9744 return io_uring_setup(entries
, params
);
9747 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
9749 struct io_uring_probe
*p
;
9753 size
= struct_size(p
, ops
, nr_args
);
9754 if (size
== SIZE_MAX
)
9756 p
= kzalloc(size
, GFP_KERNEL
);
9761 if (copy_from_user(p
, arg
, size
))
9764 if (memchr_inv(p
, 0, size
))
9767 p
->last_op
= IORING_OP_LAST
- 1;
9768 if (nr_args
> IORING_OP_LAST
)
9769 nr_args
= IORING_OP_LAST
;
9771 for (i
= 0; i
< nr_args
; i
++) {
9773 if (!io_op_defs
[i
].not_supported
)
9774 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
9779 if (copy_to_user(arg
, p
, size
))
9786 static int io_register_personality(struct io_ring_ctx
*ctx
)
9788 const struct cred
*creds
;
9792 creds
= get_current_cred();
9794 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
9795 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
9802 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
9803 unsigned int nr_args
)
9805 struct io_uring_restriction
*res
;
9809 /* Restrictions allowed only if rings started disabled */
9810 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9813 /* We allow only a single restrictions registration */
9814 if (ctx
->restrictions
.registered
)
9817 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
9820 size
= array_size(nr_args
, sizeof(*res
));
9821 if (size
== SIZE_MAX
)
9824 res
= memdup_user(arg
, size
);
9826 return PTR_ERR(res
);
9830 for (i
= 0; i
< nr_args
; i
++) {
9831 switch (res
[i
].opcode
) {
9832 case IORING_RESTRICTION_REGISTER_OP
:
9833 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
9838 __set_bit(res
[i
].register_op
,
9839 ctx
->restrictions
.register_op
);
9841 case IORING_RESTRICTION_SQE_OP
:
9842 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
9847 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
9849 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
9850 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
9852 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
9853 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
9862 /* Reset all restrictions if an error happened */
9864 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
9866 ctx
->restrictions
.registered
= true;
9872 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
9874 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9877 if (ctx
->restrictions
.registered
)
9878 ctx
->restricted
= 1;
9880 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
9881 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
9882 wake_up(&ctx
->sq_data
->wait
);
9886 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
9887 struct io_uring_rsrc_update2
*up
,
9895 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
9897 err
= io_rsrc_node_switch_start(ctx
);
9902 case IORING_RSRC_FILE
:
9903 return __io_sqe_files_update(ctx
, up
, nr_args
);
9904 case IORING_RSRC_BUFFER
:
9905 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
9910 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
9913 struct io_uring_rsrc_update2 up
;
9917 memset(&up
, 0, sizeof(up
));
9918 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
9920 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
9923 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
9924 unsigned size
, unsigned type
)
9926 struct io_uring_rsrc_update2 up
;
9928 if (size
!= sizeof(up
))
9930 if (copy_from_user(&up
, arg
, sizeof(up
)))
9932 if (!up
.nr
|| up
.resv
)
9934 return __io_register_rsrc_update(ctx
, type
, &up
, up
.nr
);
9937 static int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
9938 unsigned int size
, unsigned int type
)
9940 struct io_uring_rsrc_register rr
;
9942 /* keep it extendible */
9943 if (size
!= sizeof(rr
))
9946 memset(&rr
, 0, sizeof(rr
));
9947 if (copy_from_user(&rr
, arg
, size
))
9949 if (!rr
.nr
|| rr
.resv
|| rr
.resv2
)
9953 case IORING_RSRC_FILE
:
9954 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
9955 rr
.nr
, u64_to_user_ptr(rr
.tags
));
9956 case IORING_RSRC_BUFFER
:
9957 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
9958 rr
.nr
, u64_to_user_ptr(rr
.tags
));
9963 static int io_register_iowq_aff(struct io_ring_ctx
*ctx
, void __user
*arg
,
9966 struct io_uring_task
*tctx
= current
->io_uring
;
9967 cpumask_var_t new_mask
;
9970 if (!tctx
|| !tctx
->io_wq
)
9973 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
9976 cpumask_clear(new_mask
);
9977 if (len
> cpumask_size())
9978 len
= cpumask_size();
9980 if (copy_from_user(new_mask
, arg
, len
)) {
9981 free_cpumask_var(new_mask
);
9985 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
9986 free_cpumask_var(new_mask
);
9990 static int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
9992 struct io_uring_task
*tctx
= current
->io_uring
;
9994 if (!tctx
|| !tctx
->io_wq
)
9997 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
10000 static bool io_register_op_must_quiesce(int op
)
10003 case IORING_REGISTER_BUFFERS
:
10004 case IORING_UNREGISTER_BUFFERS
:
10005 case IORING_REGISTER_FILES
:
10006 case IORING_UNREGISTER_FILES
:
10007 case IORING_REGISTER_FILES_UPDATE
:
10008 case IORING_REGISTER_PROBE
:
10009 case IORING_REGISTER_PERSONALITY
:
10010 case IORING_UNREGISTER_PERSONALITY
:
10011 case IORING_REGISTER_FILES2
:
10012 case IORING_REGISTER_FILES_UPDATE2
:
10013 case IORING_REGISTER_BUFFERS2
:
10014 case IORING_REGISTER_BUFFERS_UPDATE
:
10015 case IORING_REGISTER_IOWQ_AFF
:
10016 case IORING_UNREGISTER_IOWQ_AFF
:
10023 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10024 void __user
*arg
, unsigned nr_args
)
10025 __releases(ctx
->uring_lock
)
10026 __acquires(ctx
->uring_lock
)
10031 * We're inside the ring mutex, if the ref is already dying, then
10032 * someone else killed the ctx or is already going through
10033 * io_uring_register().
