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>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp
;
116 u32 tail ____cacheline_aligned_in_smp
;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq
, cq
;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask
, cq_ring_mask
;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries
, cq_ring_entries
;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
197 enum io_uring_cmd_flags
{
198 IO_URING_F_NONBLOCK
= 1,
199 IO_URING_F_COMPLETE_DEFER
= 2,
202 struct io_mapped_ubuf
{
205 unsigned int nr_bvecs
;
206 unsigned long acct_pages
;
207 struct bio_vec bvec
[];
212 struct io_overflow_cqe
{
213 struct io_uring_cqe cqe
;
214 struct list_head list
;
217 struct io_fixed_file
{
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr
;
223 struct list_head list
;
228 struct io_mapped_ubuf
*buf
;
232 struct io_file_table
{
233 struct io_fixed_file
*files
;
236 struct io_rsrc_node
{
237 struct percpu_ref refs
;
238 struct list_head node
;
239 struct list_head rsrc_list
;
240 struct io_rsrc_data
*rsrc_data
;
241 struct llist_node llist
;
245 typedef void (rsrc_put_fn
)(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
);
247 struct io_rsrc_data
{
248 struct io_ring_ctx
*ctx
;
254 struct completion done
;
259 struct list_head list
;
265 struct io_restriction
{
266 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
267 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
268 u8 sqe_flags_allowed
;
269 u8 sqe_flags_required
;
274 IO_SQ_THREAD_SHOULD_STOP
= 0,
275 IO_SQ_THREAD_SHOULD_PARK
,
280 atomic_t park_pending
;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list
;
286 struct task_struct
*thread
;
287 struct wait_queue_head wait
;
289 unsigned sq_thread_idle
;
295 struct completion exited
;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link
{
303 struct io_kiocb
*head
;
304 struct io_kiocb
*last
;
307 struct io_submit_state
{
308 struct blk_plug plug
;
309 struct io_submit_link link
;
312 * io_kiocb alloc cache
314 void *reqs
[IO_REQ_CACHE_SIZE
];
315 unsigned int free_reqs
;
320 * Batch completion logic
322 struct io_kiocb
*compl_reqs
[IO_COMPL_BATCH
];
323 unsigned int compl_nr
;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list
;
327 unsigned int ios_left
;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs
;
335 struct io_rings
*rings
;
337 unsigned int compat
: 1;
338 unsigned int drain_next
: 1;
339 unsigned int eventfd_async
: 1;
340 unsigned int restricted
: 1;
341 unsigned int off_timeout_used
: 1;
342 unsigned int drain_active
: 1;
343 } ____cacheline_aligned_in_smp
;
345 /* submission data */
347 struct mutex uring_lock
;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe
*sq_sqes
;
362 unsigned cached_sq_head
;
364 struct list_head defer_list
;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node
*rsrc_node
;
371 struct io_file_table file_table
;
372 unsigned nr_user_files
;
373 unsigned nr_user_bufs
;
374 struct io_mapped_ubuf
**user_bufs
;
376 struct io_submit_state submit_state
;
377 struct list_head timeout_list
;
378 struct list_head ltimeout_list
;
379 struct list_head cq_overflow_list
;
380 struct xarray io_buffers
;
381 struct xarray personalities
;
383 unsigned sq_thread_idle
;
384 } ____cacheline_aligned_in_smp
;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list
;
388 unsigned int locked_free_nr
;
390 const struct cred
*sq_creds
; /* cred used for __io_sq_thread() */
391 struct io_sq_data
*sq_data
; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait
;
394 struct list_head sqd_list
;
396 unsigned long check_cq_overflow
;
399 unsigned cached_cq_tail
;
401 struct eventfd_ctx
*cq_ev_fd
;
402 struct wait_queue_head poll_wait
;
403 struct wait_queue_head cq_wait
;
405 atomic_t cq_timeouts
;
406 unsigned cq_last_tm_flush
;
407 } ____cacheline_aligned_in_smp
;
410 spinlock_t completion_lock
;
412 spinlock_t timeout_lock
;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list
;
421 struct hlist_head
*cancel_hash
;
422 unsigned cancel_hash_bits
;
423 bool poll_multi_queue
;
424 } ____cacheline_aligned_in_smp
;
426 struct io_restriction restrictions
;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node
*rsrc_backup_node
;
431 struct io_mapped_ubuf
*dummy_ubuf
;
432 struct io_rsrc_data
*file_data
;
433 struct io_rsrc_data
*buf_data
;
435 struct delayed_work rsrc_put_work
;
436 struct llist_head rsrc_put_llist
;
437 struct list_head rsrc_ref_list
;
438 spinlock_t rsrc_ref_lock
;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket
*ring_sock
;
446 /* hashed buffered write serialization */
447 struct io_wq_hash
*hash_map
;
449 /* Only used for accounting purposes */
450 struct user_struct
*user
;
451 struct mm_struct
*mm_account
;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist
;
455 struct delayed_work fallback_work
;
456 struct work_struct exit_work
;
457 struct list_head tctx_list
;
458 struct completion ref_comp
;
460 bool iowq_limits_set
;
464 struct io_uring_task
{
465 /* submission side */
468 struct wait_queue_head wait
;
469 const struct io_ring_ctx
*last
;
471 struct percpu_counter inflight
;
472 atomic_t inflight_tracked
;
475 spinlock_t task_lock
;
476 struct io_wq_work_list task_list
;
477 struct callback_head task_work
;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb
{
487 struct wait_queue_head
*head
;
491 struct wait_queue_entry wait
;
494 struct io_poll_update
{
500 bool update_user_data
;
509 struct io_timeout_data
{
510 struct io_kiocb
*req
;
511 struct hrtimer timer
;
512 struct timespec64 ts
;
513 enum hrtimer_mode mode
;
519 struct sockaddr __user
*addr
;
520 int __user
*addr_len
;
523 unsigned long nofile
;
543 struct list_head list
;
544 /* head of the link, used by linked timeouts only */
545 struct io_kiocb
*head
;
546 /* for linked completions */
547 struct io_kiocb
*prev
;
550 struct io_timeout_rem
{
555 struct timespec64 ts
;
561 /* NOTE: kiocb has the file as the first member, so don't do it here */
569 struct sockaddr __user
*addr
;
576 struct compat_msghdr __user
*umsg_compat
;
577 struct user_msghdr __user
*umsg
;
583 struct io_buffer
*kbuf
;
590 struct filename
*filename
;
592 unsigned long nofile
;
595 struct io_rsrc_update
{
621 struct epoll_event event
;
625 struct file
*file_out
;
633 struct io_provide_buf
{
647 const char __user
*filename
;
648 struct statx __user
*buffer
;
660 struct filename
*oldpath
;
661 struct filename
*newpath
;
669 struct filename
*filename
;
676 struct filename
*filename
;
682 struct filename
*oldpath
;
683 struct filename
*newpath
;
690 struct filename
*oldpath
;
691 struct filename
*newpath
;
695 struct io_completion
{
700 struct io_async_connect
{
701 struct sockaddr_storage address
;
704 struct io_async_msghdr
{
705 struct iovec fast_iov
[UIO_FASTIOV
];
706 /* points to an allocated iov, if NULL we use fast_iov instead */
707 struct iovec
*free_iov
;
708 struct sockaddr __user
*uaddr
;
710 struct sockaddr_storage addr
;
714 struct iovec fast_iov
[UIO_FASTIOV
];
715 const struct iovec
*free_iovec
;
716 struct iov_iter iter
;
717 struct iov_iter_state iter_state
;
719 struct wait_page_queue wpq
;
723 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
724 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
725 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
726 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
727 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
728 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
730 /* first byte is taken by user flags, shift it to not overlap */
735 REQ_F_LINK_TIMEOUT_BIT
,
736 REQ_F_NEED_CLEANUP_BIT
,
738 REQ_F_BUFFER_SELECTED_BIT
,
739 REQ_F_COMPLETE_INLINE_BIT
,
743 REQ_F_ARM_LTIMEOUT_BIT
,
744 /* keep async read/write and isreg together and in order */
745 REQ_F_NOWAIT_READ_BIT
,
746 REQ_F_NOWAIT_WRITE_BIT
,
749 /* not a real bit, just to check we're not overflowing the space */
755 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
756 /* drain existing IO first */
757 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
759 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
760 /* doesn't sever on completion < 0 */
761 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
763 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
764 /* IOSQE_BUFFER_SELECT */
765 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
767 /* fail rest of links */
768 REQ_F_FAIL
= BIT(REQ_F_FAIL_BIT
),
769 /* on inflight list, should be cancelled and waited on exit reliably */
770 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
771 /* read/write uses file position */
772 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
773 /* must not punt to workers */
774 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
775 /* has or had linked timeout */
776 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
778 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
779 /* already went through poll handler */
780 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
781 /* buffer already selected */
782 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
783 /* completion is deferred through io_comp_state */
784 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
785 /* caller should reissue async */
786 REQ_F_REISSUE
= BIT(REQ_F_REISSUE_BIT
),
787 /* supports async reads */
788 REQ_F_NOWAIT_READ
= BIT(REQ_F_NOWAIT_READ_BIT
),
789 /* supports async writes */
790 REQ_F_NOWAIT_WRITE
= BIT(REQ_F_NOWAIT_WRITE_BIT
),
792 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
793 /* has creds assigned */
794 REQ_F_CREDS
= BIT(REQ_F_CREDS_BIT
),
795 /* skip refcounting if not set */
796 REQ_F_REFCOUNT
= BIT(REQ_F_REFCOUNT_BIT
),
797 /* there is a linked timeout that has to be armed */
798 REQ_F_ARM_LTIMEOUT
= BIT(REQ_F_ARM_LTIMEOUT_BIT
),
802 struct io_poll_iocb poll
;
803 struct io_poll_iocb
*double_poll
;
806 typedef void (*io_req_tw_func_t
)(struct io_kiocb
*req
, bool *locked
);
808 struct io_task_work
{
810 struct io_wq_work_node node
;
811 struct llist_node fallback_node
;
813 io_req_tw_func_t func
;
817 IORING_RSRC_FILE
= 0,
818 IORING_RSRC_BUFFER
= 1,
822 * NOTE! Each of the iocb union members has the file pointer
823 * as the first entry in their struct definition. So you can
824 * access the file pointer through any of the sub-structs,
825 * or directly as just 'ki_filp' in this struct.
831 struct io_poll_iocb poll
;
832 struct io_poll_update poll_update
;
833 struct io_accept accept
;
835 struct io_cancel cancel
;
836 struct io_timeout timeout
;
837 struct io_timeout_rem timeout_rem
;
838 struct io_connect connect
;
839 struct io_sr_msg sr_msg
;
841 struct io_close close
;
842 struct io_rsrc_update rsrc_update
;
843 struct io_fadvise fadvise
;
844 struct io_madvise madvise
;
845 struct io_epoll epoll
;
846 struct io_splice splice
;
847 struct io_provide_buf pbuf
;
848 struct io_statx statx
;
849 struct io_shutdown shutdown
;
850 struct io_rename rename
;
851 struct io_unlink unlink
;
852 struct io_mkdir mkdir
;
853 struct io_symlink symlink
;
854 struct io_hardlink hardlink
;
855 /* use only after cleaning per-op data, see io_clean_op() */
856 struct io_completion
compl;
859 /* opcode allocated if it needs to store data for async defer */
862 /* polled IO has completed */
868 struct io_ring_ctx
*ctx
;
871 struct task_struct
*task
;
874 struct io_kiocb
*link
;
875 struct percpu_ref
*fixed_rsrc_refs
;
877 /* used with ctx->iopoll_list with reads/writes */
878 struct list_head inflight_entry
;
879 struct io_task_work io_task_work
;
880 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
881 struct hlist_node hash_node
;
882 struct async_poll
*apoll
;
883 struct io_wq_work work
;
884 const struct cred
*creds
;
886 /* store used ubuf, so we can prevent reloading */
887 struct io_mapped_ubuf
*imu
;
890 struct io_tctx_node
{
891 struct list_head ctx_node
;
892 struct task_struct
*task
;
893 struct io_ring_ctx
*ctx
;
896 struct io_defer_entry
{
897 struct list_head list
;
898 struct io_kiocb
*req
;
903 /* needs req->file assigned */
904 unsigned needs_file
: 1;
905 /* hash wq insertion if file is a regular file */
906 unsigned hash_reg_file
: 1;
907 /* unbound wq insertion if file is a non-regular file */
908 unsigned unbound_nonreg_file
: 1;
909 /* opcode is not supported by this kernel */
910 unsigned not_supported
: 1;
911 /* set if opcode supports polled "wait" */
913 unsigned pollout
: 1;
914 /* op supports buffer selection */
915 unsigned buffer_select
: 1;
916 /* do prep async if is going to be punted */
917 unsigned needs_async_setup
: 1;
918 /* should block plug */
920 /* size of async data needed, if any */
921 unsigned short async_size
;
924 static const struct io_op_def io_op_defs
[] = {
925 [IORING_OP_NOP
] = {},
926 [IORING_OP_READV
] = {
928 .unbound_nonreg_file
= 1,
931 .needs_async_setup
= 1,
933 .async_size
= sizeof(struct io_async_rw
),
935 [IORING_OP_WRITEV
] = {
938 .unbound_nonreg_file
= 1,
940 .needs_async_setup
= 1,
942 .async_size
= sizeof(struct io_async_rw
),
944 [IORING_OP_FSYNC
] = {
947 [IORING_OP_READ_FIXED
] = {
949 .unbound_nonreg_file
= 1,
952 .async_size
= sizeof(struct io_async_rw
),
954 [IORING_OP_WRITE_FIXED
] = {
957 .unbound_nonreg_file
= 1,
960 .async_size
= sizeof(struct io_async_rw
),
962 [IORING_OP_POLL_ADD
] = {
964 .unbound_nonreg_file
= 1,
966 [IORING_OP_POLL_REMOVE
] = {},
967 [IORING_OP_SYNC_FILE_RANGE
] = {
970 [IORING_OP_SENDMSG
] = {
972 .unbound_nonreg_file
= 1,
974 .needs_async_setup
= 1,
975 .async_size
= sizeof(struct io_async_msghdr
),
977 [IORING_OP_RECVMSG
] = {
979 .unbound_nonreg_file
= 1,
982 .needs_async_setup
= 1,
983 .async_size
= sizeof(struct io_async_msghdr
),
985 [IORING_OP_TIMEOUT
] = {
986 .async_size
= sizeof(struct io_timeout_data
),
988 [IORING_OP_TIMEOUT_REMOVE
] = {
989 /* used by timeout updates' prep() */
991 [IORING_OP_ACCEPT
] = {
993 .unbound_nonreg_file
= 1,
996 [IORING_OP_ASYNC_CANCEL
] = {},
997 [IORING_OP_LINK_TIMEOUT
] = {
998 .async_size
= sizeof(struct io_timeout_data
),
1000 [IORING_OP_CONNECT
] = {
1002 .unbound_nonreg_file
= 1,
1004 .needs_async_setup
= 1,
1005 .async_size
= sizeof(struct io_async_connect
),
1007 [IORING_OP_FALLOCATE
] = {
1010 [IORING_OP_OPENAT
] = {},
1011 [IORING_OP_CLOSE
] = {},
1012 [IORING_OP_FILES_UPDATE
] = {},
1013 [IORING_OP_STATX
] = {},
1014 [IORING_OP_READ
] = {
1016 .unbound_nonreg_file
= 1,
1020 .async_size
= sizeof(struct io_async_rw
),
1022 [IORING_OP_WRITE
] = {
1025 .unbound_nonreg_file
= 1,
1028 .async_size
= sizeof(struct io_async_rw
),
1030 [IORING_OP_FADVISE
] = {
1033 [IORING_OP_MADVISE
] = {},
1034 [IORING_OP_SEND
] = {
1036 .unbound_nonreg_file
= 1,
1039 [IORING_OP_RECV
] = {
1041 .unbound_nonreg_file
= 1,
1045 [IORING_OP_OPENAT2
] = {
1047 [IORING_OP_EPOLL_CTL
] = {
1048 .unbound_nonreg_file
= 1,
1050 [IORING_OP_SPLICE
] = {
1053 .unbound_nonreg_file
= 1,
1055 [IORING_OP_PROVIDE_BUFFERS
] = {},
1056 [IORING_OP_REMOVE_BUFFERS
] = {},
1060 .unbound_nonreg_file
= 1,
1062 [IORING_OP_SHUTDOWN
] = {
1065 [IORING_OP_RENAMEAT
] = {},
1066 [IORING_OP_UNLINKAT
] = {},
1067 [IORING_OP_MKDIRAT
] = {},
1068 [IORING_OP_SYMLINKAT
] = {},
1069 [IORING_OP_LINKAT
] = {},
1072 /* requests with any of those set should undergo io_disarm_next() */
1073 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1075 static bool io_disarm_next(struct io_kiocb
*req
);
1076 static void io_uring_del_tctx_node(unsigned long index
);
1077 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1078 struct task_struct
*task
,
1080 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
);
1082 static bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1083 long res
, unsigned int cflags
);
1084 static void io_put_req(struct io_kiocb
*req
);
1085 static void io_put_req_deferred(struct io_kiocb
*req
);
1086 static void io_dismantle_req(struct io_kiocb
*req
);
1087 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1088 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
1089 struct io_uring_rsrc_update2
*up
,
1091 static void io_clean_op(struct io_kiocb
*req
);
1092 static struct file
*io_file_get(struct io_ring_ctx
*ctx
,
1093 struct io_kiocb
*req
, int fd
, bool fixed
);
1094 static void __io_queue_sqe(struct io_kiocb
*req
);
1095 static void io_rsrc_put_work(struct work_struct
*work
);
1097 static void io_req_task_queue(struct io_kiocb
*req
);
1098 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
);
1099 static int io_req_prep_async(struct io_kiocb
*req
);
1101 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
1102 unsigned int issue_flags
, u32 slot_index
);
1103 static int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
);
1105 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
);
1107 static struct kmem_cache
*req_cachep
;
1109 static const struct file_operations io_uring_fops
;
1111 struct sock
*io_uring_get_socket(struct file
*file
)
1113 #if defined(CONFIG_UNIX)
1114 if (file
->f_op
== &io_uring_fops
) {
1115 struct io_ring_ctx
*ctx
= file
->private_data
;
1117 return ctx
->ring_sock
->sk
;
1122 EXPORT_SYMBOL(io_uring_get_socket
);
1124 static inline void io_tw_lock(struct io_ring_ctx
*ctx
, bool *locked
)
1127 mutex_lock(&ctx
->uring_lock
);
1132 #define io_for_each_link(pos, head) \
1133 for (pos = (head); pos; pos = pos->link)
1136 * Shamelessly stolen from the mm implementation of page reference checking,
1137 * see commit f958d7b528b1 for details.
1139 #define req_ref_zero_or_close_to_overflow(req) \
1140 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1142 static inline bool req_ref_inc_not_zero(struct io_kiocb
*req
)
1144 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1145 return atomic_inc_not_zero(&req
->refs
);
1148 static inline bool req_ref_put_and_test(struct io_kiocb
*req
)
1150 if (likely(!(req
->flags
& REQ_F_REFCOUNT
)))
1153 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1154 return atomic_dec_and_test(&req
->refs
);
1157 static inline void req_ref_put(struct io_kiocb
*req
)
1159 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1160 WARN_ON_ONCE(req_ref_put_and_test(req
));
1163 static inline void req_ref_get(struct io_kiocb
*req
)
1165 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1166 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1167 atomic_inc(&req
->refs
);
1170 static inline void __io_req_set_refcount(struct io_kiocb
*req
, int nr
)
1172 if (!(req
->flags
& REQ_F_REFCOUNT
)) {
1173 req
->flags
|= REQ_F_REFCOUNT
;
1174 atomic_set(&req
->refs
, nr
);
1178 static inline void io_req_set_refcount(struct io_kiocb
*req
)
1180 __io_req_set_refcount(req
, 1);
1183 static inline void io_req_set_rsrc_node(struct io_kiocb
*req
)
1185 struct io_ring_ctx
*ctx
= req
->ctx
;
1187 if (!req
->fixed_rsrc_refs
) {
1188 req
->fixed_rsrc_refs
= &ctx
->rsrc_node
->refs
;
1189 percpu_ref_get(req
->fixed_rsrc_refs
);
1193 static void io_refs_resurrect(struct percpu_ref
*ref
, struct completion
*compl)
1195 bool got
= percpu_ref_tryget(ref
);
1197 /* already at zero, wait for ->release() */
1199 wait_for_completion(compl);
1200 percpu_ref_resurrect(ref
);
1202 percpu_ref_put(ref
);
1205 static bool io_match_task(struct io_kiocb
*head
, struct task_struct
*task
,
1207 __must_hold(&req
->ctx
->timeout_lock
)
1209 struct io_kiocb
*req
;
1211 if (task
&& head
->task
!= task
)
1216 io_for_each_link(req
, head
) {
1217 if (req
->flags
& REQ_F_INFLIGHT
)
1223 static bool io_match_linked(struct io_kiocb
*head
)
1225 struct io_kiocb
*req
;
1227 io_for_each_link(req
, head
) {
1228 if (req
->flags
& REQ_F_INFLIGHT
)
1235 * As io_match_task() but protected against racing with linked timeouts.
1236 * User must not hold timeout_lock.
1238 static bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
1243 if (task
&& head
->task
!= task
)
1248 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
1249 struct io_ring_ctx
*ctx
= head
->ctx
;
1251 /* protect against races with linked timeouts */
1252 spin_lock_irq(&ctx
->timeout_lock
);
1253 matched
= io_match_linked(head
);
1254 spin_unlock_irq(&ctx
->timeout_lock
);
1256 matched
= io_match_linked(head
);
1261 static inline void req_set_fail(struct io_kiocb
*req
)
1263 req
->flags
|= REQ_F_FAIL
;
1266 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
1272 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1274 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1276 complete(&ctx
->ref_comp
);
1279 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1281 return !req
->timeout
.off
;
1284 static void io_fallback_req_func(struct work_struct
*work
)
1286 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
1287 fallback_work
.work
);
1288 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
1289 struct io_kiocb
*req
, *tmp
;
1290 bool locked
= false;
1292 percpu_ref_get(&ctx
->refs
);
1293 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.fallback_node
)
1294 req
->io_task_work
.func(req
, &locked
);
1297 if (ctx
->submit_state
.compl_nr
)
1298 io_submit_flush_completions(ctx
);
1299 mutex_unlock(&ctx
->uring_lock
);
1301 percpu_ref_put(&ctx
->refs
);
1305 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1307 struct io_ring_ctx
*ctx
;
1310 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1315 * Use 5 bits less than the max cq entries, that should give us around
1316 * 32 entries per hash list if totally full and uniformly spread.
1318 hash_bits
= ilog2(p
->cq_entries
);
1322 ctx
->cancel_hash_bits
= hash_bits
;
1323 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1325 if (!ctx
->cancel_hash
)
1327 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1329 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
1330 if (!ctx
->dummy_ubuf
)
1332 /* set invalid range, so io_import_fixed() fails meeting it */
1333 ctx
->dummy_ubuf
->ubuf
= -1UL;
1335 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1336 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1339 ctx
->flags
= p
->flags
;
1340 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1341 INIT_LIST_HEAD(&ctx
->sqd_list
);
1342 init_waitqueue_head(&ctx
->poll_wait
);
1343 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1344 init_completion(&ctx
->ref_comp
);
1345 xa_init_flags(&ctx
->io_buffers
, XA_FLAGS_ALLOC1
);
1346 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1347 mutex_init(&ctx
->uring_lock
);
1348 init_waitqueue_head(&ctx
->cq_wait
);
1349 spin_lock_init(&ctx
->completion_lock
);
1350 spin_lock_init(&ctx
->timeout_lock
);
1351 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1352 INIT_LIST_HEAD(&ctx
->defer_list
);
1353 INIT_LIST_HEAD(&ctx
->timeout_list
);
1354 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
1355 spin_lock_init(&ctx
->rsrc_ref_lock
);
1356 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1357 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1358 init_llist_head(&ctx
->rsrc_put_llist
);
1359 INIT_LIST_HEAD(&ctx
->tctx_list
);
1360 INIT_LIST_HEAD(&ctx
->submit_state
.free_list
);
1361 INIT_LIST_HEAD(&ctx
->locked_free_list
);
1362 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
1365 kfree(ctx
->dummy_ubuf
);
1366 kfree(ctx
->cancel_hash
);
1371 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
1373 struct io_rings
*r
= ctx
->rings
;
1375 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
1379 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1381 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1382 struct io_ring_ctx
*ctx
= req
->ctx
;
1384 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
1390 #define FFS_ASYNC_READ 0x1UL
1391 #define FFS_ASYNC_WRITE 0x2UL
1393 #define FFS_ISREG 0x4UL
1395 #define FFS_ISREG 0x0UL
1397 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1399 static inline bool io_req_ffs_set(struct io_kiocb
*req
)
1401 return IS_ENABLED(CONFIG_64BIT
) && (req
->flags
& REQ_F_FIXED_FILE
);
1404 static void io_req_track_inflight(struct io_kiocb
*req
)
1406 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1407 req
->flags
|= REQ_F_INFLIGHT
;
1408 atomic_inc(¤t
->io_uring
->inflight_tracked
);
1412 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
1414 if (WARN_ON_ONCE(!req
->link
))
1417 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
1418 req
->flags
|= REQ_F_LINK_TIMEOUT
;
1420 /* linked timeouts should have two refs once prep'ed */
1421 io_req_set_refcount(req
);
1422 __io_req_set_refcount(req
->link
, 2);
1426 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
1428 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
1430 return __io_prep_linked_timeout(req
);
1433 static void io_prep_async_work(struct io_kiocb
*req
)
1435 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1436 struct io_ring_ctx
*ctx
= req
->ctx
;
1438 if (!(req
->flags
& REQ_F_CREDS
)) {
1439 req
->flags
|= REQ_F_CREDS
;
1440 req
->creds
= get_current_cred();
1443 req
->work
.list
.next
= NULL
;
1444 req
->work
.flags
= 0;
1445 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1446 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1448 if (req
->flags
& REQ_F_ISREG
) {
1449 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1450 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1451 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
1452 if (def
->unbound_nonreg_file
)
1453 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1457 static void io_prep_async_link(struct io_kiocb
*req
)
1459 struct io_kiocb
*cur
;
1461 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
1462 struct io_ring_ctx
*ctx
= req
->ctx
;
1464 spin_lock_irq(&ctx
->timeout_lock
);
1465 io_for_each_link(cur
, req
)
1466 io_prep_async_work(cur
);
1467 spin_unlock_irq(&ctx
->timeout_lock
);
1469 io_for_each_link(cur
, req
)
1470 io_prep_async_work(cur
);
1474 static void io_queue_async_work(struct io_kiocb
*req
, bool *locked
)
1476 struct io_ring_ctx
*ctx
= req
->ctx
;
1477 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1478 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1480 /* must not take the lock, NULL it as a precaution */
1484 BUG_ON(!tctx
->io_wq
);
1486 /* init ->work of the whole link before punting */
1487 io_prep_async_link(req
);
1490 * Not expected to happen, but if we do have a bug where this _can_
1491 * happen, catch it here and ensure the request is marked as
1492 * canceled. That will make io-wq go through the usual work cancel
1493 * procedure rather than attempt to run this request (or create a new
1496 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
1497 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1499 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1500 &req
->work
, req
->flags
);
1501 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
1503 io_queue_linked_timeout(link
);
1506 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1507 __must_hold(&req
->ctx
->completion_lock
)
1508 __must_hold(&req
->ctx
->timeout_lock
)
1510 struct io_timeout_data
*io
= req
->async_data
;
1512 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1515 atomic_set(&req
->ctx
->cq_timeouts
,
1516 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1517 list_del_init(&req
->timeout
.list
);
1518 io_cqring_fill_event(req
->ctx
, req
->user_data
, status
, 0);
1519 io_put_req_deferred(req
);
1523 static void io_queue_deferred(struct io_ring_ctx
*ctx
)
1525 while (!list_empty(&ctx
->defer_list
)) {
1526 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1527 struct io_defer_entry
, list
);
1529 if (req_need_defer(de
->req
, de
->seq
))
1531 list_del_init(&de
->list
);
1532 io_req_task_queue(de
->req
);
1537 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1538 __must_hold(&ctx
->completion_lock
)
1540 u32 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1541 struct io_kiocb
*req
, *tmp
;
1543 spin_lock_irq(&ctx
->timeout_lock
);
1544 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1545 u32 events_needed
, events_got
;
1547 if (io_is_timeout_noseq(req
))
1551 * Since seq can easily wrap around over time, subtract
1552 * the last seq at which timeouts were flushed before comparing.
1553 * Assuming not more than 2^31-1 events have happened since,
1554 * these subtractions won't have wrapped, so we can check if
1555 * target is in [last_seq, current_seq] by comparing the two.
1557 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1558 events_got
= seq
- ctx
->cq_last_tm_flush
;
1559 if (events_got
< events_needed
)
1562 io_kill_timeout(req
, 0);
1564 ctx
->cq_last_tm_flush
= seq
;
1565 spin_unlock_irq(&ctx
->timeout_lock
);
1568 static void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
1570 if (ctx
->off_timeout_used
)
1571 io_flush_timeouts(ctx
);
1572 if (ctx
->drain_active
)
1573 io_queue_deferred(ctx
);
1576 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
1578 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
))
1579 __io_commit_cqring_flush(ctx
);
1580 /* order cqe stores with ring update */
1581 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1584 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1586 struct io_rings
*r
= ctx
->rings
;
1588 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
1591 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1593 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1596 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
1598 struct io_rings
*rings
= ctx
->rings
;
1599 unsigned tail
, mask
= ctx
->cq_entries
- 1;
1602 * writes to the cq entry need to come after reading head; the
1603 * control dependency is enough as we're using WRITE_ONCE to
1606 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
1609 tail
= ctx
->cached_cq_tail
++;
1610 return &rings
->cqes
[tail
& mask
];
1613 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1615 if (likely(!ctx
->cq_ev_fd
))
1617 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1619 return !ctx
->eventfd_async
|| io_wq_current_is_worker();
1623 * This should only get called when at least one event has been posted.
1624 * Some applications rely on the eventfd notification count only changing
1625 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1626 * 1:1 relationship between how many times this function is called (and
1627 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1629 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1632 * wake_up_all() may seem excessive, but io_wake_function() and
1633 * io_should_wake() handle the termination of the loop and only
1634 * wake as many waiters as we need to.
1636 if (wq_has_sleeper(&ctx
->cq_wait
))
1637 wake_up_all(&ctx
->cq_wait
);
1638 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1639 wake_up(&ctx
->sq_data
->wait
);
1640 if (io_should_trigger_evfd(ctx
))
1641 eventfd_signal(ctx
->cq_ev_fd
, 1);
1642 if (waitqueue_active(&ctx
->poll_wait
))
1643 wake_up_interruptible(&ctx
->poll_wait
);
1646 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1648 /* see waitqueue_active() comment */
1651 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1652 if (waitqueue_active(&ctx
->cq_wait
))
1653 wake_up_all(&ctx
->cq_wait
);
1655 if (io_should_trigger_evfd(ctx
))
1656 eventfd_signal(ctx
->cq_ev_fd
, 1);
1657 if (waitqueue_active(&ctx
->poll_wait
))
1658 wake_up_interruptible(&ctx
->poll_wait
);
1661 /* Returns true if there are no backlogged entries after the flush */
1662 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1664 bool all_flushed
, posted
;
1666 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
1670 spin_lock(&ctx
->completion_lock
);
1671 while (!list_empty(&ctx
->cq_overflow_list
)) {
1672 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
1673 struct io_overflow_cqe
*ocqe
;
1677 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
1678 struct io_overflow_cqe
, list
);
1680 memcpy(cqe
, &ocqe
->cqe
, sizeof(*cqe
));
1682 io_account_cq_overflow(ctx
);
1685 list_del(&ocqe
->list
);
1689 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1691 clear_bit(0, &ctx
->check_cq_overflow
);
1692 WRITE_ONCE(ctx
->rings
->sq_flags
,
1693 ctx
->rings
->sq_flags
& ~IORING_SQ_CQ_OVERFLOW
);
1697 io_commit_cqring(ctx
);
1698 spin_unlock(&ctx
->completion_lock
);
1700 io_cqring_ev_posted(ctx
);
1704 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
1708 if (test_bit(0, &ctx
->check_cq_overflow
)) {
1709 /* iopoll syncs against uring_lock, not completion_lock */
1710 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1711 mutex_lock(&ctx
->uring_lock
);
1712 ret
= __io_cqring_overflow_flush(ctx
, false);
1713 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1714 mutex_unlock(&ctx
->uring_lock
);
1720 /* must to be called somewhat shortly after putting a request */
1721 static inline void io_put_task(struct task_struct
*task
, int nr
)
1723 struct io_uring_task
*tctx
= task
->io_uring
;
1725 if (likely(task
== current
)) {
1726 tctx
->cached_refs
+= nr
;
1728 percpu_counter_sub(&tctx
->inflight
, nr
);
1729 if (unlikely(atomic_read(&tctx
->in_idle
)))
1730 wake_up(&tctx
->wait
);
1731 put_task_struct_many(task
, nr
);
1735 static void io_task_refs_refill(struct io_uring_task
*tctx
)
1737 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
1739 percpu_counter_add(&tctx
->inflight
, refill
);
1740 refcount_add(refill
, ¤t
->usage
);
1741 tctx
->cached_refs
+= refill
;
1744 static inline void io_get_task_refs(int nr
)
1746 struct io_uring_task
*tctx
= current
->io_uring
;
1748 tctx
->cached_refs
-= nr
;
1749 if (unlikely(tctx
->cached_refs
< 0))
1750 io_task_refs_refill(tctx
);
1753 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
1755 struct io_uring_task
*tctx
= task
->io_uring
;
1756 unsigned int refs
= tctx
->cached_refs
;
1759 tctx
->cached_refs
= 0;
1760 percpu_counter_sub(&tctx
->inflight
, refs
);
1761 put_task_struct_many(task
, refs
);
1765 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
1766 long res
, unsigned int cflags
)
1768 struct io_overflow_cqe
*ocqe
;
1770 ocqe
= kmalloc(sizeof(*ocqe
), GFP_ATOMIC
| __GFP_ACCOUNT
);
1773 * If we're in ring overflow flush mode, or in task cancel mode,
1774 * or cannot allocate an overflow entry, then we need to drop it
1777 io_account_cq_overflow(ctx
);
1780 if (list_empty(&ctx
->cq_overflow_list
)) {
1781 set_bit(0, &ctx
->check_cq_overflow
);
1782 WRITE_ONCE(ctx
->rings
->sq_flags
,
1783 ctx
->rings
->sq_flags
| IORING_SQ_CQ_OVERFLOW
);
1786 ocqe
->cqe
.user_data
= user_data
;
1787 ocqe
->cqe
.res
= res
;
1788 ocqe
->cqe
.flags
= cflags
;
1789 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
1793 static inline bool __io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1794 long res
, unsigned int cflags
)
1796 struct io_uring_cqe
*cqe
;
1798 trace_io_uring_complete(ctx
, user_data
, res
, cflags
);
1801 * If we can't get a cq entry, userspace overflowed the
1802 * submission (by quite a lot). Increment the overflow count in
1805 cqe
= io_get_cqe(ctx
);
1807 WRITE_ONCE(cqe
->user_data
, user_data
);
1808 WRITE_ONCE(cqe
->res
, res
);
1809 WRITE_ONCE(cqe
->flags
, cflags
);
1812 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
);
1815 /* not as hot to bloat with inlining */
1816 static noinline
bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1817 long res
, unsigned int cflags
)
1819 return __io_cqring_fill_event(ctx
, user_data
, res
, cflags
);
1822 static void io_req_complete_post(struct io_kiocb
*req
, long res
,
1823 unsigned int cflags
)
1825 struct io_ring_ctx
*ctx
= req
->ctx
;
1827 spin_lock(&ctx
->completion_lock
);
1828 __io_cqring_fill_event(ctx
, req
->user_data
, res
, cflags
);
1830 * If we're the last reference to this request, add to our locked
1833 if (req_ref_put_and_test(req
)) {
1834 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1835 if (req
->flags
& IO_DISARM_MASK
)
1836 io_disarm_next(req
);
1838 io_req_task_queue(req
->link
);
1842 io_dismantle_req(req
);
1843 io_put_task(req
->task
, 1);
1844 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
1845 ctx
->locked_free_nr
++;
1847 if (!percpu_ref_tryget(&ctx
->refs
))
1850 io_commit_cqring(ctx
);
1851 spin_unlock(&ctx
->completion_lock
);
1854 io_cqring_ev_posted(ctx
);
1855 percpu_ref_put(&ctx
->refs
);
1859 static inline bool io_req_needs_clean(struct io_kiocb
*req
)
1861 return req
->flags
& IO_REQ_CLEAN_FLAGS
;
1864 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1865 unsigned int cflags
)
1867 if (io_req_needs_clean(req
))
1870 req
->compl.cflags
= cflags
;
1871 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1874 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1875 long res
, unsigned cflags
)
1877 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1878 io_req_complete_state(req
, res
, cflags
);
1880 io_req_complete_post(req
, res
, cflags
);
1883 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1885 __io_req_complete(req
, 0, res
, 0);
1888 static void io_req_complete_failed(struct io_kiocb
*req
, long res
)
1891 io_req_complete_post(req
, res
, 0);
1894 static void io_req_complete_fail_submit(struct io_kiocb
*req
)
1897 * We don't submit, fail them all, for that replace hardlinks with
1898 * normal links. Extra REQ_F_LINK is tolerated.
1900 req
->flags
&= ~REQ_F_HARDLINK
;
1901 req
->flags
|= REQ_F_LINK
;
1902 io_req_complete_failed(req
, req
->result
);
1906 * Don't initialise the fields below on every allocation, but do that in
1907 * advance and keep them valid across allocations.
1909 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1913 req
->async_data
= NULL
;
1914 /* not necessary, but safer to zero */
1918 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1919 struct io_submit_state
*state
)
1921 spin_lock(&ctx
->completion_lock
);
1922 list_splice_init(&ctx
->locked_free_list
, &state
->free_list
);
1923 ctx
->locked_free_nr
= 0;
1924 spin_unlock(&ctx
->completion_lock
);
1927 /* Returns true IFF there are requests in the cache */
1928 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1930 struct io_submit_state
*state
= &ctx
->submit_state
;
1934 * If we have more than a batch's worth of requests in our IRQ side
1935 * locked cache, grab the lock and move them over to our submission
1938 if (READ_ONCE(ctx
->locked_free_nr
) > IO_COMPL_BATCH
)
1939 io_flush_cached_locked_reqs(ctx
, state
);
1941 nr
= state
->free_reqs
;
1942 while (!list_empty(&state
->free_list
)) {
1943 struct io_kiocb
*req
= list_first_entry(&state
->free_list
,
1944 struct io_kiocb
, inflight_entry
);
1946 list_del(&req
->inflight_entry
);
1947 state
->reqs
[nr
++] = req
;
1948 if (nr
== ARRAY_SIZE(state
->reqs
))
1952 state
->free_reqs
= nr
;
1957 * A request might get retired back into the request caches even before opcode
1958 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1959 * Because of that, io_alloc_req() should be called only under ->uring_lock
1960 * and with extra caution to not get a request that is still worked on.
1962 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1963 __must_hold(&ctx
->uring_lock
)
1965 struct io_submit_state
*state
= &ctx
->submit_state
;
1966 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1969 BUILD_BUG_ON(ARRAY_SIZE(state
->reqs
) < IO_REQ_ALLOC_BATCH
);
1971 if (likely(state
->free_reqs
|| io_flush_cached_reqs(ctx
)))
1974 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1978 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1979 * retry single alloc to be on the safe side.
1981 if (unlikely(ret
<= 0)) {
1982 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1983 if (!state
->reqs
[0])
1988 for (i
= 0; i
< ret
; i
++)
1989 io_preinit_req(state
->reqs
[i
], ctx
);
1990 state
->free_reqs
= ret
;
1993 return state
->reqs
[state
->free_reqs
];
1996 static inline void io_put_file(struct file
*file
)
2002 static void io_dismantle_req(struct io_kiocb
*req
)
2004 unsigned int flags
= req
->flags
;
2006 if (io_req_needs_clean(req
))
2008 if (!(flags
& REQ_F_FIXED_FILE
))
2009 io_put_file(req
->file
);
2010 if (req
->fixed_rsrc_refs
)
2011 percpu_ref_put(req
->fixed_rsrc_refs
);
2012 if (req
->async_data
) {
2013 kfree(req
->async_data
);
2014 req
->async_data
= NULL
;
2018 static void __io_free_req(struct io_kiocb
*req
)
2020 struct io_ring_ctx
*ctx
= req
->ctx
;
2022 io_dismantle_req(req
);
2023 io_put_task(req
->task
, 1);
2025 spin_lock(&ctx
->completion_lock
);
2026 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
2027 ctx
->locked_free_nr
++;
2028 spin_unlock(&ctx
->completion_lock
);
2030 percpu_ref_put(&ctx
->refs
);
2033 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2035 struct io_kiocb
*nxt
= req
->link
;
2037 req
->link
= nxt
->link
;
2041 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
2042 __must_hold(&req
->ctx
->completion_lock
)
2043 __must_hold(&req
->ctx
->timeout_lock
)
2045 struct io_kiocb
*link
= req
->link
;
2047 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2048 struct io_timeout_data
*io
= link
->async_data
;
2050 io_remove_next_linked(req
);
2051 link
->timeout
.head
= NULL
;
2052 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
2053 list_del(&link
->timeout
.list
);
2054 io_cqring_fill_event(link
->ctx
, link
->user_data
,
2056 io_put_req_deferred(link
);
2063 static void io_fail_links(struct io_kiocb
*req
)
2064 __must_hold(&req
->ctx
->completion_lock
)
2066 struct io_kiocb
*nxt
, *link
= req
->link
;
2070 long res
= -ECANCELED
;
2072 if (link
->flags
& REQ_F_FAIL
)
2078 trace_io_uring_fail_link(req
, link
);
2079 io_cqring_fill_event(link
->ctx
, link
->user_data
, res
, 0);
2080 io_put_req_deferred(link
);
2085 static bool io_disarm_next(struct io_kiocb
*req
)
2086 __must_hold(&req
->ctx
->completion_lock
)
2088 bool posted
= false;
2090 if (req
->flags
& REQ_F_ARM_LTIMEOUT
) {
2091 struct io_kiocb
*link
= req
->link
;
2093 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
2094 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2095 io_remove_next_linked(req
);
2096 io_cqring_fill_event(link
->ctx
, link
->user_data
,
2098 io_put_req_deferred(link
);
2101 } else if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
2102 struct io_ring_ctx
*ctx
= req
->ctx
;
2104 spin_lock_irq(&ctx
->timeout_lock
);
2105 posted
= io_kill_linked_timeout(req
);
2106 spin_unlock_irq(&ctx
->timeout_lock
);
2108 if (unlikely((req
->flags
& REQ_F_FAIL
) &&
2109 !(req
->flags
& REQ_F_HARDLINK
))) {
2110 posted
|= (req
->link
!= NULL
);
2116 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2118 struct io_kiocb
*nxt
;
2121 * If LINK is set, we have dependent requests in this chain. If we
2122 * didn't fail this request, queue the first one up, moving any other
2123 * dependencies to the next request. In case of failure, fail the rest
2126 if (req
->flags
& IO_DISARM_MASK
) {
2127 struct io_ring_ctx
*ctx
= req
->ctx
;
2130 spin_lock(&ctx
->completion_lock
);
2131 posted
= io_disarm_next(req
);
2133 io_commit_cqring(req
->ctx
);
2134 spin_unlock(&ctx
->completion_lock
);
2136 io_cqring_ev_posted(ctx
);
2143 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2145 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
2147 return __io_req_find_next(req
);
2150 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, bool *locked
)
2155 if (ctx
->submit_state
.compl_nr
)
2156 io_submit_flush_completions(ctx
);
2157 mutex_unlock(&ctx
->uring_lock
);
2160 percpu_ref_put(&ctx
->refs
);
2163 static void tctx_task_work(struct callback_head
*cb
)
2165 bool locked
= false;
2166 struct io_ring_ctx
*ctx
= NULL
;
2167 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
2171 struct io_wq_work_node
*node
;
2173 if (!tctx
->task_list
.first
&& locked
&& ctx
->submit_state
.compl_nr
)
2174 io_submit_flush_completions(ctx
);
2176 spin_lock_irq(&tctx
->task_lock
);
2177 node
= tctx
->task_list
.first
;
2178 INIT_WQ_LIST(&tctx
->task_list
);
2180 tctx
->task_running
= false;
2181 spin_unlock_irq(&tctx
->task_lock
);
2186 struct io_wq_work_node
*next
= node
->next
;
2187 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
2190 if (req
->ctx
!= ctx
) {
2191 ctx_flush_and_put(ctx
, &locked
);
2193 /* if not contended, grab and improve batching */
2194 locked
= mutex_trylock(&ctx
->uring_lock
);
2195 percpu_ref_get(&ctx
->refs
);
2197 req
->io_task_work
.func(req
, &locked
);
2204 ctx_flush_and_put(ctx
, &locked
);
2206 /* relaxed read is enough as only the task itself sets ->in_idle */
2207 if (unlikely(atomic_read(&tctx
->in_idle
)))
2208 io_uring_drop_tctx_refs(current
);
2211 static void io_req_task_work_add(struct io_kiocb
*req
)
2213 struct task_struct
*tsk
= req
->task
;
2214 struct io_uring_task
*tctx
= tsk
->io_uring
;
2215 enum task_work_notify_mode notify
;
2216 struct io_wq_work_node
*node
;
2217 unsigned long flags
;
2220 WARN_ON_ONCE(!tctx
);
2222 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2223 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
2224 running
= tctx
->task_running
;
2226 tctx
->task_running
= true;
2227 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2229 /* task_work already pending, we're done */
2234 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2235 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2236 * processing task_work. There's no reliable way to tell if TWA_RESUME
2239 notify
= (req
->ctx
->flags
& IORING_SETUP_SQPOLL
) ? TWA_NONE
: TWA_SIGNAL
;
2240 if (!task_work_add(tsk
, &tctx
->task_work
, notify
)) {
2241 wake_up_process(tsk
);
2245 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2246 tctx
->task_running
= false;
2247 node
= tctx
->task_list
.first
;
2248 INIT_WQ_LIST(&tctx
->task_list
);
2249 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2252 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
2254 if (llist_add(&req
->io_task_work
.fallback_node
,
2255 &req
->ctx
->fallback_llist
))
2256 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
2260 static void io_req_task_cancel(struct io_kiocb
*req
, bool *locked
)
2262 struct io_ring_ctx
*ctx
= req
->ctx
;
2264 /* not needed for normal modes, but SQPOLL depends on it */
2265 io_tw_lock(ctx
, locked
);
2266 io_req_complete_failed(req
, req
->result
);
2269 static void io_req_task_submit(struct io_kiocb
*req
, bool *locked
)
2271 struct io_ring_ctx
*ctx
= req
->ctx
;
2273 io_tw_lock(ctx
, locked
);
2274 /* req->task == current here, checking PF_EXITING is safe */
2275 if (likely(!(req
->task
->flags
& PF_EXITING
)))
2276 __io_queue_sqe(req
);
2278 io_req_complete_failed(req
, -EFAULT
);
2281 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2284 req
->io_task_work
.func
= io_req_task_cancel
;
2285 io_req_task_work_add(req
);
2288 static void io_req_task_queue(struct io_kiocb
*req
)
2290 req
->io_task_work
.func
= io_req_task_submit
;
2291 io_req_task_work_add(req
);
2294 static void io_req_task_queue_reissue(struct io_kiocb
*req
)
2296 req
->io_task_work
.func
= io_queue_async_work
;
2297 io_req_task_work_add(req
);
2300 static inline void io_queue_next(struct io_kiocb
*req
)
2302 struct io_kiocb
*nxt
= io_req_find_next(req
);
2305 io_req_task_queue(nxt
);
2308 static void io_free_req(struct io_kiocb
*req
)
2314 static void io_free_req_work(struct io_kiocb
*req
, bool *locked
)
2320 struct task_struct
*task
;
2325 static inline void io_init_req_batch(struct req_batch
*rb
)
2332 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2333 struct req_batch
*rb
)
2336 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2338 io_put_task(rb
->task
, rb
->task_refs
);
2341 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2342 struct io_submit_state
*state
)
2345 io_dismantle_req(req
);
2347 if (req
->task
!= rb
->task
) {
2349 io_put_task(rb
->task
, rb
->task_refs
);
2350 rb
->task
= req
->task
;
2356 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2357 state
->reqs
[state
->free_reqs
++] = req
;
2359 list_add(&req
->inflight_entry
, &state
->free_list
);
2362 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
2363 __must_hold(&ctx
->uring_lock
)
2365 struct io_submit_state
*state
= &ctx
->submit_state
;
2366 int i
, nr
= state
->compl_nr
;
2367 struct req_batch rb
;
2369 spin_lock(&ctx
->completion_lock
);
2370 for (i
= 0; i
< nr
; i
++) {
2371 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2373 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2376 io_commit_cqring(ctx
);
2377 spin_unlock(&ctx
->completion_lock
);
2378 io_cqring_ev_posted(ctx
);
2380 io_init_req_batch(&rb
);
2381 for (i
= 0; i
< nr
; i
++) {
2382 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2384 if (req_ref_put_and_test(req
))
2385 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2388 io_req_free_batch_finish(ctx
, &rb
);
2389 state
->compl_nr
= 0;
2393 * Drop reference to request, return next in chain (if there is one) if this
2394 * was the last reference to this request.
2396 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2398 struct io_kiocb
*nxt
= NULL
;
2400 if (req_ref_put_and_test(req
)) {
2401 nxt
= io_req_find_next(req
);
2407 static inline void io_put_req(struct io_kiocb
*req
)
2409 if (req_ref_put_and_test(req
))
2413 static inline void io_put_req_deferred(struct io_kiocb
*req
)
2415 if (req_ref_put_and_test(req
)) {
2416 req
->io_task_work
.func
= io_free_req_work
;
2417 io_req_task_work_add(req
);
2421 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2423 /* See comment at the top of this file */
2425 return __io_cqring_events(ctx
);
2428 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2430 struct io_rings
*rings
= ctx
->rings
;
2432 /* make sure SQ entry isn't read before tail */
2433 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2436 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2438 unsigned int cflags
;
2440 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2441 cflags
|= IORING_CQE_F_BUFFER
;
2442 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2447 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2449 struct io_buffer
*kbuf
;
2451 if (likely(!(req
->flags
& REQ_F_BUFFER_SELECTED
)))
2453 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2454 return io_put_kbuf(req
, kbuf
);
2457 static inline bool io_run_task_work(void)
2459 if (test_thread_flag(TIF_NOTIFY_SIGNAL
) || current
->task_works
) {
2460 __set_current_state(TASK_RUNNING
);
2461 tracehook_notify_signal();
2469 * Find and free completed poll iocbs
2471 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2472 struct list_head
*done
)
2474 struct req_batch rb
;
2475 struct io_kiocb
*req
;
2477 /* order with ->result store in io_complete_rw_iopoll() */
2480 io_init_req_batch(&rb
);
2481 while (!list_empty(done
)) {
2482 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2483 list_del(&req
->inflight_entry
);
2485 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2486 io_put_rw_kbuf(req
));
2489 if (req_ref_put_and_test(req
))
2490 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2493 io_commit_cqring(ctx
);
2494 io_cqring_ev_posted_iopoll(ctx
);
2495 io_req_free_batch_finish(ctx
, &rb
);
2498 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2501 struct io_kiocb
*req
, *tmp
;
2506 * Only spin for completions if we don't have multiple devices hanging
2507 * off our complete list, and we're under the requested amount.