10035 if (percpu_ref_is_dying(&ctx
->refs
))
10038 if (ctx
->restricted
) {
10039 if (opcode
>= IORING_REGISTER_LAST
)
10041 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
10042 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
10046 if (io_register_op_must_quiesce(opcode
)) {
10047 percpu_ref_kill(&ctx
->refs
);
10050 * Drop uring mutex before waiting for references to exit. If
10051 * another thread is currently inside io_uring_enter() it might
10052 * need to grab the uring_lock to make progress. If we hold it
10053 * here across the drain wait, then we can deadlock. It's safe
10054 * to drop the mutex here, since no new references will come in
10055 * after we've killed the percpu ref.
10057 mutex_unlock(&ctx
->uring_lock
);
10059 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10062 ret
= io_run_task_work_sig();
10066 mutex_lock(&ctx
->uring_lock
);
10069 io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
);
10075 case IORING_REGISTER_BUFFERS
:
10076 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
10078 case IORING_UNREGISTER_BUFFERS
:
10080 if (arg
|| nr_args
)
10082 ret
= io_sqe_buffers_unregister(ctx
);
10084 case IORING_REGISTER_FILES
:
10085 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
10087 case IORING_UNREGISTER_FILES
:
10089 if (arg
|| nr_args
)
10091 ret
= io_sqe_files_unregister(ctx
);
10093 case IORING_REGISTER_FILES_UPDATE
:
10094 ret
= io_register_files_update(ctx
, arg
, nr_args
);
10096 case IORING_REGISTER_EVENTFD
:
10097 case IORING_REGISTER_EVENTFD_ASYNC
:
10101 ret
= io_eventfd_register(ctx
, arg
);
10104 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10105 ctx
->eventfd_async
= 1;
10107 ctx
->eventfd_async
= 0;
10109 case IORING_UNREGISTER_EVENTFD
:
10111 if (arg
|| nr_args
)
10113 ret
= io_eventfd_unregister(ctx
);
10115 case IORING_REGISTER_PROBE
:
10117 if (!arg
|| nr_args
> 256)
10119 ret
= io_probe(ctx
, arg
, nr_args
);
10121 case IORING_REGISTER_PERSONALITY
:
10123 if (arg
|| nr_args
)
10125 ret
= io_register_personality(ctx
);
10127 case IORING_UNREGISTER_PERSONALITY
:
10131 ret
= io_unregister_personality(ctx
, nr_args
);
10133 case IORING_REGISTER_ENABLE_RINGS
:
10135 if (arg
|| nr_args
)
10137 ret
= io_register_enable_rings(ctx
);
10139 case IORING_REGISTER_RESTRICTIONS
:
10140 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10142 case IORING_REGISTER_FILES2
:
10143 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
10145 case IORING_REGISTER_FILES_UPDATE2
:
10146 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10149 case IORING_REGISTER_BUFFERS2
:
10150 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
10152 case IORING_REGISTER_BUFFERS_UPDATE
:
10153 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10154 IORING_RSRC_BUFFER
);
10156 case IORING_REGISTER_IOWQ_AFF
:
10158 if (!arg
|| !nr_args
)
10160 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
10162 case IORING_UNREGISTER_IOWQ_AFF
:
10164 if (arg
|| nr_args
)
10166 ret
= io_unregister_iowq_aff(ctx
);
10173 if (io_register_op_must_quiesce(opcode
)) {
10174 /* bring the ctx back to life */
10175 percpu_ref_reinit(&ctx
->refs
);
10176 reinit_completion(&ctx
->ref_comp
);
10181 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10182 void __user
*, arg
, unsigned int, nr_args
)
10184 struct io_ring_ctx
*ctx
;
10193 if (f
.file
->f_op
!= &io_uring_fops
)
10196 ctx
= f
.file
->private_data
;
10198 io_run_task_work();
10200 mutex_lock(&ctx
->uring_lock
);
10201 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10202 mutex_unlock(&ctx
->uring_lock
);
10203 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10204 ctx
->cq_ev_fd
!= NULL
, ret
);
10210 static int __init
io_uring_init(void)
10212 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10213 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10214 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10217 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10218 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10219 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10220 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10221 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10222 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10223 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10224 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10225 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10226 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10227 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10228 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10229 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10230 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10231 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10232 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10233 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10234 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10235 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10236 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10237 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10238 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10239 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10240 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10241 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10242 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10243 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10244 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10245 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10246 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
10247 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10248 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10250 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
10251 sizeof(struct io_uring_rsrc_update
));
10252 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
10253 sizeof(struct io_uring_rsrc_update2
));
10254 /* should fit into one byte */
10255 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
10257 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
10258 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
10260 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
10264 __initcall(io_uring_init
);