2509 spin
= !ctx
->poll_multi_queue
&& *nr_events
< min
;
2511 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2512 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2516 * Move completed and retryable entries to our local lists.
2517 * If we find a request that requires polling, break out
2518 * and complete those lists first, if we have entries there.
2520 if (READ_ONCE(req
->iopoll_completed
)) {
2521 list_move_tail(&req
->inflight_entry
, &done
);
2524 if (!list_empty(&done
))
2527 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2528 if (unlikely(ret
< 0))
2533 /* iopoll may have completed current req */
2534 if (READ_ONCE(req
->iopoll_completed
))
2535 list_move_tail(&req
->inflight_entry
, &done
);
2538 if (!list_empty(&done
))
2539 io_iopoll_complete(ctx
, nr_events
, &done
);
2545 * We can't just wait for polled events to come to us, we have to actively
2546 * find and complete them.
2548 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2550 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2553 mutex_lock(&ctx
->uring_lock
);
2554 while (!list_empty(&ctx
->iopoll_list
)) {
2555 unsigned int nr_events
= 0;
2557 io_do_iopoll(ctx
, &nr_events
, 0);
2559 /* let it sleep and repeat later if can't complete a request */
2563 * Ensure we allow local-to-the-cpu processing to take place,
2564 * in this case we need to ensure that we reap all events.
2565 * Also let task_work, etc. to progress by releasing the mutex
2567 if (need_resched()) {
2568 mutex_unlock(&ctx
->uring_lock
);
2570 mutex_lock(&ctx
->uring_lock
);
2573 mutex_unlock(&ctx
->uring_lock
);
2576 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2578 unsigned int nr_events
= 0;
2582 * We disallow the app entering submit/complete with polling, but we
2583 * still need to lock the ring to prevent racing with polled issue
2584 * that got punted to a workqueue.
2586 mutex_lock(&ctx
->uring_lock
);
2588 * Don't enter poll loop if we already have events pending.
2589 * If we do, we can potentially be spinning for commands that
2590 * already triggered a CQE (eg in error).
2592 if (test_bit(0, &ctx
->check_cq_overflow
))
2593 __io_cqring_overflow_flush(ctx
, false);
2594 if (io_cqring_events(ctx
))
2598 * If a submit got punted to a workqueue, we can have the
2599 * application entering polling for a command before it gets
2600 * issued. That app will hold the uring_lock for the duration
2601 * of the poll right here, so we need to take a breather every
2602 * now and then to ensure that the issue has a chance to add
2603 * the poll to the issued list. Otherwise we can spin here
2604 * forever, while the workqueue is stuck trying to acquire the
2607 if (list_empty(&ctx
->iopoll_list
)) {
2608 u32 tail
= ctx
->cached_cq_tail
;
2610 mutex_unlock(&ctx
->uring_lock
);
2612 mutex_lock(&ctx
->uring_lock
);
2614 /* some requests don't go through iopoll_list */
2615 if (tail
!= ctx
->cached_cq_tail
||
2616 list_empty(&ctx
->iopoll_list
))
2619 ret
= io_do_iopoll(ctx
, &nr_events
, min
);
2620 } while (!ret
&& nr_events
< min
&& !need_resched());
2622 mutex_unlock(&ctx
->uring_lock
);
2626 static void kiocb_end_write(struct io_kiocb
*req
)
2629 * Tell lockdep we inherited freeze protection from submission
2632 if (req
->flags
& REQ_F_ISREG
) {
2633 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
2635 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
2641 static bool io_resubmit_prep(struct io_kiocb
*req
)
2643 struct io_async_rw
*rw
= req
->async_data
;
2646 return !io_req_prep_async(req
);
2647 iov_iter_restore(&rw
->iter
, &rw
->iter_state
);
2651 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2653 umode_t mode
= file_inode(req
->file
)->i_mode
;
2654 struct io_ring_ctx
*ctx
= req
->ctx
;
2656 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2658 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2659 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2662 * If ref is dying, we might be running poll reap from the exit work.
2663 * Don't attempt to reissue from that path, just let it fail with
2666 if (percpu_ref_is_dying(&ctx
->refs
))
2669 * Play it safe and assume not safe to re-import and reissue if we're
2670 * not in the original thread group (or in task context).
2672 if (!same_thread_group(req
->task
, current
) || !in_task())
2677 static bool io_resubmit_prep(struct io_kiocb
*req
)
2681 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2687 static bool __io_complete_rw_common(struct io_kiocb
*req
, long res
)
2689 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
)
2690 kiocb_end_write(req
);
2691 if (res
!= req
->result
) {
2692 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
2693 io_rw_should_reissue(req
)) {
2694 req
->flags
|= REQ_F_REISSUE
;
2703 static void io_req_task_complete(struct io_kiocb
*req
, bool *locked
)
2705 unsigned int cflags
= io_put_rw_kbuf(req
);
2706 long res
= req
->result
;
2709 struct io_ring_ctx
*ctx
= req
->ctx
;
2710 struct io_submit_state
*state
= &ctx
->submit_state
;
2712 io_req_complete_state(req
, res
, cflags
);
2713 state
->compl_reqs
[state
->compl_nr
++] = req
;
2714 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
2715 io_submit_flush_completions(ctx
);
2717 io_req_complete_post(req
, res
, cflags
);
2721 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2722 unsigned int issue_flags
)
2724 if (__io_complete_rw_common(req
, res
))
2726 __io_req_complete(req
, issue_flags
, req
->result
, io_put_rw_kbuf(req
));
2729 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2731 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2733 if (__io_complete_rw_common(req
, res
))
2736 req
->io_task_work
.func
= io_req_task_complete
;
2737 io_req_task_work_add(req
);
2740 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2742 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2744 if (kiocb
->ki_flags
& IOCB_WRITE
)
2745 kiocb_end_write(req
);
2746 if (unlikely(res
!= req
->result
)) {
2747 if (res
== -EAGAIN
&& io_rw_should_reissue(req
)) {
2748 req
->flags
|= REQ_F_REISSUE
;
2753 WRITE_ONCE(req
->result
, res
);
2754 /* order with io_iopoll_complete() checking ->result */
2756 WRITE_ONCE(req
->iopoll_completed
, 1);
2760 * After the iocb has been issued, it's safe to be found on the poll list.
2761 * Adding the kiocb to the list AFTER submission ensures that we don't
2762 * find it from a io_do_iopoll() thread before the issuer is done
2763 * accessing the kiocb cookie.
2765 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2767 struct io_ring_ctx
*ctx
= req
->ctx
;
2768 const bool in_async
= io_wq_current_is_worker();
2770 /* workqueue context doesn't hold uring_lock, grab it now */
2771 if (unlikely(in_async
))
2772 mutex_lock(&ctx
->uring_lock
);
2775 * Track whether we have multiple files in our lists. This will impact
2776 * how we do polling eventually, not spinning if we're on potentially
2777 * different devices.
2779 if (list_empty(&ctx
->iopoll_list
)) {
2780 ctx
->poll_multi_queue
= false;
2781 } else if (!ctx
->poll_multi_queue
) {
2782 struct io_kiocb
*list_req
;
2783 unsigned int queue_num0
, queue_num1
;
2785 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2788 if (list_req
->file
!= req
->file
) {
2789 ctx
->poll_multi_queue
= true;
2791 queue_num0
= blk_qc_t_to_queue_num(list_req
->rw
.kiocb
.ki_cookie
);
2792 queue_num1
= blk_qc_t_to_queue_num(req
->rw
.kiocb
.ki_cookie
);
2793 if (queue_num0
!= queue_num1
)
2794 ctx
->poll_multi_queue
= true;
2799 * For fast devices, IO may have already completed. If it has, add
2800 * it to the front so we find it first.
2802 if (READ_ONCE(req
->iopoll_completed
))
2803 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2805 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2807 if (unlikely(in_async
)) {
2809 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2810 * in sq thread task context or in io worker task context. If
2811 * current task context is sq thread, we don't need to check
2812 * whether should wake up sq thread.
2814 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2815 wq_has_sleeper(&ctx
->sq_data
->wait
))
2816 wake_up(&ctx
->sq_data
->wait
);
2818 mutex_unlock(&ctx
->uring_lock
);
2822 static bool io_bdev_nowait(struct block_device
*bdev
)
2824 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2828 * If we tracked the file through the SCM inflight mechanism, we could support
2829 * any file. For now, just ensure that anything potentially problematic is done
2832 static bool __io_file_supports_nowait(struct file
*file
, int rw
)
2834 umode_t mode
= file_inode(file
)->i_mode
;
2836 if (S_ISBLK(mode
)) {
2837 if (IS_ENABLED(CONFIG_BLOCK
) &&
2838 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2844 if (S_ISREG(mode
)) {
2845 if (IS_ENABLED(CONFIG_BLOCK
) &&
2846 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2847 file
->f_op
!= &io_uring_fops
)
2852 /* any ->read/write should understand O_NONBLOCK */
2853 if (file
->f_flags
& O_NONBLOCK
)
2856 if (!(file
->f_mode
& FMODE_NOWAIT
))
2860 return file
->f_op
->read_iter
!= NULL
;
2862 return file
->f_op
->write_iter
!= NULL
;
2865 static bool io_file_supports_nowait(struct io_kiocb
*req
, int rw
)
2867 if (rw
== READ
&& (req
->flags
& REQ_F_NOWAIT_READ
))
2869 else if (rw
== WRITE
&& (req
->flags
& REQ_F_NOWAIT_WRITE
))
2872 return __io_file_supports_nowait(req
->file
, rw
);
2875 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2878 struct io_ring_ctx
*ctx
= req
->ctx
;
2879 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2880 struct file
*file
= req
->file
;
2884 if (!io_req_ffs_set(req
) && S_ISREG(file_inode(file
)->i_mode
))
2885 req
->flags
|= REQ_F_ISREG
;
2887 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2888 if (kiocb
->ki_pos
== -1) {
2889 if (!(file
->f_mode
& FMODE_STREAM
)) {
2890 req
->flags
|= REQ_F_CUR_POS
;
2891 kiocb
->ki_pos
= file
->f_pos
;
2896 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2897 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2898 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2903 * If the file is marked O_NONBLOCK, still allow retry for it if it
2904 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2905 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2907 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) ||
2908 ((file
->f_flags
& O_NONBLOCK
) && !io_file_supports_nowait(req
, rw
)))
2909 req
->flags
|= REQ_F_NOWAIT
;
2911 ioprio
= READ_ONCE(sqe
->ioprio
);
2913 ret
= ioprio_check_cap(ioprio
);
2917 kiocb
->ki_ioprio
= ioprio
;
2919 kiocb
->ki_ioprio
= get_current_ioprio();
2921 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2922 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2923 !kiocb
->ki_filp
->f_op
->iopoll
)
2926 kiocb
->ki_flags
|= IOCB_HIPRI
| IOCB_ALLOC_CACHE
;
2927 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2928 req
->iopoll_completed
= 0;
2930 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2932 kiocb
->ki_complete
= io_complete_rw
;
2935 if (req
->opcode
== IORING_OP_READ_FIXED
||
2936 req
->opcode
== IORING_OP_WRITE_FIXED
) {
2938 io_req_set_rsrc_node(req
);
2941 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2942 req
->rw
.len
= READ_ONCE(sqe
->len
);
2943 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2947 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2953 case -ERESTARTNOINTR
:
2954 case -ERESTARTNOHAND
:
2955 case -ERESTART_RESTARTBLOCK
:
2957 * We can't just restart the syscall, since previously
2958 * submitted sqes may already be in progress. Just fail this
2964 kiocb
->ki_complete(kiocb
, ret
, 0);
2968 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2969 unsigned int issue_flags
)
2971 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2972 struct io_async_rw
*io
= req
->async_data
;
2974 /* add previously done IO, if any */
2975 if (io
&& io
->bytes_done
> 0) {
2977 ret
= io
->bytes_done
;
2979 ret
+= io
->bytes_done
;
2982 if (req
->flags
& REQ_F_CUR_POS
)
2983 req
->file
->f_pos
= kiocb
->ki_pos
;
2984 if (ret
>= 0 && (kiocb
->ki_complete
== io_complete_rw
))
2985 __io_complete_rw(req
, ret
, 0, issue_flags
);
2987 io_rw_done(kiocb
, ret
);
2989 if (req
->flags
& REQ_F_REISSUE
) {
2990 req
->flags
&= ~REQ_F_REISSUE
;
2991 if (io_resubmit_prep(req
)) {
2992 io_req_task_queue_reissue(req
);
2994 unsigned int cflags
= io_put_rw_kbuf(req
);
2995 struct io_ring_ctx
*ctx
= req
->ctx
;
2998 if (!(issue_flags
& IO_URING_F_NONBLOCK
)) {
2999 mutex_lock(&ctx
->uring_lock
);
3000 __io_req_complete(req
, issue_flags
, ret
, cflags
);
3001 mutex_unlock(&ctx
->uring_lock
);
3003 __io_req_complete(req
, issue_flags
, ret
, cflags
);
3009 static int __io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
3010 struct io_mapped_ubuf
*imu
)
3012 size_t len
= req
->rw
.len
;
3013 u64 buf_end
, buf_addr
= req
->rw
.addr
;
3016 if (unlikely(check_add_overflow(buf_addr
, (u64
)len
, &buf_end
)))
3018 /* not inside the mapped region */
3019 if (unlikely(buf_addr
< imu
->ubuf
|| buf_end
> imu
->ubuf_end
))
3023 * May not be a start of buffer, set size appropriately
3024 * and advance us to the beginning.
3026 offset
= buf_addr
- imu
->ubuf
;
3027 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3031 * Don't use iov_iter_advance() here, as it's really slow for
3032 * using the latter parts of a big fixed buffer - it iterates
3033 * over each segment manually. We can cheat a bit here, because
3036 * 1) it's a BVEC iter, we set it up
3037 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3038 * first and last bvec
3040 * So just find our index, and adjust the iterator afterwards.
3041 * If the offset is within the first bvec (or the whole first
3042 * bvec, just use iov_iter_advance(). This makes it easier
3043 * since we can just skip the first segment, which may not
3044 * be PAGE_SIZE aligned.
3046 const struct bio_vec
*bvec
= imu
->bvec
;
3048 if (offset
<= bvec
->bv_len
) {
3049 iov_iter_advance(iter
, offset
);
3051 unsigned long seg_skip
;
3053 /* skip first vec */
3054 offset
-= bvec
->bv_len
;
3055 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3057 iter
->bvec
= bvec
+ seg_skip
;
3058 iter
->nr_segs
-= seg_skip
;
3059 iter
->count
-= bvec
->bv_len
+ offset
;
3060 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3067 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
3069 struct io_ring_ctx
*ctx
= req
->ctx
;
3070 struct io_mapped_ubuf
*imu
= req
->imu
;
3071 u16 index
, buf_index
= req
->buf_index
;
3074 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3076 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3077 imu
= READ_ONCE(ctx
->user_bufs
[index
]);
3080 return __io_import_fixed(req
, rw
, iter
, imu
);
3083 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3086 mutex_unlock(&ctx
->uring_lock
);
3089 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3092 * "Normal" inline submissions always hold the uring_lock, since we
3093 * grab it from the system call. Same is true for the SQPOLL offload.
3094 * The only exception is when we've detached the request and issue it
3095 * from an async worker thread, grab the lock for that case.
3098 mutex_lock(&ctx
->uring_lock
);
3101 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3102 int bgid
, struct io_buffer
*kbuf
,
3105 struct io_buffer
*head
;
3107 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3110 io_ring_submit_lock(req
->ctx
, needs_lock
);
3112 lockdep_assert_held(&req
->ctx
->uring_lock
);
3114 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
3116 if (!list_empty(&head
->list
)) {
3117 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3119 list_del(&kbuf
->list
);
3122 xa_erase(&req
->ctx
->io_buffers
, bgid
);
3124 if (*len
> kbuf
->len
)
3127 kbuf
= ERR_PTR(-ENOBUFS
);
3130 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3135 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3138 struct io_buffer
*kbuf
;
3141 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3142 bgid
= req
->buf_index
;
3143 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3146 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3147 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3148 return u64_to_user_ptr(kbuf
->addr
);
3151 #ifdef CONFIG_COMPAT
3152 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3155 struct compat_iovec __user
*uiov
;
3156 compat_ssize_t clen
;
3160 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3161 if (!access_ok(uiov
, sizeof(*uiov
)))
3163 if (__get_user(clen
, &uiov
->iov_len
))
3169 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3171 return PTR_ERR(buf
);
3172 iov
[0].iov_base
= buf
;
3173 iov
[0].iov_len
= (compat_size_t
) len
;
3178 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3181 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3185 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3188 len
= iov
[0].iov_len
;
3191 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3193 return PTR_ERR(buf
);
3194 iov
[0].iov_base
= buf
;
3195 iov
[0].iov_len
= len
;
3199 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3202 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3203 struct io_buffer
*kbuf
;
3205 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3206 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3207 iov
[0].iov_len
= kbuf
->len
;
3210 if (req
->rw
.len
!= 1)
3213 #ifdef CONFIG_COMPAT
3214 if (req
->ctx
->compat
)
3215 return io_compat_import(req
, iov
, needs_lock
);
3218 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3221 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
3222 struct iov_iter
*iter
, bool needs_lock
)
3224 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3225 size_t sqe_len
= req
->rw
.len
;
3226 u8 opcode
= req
->opcode
;
3229 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3231 return io_import_fixed(req
, rw
, iter
);
3234 /* buffer index only valid with fixed read/write, or buffer select */
3235 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3238 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3239 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3240 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3242 return PTR_ERR(buf
);
3243 req
->rw
.len
= sqe_len
;
3246 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3251 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3252 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3254 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3259 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3263 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3265 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3269 * For files that don't have ->read_iter() and ->write_iter(), handle them
3270 * by looping over ->read() or ->write() manually.
3272 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3274 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3275 struct file
*file
= req
->file
;
3279 * Don't support polled IO through this interface, and we can't
3280 * support non-blocking either. For the latter, this just causes
3281 * the kiocb to be handled from an async context.
3283 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3285 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3288 while (iov_iter_count(iter
)) {
3292 if (!iov_iter_is_bvec(iter
)) {
3293 iovec
= iov_iter_iovec(iter
);
3295 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3296 iovec
.iov_len
= req
->rw
.len
;
3300 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3301 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3303 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3304 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3313 if (!iov_iter_is_bvec(iter
)) {
3314 iov_iter_advance(iter
, nr
);
3321 if (nr
!= iovec
.iov_len
)
3328 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3329 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3331 struct io_async_rw
*rw
= req
->async_data
;
3333 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3334 rw
->free_iovec
= iovec
;
3336 /* can only be fixed buffers, no need to do anything */
3337 if (iov_iter_is_bvec(iter
))
3340 unsigned iov_off
= 0;
3342 rw
->iter
.iov
= rw
->fast_iov
;
3343 if (iter
->iov
!= fast_iov
) {
3344 iov_off
= iter
->iov
- fast_iov
;
3345 rw
->iter
.iov
+= iov_off
;
3347 if (rw
->fast_iov
!= fast_iov
)
3348 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3349 sizeof(struct iovec
) * iter
->nr_segs
);
3351 req
->flags
|= REQ_F_NEED_CLEANUP
;
3355 static inline int io_alloc_async_data(struct io_kiocb
*req
)
3357 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3358 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3359 return req
->async_data
== NULL
;
3362 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3363 const struct iovec
*fast_iov
,
3364 struct iov_iter
*iter
, bool force
)
3366 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_setup
)
3368 if (!req
->async_data
) {
3369 struct io_async_rw
*iorw
;
3371 if (io_alloc_async_data(req
)) {
3376 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3377 iorw
= req
->async_data
;
3378 /* we've copied and mapped the iter, ensure state is saved */
3379 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3384 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3386 struct io_async_rw
*iorw
= req
->async_data
;
3387 struct iovec
*iov
= iorw
->fast_iov
;
3390 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3391 if (unlikely(ret
< 0))
3394 iorw
->bytes_done
= 0;
3395 iorw
->free_iovec
= iov
;
3397 req
->flags
|= REQ_F_NEED_CLEANUP
;
3398 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3402 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3404 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3406 return io_prep_rw(req
, sqe
, READ
);
3410 * This is our waitqueue callback handler, registered through lock_page_async()
3411 * when we initially tried to do the IO with the iocb armed our waitqueue.
3412 * This gets called when the page is unlocked, and we generally expect that to
3413 * happen when the page IO is completed and the page is now uptodate. This will
3414 * queue a task_work based retry of the operation, attempting to copy the data
3415 * again. If the latter fails because the page was NOT uptodate, then we will
3416 * do a thread based blocking retry of the operation. That's the unexpected
3419 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3420 int sync
, void *arg
)
3422 struct wait_page_queue
*wpq
;
3423 struct io_kiocb
*req
= wait
->private;
3424 struct wait_page_key
*key
= arg
;
3426 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3428 if (!wake_page_match(wpq
, key
))
3431 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3432 list_del_init(&wait
->entry
);
3433 io_req_task_queue(req
);
3438 * This controls whether a given IO request should be armed for async page
3439 * based retry. If we return false here, the request is handed to the async
3440 * worker threads for retry. If we're doing buffered reads on a regular file,
3441 * we prepare a private wait_page_queue entry and retry the operation. This
3442 * will either succeed because the page is now uptodate and unlocked, or it
3443 * will register a callback when the page is unlocked at IO completion. Through
3444 * that callback, io_uring uses task_work to setup a retry of the operation.
3445 * That retry will attempt the buffered read again. The retry will generally
3446 * succeed, or in rare cases where it fails, we then fall back to using the
3447 * async worker threads for a blocking retry.
3449 static bool io_rw_should_retry(struct io_kiocb
*req
)
3451 struct io_async_rw
*rw
= req
->async_data
;
3452 struct wait_page_queue
*wait
= &rw
->wpq
;
3453 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3455 /* never retry for NOWAIT, we just complete with -EAGAIN */
3456 if (req
->flags
& REQ_F_NOWAIT
)
3459 /* Only for buffered IO */
3460 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3464 * just use poll if we can, and don't attempt if the fs doesn't
3465 * support callback based unlocks
3467 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3470 wait
->wait
.func
= io_async_buf_func
;
3471 wait
->wait
.private = req
;
3472 wait
->wait
.flags
= 0;
3473 INIT_LIST_HEAD(&wait
->wait
.entry
);
3474 kiocb
->ki_flags
|= IOCB_WAITQ
;
3475 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3476 kiocb
->ki_waitq
= wait
;
3480 static inline int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3482 if (req
->file
->f_op
->read_iter
)
3483 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3484 else if (req
->file
->f_op
->read
)
3485 return loop_rw_iter(READ
, req
, iter
);
3490 static bool need_read_all(struct io_kiocb
*req
)
3492 return req
->flags
& REQ_F_ISREG
||
3493 S_ISBLK(file_inode(req
->file
)->i_mode
);
3496 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3498 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3499 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3500 struct iov_iter __iter
, *iter
= &__iter
;
3501 struct io_async_rw
*rw
= req
->async_data
;
3502 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3503 struct iov_iter_state __state
, *state
;
3508 state
= &rw
->iter_state
;
3510 * We come here from an earlier attempt, restore our state to
3511 * match in case it doesn't. It's cheap enough that we don't
3512 * need to make this conditional.
3514 iov_iter_restore(iter
, state
);
3517 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3521 iov_iter_save_state(iter
, state
);
3523 req
->result
= iov_iter_count(iter
);
3525 /* Ensure we clear previously set non-block flag */
3526 if (!force_nonblock
)
3527 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3529 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3531 /* If the file doesn't support async, just async punt */
3532 if (force_nonblock
&& !io_file_supports_nowait(req
, READ
)) {
3533 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3534 return ret
?: -EAGAIN
;
3537 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3538 if (unlikely(ret
)) {
3543 ret
= io_iter_do_read(req
, iter
);
3545 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
3546 req
->flags
&= ~REQ_F_REISSUE
;
3547 /* IOPOLL retry should happen for io-wq threads */
3548 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3550 /* no retry on NONBLOCK nor RWF_NOWAIT */
3551 if (req
->flags
& REQ_F_NOWAIT
)
3554 } else if (ret
== -EIOCBQUEUED
) {
3556 } else if (ret
<= 0 || ret
== req
->result
|| !force_nonblock
||
3557 (req
->flags
& REQ_F_NOWAIT
) || !need_read_all(req
)) {
3558 /* read all, failed, already did sync or don't want to retry */
3563 * Don't depend on the iter state matching what was consumed, or being
3564 * untouched in case of error. Restore it and we'll advance it
3565 * manually if we need to.
3567 iov_iter_restore(iter
, state
);
3569 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3574 rw
= req
->async_data
;
3576 * Now use our persistent iterator and state, if we aren't already.
3577 * We've restored and mapped the iter to match.
3579 if (iter
!= &rw
->iter
) {
3581 state
= &rw
->iter_state
;
3586 * We end up here because of a partial read, either from
3587 * above or inside this loop. Advance the iter by the bytes
3588 * that were consumed.
3590 iov_iter_advance(iter
, ret
);
3591 if (!iov_iter_count(iter
))
3593 rw
->bytes_done
+= ret
;
3594 iov_iter_save_state(iter
, state
);
3596 /* if we can retry, do so with the callbacks armed */
3597 if (!io_rw_should_retry(req
)) {
3598 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3603 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3604 * we get -EIOCBQUEUED, then we'll get a notification when the
3605 * desired page gets unlocked. We can also get a partial read
3606 * here, and if we do, then just retry at the new offset.
3608 ret
= io_iter_do_read(req
, iter
);
3609 if (ret
== -EIOCBQUEUED
)
3611 /* we got some bytes, but not all. retry. */
3612 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3613 iov_iter_restore(iter
, state
);
3616 kiocb_done(kiocb
, ret
, issue_flags
);
3618 /* it's faster to check here then delegate to kfree */
3624 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3626 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3628 return io_prep_rw(req
, sqe
, WRITE
);
3631 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3633 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3634 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3635 struct iov_iter __iter
, *iter
= &__iter
;
3636 struct io_async_rw
*rw
= req
->async_data
;
3637 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3638 struct iov_iter_state __state
, *state
;
3643 state
= &rw
->iter_state
;
3644 iov_iter_restore(iter
, state
);
3647 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3651 iov_iter_save_state(iter
, state
);
3653 req
->result
= iov_iter_count(iter
);
3655 /* Ensure we clear previously set non-block flag */
3656 if (!force_nonblock
)
3657 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3659 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3661 /* If the file doesn't support async, just async punt */
3662 if (force_nonblock
&& !io_file_supports_nowait(req
, WRITE
))
3665 /* file path doesn't support NOWAIT for non-direct_IO */
3666 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3667 (req
->flags
& REQ_F_ISREG
))
3670 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3675 * Open-code file_start_write here to grab freeze protection,
3676 * which will be released by another thread in
3677 * io_complete_rw(). Fool lockdep by telling it the lock got
3678 * released so that it doesn't complain about the held lock when
3679 * we return to userspace.
3681 if (req
->flags
& REQ_F_ISREG
) {
3682 sb_start_write(file_inode(req
->file
)->i_sb
);
3683 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3686 kiocb
->ki_flags
|= IOCB_WRITE
;
3688 if (req
->file
->f_op
->write_iter
)
3689 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3690 else if (req
->file
->f_op
->write
)
3691 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3695 if (req
->flags
& REQ_F_REISSUE
) {
3696 req
->flags
&= ~REQ_F_REISSUE
;
3701 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3702 * retry them without IOCB_NOWAIT.
3704 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3706 /* no retry on NONBLOCK nor RWF_NOWAIT */
3707 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3709 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3710 /* IOPOLL retry should happen for io-wq threads */
3711 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3714 kiocb_done(kiocb
, ret2
, issue_flags
);
3717 iov_iter_restore(iter
, state
);
3718 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3719 return ret
?: -EAGAIN
;
3722 /* it's reportedly faster than delegating the null check to kfree() */
3728 static int io_renameat_prep(struct io_kiocb
*req
,
3729 const struct io_uring_sqe
*sqe
)
3731 struct io_rename
*ren
= &req
->rename
;
3732 const char __user
*oldf
, *newf
;
3734 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3736 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3738 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3741 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3742 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3743 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3744 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3745 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3747 ren
->oldpath
= getname(oldf
);
3748 if (IS_ERR(ren
->oldpath
))
3749 return PTR_ERR(ren
->oldpath
);
3751 ren
->newpath
= getname(newf
);
3752 if (IS_ERR(ren
->newpath
)) {
3753 putname(ren
->oldpath
);
3754 return PTR_ERR(ren
->newpath
);
3757 req
->flags
|= REQ_F_NEED_CLEANUP
;
3761 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3763 struct io_rename
*ren
= &req
->rename
;
3766 if (issue_flags
& IO_URING_F_NONBLOCK
)
3769 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3770 ren
->newpath
, ren
->flags
);
3772 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3775 io_req_complete(req
, ret
);
3779 static int io_unlinkat_prep(struct io_kiocb
*req
,
3780 const struct io_uring_sqe
*sqe
)
3782 struct io_unlink
*un
= &req
->unlink
;
3783 const char __user
*fname
;
3785 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3787 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
3790 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3793 un
->dfd
= READ_ONCE(sqe
->fd
);
3795 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3796 if (un
->flags
& ~AT_REMOVEDIR
)
3799 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3800 un
->filename
= getname(fname
);
3801 if (IS_ERR(un
->filename
))
3802 return PTR_ERR(un
->filename
);
3804 req
->flags
|= REQ_F_NEED_CLEANUP
;
3808 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3810 struct io_unlink
*un
= &req
->unlink
;
3813 if (issue_flags
& IO_URING_F_NONBLOCK
)
3816 if (un
->flags
& AT_REMOVEDIR
)
3817 ret
= do_rmdir(un
->dfd
, un
->filename
);
3819 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3821 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3824 io_req_complete(req
, ret
);
3828 static int io_mkdirat_prep(struct io_kiocb
*req
,
3829 const struct io_uring_sqe
*sqe
)
3831 struct io_mkdir
*mkd
= &req
->mkdir
;
3832 const char __user
*fname
;
3834 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3836 if (sqe
->ioprio
|| sqe
->off
|| sqe
->rw_flags
|| sqe
->buf_index
||
3839 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3842 mkd
->dfd
= READ_ONCE(sqe
->fd
);
3843 mkd
->mode
= READ_ONCE(sqe
->len
);
3845 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3846 mkd
->filename
= getname(fname
);
3847 if (IS_ERR(mkd
->filename
))
3848 return PTR_ERR(mkd
->filename
);
3850 req
->flags
|= REQ_F_NEED_CLEANUP
;
3854 static int io_mkdirat(struct io_kiocb
*req
, int issue_flags
)
3856 struct io_mkdir
*mkd
= &req
->mkdir
;
3859 if (issue_flags
& IO_URING_F_NONBLOCK
)
3862 ret
= do_mkdirat(mkd
->dfd
, mkd
->filename
, mkd
->mode
);
3864 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3867 io_req_complete(req
, ret
);
3871 static int io_symlinkat_prep(struct io_kiocb
*req
,
3872 const struct io_uring_sqe
*sqe
)
3874 struct io_symlink
*sl
= &req
->symlink
;
3875 const char __user
*oldpath
, *newpath
;
3877 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3879 if (sqe
->ioprio
|| sqe
->len
|| sqe
->rw_flags
|| sqe
->buf_index
||
3882 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3885 sl
->new_dfd
= READ_ONCE(sqe
->fd
);
3886 oldpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3887 newpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3889 sl
->oldpath
= getname(oldpath
);
3890 if (IS_ERR(sl
->oldpath
))
3891 return PTR_ERR(sl
->oldpath
);
3893 sl
->newpath
= getname(newpath
);
3894 if (IS_ERR(sl
->newpath
)) {
3895 putname(sl
->oldpath
);
3896 return PTR_ERR(sl
->newpath
);
3899 req
->flags
|= REQ_F_NEED_CLEANUP
;
3903 static int io_symlinkat(struct io_kiocb
*req
, int issue_flags
)
3905 struct io_symlink
*sl
= &req
->symlink
;
3908 if (issue_flags
& IO_URING_F_NONBLOCK
)
3911 ret
= do_symlinkat(sl
->oldpath
, sl
->new_dfd
, sl
->newpath
);
3913 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3916 io_req_complete(req
, ret
);
3920 static int io_linkat_prep(struct io_kiocb
*req
,
3921 const struct io_uring_sqe
*sqe
)
3923 struct io_hardlink
*lnk
= &req
->hardlink
;
3924 const char __user
*oldf
, *newf
;
3926 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3928 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3930 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3933 lnk
->old_dfd
= READ_ONCE(sqe
->fd
);
3934 lnk
->new_dfd
= READ_ONCE(sqe
->len
);
3935 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3936 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3937 lnk
->flags
= READ_ONCE(sqe
->hardlink_flags
);
3939 lnk
->oldpath
= getname(oldf
);
3940 if (IS_ERR(lnk
->oldpath
))
3941 return PTR_ERR(lnk
->oldpath
);
3943 lnk
->newpath
= getname(newf
);
3944 if (IS_ERR(lnk
->newpath
)) {
3945 putname(lnk
->oldpath
);
3946 return PTR_ERR(lnk
->newpath
);
3949 req
->flags
|= REQ_F_NEED_CLEANUP
;
3953 static int io_linkat(struct io_kiocb
*req
, int issue_flags
)
3955 struct io_hardlink
*lnk
= &req
->hardlink
;
3958 if (issue_flags
& IO_URING_F_NONBLOCK
)
3961 ret
= do_linkat(lnk
->old_dfd
, lnk
->oldpath
, lnk
->new_dfd
,
3962 lnk
->newpath
, lnk
->flags
);
3964 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3967 io_req_complete(req
, ret
);
3971 static int io_shutdown_prep(struct io_kiocb
*req
,
3972 const struct io_uring_sqe
*sqe
)
3974 #if defined(CONFIG_NET)
3975 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3977 if (unlikely(sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3978 sqe
->buf_index
|| sqe
->splice_fd_in
))
3981 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3988 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3990 #if defined(CONFIG_NET)
3991 struct socket
*sock
;
3994 if (issue_flags
& IO_URING_F_NONBLOCK
)
3997 sock
= sock_from_file(req
->file
);
3998 if (unlikely(!sock
))
4001 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
4004 io_req_complete(req
, ret
);
4011 static int __io_splice_prep(struct io_kiocb
*req
,
4012 const struct io_uring_sqe
*sqe
)
4014 struct io_splice
*sp
= &req
->splice
;
4015 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
4017 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4020 sp
->len
= READ_ONCE(sqe
->len
);
4021 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
4022 if (unlikely(sp
->flags
& ~valid_flags
))
4024 sp
->splice_fd_in
= READ_ONCE(sqe
->splice_fd_in
);
4028 static int io_tee_prep(struct io_kiocb
*req
,
4029 const struct io_uring_sqe
*sqe
)
4031 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
4033 return __io_splice_prep(req
, sqe
);
4036 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
4038 struct io_splice
*sp
= &req
->splice
;
4039 struct file
*out
= sp
->file_out
;
4040 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4044 if (issue_flags
& IO_URING_F_NONBLOCK
)
4047 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4048 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4055 ret
= do_tee(in
, out
, sp
->len
, flags
);
4057 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4062 io_req_complete(req
, ret
);
4066 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4068 struct io_splice
*sp
= &req
->splice
;
4070 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
4071 sp
->off_out
= READ_ONCE(sqe
->off
);
4072 return __io_splice_prep(req
, sqe
);
4075 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
4077 struct io_splice
*sp
= &req
->splice
;
4078 struct file
*out
= sp
->file_out
;
4079 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4080 loff_t
*poff_in
, *poff_out
;
4084 if (issue_flags
& IO_URING_F_NONBLOCK
)
4087 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4088 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4094 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
4095 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
4098 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
4100 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4105 io_req_complete(req
, ret
);
4110 * IORING_OP_NOP just posts a completion event, nothing else.
4112 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
4114 struct io_ring_ctx
*ctx
= req
->ctx
;
4116 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4119 __io_req_complete(req
, issue_flags
, 0, 0);
4123 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4125 struct io_ring_ctx
*ctx
= req
->ctx
;
4127 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4129 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4133 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
4134 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
4137 req
->sync
.off
= READ_ONCE(sqe
->off
);
4138 req
->sync
.len
= READ_ONCE(sqe
->len
);
4142 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
4144 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
4147 /* fsync always requires a blocking context */
4148 if (issue_flags
& IO_URING_F_NONBLOCK
)
4151 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
4152 end
> 0 ? end
: LLONG_MAX
,
4153 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
4156 io_req_complete(req
, ret
);
4160 static int io_fallocate_prep(struct io_kiocb
*req
,
4161 const struct io_uring_sqe
*sqe
)
4163 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
||
4166 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4169 req
->sync
.off
= READ_ONCE(sqe
->off
);
4170 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4171 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4175 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
4179 /* fallocate always requiring blocking context */
4180 if (issue_flags
& IO_URING_F_NONBLOCK
)
4182 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4186 io_req_complete(req
, ret
);
4190 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4192 const char __user
*fname
;
4195 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4197 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4199 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4202 /* open.how should be already initialised */
4203 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4204 req
->open
.how
.flags
|= O_LARGEFILE
;
4206 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4207 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4208 req
->open
.filename
= getname(fname
);
4209 if (IS_ERR(req
->open
.filename
)) {
4210 ret
= PTR_ERR(req
->open
.filename
);
4211 req
->open
.filename
= NULL
;
4215 req
->open
.file_slot
= READ_ONCE(sqe
->file_index
);
4216 if (req
->open
.file_slot
&& (req
->open
.how
.flags
& O_CLOEXEC
))
4219 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4220 req
->flags
|= REQ_F_NEED_CLEANUP
;
4224 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4226 u64 mode
= READ_ONCE(sqe
->len
);
4227 u64 flags
= READ_ONCE(sqe
->open_flags
);
4229 req
->open
.how
= build_open_how(flags
, mode
);
4230 return __io_openat_prep(req
, sqe
);
4233 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4235 struct open_how __user
*how
;
4239 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4240 len
= READ_ONCE(sqe
->len
);
4241 if (len
< OPEN_HOW_SIZE_VER0
)
4244 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4249 return __io_openat_prep(req
, sqe
);
4252 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
4254 struct open_flags op
;
4256 bool resolve_nonblock
, nonblock_set
;
4257 bool fixed
= !!req
->open
.file_slot
;
4260 ret
= build_open_flags(&req
->open
.how
, &op
);
4263 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
4264 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
4265 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4267 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4268 * it'll always -EAGAIN
4270 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
4272 op
.lookup_flags
|= LOOKUP_CACHED
;
4273 op
.open_flag
|= O_NONBLOCK
;
4277 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4282 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4285 * We could hang on to this 'fd' on retrying, but seems like
4286 * marginal gain for something that is now known to be a slower
4287 * path. So just put it, and we'll get a new one when we retry.
4292 ret
= PTR_ERR(file
);
4293 /* only retry if RESOLVE_CACHED wasn't already set by application */
4294 if (ret
== -EAGAIN
&&
4295 (!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)))
4300 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
4301 file
->f_flags
&= ~O_NONBLOCK
;
4302 fsnotify_open(file
);
4305 fd_install(ret
, file
);
4307 ret
= io_install_fixed_file(req
, file
, issue_flags
,
4308 req
->open
.file_slot
- 1);
4310 putname(req
->open
.filename
);
4311 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4314 __io_req_complete(req
, issue_flags
, ret
, 0);
4318 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
4320 return io_openat2(req
, issue_flags
);
4323 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4324 const struct io_uring_sqe
*sqe
)
4326 struct io_provide_buf
*p
= &req
->pbuf
;
4329 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
4333 tmp
= READ_ONCE(sqe
->fd
);
4334 if (!tmp
|| tmp
> USHRT_MAX
)
4337 memset(p
, 0, sizeof(*p
));
4339 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4343 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4344 int bgid
, unsigned nbufs
)
4348 /* shouldn't happen */
4352 /* the head kbuf is the list itself */
4353 while (!list_empty(&buf
->list
)) {
4354 struct io_buffer
*nxt
;
4356 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4357 list_del(&nxt
->list
);
4365 xa_erase(&ctx
->io_buffers
, bgid
);
4370 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4372 struct io_provide_buf
*p
= &req
->pbuf
;
4373 struct io_ring_ctx
*ctx
= req
->ctx
;
4374 struct io_buffer
*head
;
4376 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4378 io_ring_submit_lock(ctx
, !force_nonblock
);
4380 lockdep_assert_held(&ctx
->uring_lock
);
4383 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4385 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4389 /* complete before unlock, IOPOLL may need the lock */
4390 __io_req_complete(req
, issue_flags
, ret
, 0);
4391 io_ring_submit_unlock(ctx
, !force_nonblock
);
4395 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4396 const struct io_uring_sqe
*sqe
)
4398 unsigned long size
, tmp_check
;
4399 struct io_provide_buf
*p
= &req
->pbuf
;
4402 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
4405 tmp
= READ_ONCE(sqe
->fd
);
4406 if (!tmp
|| tmp
> USHRT_MAX
)
4409 p
->addr
= READ_ONCE(sqe
->addr
);
4410 p
->len
= READ_ONCE(sqe
->len
);
4412 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
4415 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
4418 size
= (unsigned long)p
->len
* p
->nbufs
;
4419 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4422 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4423 tmp
= READ_ONCE(sqe
->off
);
4424 if (tmp
> USHRT_MAX
)
4430 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4432 struct io_buffer
*buf
;
4433 u64 addr
= pbuf
->addr
;
4434 int i
, bid
= pbuf
->bid
;
4436 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4437 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL_ACCOUNT
);
4442 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
4447 INIT_LIST_HEAD(&buf
->list
);
4450 list_add_tail(&buf
->list
, &(*head
)->list
);
4455 return i
? i
: -ENOMEM
;
4458 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4460 struct io_provide_buf
*p
= &req
->pbuf
;
4461 struct io_ring_ctx
*ctx
= req
->ctx
;
4462 struct io_buffer
*head
, *list
;
4464 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4466 io_ring_submit_lock(ctx
, !force_nonblock
);
4468 lockdep_assert_held(&ctx
->uring_lock
);
4470 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4472 ret
= io_add_buffers(p
, &head
);
4473 if (ret
>= 0 && !list
) {
4474 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4476 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4480 /* complete before unlock, IOPOLL may need the lock */
4481 __io_req_complete(req
, issue_flags
, ret
, 0);
4482 io_ring_submit_unlock(ctx
, !force_nonblock
);
4486 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4487 const struct io_uring_sqe
*sqe
)
4489 #if defined(CONFIG_EPOLL)
4490 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4492 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4495 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4496 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4497 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4499 if (ep_op_has_event(req
->epoll
.op
)) {
4500 struct epoll_event __user
*ev
;
4502 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4503 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4513 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4515 #if defined(CONFIG_EPOLL)
4516 struct io_epoll
*ie
= &req
->epoll
;
4518 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4520 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4521 if (force_nonblock
&& ret
== -EAGAIN
)
4526 __io_req_complete(req
, issue_flags
, ret
, 0);
4533 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4535 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4536 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->splice_fd_in
)
4538 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4541 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4542 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4543 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4550 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4552 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4553 struct io_madvise
*ma
= &req
->madvise
;
4556 if (issue_flags
& IO_URING_F_NONBLOCK
)
4559 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4562 io_req_complete(req
, ret
);
4569 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4571 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
|| sqe
->splice_fd_in
)
4573 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4576 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4577 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4578 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4582 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4584 struct io_fadvise
*fa
= &req
->fadvise
;
4587 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4588 switch (fa
->advice
) {
4589 case POSIX_FADV_NORMAL
:
4590 case POSIX_FADV_RANDOM
:
4591 case POSIX_FADV_SEQUENTIAL
:
4598 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4601 __io_req_complete(req
, issue_flags
, ret
, 0);
4605 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4607 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4609 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4611 if (req
->flags
& REQ_F_FIXED_FILE
)
4614 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4615 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4616 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4617 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4618 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4623 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4625 struct io_statx
*ctx
= &req
->statx
;
4628 if (issue_flags
& IO_URING_F_NONBLOCK
)
4631 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4636 io_req_complete(req
, ret
);
4640 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4642 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4644 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4645 sqe
->rw_flags
|| sqe
->buf_index
)
4647 if (req
->flags
& REQ_F_FIXED_FILE
)
4650 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4651 req
->close
.file_slot
= READ_ONCE(sqe
->file_index
);
4652 if (req
->close
.file_slot
&& req
->close
.fd
)
4658 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4660 struct files_struct
*files
= current
->files
;
4661 struct io_close
*close
= &req
->close
;
4662 struct fdtable
*fdt
;
4663 struct file
*file
= NULL
;
4666 if (req
->close
.file_slot
) {
4667 ret
= io_close_fixed(req
, issue_flags
);
4671 spin_lock(&files
->file_lock
);
4672 fdt
= files_fdtable(files
);
4673 if (close
->fd
>= fdt
->max_fds
) {
4674 spin_unlock(&files
->file_lock
);
4677 file
= fdt
->fd
[close
->fd
];
4678 if (!file
|| file
->f_op
== &io_uring_fops
) {
4679 spin_unlock(&files
->file_lock
);
4684 /* if the file has a flush method, be safe and punt to async */
4685 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4686 spin_unlock(&files
->file_lock
);
4690 ret
= __close_fd_get_file(close
->fd
, &file
);
4691 spin_unlock(&files
->file_lock
);
4698 /* No ->flush() or already async, safely close from here */
4699 ret
= filp_close(file
, current
->files
);
4705 __io_req_complete(req
, issue_flags
, ret
, 0);
4709 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4711 struct io_ring_ctx
*ctx
= req
->ctx
;
4713 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4715 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4719 req
->sync
.off
= READ_ONCE(sqe
->off
);
4720 req
->sync
.len
= READ_ONCE(sqe
->len
);
4721 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4725 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4729 /* sync_file_range always requires a blocking context */
4730 if (issue_flags
& IO_URING_F_NONBLOCK
)
4733 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4737 io_req_complete(req
, ret
);
4741 #if defined(CONFIG_NET)
4742 static int io_setup_async_msg(struct io_kiocb
*req
,
4743 struct io_async_msghdr
*kmsg
)
4745 struct io_async_msghdr
*async_msg
= req
->async_data
;
4749 if (io_alloc_async_data(req
)) {
4750 kfree(kmsg
->free_iov
);
4753 async_msg
= req
->async_data
;
4754 req
->flags
|= REQ_F_NEED_CLEANUP
;
4755 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4756 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4757 /* if were using fast_iov, set it to the new one */
4758 if (!async_msg
->free_iov
)
4759 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4764 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4765 struct io_async_msghdr
*iomsg
)
4767 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4768 iomsg
->free_iov
= iomsg
->fast_iov
;
4769 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4770 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4773 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4777 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4779 req
->flags
|= REQ_F_NEED_CLEANUP
;
4783 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4785 struct io_sr_msg
*sr
= &req
->sr_msg
;
4787 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4789 if (unlikely(sqe
->addr2
|| sqe
->file_index
))
4792 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4793 sr
->len
= READ_ONCE(sqe
->len
);
4794 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4795 if (sr
->msg_flags
& MSG_DONTWAIT
)
4796 req
->flags
|= REQ_F_NOWAIT
;
4798 #ifdef CONFIG_COMPAT
4799 if (req
->ctx
->compat
)
4800 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4805 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4807 struct io_async_msghdr iomsg
, *kmsg
;
4808 struct socket
*sock
;
4813 sock
= sock_from_file(req
->file
);
4814 if (unlikely(!sock
))
4817 kmsg
= req
->async_data
;
4819 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4825 flags
= req
->sr_msg
.msg_flags
;
4826 if (issue_flags
& IO_URING_F_NONBLOCK
)
4827 flags
|= MSG_DONTWAIT
;
4828 if (flags
& MSG_WAITALL
)
4829 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4831 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4832 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4833 return io_setup_async_msg(req
, kmsg
);
4834 if (ret
== -ERESTARTSYS
)
4837 /* fast path, check for non-NULL to avoid function call */
4839 kfree(kmsg
->free_iov
);
4840 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4843 __io_req_complete(req
, issue_flags
, ret
, 0);
4847 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4849 struct io_sr_msg
*sr
= &req
->sr_msg
;
4852 struct socket
*sock
;
4857 sock
= sock_from_file(req
->file
);
4858 if (unlikely(!sock
))
4861 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4865 msg
.msg_name
= NULL
;
4866 msg
.msg_control
= NULL
;
4867 msg
.msg_controllen
= 0;
4868 msg
.msg_namelen
= 0;
4870 flags
= req
->sr_msg
.msg_flags
;
4871 if (issue_flags
& IO_URING_F_NONBLOCK
)
4872 flags
|= MSG_DONTWAIT
;
4873 if (flags
& MSG_WAITALL
)
4874 min_ret
= iov_iter_count(&msg
.msg_iter
);
4876 msg
.msg_flags
= flags
;
4877 ret
= sock_sendmsg(sock
, &msg
);
4878 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4880 if (ret
== -ERESTARTSYS
)
4885 __io_req_complete(req
, issue_flags
, ret
, 0);
4889 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4890 struct io_async_msghdr
*iomsg
)
4892 struct io_sr_msg
*sr
= &req
->sr_msg
;
4893 struct iovec __user
*uiov
;
4897 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4898 &iomsg
->uaddr
, &uiov
, &iov_len
);
4902 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4905 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4907 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4908 iomsg
->free_iov
= NULL
;
4910 iomsg
->free_iov
= iomsg
->fast_iov
;
4911 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4912 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4921 #ifdef CONFIG_COMPAT
4922 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4923 struct io_async_msghdr
*iomsg
)
4925 struct io_sr_msg
*sr
= &req
->sr_msg
;
4926 struct compat_iovec __user
*uiov
;
4931 ret
= __get_compat_msghdr(&iomsg
->msg
, sr
->umsg_compat
, &iomsg
->uaddr
,
4936 uiov
= compat_ptr(ptr
);
4937 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4938 compat_ssize_t clen
;
4942 if (!access_ok(uiov
, sizeof(*uiov
)))
4944 if (__get_user(clen
, &uiov
->iov_len
))
4949 iomsg
->free_iov
= NULL
;
4951 iomsg
->free_iov
= iomsg
->fast_iov
;
4952 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4953 UIO_FASTIOV
, &iomsg
->free_iov
,
4954 &iomsg
->msg
.msg_iter
, true);
4963 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4964 struct io_async_msghdr
*iomsg
)
4966 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4968 #ifdef CONFIG_COMPAT
4969 if (req
->ctx
->compat
)
4970 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4973 return __io_recvmsg_copy_hdr(req
, iomsg
);
4976 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4979 struct io_sr_msg
*sr
= &req
->sr_msg
;
4980 struct io_buffer
*kbuf
;
4982 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4987 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4991 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4993 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4996 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
5000 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
5002 req
->flags
|= REQ_F_NEED_CLEANUP
;
5006 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5008 struct io_sr_msg
*sr
= &req
->sr_msg
;
5010 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5012 if (unlikely(sqe
->addr2
|| sqe
->file_index
))
5015 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5016 sr
->len
= READ_ONCE(sqe
->len
);
5017 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
5018 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
5019 if (sr
->msg_flags
& MSG_DONTWAIT
)
5020 req
->flags
|= REQ_F_NOWAIT
;
5022 #ifdef CONFIG_COMPAT
5023 if (req
->ctx
->compat
)
5024 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
5029 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
5031 struct io_async_msghdr iomsg
, *kmsg
;
5032 struct socket
*sock
;
5033 struct io_buffer
*kbuf
;
5036 int ret
, cflags
= 0;
5037 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5039 sock
= sock_from_file(req
->file
);
5040 if (unlikely(!sock
))
5043 kmsg
= req
->async_data
;
5045 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
5051 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5052 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5054 return PTR_ERR(kbuf
);
5055 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
5056 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
5057 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
5058 1, req
->sr_msg
.len
);
5061 flags
= req
->sr_msg
.msg_flags
;
5063 flags
|= MSG_DONTWAIT
;
5064 if (flags
& MSG_WAITALL
)
5065 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
5067 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
5068 kmsg
->uaddr
, flags
);
5069 if (force_nonblock
&& ret
== -EAGAIN
)
5070 return io_setup_async_msg(req
, kmsg
);
5071 if (ret
== -ERESTARTSYS
)
5074 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5075 cflags
= io_put_recv_kbuf(req
);
5076 /* fast path, check for non-NULL to avoid function call */
5078 kfree(kmsg
->free_iov
);
5079 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5080 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5082 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5086 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
5088 struct io_buffer
*kbuf
;
5089 struct io_sr_msg
*sr
= &req
->sr_msg
;
5091 void __user
*buf
= sr
->buf
;
5092 struct socket
*sock
;
5096 int ret
, cflags
= 0;
5097 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5099 sock
= sock_from_file(req
->file
);
5100 if (unlikely(!sock
))
5103 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5104 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5106 return PTR_ERR(kbuf
);
5107 buf
= u64_to_user_ptr(kbuf
->addr
);
5110 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
5114 msg
.msg_name
= NULL
;
5115 msg
.msg_control
= NULL
;
5116 msg
.msg_controllen
= 0;
5117 msg
.msg_namelen
= 0;
5118 msg
.msg_iocb
= NULL
;
5121 flags
= req
->sr_msg
.msg_flags
;
5123 flags
|= MSG_DONTWAIT
;
5124 if (flags
& MSG_WAITALL
)
5125 min_ret
= iov_iter_count(&msg
.msg_iter
);
5127 ret
= sock_recvmsg(sock
, &msg
, flags
);
5128 if (force_nonblock
&& ret
== -EAGAIN
)
5130 if (ret
== -ERESTARTSYS
)
5133 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5134 cflags
= io_put_recv_kbuf(req
);
5135 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5137 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5141 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5143 struct io_accept
*accept
= &req
->accept
;
5145 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5147 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
5150 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5151 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5152 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
5153 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
5155 accept
->file_slot
= READ_ONCE(sqe
->file_index
);
5156 if (accept
->file_slot
&& (accept
->flags
& SOCK_CLOEXEC
))
5158 if (accept
->flags
& ~(SOCK_CLOEXEC
| SOCK_NONBLOCK
))
5160 if (SOCK_NONBLOCK
!= O_NONBLOCK
&& (accept
->flags
& SOCK_NONBLOCK
))
5161 accept
->flags
= (accept
->flags
& ~SOCK_NONBLOCK
) | O_NONBLOCK
;
5165 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
5167 struct io_accept
*accept
= &req
->accept
;
5168 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5169 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5170 bool fixed
= !!accept
->file_slot
;
5174 if (req
->file
->f_flags
& O_NONBLOCK
)
5175 req
->flags
|= REQ_F_NOWAIT
;
5178 fd
= __get_unused_fd_flags(accept
->flags
, accept
->nofile
);
5179 if (unlikely(fd
< 0))
5182 file
= do_accept(req
->file
, file_flags
, accept
->addr
, accept
->addr_len
,
5187 ret
= PTR_ERR(file
);
5188 if (ret
== -EAGAIN
&& force_nonblock
)
5190 if (ret
== -ERESTARTSYS
)
5193 } else if (!fixed
) {
5194 fd_install(fd
, file
);
5197 ret
= io_install_fixed_file(req
, file
, issue_flags
,
5198 accept
->file_slot
- 1);
5200 __io_req_complete(req
, issue_flags
, ret
, 0);
5204 static int io_connect_prep_async(struct io_kiocb
*req
)
5206 struct io_async_connect
*io
= req
->async_data
;
5207 struct io_connect
*conn
= &req
->connect
;
5209 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
5212 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5214 struct io_connect
*conn
= &req
->connect
;
5216 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5218 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
||
5222 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5223 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5227 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
5229 struct io_async_connect __io
, *io
;
5230 unsigned file_flags
;
5232 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5234 if (req
->async_data
) {
5235 io
= req
->async_data
;
5237 ret
= move_addr_to_kernel(req
->connect
.addr
,
5238 req
->connect
.addr_len
,
5245 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5247 ret
= __sys_connect_file(req
->file
, &io
->address
,
5248 req
->connect
.addr_len
, file_flags
);
5249 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5250 if (req
->async_data
)
5252 if (io_alloc_async_data(req
)) {
5256 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5259 if (ret
== -ERESTARTSYS
)
5264 __io_req_complete(req
, issue_flags
, ret
, 0);
5267 #else /* !CONFIG_NET */
5268 #define IO_NETOP_FN(op) \
5269 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5271 return -EOPNOTSUPP; \
5274 #define IO_NETOP_PREP(op) \
5276 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5278 return -EOPNOTSUPP; \
5281 #define IO_NETOP_PREP_ASYNC(op) \
5283 static int io_##op##_prep_async(struct io_kiocb *req) \
5285 return -EOPNOTSUPP; \
5288 IO_NETOP_PREP_ASYNC(sendmsg
);
5289 IO_NETOP_PREP_ASYNC(recvmsg
);
5290 IO_NETOP_PREP_ASYNC(connect
);
5291 IO_NETOP_PREP(accept
);
5294 #endif /* CONFIG_NET */
5296 struct io_poll_table
{
5297 struct poll_table_struct pt
;
5298 struct io_kiocb
*req
;
5303 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5304 __poll_t mask
, io_req_tw_func_t func
)
5306 /* for instances that support it check for an event match first: */
5307 if (mask
&& !(mask
& poll
->events
))
5310 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5312 list_del_init(&poll
->wait
.entry
);
5315 req
->io_task_work
.func
= func
;
5318 * If this fails, then the task is exiting. When a task exits, the
5319 * work gets canceled, so just cancel this request as well instead
5320 * of executing it. We can't safely execute it anyway, as we may not
5321 * have the needed state needed for it anyway.
5323 io_req_task_work_add(req
);
5327 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5328 __acquires(&req
->ctx
->completion_lock
)
5330 struct io_ring_ctx
*ctx
= req
->ctx
;
5332 /* req->task == current here, checking PF_EXITING is safe */
5333 if (unlikely(req
->task
->flags
& PF_EXITING
))
5334 WRITE_ONCE(poll
->canceled
, true);
5336 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5337 struct poll_table_struct pt
= { ._key
= poll
->events
};
5339 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5342 spin_lock(&ctx
->completion_lock
);
5343 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5344 add_wait_queue(poll
->head
, &poll
->wait
);
5351 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5353 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5354 if (req
->opcode
== IORING_OP_POLL_ADD
)
5355 return req
->async_data
;
5356 return req
->apoll
->double_poll
;
5359 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5361 if (req
->opcode
== IORING_OP_POLL_ADD
)
5363 return &req
->apoll
->poll
;
5366 static void io_poll_remove_double(struct io_kiocb
*req
)
5367 __must_hold(&req
->ctx
->completion_lock
)
5369 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5371 lockdep_assert_held(&req
->ctx
->completion_lock
);
5373 if (poll
&& poll
->head
) {
5374 struct wait_queue_head
*head
= poll
->head
;
5376 spin_lock_irq(&head
->lock
);
5377 list_del_init(&poll
->wait
.entry
);
5378 if (poll
->wait
.private)
5381 spin_unlock_irq(&head
->lock
);
5385 static bool __io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
5386 __must_hold(&req
->ctx
->completion_lock
)
5388 struct io_ring_ctx
*ctx
= req
->ctx
;
5389 unsigned flags
= IORING_CQE_F_MORE
;
5392 if (READ_ONCE(req
->poll
.canceled
)) {
5394 req
->poll
.events
|= EPOLLONESHOT
;
5396 error
= mangle_poll(mask
);
5398 if (req
->poll
.events
& EPOLLONESHOT
)
5400 if (!io_cqring_fill_event(ctx
, req
->user_data
, error
, flags
)) {
5401 req
->poll
.events
|= EPOLLONESHOT
;
5404 if (flags
& IORING_CQE_F_MORE
)
5407 return !(flags
& IORING_CQE_F_MORE
);
5410 static inline bool io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
5411 __must_hold(&req
->ctx
->completion_lock
)
5415 done
= __io_poll_complete(req
, mask
);
5416 io_commit_cqring(req
->ctx
);
5420 static void io_poll_task_func(struct io_kiocb
*req
, bool *locked
)
5422 struct io_ring_ctx
*ctx
= req
->ctx
;
5423 struct io_kiocb
*nxt
;
5425 if (io_poll_rewait(req
, &req
->poll
)) {
5426 spin_unlock(&ctx
->completion_lock
);
5430 if (req
->poll
.done
) {
5431 spin_unlock(&ctx
->completion_lock
);
5434 done
= __io_poll_complete(req
, req
->result
);
5436 io_poll_remove_double(req
);
5437 hash_del(&req
->hash_node
);
5438 req
->poll
.done
= true;
5441 add_wait_queue(req
->poll
.head
, &req
->poll
.wait
);
5443 io_commit_cqring(ctx
);
5444 spin_unlock(&ctx
->completion_lock
);
5445 io_cqring_ev_posted(ctx
);
5448 nxt
= io_put_req_find_next(req
);
5450 io_req_task_submit(nxt
, locked
);
5455 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5456 int sync
, void *key
)
5458 struct io_kiocb
*req
= wait
->private;
5459 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5460 __poll_t mask
= key_to_poll(key
);
5461 unsigned long flags
;
5463 /* for instances that support it check for an event match first: */
5464 if (mask
&& !(mask
& poll
->events
))
5466 if (!(poll
->events
& EPOLLONESHOT
))
5467 return poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5469 list_del_init(&wait
->entry
);
5474 spin_lock_irqsave(&poll
->head
->lock
, flags
);
5475 done
= list_empty(&poll
->wait
.entry
);
5477 list_del_init(&poll
->wait
.entry
);
5478 /* make sure double remove sees this as being gone */
5479 wait
->private = NULL
;
5480 spin_unlock_irqrestore(&poll
->head
->lock
, flags
);
5482 /* use wait func handler, so it matches the rq type */
5483 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5490 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5491 wait_queue_func_t wake_func
)
5495 poll
->canceled
= false;
5496 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5497 /* mask in events that we always want/need */
5498 poll
->events
= events
| IO_POLL_UNMASK
;
5499 INIT_LIST_HEAD(&poll
->wait
.entry
);
5500 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5503 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5504 struct wait_queue_head
*head
,
5505 struct io_poll_iocb
**poll_ptr
)
5507 struct io_kiocb
*req
= pt
->req
;
5510 * The file being polled uses multiple waitqueues for poll handling
5511 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5514 if (unlikely(pt
->nr_entries
)) {
5515 struct io_poll_iocb
*poll_one
= poll
;
5517 /* double add on the same waitqueue head, ignore */
5518 if (poll_one
->head
== head
)
5520 /* already have a 2nd entry, fail a third attempt */
5522 if ((*poll_ptr
)->head
== head
)
5524 pt
->error
= -EINVAL
;
5528 * Can't handle multishot for double wait for now, turn it
5529 * into one-shot mode.
5531 if (!(poll_one
->events
& EPOLLONESHOT
))
5532 poll_one
->events
|= EPOLLONESHOT
;
5533 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5535 pt
->error
= -ENOMEM
;
5538 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5540 poll
->wait
.private = req
;
5547 if (poll
->events
& EPOLLEXCLUSIVE
)
5548 add_wait_queue_exclusive(head
, &poll
->wait
);
5550 add_wait_queue(head
, &poll
->wait
);
5553 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5554 struct poll_table_struct
*p
)
5556 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5557 struct async_poll
*apoll
= pt
->req
->apoll
;
5559 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5562 static void io_async_task_func(struct io_kiocb
*req
, bool *locked
)
5564 struct async_poll
*apoll
= req
->apoll
;
5565 struct io_ring_ctx
*ctx
= req
->ctx
;
5567 trace_io_uring_task_run(req
->ctx
, req
, req
->opcode
, req
->user_data
);
5569 if (io_poll_rewait(req
, &apoll
->poll
)) {
5570 spin_unlock(&ctx
->completion_lock
);
5574 hash_del(&req
->hash_node
);
5575 io_poll_remove_double(req
);
5576 apoll
->poll
.done
= true;
5577 spin_unlock(&ctx
->completion_lock
);
5579 if (!READ_ONCE(apoll
->poll
.canceled
))
5580 io_req_task_submit(req
, locked
);
5582 io_req_complete_failed(req
, -ECANCELED
);
5585 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5588 struct io_kiocb
*req
= wait
->private;
5589 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5591 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5594 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5597 static void io_poll_req_insert(struct io_kiocb
*req
)
5599 struct io_ring_ctx
*ctx
= req
->ctx
;
5600 struct hlist_head
*list
;
5602 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5603 hlist_add_head(&req
->hash_node
, list
);
5606 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5607 struct io_poll_iocb
*poll
,
5608 struct io_poll_table
*ipt
, __poll_t mask
,
5609 wait_queue_func_t wake_func
)
5610 __acquires(&ctx
->completion_lock
)
5612 struct io_ring_ctx
*ctx
= req
->ctx
;
5613 bool cancel
= false;
5615 INIT_HLIST_NODE(&req
->hash_node
);
5616 io_init_poll_iocb(poll
, mask
, wake_func
);
5617 poll
->file
= req
->file
;
5618 poll
->wait
.private = req
;
5620 ipt
->pt
._key
= mask
;
5623 ipt
->nr_entries
= 0;
5625 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5626 if (unlikely(!ipt
->nr_entries
) && !ipt
->error
)
5627 ipt
->error
= -EINVAL
;
5629 spin_lock(&ctx
->completion_lock
);
5630 if (ipt
->error
|| (mask
&& (poll
->events
& EPOLLONESHOT
)))
5631 io_poll_remove_double(req
);
5632 if (likely(poll
->head
)) {
5633 spin_lock_irq(&poll
->head
->lock
);
5634 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5640 if ((mask
&& (poll
->events
& EPOLLONESHOT
)) || ipt
->error
)
5641 list_del_init(&poll
->wait
.entry
);
5643 WRITE_ONCE(poll
->canceled
, true);
5644 else if (!poll
->done
) /* actually waiting for an event */
5645 io_poll_req_insert(req
);
5646 spin_unlock_irq(&poll
->head
->lock
);
5658 static int io_arm_poll_handler(struct io_kiocb
*req
)
5660 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5661 struct io_ring_ctx
*ctx
= req
->ctx
;
5662 struct async_poll
*apoll
;
5663 struct io_poll_table ipt
;
5664 __poll_t ret
, mask
= EPOLLONESHOT
| POLLERR
| POLLPRI
;
5666 if (!req
->file
|| !file_can_poll(req
->file
))
5667 return IO_APOLL_ABORTED
;
5668 if (req
->flags
& REQ_F_POLLED
)
5669 return IO_APOLL_ABORTED
;
5670 if (!def
->pollin
&& !def
->pollout
)
5671 return IO_APOLL_ABORTED
;
5674 mask
|= POLLIN
| POLLRDNORM
;
5676 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5677 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5678 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5681 mask
|= POLLOUT
| POLLWRNORM
;
5684 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5685 if (unlikely(!apoll
))
5686 return IO_APOLL_ABORTED
;
5687 apoll
->double_poll
= NULL
;
5689 req
->flags
|= REQ_F_POLLED
;
5690 ipt
.pt
._qproc
= io_async_queue_proc
;
5691 io_req_set_refcount(req
);
5693 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5695 spin_unlock(&ctx
->completion_lock
);
5696 if (ret
|| ipt
.error
)
5697 return ret
? IO_APOLL_READY
: IO_APOLL_ABORTED
;
5699 trace_io_uring_poll_arm(ctx
, req
, req
->opcode
, req
->user_data
,
5700 mask
, apoll
->poll
.events
);
5704 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5705 struct io_poll_iocb
*poll
, bool do_cancel
)
5706 __must_hold(&req
->ctx
->completion_lock
)
5708 bool do_complete
= false;
5712 spin_lock_irq(&poll
->head
->lock
);
5714 WRITE_ONCE(poll
->canceled
, true);
5715 if (!list_empty(&poll
->wait
.entry
)) {
5716 list_del_init(&poll
->wait
.entry
);
5719 spin_unlock_irq(&poll
->head
->lock
);
5720 hash_del(&req
->hash_node
);
5724 static bool io_poll_remove_one(struct io_kiocb
*req
)
5725 __must_hold(&req
->ctx
->completion_lock
)
5729 io_poll_remove_double(req
);
5730 do_complete
= __io_poll_remove_one(req
, io_poll_get_single(req
), true);
5733 io_cqring_fill_event(req
->ctx
, req
->user_data
, -ECANCELED
, 0);
5734 io_commit_cqring(req
->ctx
);
5736 io_put_req_deferred(req
);
5742 * Returns true if we found and killed one or more poll requests
5744 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5747 struct hlist_node
*tmp
;
5748 struct io_kiocb
*req
;
5751 spin_lock(&ctx
->completion_lock
);
5752 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5753 struct hlist_head
*list
;
5755 list
= &ctx
->cancel_hash
[i
];
5756 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5757 if (io_match_task_safe(req
, tsk
, cancel_all
))
5758 posted
+= io_poll_remove_one(req
);
5761 spin_unlock(&ctx
->completion_lock
);
5764 io_cqring_ev_posted(ctx
);
5769 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5771 __must_hold(&ctx
->completion_lock
)
5773 struct hlist_head
*list
;
5774 struct io_kiocb
*req
;
5776 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5777 hlist_for_each_entry(req
, list
, hash_node
) {
5778 if (sqe_addr
!= req
->user_data
)
5780 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
5787 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5789 __must_hold(&ctx
->completion_lock
)
5791 struct io_kiocb
*req
;
5793 req
= io_poll_find(ctx
, sqe_addr
, poll_only
);
5796 if (io_poll_remove_one(req
))
5802 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
5807 events
= READ_ONCE(sqe
->poll32_events
);
5809 events
= swahw32(events
);
5811 if (!(flags
& IORING_POLL_ADD_MULTI
))
5812 events
|= EPOLLONESHOT
;
5813 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
5816 static int io_poll_update_prep(struct io_kiocb
*req
,
5817 const struct io_uring_sqe
*sqe
)
5819 struct io_poll_update
*upd
= &req
->poll_update
;
5822 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5824 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
5826 flags
= READ_ONCE(sqe
->len
);
5827 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
5828 IORING_POLL_ADD_MULTI
))
5830 /* meaningless without update */
5831 if (flags
== IORING_POLL_ADD_MULTI
)
5834 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
5835 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
5836 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
5838 upd
->new_user_data
= READ_ONCE(sqe
->off
);
5839 if (!upd
->update_user_data
&& upd
->new_user_data
)
5841 if (upd
->update_events
)
5842 upd
->events
= io_poll_parse_events(sqe
, flags
);
5843 else if (sqe
->poll32_events
)
5849 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5852 struct io_kiocb
*req
= wait
->private;
5853 struct io_poll_iocb
*poll
= &req
->poll
;
5855 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5858 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5859 struct poll_table_struct
*p
)
5861 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5863 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5866 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5868 struct io_poll_iocb
*poll
= &req
->poll
;
5871 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5873 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
5875 flags
= READ_ONCE(sqe
->len
);
5876 if (flags
& ~IORING_POLL_ADD_MULTI
)
5879 io_req_set_refcount(req
);
5880 poll
->events
= io_poll_parse_events(sqe
, flags
);
5884 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5886 struct io_poll_iocb
*poll
= &req
->poll
;
5887 struct io_ring_ctx
*ctx
= req
->ctx
;
5888 struct io_poll_table ipt
;
5892 ipt
.pt
._qproc
= io_poll_queue_proc
;
5894 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5897 if (mask
) { /* no async, we'd stolen it */
5899 done
= io_poll_complete(req
, mask
);
5901 spin_unlock(&ctx
->completion_lock
);
5904 io_cqring_ev_posted(ctx
);
5911 static int io_poll_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5913 struct io_ring_ctx
*ctx
= req
->ctx
;
5914 struct io_kiocb
*preq
;
5918 spin_lock(&ctx
->completion_lock
);
5919 preq
= io_poll_find(ctx
, req
->poll_update
.old_user_data
, true);
5925 if (!req
->poll_update
.update_events
&& !req
->poll_update
.update_user_data
) {
5927 ret
= io_poll_remove_one(preq
) ? 0 : -EALREADY
;
5932 * Don't allow racy completion with singleshot, as we cannot safely
5933 * update those. For multishot, if we're racing with completion, just
5934 * let completion re-add it.
5936 io_poll_remove_double(preq
);
5937 completing
= !__io_poll_remove_one(preq
, &preq
->poll
, false);
5938 if (completing
&& (preq
->poll
.events
& EPOLLONESHOT
)) {
5942 /* we now have a detached poll request. reissue. */
5946 spin_unlock(&ctx
->completion_lock
);
5948 io_req_complete(req
, ret
);
5951 /* only mask one event flags, keep behavior flags */
5952 if (req
->poll_update
.update_events
) {
5953 preq
->poll
.events
&= ~0xffff;
5954 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
5955 preq
->poll
.events
|= IO_POLL_UNMASK
;
5957 if (req
->poll_update
.update_user_data
)
5958 preq
->user_data
= req
->poll_update
.new_user_data
;
5959 spin_unlock(&ctx
->completion_lock
);
5961 /* complete update request, we're done with it */
5962 io_req_complete(req
, ret
);
5965 ret
= io_poll_add(preq
, issue_flags
);
5968 io_req_complete(preq
, ret
);
5974 static void io_req_task_timeout(struct io_kiocb
*req
, bool *locked
)
5977 io_req_complete_post(req
, -ETIME
, 0);
5980 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5982 struct io_timeout_data
*data
= container_of(timer
,
5983 struct io_timeout_data
, timer
);
5984 struct io_kiocb
*req
= data
->req
;
5985 struct io_ring_ctx
*ctx
= req
->ctx
;
5986 unsigned long flags
;
5988 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
5989 list_del_init(&req
->timeout
.list
);
5990 atomic_set(&req
->ctx
->cq_timeouts
,
5991 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5992 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
5994 req
->io_task_work
.func
= io_req_task_timeout
;
5995 io_req_task_work_add(req
);
5996 return HRTIMER_NORESTART
;
5999 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
6001 __must_hold(&ctx
->timeout_lock
)
6003 struct io_timeout_data
*io
;
6004 struct io_kiocb
*req
;
6007 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
6008 found
= user_data
== req
->user_data
;
6013 return ERR_PTR(-ENOENT
);
6015 io
= req
->async_data
;
6016 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6017 return ERR_PTR(-EALREADY
);
6018 list_del_init(&req
->timeout
.list
);
6022 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
6023 __must_hold(&ctx
->completion_lock
)
6024 __must_hold(&ctx
->timeout_lock
)
6026 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6029 return PTR_ERR(req
);
6032 io_cqring_fill_event(ctx
, req
->user_data
, -ECANCELED
, 0);
6033 io_put_req_deferred(req
);
6037 static clockid_t
io_timeout_get_clock(struct io_timeout_data
*data
)
6039 switch (data
->flags
& IORING_TIMEOUT_CLOCK_MASK
) {
6040 case IORING_TIMEOUT_BOOTTIME
:
6041 return CLOCK_BOOTTIME
;
6042 case IORING_TIMEOUT_REALTIME
:
6043 return CLOCK_REALTIME
;
6045 /* can't happen, vetted at prep time */
6049 return CLOCK_MONOTONIC
;
6053 static int io_linked_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6054 struct timespec64
*ts
, enum hrtimer_mode mode
)
6055 __must_hold(&ctx
->timeout_lock
)
6057 struct io_timeout_data
*io
;
6058 struct io_kiocb
*req
;
6061 list_for_each_entry(req
, &ctx
->ltimeout_list
, timeout
.list
) {
6062 found
= user_data
== req
->user_data
;
6069 io
= req
->async_data
;
6070 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6072 hrtimer_init(&io
->timer
, io_timeout_get_clock(io
), mode
);
6073 io
->timer
.function
= io_link_timeout_fn
;
6074 hrtimer_start(&io
->timer
, timespec64_to_ktime(*ts
), mode
);
6078 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6079 struct timespec64
*ts
, enum hrtimer_mode mode
)
6080 __must_hold(&ctx
->timeout_lock
)
6082 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6083 struct io_timeout_data
*data
;
6086 return PTR_ERR(req
);
6088 req
->timeout
.off
= 0; /* noseq */
6089 data
= req
->async_data
;
6090 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
6091 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), mode
);
6092 data
->timer
.function
= io_timeout_fn
;
6093 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
6097 static int io_timeout_remove_prep(struct io_kiocb
*req
,
6098 const struct io_uring_sqe
*sqe
)
6100 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6102 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6104 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6106 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->splice_fd_in
)
6109 tr
->ltimeout
= false;
6110 tr
->addr
= READ_ONCE(sqe
->addr
);
6111 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
6112 if (tr
->flags
& IORING_TIMEOUT_UPDATE_MASK
) {
6113 if (hweight32(tr
->flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6115 if (tr
->flags
& IORING_LINK_TIMEOUT_UPDATE
)
6116 tr
->ltimeout
= true;
6117 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE_MASK
|IORING_TIMEOUT_ABS
))
6119 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
6121 } else if (tr
->flags
) {
6122 /* timeout removal doesn't support flags */
6129 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
6131 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
6136 * Remove or update an existing timeout command
6138 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
6140 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6141 struct io_ring_ctx
*ctx
= req
->ctx
;
6144 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
)) {
6145 spin_lock(&ctx
->completion_lock
);
6146 spin_lock_irq(&ctx
->timeout_lock
);
6147 ret
= io_timeout_cancel(ctx
, tr
->addr
);
6148 spin_unlock_irq(&ctx
->timeout_lock
);
6149 spin_unlock(&ctx
->completion_lock
);
6151 enum hrtimer_mode mode
= io_translate_timeout_mode(tr
->flags
);
6153 spin_lock_irq(&ctx
->timeout_lock
);
6155 ret
= io_linked_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6157 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6158 spin_unlock_irq(&ctx
->timeout_lock
);
6163 io_req_complete_post(req
, ret
, 0);
6167 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6168 bool is_timeout_link
)
6170 struct io_timeout_data
*data
;
6172 u32 off
= READ_ONCE(sqe
->off
);
6174 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6176 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1 ||
6179 if (off
&& is_timeout_link
)
6181 flags
= READ_ONCE(sqe
->timeout_flags
);
6182 if (flags
& ~(IORING_TIMEOUT_ABS
| IORING_TIMEOUT_CLOCK_MASK
))
6184 /* more than one clock specified is invalid, obviously */
6185 if (hweight32(flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6188 INIT_LIST_HEAD(&req
->timeout
.list
);
6189 req
->timeout
.off
= off
;
6190 if (unlikely(off
&& !req
->ctx
->off_timeout_used
))
6191 req
->ctx
->off_timeout_used
= true;
6193 if (!req
->async_data
&& io_alloc_async_data(req
))
6196 data
= req
->async_data
;
6198 data
->flags
= flags
;
6200 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
6203 INIT_LIST_HEAD(&req
->timeout
.list
);
6204 data
->mode
= io_translate_timeout_mode(flags
);
6205 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), data
->mode
);
6207 if (is_timeout_link
) {
6208 struct io_submit_link
*link
= &req
->ctx
->submit_state
.link
;
6212 if (link
->last
->opcode
== IORING_OP_LINK_TIMEOUT
)
6214 req
->timeout
.head
= link
->last
;
6215 link
->last
->flags
|= REQ_F_ARM_LTIMEOUT
;
6220 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
6222 struct io_ring_ctx
*ctx
= req
->ctx
;
6223 struct io_timeout_data
*data
= req
->async_data
;
6224 struct list_head
*entry
;
6225 u32 tail
, off
= req
->timeout
.off
;
6227 spin_lock_irq(&ctx
->timeout_lock
);
6230 * sqe->off holds how many events that need to occur for this
6231 * timeout event to be satisfied. If it isn't set, then this is
6232 * a pure timeout request, sequence isn't used.
6234 if (io_is_timeout_noseq(req
)) {
6235 entry
= ctx
->timeout_list
.prev
;
6239 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
6240 req
->timeout
.target_seq
= tail
+ off
;
6242 /* Update the last seq here in case io_flush_timeouts() hasn't.
6243 * This is safe because ->completion_lock is held, and submissions
6244 * and completions are never mixed in the same ->completion_lock section.
6246 ctx
->cq_last_tm_flush
= tail
;
6249 * Insertion sort, ensuring the first entry in the list is always
6250 * the one we need first.
6252 list_for_each_prev(entry
, &ctx
->timeout_list
) {
6253 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
6256 if (io_is_timeout_noseq(nxt
))
6258 /* nxt.seq is behind @tail, otherwise would've been completed */
6259 if (off
>= nxt
->timeout
.target_seq
- tail
)
6263 list_add(&req
->timeout
.list
, entry
);
6264 data
->timer
.function
= io_timeout_fn
;
6265 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
6266 spin_unlock_irq(&ctx
->timeout_lock
);
6270 struct io_cancel_data
{
6271 struct io_ring_ctx
*ctx
;
6275 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
6277 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6278 struct io_cancel_data
*cd
= data
;
6280 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
6283 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
6284 struct io_ring_ctx
*ctx
)
6286 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
6287 enum io_wq_cancel cancel_ret
;
6290 if (!tctx
|| !tctx
->io_wq
)
6293 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
6294 switch (cancel_ret
) {
6295 case IO_WQ_CANCEL_OK
:
6298 case IO_WQ_CANCEL_RUNNING
:
6301 case IO_WQ_CANCEL_NOTFOUND
:
6309 static int io_try_cancel_userdata(struct io_kiocb
*req
, u64 sqe_addr
)
6311 struct io_ring_ctx
*ctx
= req
->ctx
;
6314 WARN_ON_ONCE(!io_wq_current_is_worker() && req
->task
!= current
);
6316 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
6320 spin_lock(&ctx
->completion_lock
);
6321 spin_lock_irq(&ctx
->timeout_lock
);
6322 ret
= io_timeout_cancel(ctx
, sqe_addr
);
6323 spin_unlock_irq(&ctx
->timeout_lock
);
6326 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
6328 spin_unlock(&ctx
->completion_lock
);
6332 static int io_async_cancel_prep(struct io_kiocb
*req
,
6333 const struct io_uring_sqe
*sqe
)
6335 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6337 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6339 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
||
6343 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
6347 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
6349 struct io_ring_ctx
*ctx
= req
->ctx
;
6350 u64 sqe_addr
= req
->cancel
.addr
;
6351 struct io_tctx_node
*node
;
6354 ret
= io_try_cancel_userdata(req
, sqe_addr
);
6358 /* slow path, try all io-wq's */
6359 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6361 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
6362 struct io_uring_task
*tctx
= node
->task
->io_uring
;
6364 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
6368 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6372 io_req_complete_post(req
, ret
, 0);
6376 static int io_rsrc_update_prep(struct io_kiocb
*req
,
6377 const struct io_uring_sqe
*sqe
)
6379 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6381 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
6384 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
6385 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
6386 if (!req
->rsrc_update
.nr_args
)
6388 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
6392 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
6394 struct io_ring_ctx
*ctx
= req
->ctx
;
6395 struct io_uring_rsrc_update2 up
;
6398 up
.offset
= req
->rsrc_update
.offset
;
6399 up
.data
= req
->rsrc_update
.arg
;
6405 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6406 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
6407 &up
, req
->rsrc_update
.nr_args
);
6408 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6412 __io_req_complete(req
, issue_flags
, ret
, 0);
6416 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6418 switch (req
->opcode
) {
6421 case IORING_OP_READV
:
6422 case IORING_OP_READ_FIXED
:
6423 case IORING_OP_READ
:
6424 return io_read_prep(req
, sqe
);
6425 case IORING_OP_WRITEV
:
6426 case IORING_OP_WRITE_FIXED
:
6427 case IORING_OP_WRITE
:
6428 return io_write_prep(req
, sqe
);
6429 case IORING_OP_POLL_ADD
:
6430 return io_poll_add_prep(req
, sqe
);
6431 case IORING_OP_POLL_REMOVE
:
6432 return io_poll_update_prep(req
, sqe
);
6433 case IORING_OP_FSYNC
:
6434 return io_fsync_prep(req
, sqe
);
6435 case IORING_OP_SYNC_FILE_RANGE
:
6436 return io_sfr_prep(req
, sqe
);
6437 case IORING_OP_SENDMSG
:
6438 case IORING_OP_SEND
:
6439 return io_sendmsg_prep(req
, sqe
);
6440 case IORING_OP_RECVMSG
:
6441 case IORING_OP_RECV
:
6442 return io_recvmsg_prep(req
, sqe
);
6443 case IORING_OP_CONNECT
:
6444 return io_connect_prep(req
, sqe
);
6445 case IORING_OP_TIMEOUT
:
6446 return io_timeout_prep(req
, sqe
, false);
6447 case IORING_OP_TIMEOUT_REMOVE
:
6448 return io_timeout_remove_prep(req
, sqe
);
6449 case IORING_OP_ASYNC_CANCEL
:
6450 return io_async_cancel_prep(req
, sqe
);
6451 case IORING_OP_LINK_TIMEOUT
:
6452 return io_timeout_prep(req
, sqe
, true);
6453 case IORING_OP_ACCEPT
:
6454 return io_accept_prep(req
, sqe
);
6455 case IORING_OP_FALLOCATE
:
6456 return io_fallocate_prep(req
, sqe
);
6457 case IORING_OP_OPENAT
:
6458 return io_openat_prep(req
, sqe
);
6459 case IORING_OP_CLOSE
:
6460 return io_close_prep(req
, sqe
);
6461 case IORING_OP_FILES_UPDATE
:
6462 return io_rsrc_update_prep(req
, sqe
);
6463 case IORING_OP_STATX
:
6464 return io_statx_prep(req
, sqe
);
6465 case IORING_OP_FADVISE
:
6466 return io_fadvise_prep(req
, sqe
);
6467 case IORING_OP_MADVISE
:
6468 return io_madvise_prep(req
, sqe
);
6469 case IORING_OP_OPENAT2
:
6470 return io_openat2_prep(req
, sqe
);
6471 case IORING_OP_EPOLL_CTL
:
6472 return io_epoll_ctl_prep(req
, sqe
);
6473 case IORING_OP_SPLICE
:
6474 return io_splice_prep(req
, sqe
);
6475 case IORING_OP_PROVIDE_BUFFERS
:
6476 return io_provide_buffers_prep(req
, sqe
);
6477 case IORING_OP_REMOVE_BUFFERS
:
6478 return io_remove_buffers_prep(req
, sqe
);
6480 return io_tee_prep(req
, sqe
);
6481 case IORING_OP_SHUTDOWN
:
6482 return io_shutdown_prep(req
, sqe
);
6483 case IORING_OP_RENAMEAT
:
6484 return io_renameat_prep(req
, sqe
);
6485 case IORING_OP_UNLINKAT
:
6486 return io_unlinkat_prep(req
, sqe
);
6487 case IORING_OP_MKDIRAT
:
6488 return io_mkdirat_prep(req
, sqe
);
6489 case IORING_OP_SYMLINKAT
:
6490 return io_symlinkat_prep(req
, sqe
);
6491 case IORING_OP_LINKAT
:
6492 return io_linkat_prep(req
, sqe
);
6495 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6500 static int io_req_prep_async(struct io_kiocb
*req
)
6502 if (!io_op_defs
[req
->opcode
].needs_async_setup
)
6504 if (WARN_ON_ONCE(req
->async_data
))
6506 if (io_alloc_async_data(req
))
6509 switch (req
->opcode
) {
6510 case IORING_OP_READV
:
6511 return io_rw_prep_async(req
, READ
);
6512 case IORING_OP_WRITEV
:
6513 return io_rw_prep_async(req
, WRITE
);
6514 case IORING_OP_SENDMSG
:
6515 return io_sendmsg_prep_async(req
);
6516 case IORING_OP_RECVMSG
:
6517 return io_recvmsg_prep_async(req
);
6518 case IORING_OP_CONNECT
:
6519 return io_connect_prep_async(req
);
6521 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
6526 static u32
io_get_sequence(struct io_kiocb
*req
)
6528 u32 seq
= req
->ctx
->cached_sq_head
;
6530 /* need original cached_sq_head, but it was increased for each req */
6531 io_for_each_link(req
, req
)
6536 static bool io_drain_req(struct io_kiocb
*req
)
6538 struct io_kiocb
*pos
;
6539 struct io_ring_ctx
*ctx
= req
->ctx
;
6540 struct io_defer_entry
*de
;
6544 if (req
->flags
& REQ_F_FAIL
) {
6545 io_req_complete_fail_submit(req
);
6550 * If we need to drain a request in the middle of a link, drain the
6551 * head request and the next request/link after the current link.
6552 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6553 * maintained for every request of our link.
6555 if (ctx
->drain_next
) {
6556 req
->flags
|= REQ_F_IO_DRAIN
;
6557 ctx
->drain_next
= false;
6559 /* not interested in head, start from the first linked */
6560 io_for_each_link(pos
, req
->link
) {
6561 if (pos
->flags
& REQ_F_IO_DRAIN
) {
6562 ctx
->drain_next
= true;
6563 req
->flags
|= REQ_F_IO_DRAIN
;
6568 /* Still need defer if there is pending req in defer list. */
6569 spin_lock(&ctx
->completion_lock
);
6570 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6571 !(req
->flags
& REQ_F_IO_DRAIN
))) {
6572 spin_unlock(&ctx
->completion_lock
);
6573 ctx
->drain_active
= false;
6576 spin_unlock(&ctx
->completion_lock
);
6578 seq
= io_get_sequence(req
);
6579 /* Still a chance to pass the sequence check */
6580 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6583 ret
= io_req_prep_async(req
);
6586 io_prep_async_link(req
);
6587 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6591 io_req_complete_failed(req
, ret
);
6595 spin_lock(&ctx
->completion_lock
);
6596 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6597 spin_unlock(&ctx
->completion_lock
);
6599 io_queue_async_work(req
, NULL
);
6603 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6606 list_add_tail(&de
->list
, &ctx
->defer_list
);
6607 spin_unlock(&ctx
->completion_lock
);
6611 static void io_clean_op(struct io_kiocb
*req
)
6613 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6614 switch (req
->opcode
) {
6615 case IORING_OP_READV
:
6616 case IORING_OP_READ_FIXED
:
6617 case IORING_OP_READ
:
6618 kfree((void *)(unsigned long)req
->rw
.addr
);
6620 case IORING_OP_RECVMSG
:
6621 case IORING_OP_RECV
:
6622 kfree(req
->sr_msg
.kbuf
);
6627 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6628 switch (req
->opcode
) {
6629 case IORING_OP_READV
:
6630 case IORING_OP_READ_FIXED
:
6631 case IORING_OP_READ
:
6632 case IORING_OP_WRITEV
:
6633 case IORING_OP_WRITE_FIXED
:
6634 case IORING_OP_WRITE
: {
6635 struct io_async_rw
*io
= req
->async_data
;
6637 kfree(io
->free_iovec
);
6640 case IORING_OP_RECVMSG
:
6641 case IORING_OP_SENDMSG
: {
6642 struct io_async_msghdr
*io
= req
->async_data
;
6644 kfree(io
->free_iov
);
6647 case IORING_OP_OPENAT
:
6648 case IORING_OP_OPENAT2
:
6649 if (req
->open
.filename
)
6650 putname(req
->open
.filename
);
6652 case IORING_OP_RENAMEAT
:
6653 putname(req
->rename
.oldpath
);
6654 putname(req
->rename
.newpath
);
6656 case IORING_OP_UNLINKAT
:
6657 putname(req
->unlink
.filename
);
6659 case IORING_OP_MKDIRAT
:
6660 putname(req
->mkdir
.filename
);
6662 case IORING_OP_SYMLINKAT
:
6663 putname(req
->symlink
.oldpath
);
6664 putname(req
->symlink
.newpath
);
6666 case IORING_OP_LINKAT
:
6667 putname(req
->hardlink
.oldpath
);
6668 putname(req
->hardlink
.newpath
);
6672 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
6673 kfree(req
->apoll
->double_poll
);
6677 if (req
->flags
& REQ_F_INFLIGHT
) {
6678 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6680 atomic_dec(&tctx
->inflight_tracked
);
6682 if (req
->flags
& REQ_F_CREDS
)
6683 put_cred(req
->creds
);
6685 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
6688 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6690 struct io_ring_ctx
*ctx
= req
->ctx
;
6691 const struct cred
*creds
= NULL
;
6694 if ((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred())
6695 creds
= override_creds(req
->creds
);
6697 switch (req
->opcode
) {
6699 ret
= io_nop(req
, issue_flags
);
6701 case IORING_OP_READV
:
6702 case IORING_OP_READ_FIXED
:
6703 case IORING_OP_READ
:
6704 ret
= io_read(req
, issue_flags
);
6706 case IORING_OP_WRITEV
:
6707 case IORING_OP_WRITE_FIXED
:
6708 case IORING_OP_WRITE
:
6709 ret
= io_write(req
, issue_flags
);
6711 case IORING_OP_FSYNC
:
6712 ret
= io_fsync(req
, issue_flags
);
6714 case IORING_OP_POLL_ADD
:
6715 ret
= io_poll_add(req
, issue_flags
);
6717 case IORING_OP_POLL_REMOVE
:
6718 ret
= io_poll_update(req
, issue_flags
);
6720 case IORING_OP_SYNC_FILE_RANGE
:
6721 ret
= io_sync_file_range(req
, issue_flags
);
6723 case IORING_OP_SENDMSG
:
6724 ret
= io_sendmsg(req
, issue_flags
);
6726 case IORING_OP_SEND
:
6727 ret
= io_send(req
, issue_flags
);
6729 case IORING_OP_RECVMSG
:
6730 ret
= io_recvmsg(req
, issue_flags
);
6732 case IORING_OP_RECV
:
6733 ret
= io_recv(req
, issue_flags
);
6735 case IORING_OP_TIMEOUT
:
6736 ret
= io_timeout(req
, issue_flags
);
6738 case IORING_OP_TIMEOUT_REMOVE
:
6739 ret
= io_timeout_remove(req
, issue_flags
);
6741 case IORING_OP_ACCEPT
:
6742 ret
= io_accept(req
, issue_flags
);
6744 case IORING_OP_CONNECT
:
6745 ret
= io_connect(req
, issue_flags
);
6747 case IORING_OP_ASYNC_CANCEL
:
6748 ret
= io_async_cancel(req
, issue_flags
);
6750 case IORING_OP_FALLOCATE
:
6751 ret
= io_fallocate(req
, issue_flags
);
6753 case IORING_OP_OPENAT
:
6754 ret
= io_openat(req
, issue_flags
);
6756 case IORING_OP_CLOSE
:
6757 ret
= io_close(req
, issue_flags
);
6759 case IORING_OP_FILES_UPDATE
:
6760 ret
= io_files_update(req
, issue_flags
);
6762 case IORING_OP_STATX
:
6763 ret
= io_statx(req
, issue_flags
);
6765 case IORING_OP_FADVISE
:
6766 ret
= io_fadvise(req
, issue_flags
);
6768 case IORING_OP_MADVISE
:
6769 ret
= io_madvise(req
, issue_flags
);
6771 case IORING_OP_OPENAT2
:
6772 ret
= io_openat2(req
, issue_flags
);
6774 case IORING_OP_EPOLL_CTL
:
6775 ret
= io_epoll_ctl(req
, issue_flags
);
6777 case IORING_OP_SPLICE
:
6778 ret
= io_splice(req
, issue_flags
);
6780 case IORING_OP_PROVIDE_BUFFERS
:
6781 ret
= io_provide_buffers(req
, issue_flags
);
6783 case IORING_OP_REMOVE_BUFFERS
:
6784 ret
= io_remove_buffers(req
, issue_flags
);
6787 ret
= io_tee(req
, issue_flags
);
6789 case IORING_OP_SHUTDOWN
:
6790 ret
= io_shutdown(req
, issue_flags
);
6792 case IORING_OP_RENAMEAT
:
6793 ret
= io_renameat(req
, issue_flags
);
6795 case IORING_OP_UNLINKAT
:
6796 ret
= io_unlinkat(req
, issue_flags
);
6798 case IORING_OP_MKDIRAT
:
6799 ret
= io_mkdirat(req
, issue_flags
);
6801 case IORING_OP_SYMLINKAT
:
6802 ret
= io_symlinkat(req
, issue_flags
);
6804 case IORING_OP_LINKAT
:
6805 ret
= io_linkat(req
, issue_flags
);
6813 revert_creds(creds
);
6816 /* If the op doesn't have a file, we're not polling for it */
6817 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
6818 io_iopoll_req_issued(req
);
6823 static struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
6825 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6827 req
= io_put_req_find_next(req
);
6828 return req
? &req
->work
: NULL
;
6831 static void io_wq_submit_work(struct io_wq_work
*work
)
6833 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6834 struct io_kiocb
*timeout
;
6837 /* one will be dropped by ->io_free_work() after returning to io-wq */
6838 if (!(req
->flags
& REQ_F_REFCOUNT
))
6839 __io_req_set_refcount(req
, 2);
6843 timeout
= io_prep_linked_timeout(req
);
6845 io_queue_linked_timeout(timeout
);
6847 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6848 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6853 ret
= io_issue_sqe(req
, 0);
6855 * We can get EAGAIN for polled IO even though we're
6856 * forcing a sync submission from here, since we can't
6857 * wait for request slots on the block side.
6865 /* avoid locking problems by failing it from a clean context */
6867 io_req_task_queue_fail(req
, ret
);
6870 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
6873 return &table
->files
[i
];
6876 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6879 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
6881 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
6884 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
6886 unsigned long file_ptr
= (unsigned long) file
;
6888 if (__io_file_supports_nowait(file
, READ
))
6889 file_ptr
|= FFS_ASYNC_READ
;
6890 if (__io_file_supports_nowait(file
, WRITE
))
6891 file_ptr
|= FFS_ASYNC_WRITE
;
6892 if (S_ISREG(file_inode(file
)->i_mode
))
6893 file_ptr
|= FFS_ISREG
;
6894 file_slot
->file_ptr
= file_ptr
;
6897 static inline struct file
*io_file_get_fixed(struct io_ring_ctx
*ctx
,
6898 struct io_kiocb
*req
, int fd
)
6901 unsigned long file_ptr
;
6903 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6905 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6906 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
6907 file
= (struct file
*) (file_ptr
& FFS_MASK
);
6908 file_ptr
&= ~FFS_MASK
;
6909 /* mask in overlapping REQ_F and FFS bits */
6910 req
->flags
|= (file_ptr
<< REQ_F_NOWAIT_READ_BIT
);
6911 io_req_set_rsrc_node(req
);
6915 static struct file
*io_file_get_normal(struct io_ring_ctx
*ctx
,
6916 struct io_kiocb
*req
, int fd
)
6918 struct file
*file
= fget(fd
);
6920 trace_io_uring_file_get(ctx
, fd
);
6922 /* we don't allow fixed io_uring files */
6923 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6924 io_req_track_inflight(req
);
6928 static inline struct file
*io_file_get(struct io_ring_ctx
*ctx
,
6929 struct io_kiocb
*req
, int fd
, bool fixed
)
6932 return io_file_get_fixed(ctx
, req
, fd
);
6934 return io_file_get_normal(ctx
, req
, fd
);
6937 static void io_req_task_link_timeout(struct io_kiocb
*req
, bool *locked
)
6939 struct io_kiocb
*prev
= req
->timeout
.prev
;
6943 if (!(req
->task
->flags
& PF_EXITING
))
6944 ret
= io_try_cancel_userdata(req
, prev
->user_data
);
6945 io_req_complete_post(req
, ret
?: -ETIME
, 0);
6948 io_req_complete_post(req
, -ETIME
, 0);
6952 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6954 struct io_timeout_data
*data
= container_of(timer
,
6955 struct io_timeout_data
, timer
);
6956 struct io_kiocb
*prev
, *req
= data
->req
;
6957 struct io_ring_ctx
*ctx
= req
->ctx
;
6958 unsigned long flags
;
6960 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
6961 prev
= req
->timeout
.head
;
6962 req
->timeout
.head
= NULL
;
6965 * We don't expect the list to be empty, that will only happen if we
6966 * race with the completion of the linked work.
6969 io_remove_next_linked(prev
);
6970 if (!req_ref_inc_not_zero(prev
))
6973 list_del(&req
->timeout
.list
);
6974 req
->timeout
.prev
= prev
;
6975 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
6977 req
->io_task_work
.func
= io_req_task_link_timeout
;
6978 io_req_task_work_add(req
);
6979 return HRTIMER_NORESTART
;
6982 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6984 struct io_ring_ctx
*ctx
= req
->ctx
;
6986 spin_lock_irq(&ctx
->timeout_lock
);
6988 * If the back reference is NULL, then our linked request finished
6989 * before we got a chance to setup the timer
6991 if (req
->timeout
.head
) {
6992 struct io_timeout_data
*data
= req
->async_data
;
6994 data
->timer
.function
= io_link_timeout_fn
;
6995 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6997 list_add_tail(&req
->timeout
.list
, &ctx
->ltimeout_list
);
6999 spin_unlock_irq(&ctx
->timeout_lock
);
7000 /* drop submission reference */
7004 static void __io_queue_sqe(struct io_kiocb
*req
)
7005 __must_hold(&req
->ctx
->uring_lock
)
7007 struct io_kiocb
*linked_timeout
;
7011 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
7014 * We async punt it if the file wasn't marked NOWAIT, or if the file
7015 * doesn't support non-blocking read/write attempts
7018 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
7019 struct io_ring_ctx
*ctx
= req
->ctx
;
7020 struct io_submit_state
*state
= &ctx
->submit_state
;
7022 state
->compl_reqs
[state
->compl_nr
++] = req
;
7023 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
7024 io_submit_flush_completions(ctx
);
7028 linked_timeout
= io_prep_linked_timeout(req
);
7030 io_queue_linked_timeout(linked_timeout
);
7031 } else if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
7032 linked_timeout
= io_prep_linked_timeout(req
);
7034 switch (io_arm_poll_handler(req
)) {
7035 case IO_APOLL_READY
:
7037 io_queue_linked_timeout(linked_timeout
);
7039 case IO_APOLL_ABORTED
:
7041 * Queued up for async execution, worker will release
7042 * submit reference when the iocb is actually submitted.
7044 io_queue_async_work(req
, NULL
);
7049 io_queue_linked_timeout(linked_timeout
);
7051 io_req_complete_failed(req
, ret
);
7055 static inline void io_queue_sqe(struct io_kiocb
*req
)
7056 __must_hold(&req
->ctx
->uring_lock
)
7058 if (unlikely(req
->ctx
->drain_active
) && io_drain_req(req
))
7061 if (likely(!(req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
)))) {
7062 __io_queue_sqe(req
);
7063 } else if (req
->flags
& REQ_F_FAIL
) {
7064 io_req_complete_fail_submit(req
);
7066 int ret
= io_req_prep_async(req
);
7069 io_req_complete_failed(req
, ret
);
7071 io_queue_async_work(req
, NULL
);
7076 * Check SQE restrictions (opcode and flags).
7078 * Returns 'true' if SQE is allowed, 'false' otherwise.
7080 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
7081 struct io_kiocb
*req
,
7082 unsigned int sqe_flags
)
7084 if (likely(!ctx
->restricted
))
7087 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
7090 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
7091 ctx
->restrictions
.sqe_flags_required
)
7094 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
7095 ctx
->restrictions
.sqe_flags_required
))
7101 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7102 const struct io_uring_sqe
*sqe
)
7103 __must_hold(&ctx
->uring_lock
)
7105 struct io_submit_state
*state
;
7106 unsigned int sqe_flags
;
7107 int personality
, ret
= 0;
7109 /* req is partially pre-initialised, see io_preinit_req() */
7110 req
->opcode
= READ_ONCE(sqe
->opcode
);
7111 /* same numerical values with corresponding REQ_F_*, safe to copy */
7112 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
7113 req
->user_data
= READ_ONCE(sqe
->user_data
);
7115 req
->fixed_rsrc_refs
= NULL
;
7116 req
->task
= current
;
7118 /* enforce forwards compatibility on users */
7119 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
7121 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
7123 if (!io_check_restriction(ctx
, req
, sqe_flags
))
7126 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
7127 !io_op_defs
[req
->opcode
].buffer_select
)
7129 if (unlikely(sqe_flags
& IOSQE_IO_DRAIN
))
7130 ctx
->drain_active
= true;
7132 personality
= READ_ONCE(sqe
->personality
);
7134 req
->creds
= xa_load(&ctx
->personalities
, personality
);
7137 get_cred(req
->creds
);
7138 req
->flags
|= REQ_F_CREDS
;
7140 state
= &ctx
->submit_state
;
7143 * Plug now if we have more than 1 IO left after this, and the target
7144 * is potentially a read/write to block based storage.
7146 if (!state
->plug_started
&& state
->ios_left
> 1 &&
7147 io_op_defs
[req
->opcode
].plug
) {
7148 blk_start_plug(&state
->plug
);
7149 state
->plug_started
= true;
7152 if (io_op_defs
[req
->opcode
].needs_file
) {
7153 req
->file
= io_file_get(ctx
, req
, READ_ONCE(sqe
->fd
),
7154 (sqe_flags
& IOSQE_FIXED_FILE
));
7155 if (unlikely(!req
->file
))
7163 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7164 const struct io_uring_sqe
*sqe
)
7165 __must_hold(&ctx
->uring_lock
)
7167 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
7170 ret
= io_init_req(ctx
, req
, sqe
);
7171 if (unlikely(ret
)) {
7173 /* fail even hard links since we don't submit */
7176 * we can judge a link req is failed or cancelled by if
7177 * REQ_F_FAIL is set, but the head is an exception since
7178 * it may be set REQ_F_FAIL because of other req's failure
7179 * so let's leverage req->result to distinguish if a head
7180 * is set REQ_F_FAIL because of its failure or other req's
7181 * failure so that we can set the correct ret code for it.
7182 * init result here to avoid affecting the normal path.
7184 if (!(link
->head
->flags
& REQ_F_FAIL
))
7185 req_fail_link_node(link
->head
, -ECANCELED
);
7186 } else if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7188 * the current req is a normal req, we should return
7189 * error and thus break the submittion loop.
7191 io_req_complete_failed(req
, ret
);
7194 req_fail_link_node(req
, ret
);
7196 ret
= io_req_prep(req
, sqe
);
7201 /* don't need @sqe from now on */
7202 trace_io_uring_submit_sqe(ctx
, req
, req
->opcode
, req
->user_data
,
7204 ctx
->flags
& IORING_SETUP_SQPOLL
);
7207 * If we already have a head request, queue this one for async
7208 * submittal once the head completes. If we don't have a head but
7209 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7210 * submitted sync once the chain is complete. If none of those
7211 * conditions are true (normal request), then just queue it.
7214 struct io_kiocb
*head
= link
->head
;
7216 if (!(req
->flags
& REQ_F_FAIL
)) {
7217 ret
= io_req_prep_async(req
);
7218 if (unlikely(ret
)) {
7219 req_fail_link_node(req
, ret
);
7220 if (!(head
->flags
& REQ_F_FAIL
))
7221 req_fail_link_node(head
, -ECANCELED
);
7224 trace_io_uring_link(ctx
, req
, head
);
7225 link
->last
->link
= req
;
7228 /* last request of a link, enqueue the link */
7229 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7234 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
7246 * Batched submission is done, ensure local IO is flushed out.
7248 static void io_submit_state_end(struct io_submit_state
*state
,
7249 struct io_ring_ctx
*ctx
)
7251 if (state
->link
.head
)
7252 io_queue_sqe(state
->link
.head
);
7253 if (state
->compl_nr
)
7254 io_submit_flush_completions(ctx
);
7255 if (state
->plug_started
)
7256 blk_finish_plug(&state
->plug
);
7260 * Start submission side cache.
7262 static void io_submit_state_start(struct io_submit_state
*state
,
7263 unsigned int max_ios
)
7265 state
->plug_started
= false;
7266 state
->ios_left
= max_ios
;
7267 /* set only head, no need to init link_last in advance */
7268 state
->link
.head
= NULL
;
7271 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
7273 struct io_rings
*rings
= ctx
->rings
;
7276 * Ensure any loads from the SQEs are done at this point,
7277 * since once we write the new head, the application could
7278 * write new data to them.
7280 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
7284 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7285 * that is mapped by userspace. This means that care needs to be taken to
7286 * ensure that reads are stable, as we cannot rely on userspace always
7287 * being a good citizen. If members of the sqe are validated and then later
7288 * used, it's important that those reads are done through READ_ONCE() to
7289 * prevent a re-load down the line.
7291 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
7293 unsigned head
, mask
= ctx
->sq_entries
- 1;
7294 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
7297 * The cached sq head (or cq tail) serves two purposes:
7299 * 1) allows us to batch the cost of updating the user visible
7301 * 2) allows the kernel side to track the head on its own, even
7302 * though the application is the one updating it.
7304 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
7305 if (likely(head
< ctx
->sq_entries
))
7306 return &ctx
->sq_sqes
[head
];
7308 /* drop invalid entries */
7310 WRITE_ONCE(ctx
->rings
->sq_dropped
,
7311 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
7315 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
7316 __must_hold(&ctx
->uring_lock
)
7320 /* make sure SQ entry isn't read before tail */
7321 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
7322 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
7324 io_get_task_refs(nr
);
7326 io_submit_state_start(&ctx
->submit_state
, nr
);
7327 while (submitted
< nr
) {
7328 const struct io_uring_sqe
*sqe
;
7329 struct io_kiocb
*req
;
7331 req
= io_alloc_req(ctx
);
7332 if (unlikely(!req
)) {
7334 submitted
= -EAGAIN
;
7337 sqe
= io_get_sqe(ctx
);
7338 if (unlikely(!sqe
)) {
7339 list_add(&req
->inflight_entry
, &ctx
->submit_state
.free_list
);
7342 /* will complete beyond this point, count as submitted */
7344 if (io_submit_sqe(ctx
, req
, sqe
))
7348 if (unlikely(submitted
!= nr
)) {
7349 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
7350 int unused
= nr
- ref_used
;
7352 current
->io_uring
->cached_refs
+= unused
;
7353 percpu_ref_put_many(&ctx
->refs
, unused
);
7356 io_submit_state_end(&ctx
->submit_state
, ctx
);
7357 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7358 io_commit_sqring(ctx
);
7363 static inline bool io_sqd_events_pending(struct io_sq_data
*sqd
)
7365 return READ_ONCE(sqd
->state
);
7368 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
7370 /* Tell userspace we may need a wakeup call */
7371 spin_lock(&ctx
->completion_lock
);
7372 WRITE_ONCE(ctx
->rings
->sq_flags
,
7373 ctx
->rings
->sq_flags
| IORING_SQ_NEED_WAKEUP
);
7374 spin_unlock(&ctx
->completion_lock
);
7377 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
7379 spin_lock(&ctx
->completion_lock
);
7380 WRITE_ONCE(ctx
->rings
->sq_flags
,
7381 ctx
->rings
->sq_flags
& ~IORING_SQ_NEED_WAKEUP
);
7382 spin_unlock(&ctx
->completion_lock
);
7385 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
7387 unsigned int to_submit
;
7390 to_submit
= io_sqring_entries(ctx
);
7391 /* if we're handling multiple rings, cap submit size for fairness */
7392 if (cap_entries
&& to_submit
> IORING_SQPOLL_CAP_ENTRIES_VALUE
)
7393 to_submit
= IORING_SQPOLL_CAP_ENTRIES_VALUE
;
7395 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
7396 unsigned nr_events
= 0;
7397 const struct cred
*creds
= NULL
;
7399 if (ctx
->sq_creds
!= current_cred())
7400 creds
= override_creds(ctx
->sq_creds
);
7402 mutex_lock(&ctx
->uring_lock
);
7403 if (!list_empty(&ctx
->iopoll_list
))
7404 io_do_iopoll(ctx
, &nr_events
, 0);
7407 * Don't submit if refs are dying, good for io_uring_register(),
7408 * but also it is relied upon by io_ring_exit_work()
7410 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
7411 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
7412 ret
= io_submit_sqes(ctx
, to_submit
);
7413 mutex_unlock(&ctx
->uring_lock
);
7415 if (to_submit
&& wq_has_sleeper(&ctx
->sqo_sq_wait
))
7416 wake_up(&ctx
->sqo_sq_wait
);
7418 revert_creds(creds
);
7424 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7426 struct io_ring_ctx
*ctx
;
7427 unsigned sq_thread_idle
= 0;
7429 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7430 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
7431 sqd
->sq_thread_idle
= sq_thread_idle
;
7434 static bool io_sqd_handle_event(struct io_sq_data
*sqd
)
7436 bool did_sig
= false;
7437 struct ksignal ksig
;
7439 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
7440 signal_pending(current
)) {
7441 mutex_unlock(&sqd
->lock
);
7442 if (signal_pending(current
))
7443 did_sig
= get_signal(&ksig
);
7445 mutex_lock(&sqd
->lock
);
7447 return did_sig
|| test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7450 static int io_sq_thread(void *data
)
7452 struct io_sq_data
*sqd
= data
;
7453 struct io_ring_ctx
*ctx
;
7454 unsigned long timeout
= 0;
7455 char buf
[TASK_COMM_LEN
];
7458 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
7459 set_task_comm(current
, buf
);
7461 if (sqd
->sq_cpu
!= -1)
7462 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
7464 set_cpus_allowed_ptr(current
, cpu_online_mask
);
7465 current
->flags
|= PF_NO_SETAFFINITY
;
7467 mutex_lock(&sqd
->lock
);
7469 bool cap_entries
, sqt_spin
= false;
7471 if (io_sqd_events_pending(sqd
) || signal_pending(current
)) {
7472 if (io_sqd_handle_event(sqd
))
7474 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7477 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7478 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7479 int ret
= __io_sq_thread(ctx
, cap_entries
);
7481 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7484 if (io_run_task_work())
7487 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7490 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7494 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7495 if (!io_sqd_events_pending(sqd
) && !current
->task_works
) {
7496 bool needs_sched
= true;
7498 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7499 io_ring_set_wakeup_flag(ctx
);
7501 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7502 !list_empty_careful(&ctx
->iopoll_list
)) {
7503 needs_sched
= false;
7506 if (io_sqring_entries(ctx
)) {
7507 needs_sched
= false;
7513 mutex_unlock(&sqd
->lock
);
7515 mutex_lock(&sqd
->lock
);
7517 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7518 io_ring_clear_wakeup_flag(ctx
);
7521 finish_wait(&sqd
->wait
, &wait
);
7522 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7525 io_uring_cancel_generic(true, sqd
);
7527 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7528 io_ring_set_wakeup_flag(ctx
);
7530 mutex_unlock(&sqd
->lock
);
7532 complete(&sqd
->exited
);
7536 struct io_wait_queue
{
7537 struct wait_queue_entry wq
;
7538 struct io_ring_ctx
*ctx
;
7540 unsigned nr_timeouts
;
7543 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7545 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7546 int dist
= ctx
->cached_cq_tail
- (int) iowq
->cq_tail
;
7549 * Wake up if we have enough events, or if a timeout occurred since we
7550 * started waiting. For timeouts, we always want to return to userspace,
7551 * regardless of event count.
7553 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7556 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7557 int wake_flags
, void *key
)
7559 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7563 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7564 * the task, and the next invocation will do it.
7566 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->check_cq_overflow
))
7567 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7571 static int io_run_task_work_sig(void)
7573 if (io_run_task_work())
7575 if (!signal_pending(current
))
7577 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
7578 return -ERESTARTSYS
;
7582 /* when returns >0, the caller should retry */
7583 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7584 struct io_wait_queue
*iowq
,
7589 /* make sure we run task_work before checking for signals */
7590 ret
= io_run_task_work_sig();
7591 if (ret
|| io_should_wake(iowq
))
7593 /* let the caller flush overflows, retry */
7594 if (test_bit(0, &ctx
->check_cq_overflow
))
7597 if (!schedule_hrtimeout(&timeout
, HRTIMER_MODE_ABS
))
7603 * Wait until events become available, if we don't already have some. The
7604 * application must reap them itself, as they reside on the shared cq ring.
7606 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7607 const sigset_t __user
*sig
, size_t sigsz
,
7608 struct __kernel_timespec __user
*uts
)
7610 struct io_wait_queue iowq
;
7611 struct io_rings
*rings
= ctx
->rings
;
7612 ktime_t timeout
= KTIME_MAX
;
7616 io_cqring_overflow_flush(ctx
);
7617 if (io_cqring_events(ctx
) >= min_events
)
7619 if (!io_run_task_work())
7624 struct timespec64 ts
;
7626 if (get_timespec64(&ts
, uts
))
7628 timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
7632 #ifdef CONFIG_COMPAT
7633 if (in_compat_syscall())
7634 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7638 ret
= set_user_sigmask(sig
, sigsz
);
7644 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
7645 iowq
.wq
.private = current
;
7646 INIT_LIST_HEAD(&iowq
.wq
.entry
);
7648 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7649 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
7651 trace_io_uring_cqring_wait(ctx
, min_events
);
7653 /* if we can't even flush overflow, don't wait for more */
7654 if (!io_cqring_overflow_flush(ctx
)) {
7658 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
7659 TASK_INTERRUPTIBLE
);
7660 ret
= io_cqring_wait_schedule(ctx
, &iowq
, timeout
);
7661 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
7665 restore_saved_sigmask_unless(ret
== -EINTR
);
7667 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7670 static void io_free_page_table(void **table
, size_t size
)
7672 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7674 for (i
= 0; i
< nr_tables
; i
++)
7679 static void **io_alloc_page_table(size_t size
)
7681 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7682 size_t init_size
= size
;
7685 table
= kcalloc(nr_tables
, sizeof(*table
), GFP_KERNEL_ACCOUNT
);
7689 for (i
= 0; i
< nr_tables
; i
++) {
7690 unsigned int this_size
= min_t(size_t, size
, PAGE_SIZE
);
7692 table
[i
] = kzalloc(this_size
, GFP_KERNEL_ACCOUNT
);
7694 io_free_page_table(table
, init_size
);
7702 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
7704 percpu_ref_exit(&ref_node
->refs
);
7708 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7710 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
7711 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
7712 unsigned long flags
;
7713 bool first_add
= false;
7714 unsigned long delay
= HZ
;
7716 spin_lock_irqsave(&ctx
->rsrc_ref_lock
, flags
);
7719 /* if we are mid-quiesce then do not delay */
7720 if (node
->rsrc_data
->quiesce
)
7723 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7724 node
= list_first_entry(&ctx
->rsrc_ref_list
,
7725 struct io_rsrc_node
, node
);
7726 /* recycle ref nodes in order */
7729 list_del(&node
->node
);
7730 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
7732 spin_unlock_irqrestore(&ctx
->rsrc_ref_lock
, flags
);
7735 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
7738 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
)
7740 struct io_rsrc_node
*ref_node
;
7742 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7746 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7751 INIT_LIST_HEAD(&ref_node
->node
);
7752 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7753 ref_node
->done
= false;
7757 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
7758 struct io_rsrc_data
*data_to_kill
)
7760 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
7761 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
7764 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
7766 rsrc_node
->rsrc_data
= data_to_kill
;
7767 spin_lock_irq(&ctx
->rsrc_ref_lock
);
7768 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
7769 spin_unlock_irq(&ctx
->rsrc_ref_lock
);
7771 atomic_inc(&data_to_kill
->refs
);
7772 percpu_ref_kill(&rsrc_node
->refs
);
7773 ctx
->rsrc_node
= NULL
;
7776 if (!ctx
->rsrc_node
) {
7777 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
7778 ctx
->rsrc_backup_node
= NULL
;
7782 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
7784 if (ctx
->rsrc_backup_node
)
7786 ctx
->rsrc_backup_node
= io_rsrc_node_alloc(ctx
);
7787 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
7790 static int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
, struct io_ring_ctx
*ctx
)
7794 /* As we may drop ->uring_lock, other task may have started quiesce */
7798 data
->quiesce
= true;
7800 ret
= io_rsrc_node_switch_start(ctx
);
7803 io_rsrc_node_switch(ctx
, data
);
7805 /* kill initial ref, already quiesced if zero */
7806 if (atomic_dec_and_test(&data
->refs
))
7808 mutex_unlock(&ctx
->uring_lock
);
7809 flush_delayed_work(&ctx
->rsrc_put_work
);
7810 ret
= wait_for_completion_interruptible(&data
->done
);
7812 mutex_lock(&ctx
->uring_lock
);
7813 if (atomic_read(&data
->refs
) > 0) {
7815 * it has been revived by another thread while
7818 mutex_unlock(&ctx
->uring_lock
);
7824 atomic_inc(&data
->refs
);
7825 /* wait for all works potentially completing data->done */
7826 flush_delayed_work(&ctx
->rsrc_put_work
);
7827 reinit_completion(&data
->done
);
7829 ret
= io_run_task_work_sig();
7830 mutex_lock(&ctx
->uring_lock
);
7832 data
->quiesce
= false;
7837 static u64
*io_get_tag_slot(struct io_rsrc_data
*data
, unsigned int idx
)
7839 unsigned int off
= idx
& IO_RSRC_TAG_TABLE_MASK
;
7840 unsigned int table_idx
= idx
>> IO_RSRC_TAG_TABLE_SHIFT
;
7842 return &data
->tags
[table_idx
][off
];
7845 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
7847 size_t size
= data
->nr
* sizeof(data
->tags
[0][0]);
7850 io_free_page_table((void **)data
->tags
, size
);
7854 static int io_rsrc_data_alloc(struct io_ring_ctx
*ctx
, rsrc_put_fn
*do_put
,
7855 u64 __user
*utags
, unsigned nr
,
7856 struct io_rsrc_data
**pdata
)
7858 struct io_rsrc_data
*data
;
7862 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7865 data
->tags
= (u64
**)io_alloc_page_table(nr
* sizeof(data
->tags
[0][0]));
7873 data
->do_put
= do_put
;
7876 for (i
= 0; i
< nr
; i
++) {
7877 u64
*tag_slot
= io_get_tag_slot(data
, i
);
7879 if (copy_from_user(tag_slot
, &utags
[i
],
7885 atomic_set(&data
->refs
, 1);
7886 init_completion(&data
->done
);
7890 io_rsrc_data_free(data
);
7894 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7896 table
->files
= kvcalloc(nr_files
, sizeof(table
->files
[0]),
7897 GFP_KERNEL_ACCOUNT
);
7898 return !!table
->files
;
7901 static void io_free_file_tables(struct io_file_table
*table
)
7903 kvfree(table
->files
);
7904 table
->files
= NULL
;
7907 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7909 #if defined(CONFIG_UNIX)
7910 if (ctx
->ring_sock
) {
7911 struct sock
*sock
= ctx
->ring_sock
->sk
;
7912 struct sk_buff
*skb
;
7914 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7920 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7923 file
= io_file_from_index(ctx
, i
);
7928 io_free_file_tables(&ctx
->file_table
);
7929 io_rsrc_data_free(ctx
->file_data
);
7930 ctx
->file_data
= NULL
;
7931 ctx
->nr_user_files
= 0;
7934 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7936 unsigned nr
= ctx
->nr_user_files
;
7939 if (!ctx
->file_data
)
7943 * Quiesce may unlock ->uring_lock, and while it's not held
7944 * prevent new requests using the table.
7946 ctx
->nr_user_files
= 0;
7947 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
7948 ctx
->nr_user_files
= nr
;
7950 __io_sqe_files_unregister(ctx
);
7954 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7955 __releases(&sqd
->lock
)
7957 WARN_ON_ONCE(sqd
->thread
== current
);
7960 * Do the dance but not conditional clear_bit() because it'd race with
7961 * other threads incrementing park_pending and setting the bit.
7963 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7964 if (atomic_dec_return(&sqd
->park_pending
))
7965 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7966 mutex_unlock(&sqd
->lock
);
7969 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7970 __acquires(&sqd
->lock
)
7972 WARN_ON_ONCE(sqd
->thread
== current
);
7974 atomic_inc(&sqd
->park_pending
);
7975 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7976 mutex_lock(&sqd
->lock
);
7978 wake_up_process(sqd
->thread
);
7981 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7983 WARN_ON_ONCE(sqd
->thread
== current
);
7984 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
7986 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7987 mutex_lock(&sqd
->lock
);
7989 wake_up_process(sqd
->thread
);
7990 mutex_unlock(&sqd
->lock
);
7991 wait_for_completion(&sqd
->exited
);
7994 static void io_put_sq_data(struct io_sq_data
*sqd
)
7996 if (refcount_dec_and_test(&sqd
->refs
)) {
7997 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
7999 io_sq_thread_stop(sqd
);
8004 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
8006 struct io_sq_data
*sqd
= ctx
->sq_data
;
8009 io_sq_thread_park(sqd
);
8010 list_del_init(&ctx
->sqd_list
);
8011 io_sqd_update_thread_idle(sqd
);
8012 io_sq_thread_unpark(sqd
);
8014 io_put_sq_data(sqd
);
8015 ctx
->sq_data
= NULL
;
8019 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
8021 struct io_ring_ctx
*ctx_attach
;
8022 struct io_sq_data
*sqd
;
8025 f
= fdget(p
->wq_fd
);
8027 return ERR_PTR(-ENXIO
);
8028 if (f
.file
->f_op
!= &io_uring_fops
) {
8030 return ERR_PTR(-EINVAL
);
8033 ctx_attach
= f
.file
->private_data
;
8034 sqd
= ctx_attach
->sq_data
;
8037 return ERR_PTR(-EINVAL
);
8039 if (sqd
->task_tgid
!= current
->tgid
) {
8041 return ERR_PTR(-EPERM
);
8044 refcount_inc(&sqd
->refs
);
8049 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
8052 struct io_sq_data
*sqd
;
8055 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
8056 sqd
= io_attach_sq_data(p
);
8061 /* fall through for EPERM case, setup new sqd/task */
8062 if (PTR_ERR(sqd
) != -EPERM
)
8066 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
8068 return ERR_PTR(-ENOMEM
);
8070 atomic_set(&sqd
->park_pending
, 0);
8071 refcount_set(&sqd
->refs
, 1);
8072 INIT_LIST_HEAD(&sqd
->ctx_list
);
8073 mutex_init(&sqd
->lock
);
8074 init_waitqueue_head(&sqd
->wait
);
8075 init_completion(&sqd
->exited
);
8079 #if defined(CONFIG_UNIX)
8081 * Ensure the UNIX gc is aware of our file set, so we are certain that
8082 * the io_uring can be safely unregistered on process exit, even if we have
8083 * loops in the file referencing.
8085 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
8087 struct sock
*sk
= ctx
->ring_sock
->sk
;
8088 struct scm_fp_list
*fpl
;
8089 struct sk_buff
*skb
;
8092 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
8096 skb
= alloc_skb(0, GFP_KERNEL
);
8105 fpl
->user
= get_uid(current_user());
8106 for (i
= 0; i
< nr
; i
++) {
8107 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8111 fpl
->fp
[nr_files
] = get_file(file
);
8112 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
8117 fpl
->max
= SCM_MAX_FD
;
8118 fpl
->count
= nr_files
;
8119 UNIXCB(skb
).fp
= fpl
;
8120 skb
->destructor
= unix_destruct_scm
;
8121 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
8122 skb_queue_head(&sk
->sk_receive_queue
, skb
);
8124 for (i
= 0; i
< nr
; i
++) {
8125 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8132 free_uid(fpl
->user
);
8140 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8141 * causes regular reference counting to break down. We rely on the UNIX
8142 * garbage collection to take care of this problem for us.
8144 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8146 unsigned left
, total
;
8150 left
= ctx
->nr_user_files
;
8152 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
8154 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
8158 total
+= this_files
;
8164 while (total
< ctx
->nr_user_files
) {
8165 struct file
*file
= io_file_from_index(ctx
, total
);
8175 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8181 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8183 struct file
*file
= prsrc
->file
;
8184 #if defined(CONFIG_UNIX)
8185 struct sock
*sock
= ctx
->ring_sock
->sk
;
8186 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
8187 struct sk_buff
*skb
;
8190 __skb_queue_head_init(&list
);
8193 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8194 * remove this entry and rearrange the file array.
8196 skb
= skb_dequeue(head
);
8198 struct scm_fp_list
*fp
;
8200 fp
= UNIXCB(skb
).fp
;
8201 for (i
= 0; i
< fp
->count
; i
++) {
8204 if (fp
->fp
[i
] != file
)
8207 unix_notinflight(fp
->user
, fp
->fp
[i
]);
8208 left
= fp
->count
- 1 - i
;
8210 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
8211 left
* sizeof(struct file
*));
8218 __skb_queue_tail(&list
, skb
);
8228 __skb_queue_tail(&list
, skb
);
8230 skb
= skb_dequeue(head
);
8233 if (skb_peek(&list
)) {
8234 spin_lock_irq(&head
->lock
);
8235 while ((skb
= __skb_dequeue(&list
)) != NULL
)
8236 __skb_queue_tail(head
, skb
);
8237 spin_unlock_irq(&head
->lock
);
8244 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
8246 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
8247 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
8248 struct io_rsrc_put
*prsrc
, *tmp
;
8250 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
8251 list_del(&prsrc
->list
);
8254 bool lock_ring
= ctx
->flags
& IORING_SETUP_IOPOLL
;
8256 io_ring_submit_lock(ctx
, lock_ring
);
8257 spin_lock(&ctx
->completion_lock
);
8258 io_cqring_fill_event(ctx
, prsrc
->tag
, 0, 0);
8260 io_commit_cqring(ctx
);
8261 spin_unlock(&ctx
->completion_lock
);
8262 io_cqring_ev_posted(ctx
);
8263 io_ring_submit_unlock(ctx
, lock_ring
);
8266 rsrc_data
->do_put(ctx
, prsrc
);
8270 io_rsrc_node_destroy(ref_node
);
8271 if (atomic_dec_and_test(&rsrc_data
->refs
))
8272 complete(&rsrc_data
->done
);
8275 static void io_rsrc_put_work(struct work_struct
*work
)
8277 struct io_ring_ctx
*ctx
;
8278 struct llist_node
*node
;
8280 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
8281 node
= llist_del_all(&ctx
->rsrc_put_llist
);
8284 struct io_rsrc_node
*ref_node
;
8285 struct llist_node
*next
= node
->next
;
8287 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
8288 __io_rsrc_put_work(ref_node
);
8293 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8294 unsigned nr_args
, u64 __user
*tags
)
8296 __s32 __user
*fds
= (__s32 __user
*) arg
;
8305 if (nr_args
> IORING_MAX_FIXED_FILES
)
8307 if (nr_args
> rlimit(RLIMIT_NOFILE
))
8309 ret
= io_rsrc_node_switch_start(ctx
);
8312 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, tags
, nr_args
,
8318 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
))
8321 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
8322 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
8326 /* allow sparse sets */
8329 if (unlikely(*io_get_tag_slot(ctx
->file_data
, i
)))
8336 if (unlikely(!file
))
8340 * Don't allow io_uring instances to be registered. If UNIX
8341 * isn't enabled, then this causes a reference cycle and this
8342 * instance can never get freed. If UNIX is enabled we'll
8343 * handle it just fine, but there's still no point in allowing
8344 * a ring fd as it doesn't support regular read/write anyway.
8346 if (file
->f_op
== &io_uring_fops
) {
8350 io_fixed_file_set(io_fixed_file_slot(&ctx
->file_table
, i
), file
);
8353 ret
= io_sqe_files_scm(ctx
);
8355 __io_sqe_files_unregister(ctx
);
8359 io_rsrc_node_switch(ctx
, NULL
);
8362 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
8363 file
= io_file_from_index(ctx
, i
);
8367 io_free_file_tables(&ctx
->file_table
);
8368 ctx
->nr_user_files
= 0;
8370 io_rsrc_data_free(ctx
->file_data
);
8371 ctx
->file_data
= NULL
;
8375 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
8378 #if defined(CONFIG_UNIX)
8379 struct sock
*sock
= ctx
->ring_sock
->sk
;
8380 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
8381 struct sk_buff
*skb
;
8384 * See if we can merge this file into an existing skb SCM_RIGHTS
8385 * file set. If there's no room, fall back to allocating a new skb
8386 * and filling it in.
8388 spin_lock_irq(&head
->lock
);
8389 skb
= skb_peek(head
);
8391 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
8393 if (fpl
->count
< SCM_MAX_FD
) {
8394 __skb_unlink(skb
, head
);
8395 spin_unlock_irq(&head
->lock
);
8396 fpl
->fp
[fpl
->count
] = get_file(file
);
8397 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
8399 spin_lock_irq(&head
->lock
);
8400 __skb_queue_head(head
, skb
);
8405 spin_unlock_irq(&head
->lock
);
8412 return __io_sqe_files_scm(ctx
, 1, index
);
8418 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
8419 struct io_rsrc_node
*node
, void *rsrc
)
8421 u64
*tag_slot
= io_get_tag_slot(data
, idx
);
8422 struct io_rsrc_put
*prsrc
;
8424 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
8428 prsrc
->tag
= *tag_slot
;
8431 list_add(&prsrc
->list
, &node
->rsrc_list
);
8435 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
8436 unsigned int issue_flags
, u32 slot_index
)
8438 struct io_ring_ctx
*ctx
= req
->ctx
;
8439 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
8440 bool needs_switch
= false;
8441 struct io_fixed_file
*file_slot
;
8444 io_ring_submit_lock(ctx
, !force_nonblock
);
8445 if (file
->f_op
== &io_uring_fops
)
8448 if (!ctx
->file_data
)
8451 if (slot_index
>= ctx
->nr_user_files
)
8454 slot_index
= array_index_nospec(slot_index
, ctx
->nr_user_files
);
8455 file_slot
= io_fixed_file_slot(&ctx
->file_table
, slot_index
);
8457 if (file_slot
->file_ptr
) {
8458 struct file
*old_file
;
8460 ret
= io_rsrc_node_switch_start(ctx
);
8464 old_file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8465 ret
= io_queue_rsrc_removal(ctx
->file_data
, slot_index
,
8466 ctx
->rsrc_node
, old_file
);
8469 file_slot
->file_ptr
= 0;
8470 needs_switch
= true;
8473 *io_get_tag_slot(ctx
->file_data
, slot_index
) = 0;
8474 io_fixed_file_set(file_slot
, file
);
8475 ret
= io_sqe_file_register(ctx
, file
, slot_index
);
8477 file_slot
->file_ptr
= 0;
8484 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8485 io_ring_submit_unlock(ctx
, !force_nonblock
);
8491 static int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
)
8493 unsigned int offset
= req
->close
.file_slot
- 1;
8494 struct io_ring_ctx
*ctx
= req
->ctx
;
8495 struct io_fixed_file
*file_slot
;
8499 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8501 if (unlikely(!ctx
->file_data
))
8504 if (offset
>= ctx
->nr_user_files
)
8506 ret
= io_rsrc_node_switch_start(ctx
);
8510 offset
= array_index_nospec(offset
, ctx
->nr_user_files
);
8511 file_slot
= io_fixed_file_slot(&ctx
->file_table
, offset
);
8513 if (!file_slot
->file_ptr
)
8516 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8517 ret
= io_queue_rsrc_removal(ctx
->file_data
, offset
, ctx
->rsrc_node
, file
);
8521 file_slot
->file_ptr
= 0;
8522 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8525 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8529 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
8530 struct io_uring_rsrc_update2
*up
,
8533 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
8534 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
8535 struct io_rsrc_data
*data
= ctx
->file_data
;
8536 struct io_fixed_file
*file_slot
;
8540 bool needs_switch
= false;
8542 if (!ctx
->file_data
)
8544 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
8547 for (done
= 0; done
< nr_args
; done
++) {
8550 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
8551 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8555 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
8559 if (fd
== IORING_REGISTER_FILES_SKIP
)
8562 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
8563 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
8565 if (file_slot
->file_ptr
) {
8566 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8567 err
= io_queue_rsrc_removal(data
, i
, ctx
->rsrc_node
, file
);
8570 file_slot
->file_ptr
= 0;
8571 needs_switch
= true;
8580 * Don't allow io_uring instances to be registered. If
8581 * UNIX isn't enabled, then this causes a reference
8582 * cycle and this instance can never get freed. If UNIX
8583 * is enabled we'll handle it just fine, but there's
8584 * still no point in allowing a ring fd as it doesn't
8585 * support regular read/write anyway.
8587 if (file
->f_op
== &io_uring_fops
) {
8592 *io_get_tag_slot(data
, i
) = tag
;
8593 io_fixed_file_set(file_slot
, file
);
8594 err
= io_sqe_file_register(ctx
, file
, i
);
8596 file_slot
->file_ptr
= 0;
8604 io_rsrc_node_switch(ctx
, data
);
8605 return done
? done
: err
;
8608 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
8609 struct task_struct
*task
)
8611 struct io_wq_hash
*hash
;
8612 struct io_wq_data data
;
8613 unsigned int concurrency
;
8615 mutex_lock(&ctx
->uring_lock
);
8616 hash
= ctx
->hash_map
;
8618 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
8620 mutex_unlock(&ctx
->uring_lock
);
8621 return ERR_PTR(-ENOMEM
);
8623 refcount_set(&hash
->refs
, 1);
8624 init_waitqueue_head(&hash
->wait
);
8625 ctx
->hash_map
= hash
;
8627 mutex_unlock(&ctx
->uring_lock
);
8631 data
.free_work
= io_wq_free_work
;
8632 data
.do_work
= io_wq_submit_work
;
8634 /* Do QD, or 4 * CPUS, whatever is smallest */
8635 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8637 return io_wq_create(concurrency
, &data
);
8640 static int io_uring_alloc_task_context(struct task_struct
*task
,
8641 struct io_ring_ctx
*ctx
)
8643 struct io_uring_task
*tctx
;
8646 tctx
= kzalloc(sizeof(*tctx
), GFP_KERNEL
);
8647 if (unlikely(!tctx
))
8650 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8651 if (unlikely(ret
)) {
8656 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
8657 if (IS_ERR(tctx
->io_wq
)) {
8658 ret
= PTR_ERR(tctx
->io_wq
);
8659 percpu_counter_destroy(&tctx
->inflight
);
8665 init_waitqueue_head(&tctx
->wait
);
8666 atomic_set(&tctx
->in_idle
, 0);
8667 atomic_set(&tctx
->inflight_tracked
, 0);
8668 task
->io_uring
= tctx
;
8669 spin_lock_init(&tctx
->task_lock
);
8670 INIT_WQ_LIST(&tctx
->task_list
);
8671 init_task_work(&tctx
->task_work
, tctx_task_work
);
8675 void __io_uring_free(struct task_struct
*tsk
)
8677 struct io_uring_task
*tctx
= tsk
->io_uring
;
8679 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8680 WARN_ON_ONCE(tctx
->io_wq
);
8681 WARN_ON_ONCE(tctx
->cached_refs
);
8683 percpu_counter_destroy(&tctx
->inflight
);
8685 tsk
->io_uring
= NULL
;
8688 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8689 struct io_uring_params
*p
)
8693 /* Retain compatibility with failing for an invalid attach attempt */
8694 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
8695 IORING_SETUP_ATTACH_WQ
) {
8698 f
= fdget(p
->wq_fd
);
8701 if (f
.file
->f_op
!= &io_uring_fops
) {
8707 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8708 struct task_struct
*tsk
;
8709 struct io_sq_data
*sqd
;
8712 sqd
= io_get_sq_data(p
, &attached
);
8718 ctx
->sq_creds
= get_current_cred();
8720 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8721 if (!ctx
->sq_thread_idle
)
8722 ctx
->sq_thread_idle
= HZ
;
8724 io_sq_thread_park(sqd
);
8725 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
8726 io_sqd_update_thread_idle(sqd
);
8727 /* don't attach to a dying SQPOLL thread, would be racy */
8728 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
8729 io_sq_thread_unpark(sqd
);
8736 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8737 int cpu
= p
->sq_thread_cpu
;
8740 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
8747 sqd
->task_pid
= current
->pid
;
8748 sqd
->task_tgid
= current
->tgid
;
8749 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
8756 ret
= io_uring_alloc_task_context(tsk
, ctx
);
8757 wake_up_new_task(tsk
);
8760 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8761 /* Can't have SQ_AFF without SQPOLL */
8768 complete(&ctx
->sq_data
->exited
);
8770 io_sq_thread_finish(ctx
);
8774 static inline void __io_unaccount_mem(struct user_struct
*user
,
8775 unsigned long nr_pages
)
8777 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8780 static inline int __io_account_mem(struct user_struct
*user
,
8781 unsigned long nr_pages
)
8783 unsigned long page_limit
, cur_pages
, new_pages
;
8785 /* Don't allow more pages than we can safely lock */
8786 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8789 cur_pages
= atomic_long_read(&user
->locked_vm
);
8790 new_pages
= cur_pages
+ nr_pages
;
8791 if (new_pages
> page_limit
)
8793 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8794 new_pages
) != cur_pages
);
8799 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8802 __io_unaccount_mem(ctx
->user
, nr_pages
);
8804 if (ctx
->mm_account
)
8805 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8808 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8813 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8818 if (ctx
->mm_account
)
8819 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8824 static void io_mem_free(void *ptr
)
8831 page
= virt_to_head_page(ptr
);
8832 if (put_page_testzero(page
))
8833 free_compound_page(page
);
8836 static void *io_mem_alloc(size_t size
)
8838 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
8840 return (void *) __get_free_pages(gfp
, get_order(size
));
8843 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8846 struct io_rings
*rings
;
8847 size_t off
, sq_array_size
;
8849 off
= struct_size(rings
, cqes
, cq_entries
);
8850 if (off
== SIZE_MAX
)
8854 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8862 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8863 if (sq_array_size
== SIZE_MAX
)
8866 if (check_add_overflow(off
, sq_array_size
, &off
))
8872 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
8874 struct io_mapped_ubuf
*imu
= *slot
;
8877 if (imu
!= ctx
->dummy_ubuf
) {
8878 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
8879 unpin_user_page(imu
->bvec
[i
].bv_page
);
8880 if (imu
->acct_pages
)
8881 io_unaccount_mem(ctx
, imu
->acct_pages
);
8887 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8889 io_buffer_unmap(ctx
, &prsrc
->buf
);
8893 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8897 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
8898 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
8899 kfree(ctx
->user_bufs
);
8900 io_rsrc_data_free(ctx
->buf_data
);
8901 ctx
->user_bufs
= NULL
;
8902 ctx
->buf_data
= NULL
;
8903 ctx
->nr_user_bufs
= 0;
8906 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8908 unsigned nr
= ctx
->nr_user_bufs
;
8915 * Quiesce may unlock ->uring_lock, and while it's not held
8916 * prevent new requests using the table.
8918 ctx
->nr_user_bufs
= 0;
8919 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
8920 ctx
->nr_user_bufs
= nr
;
8922 __io_sqe_buffers_unregister(ctx
);
8926 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8927 void __user
*arg
, unsigned index
)
8929 struct iovec __user
*src
;
8931 #ifdef CONFIG_COMPAT
8933 struct compat_iovec __user
*ciovs
;
8934 struct compat_iovec ciov
;
8936 ciovs
= (struct compat_iovec __user
*) arg
;
8937 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8940 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8941 dst
->iov_len
= ciov
.iov_len
;
8945 src
= (struct iovec __user
*) arg
;
8946 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8952 * Not super efficient, but this is just a registration time. And we do cache
8953 * the last compound head, so generally we'll only do a full search if we don't
8956 * We check if the given compound head page has already been accounted, to
8957 * avoid double accounting it. This allows us to account the full size of the
8958 * page, not just the constituent pages of a huge page.
8960 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8961 int nr_pages
, struct page
*hpage
)
8965 /* check current page array */
8966 for (i
= 0; i
< nr_pages
; i
++) {
8967 if (!PageCompound(pages
[i
]))
8969 if (compound_head(pages
[i
]) == hpage
)
8973 /* check previously registered pages */
8974 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8975 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
8977 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8978 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8980 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8988 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8989 int nr_pages
, struct io_mapped_ubuf
*imu
,
8990 struct page
**last_hpage
)
8994 imu
->acct_pages
= 0;
8995 for (i
= 0; i
< nr_pages
; i
++) {
8996 if (!PageCompound(pages
[i
])) {
9001 hpage
= compound_head(pages
[i
]);
9002 if (hpage
== *last_hpage
)
9004 *last_hpage
= hpage
;
9005 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
9007 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
9011 if (!imu
->acct_pages
)
9014 ret
= io_account_mem(ctx
, imu
->acct_pages
);
9016 imu
->acct_pages
= 0;
9020 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
9021 struct io_mapped_ubuf
**pimu
,
9022 struct page
**last_hpage
)
9024 struct io_mapped_ubuf
*imu
= NULL
;
9025 struct vm_area_struct
**vmas
= NULL
;
9026 struct page
**pages
= NULL
;
9027 unsigned long off
, start
, end
, ubuf
;
9029 int ret
, pret
, nr_pages
, i
;
9031 if (!iov
->iov_base
) {
9032 *pimu
= ctx
->dummy_ubuf
;
9036 ubuf
= (unsigned long) iov
->iov_base
;
9037 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
9038 start
= ubuf
>> PAGE_SHIFT
;
9039 nr_pages
= end
- start
;
9044 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
9048 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
9053 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
9058 mmap_read_lock(current
->mm
);
9059 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
9061 if (pret
== nr_pages
) {
9062 /* don't support file backed memory */
9063 for (i
= 0; i
< nr_pages
; i
++) {
9064 struct vm_area_struct
*vma
= vmas
[i
];
9066 if (vma_is_shmem(vma
))
9069 !is_file_hugepages(vma
->vm_file
)) {
9075 ret
= pret
< 0 ? pret
: -EFAULT
;
9077 mmap_read_unlock(current
->mm
);
9080 * if we did partial map, or found file backed vmas,
9081 * release any pages we did get
9084 unpin_user_pages(pages
, pret
);
9088 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
9090 unpin_user_pages(pages
, pret
);
9094 off
= ubuf
& ~PAGE_MASK
;
9095 size
= iov
->iov_len
;
9096 for (i
= 0; i
< nr_pages
; i
++) {
9099 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
9100 imu
->bvec
[i
].bv_page
= pages
[i
];
9101 imu
->bvec
[i
].bv_len
= vec_len
;
9102 imu
->bvec
[i
].bv_offset
= off
;
9106 /* store original address for later verification */
9108 imu
->ubuf_end
= ubuf
+ iov
->iov_len
;
9109 imu
->nr_bvecs
= nr_pages
;
9120 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
9122 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
9123 return ctx
->user_bufs
? 0 : -ENOMEM
;
9126 static int io_buffer_validate(struct iovec
*iov
)
9128 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
9131 * Don't impose further limits on the size and buffer
9132 * constraints here, we'll -EINVAL later when IO is
9133 * submitted if they are wrong.
9136 return iov
->iov_len
? -EFAULT
: 0;
9140 /* arbitrary limit, but we need something */
9141 if (iov
->iov_len
> SZ_1G
)
9144 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
9150 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
9151 unsigned int nr_args
, u64 __user
*tags
)
9153 struct page
*last_hpage
= NULL
;
9154 struct io_rsrc_data
*data
;
9160 if (!nr_args
|| nr_args
> IORING_MAX_REG_BUFFERS
)
9162 ret
= io_rsrc_node_switch_start(ctx
);
9165 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, tags
, nr_args
, &data
);
9168 ret
= io_buffers_map_alloc(ctx
, nr_args
);
9170 io_rsrc_data_free(data
);
9174 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
9175 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
9178 ret
= io_buffer_validate(&iov
);
9181 if (!iov
.iov_base
&& *io_get_tag_slot(data
, i
)) {
9186 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
9192 WARN_ON_ONCE(ctx
->buf_data
);
9194 ctx
->buf_data
= data
;
9196 __io_sqe_buffers_unregister(ctx
);
9198 io_rsrc_node_switch(ctx
, NULL
);
9202 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
9203 struct io_uring_rsrc_update2
*up
,
9204 unsigned int nr_args
)
9206 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
9207 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
9208 struct page
*last_hpage
= NULL
;
9209 bool needs_switch
= false;
9215 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
9218 for (done
= 0; done
< nr_args
; done
++) {
9219 struct io_mapped_ubuf
*imu
;
9220 int offset
= up
->offset
+ done
;
9223 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
9226 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
9230 err
= io_buffer_validate(&iov
);
9233 if (!iov
.iov_base
&& tag
) {
9237 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
9241 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
9242 if (ctx
->user_bufs
[i
] != ctx
->dummy_ubuf
) {
9243 err
= io_queue_rsrc_removal(ctx
->buf_data
, i
,
9244 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
9245 if (unlikely(err
)) {
9246 io_buffer_unmap(ctx
, &imu
);
9249 ctx
->user_bufs
[i
] = NULL
;
9250 needs_switch
= true;
9253 ctx
->user_bufs
[i
] = imu
;
9254 *io_get_tag_slot(ctx
->buf_data
, offset
) = tag
;
9258 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
9259 return done
? done
: err
;
9262 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
9264 __s32 __user
*fds
= arg
;
9270 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
9273 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
9274 if (IS_ERR(ctx
->cq_ev_fd
)) {
9275 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
9277 ctx
->cq_ev_fd
= NULL
;
9284 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
9286 if (ctx
->cq_ev_fd
) {
9287 eventfd_ctx_put(ctx
->cq_ev_fd
);
9288 ctx
->cq_ev_fd
= NULL
;
9295 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
9297 struct io_buffer
*buf
;
9298 unsigned long index
;
9300 xa_for_each(&ctx
->io_buffers
, index
, buf
)
9301 __io_remove_buffers(ctx
, buf
, index
, -1U);
9304 static void io_req_cache_free(struct list_head
*list
)
9306 struct io_kiocb
*req
, *nxt
;
9308 list_for_each_entry_safe(req
, nxt
, list
, inflight_entry
) {
9309 list_del(&req
->inflight_entry
);
9310 kmem_cache_free(req_cachep
, req
);
9314 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
9316 struct io_submit_state
*state
= &ctx
->submit_state
;
9318 mutex_lock(&ctx
->uring_lock
);
9320 if (state
->free_reqs
) {
9321 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
9322 state
->free_reqs
= 0;
9325 io_flush_cached_locked_reqs(ctx
, state
);
9326 io_req_cache_free(&state
->free_list
);
9327 mutex_unlock(&ctx
->uring_lock
);
9330 static void io_wait_rsrc_data(struct io_rsrc_data
*data
)
9332 if (data
&& !atomic_dec_and_test(&data
->refs
))
9333 wait_for_completion(&data
->done
);
9336 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
9338 io_sq_thread_finish(ctx
);
9340 if (ctx
->mm_account
) {
9341 mmdrop(ctx
->mm_account
);
9342 ctx
->mm_account
= NULL
;
9345 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9346 io_wait_rsrc_data(ctx
->buf_data
);
9347 io_wait_rsrc_data(ctx
->file_data
);
9349 mutex_lock(&ctx
->uring_lock
);
9351 __io_sqe_buffers_unregister(ctx
);
9353 __io_sqe_files_unregister(ctx
);
9355 __io_cqring_overflow_flush(ctx
, true);
9356 mutex_unlock(&ctx
->uring_lock
);
9357 io_eventfd_unregister(ctx
);
9358 io_destroy_buffers(ctx
);
9360 put_cred(ctx
->sq_creds
);
9362 /* there are no registered resources left, nobody uses it */
9364 io_rsrc_node_destroy(ctx
->rsrc_node
);
9365 if (ctx
->rsrc_backup_node
)
9366 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
9367 flush_delayed_work(&ctx
->rsrc_put_work
);
9369 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
9370 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
9372 #if defined(CONFIG_UNIX)
9373 if (ctx
->ring_sock
) {
9374 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
9375 sock_release(ctx
->ring_sock
);
9378 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
9380 io_mem_free(ctx
->rings
);
9381 io_mem_free(ctx
->sq_sqes
);
9383 percpu_ref_exit(&ctx
->refs
);
9384 free_uid(ctx
->user
);
9385 io_req_caches_free(ctx
);
9387 io_wq_put_hash(ctx
->hash_map
);
9388 kfree(ctx
->cancel_hash
);
9389 kfree(ctx
->dummy_ubuf
);
9393 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
9395 struct io_ring_ctx
*ctx
= file
->private_data
;
9398 poll_wait(file
, &ctx
->poll_wait
, wait
);
9400 * synchronizes with barrier from wq_has_sleeper call in
9404 if (!io_sqring_full(ctx
))
9405 mask
|= EPOLLOUT
| EPOLLWRNORM
;
9408 * Don't flush cqring overflow list here, just do a simple check.
9409 * Otherwise there could possible be ABBA deadlock:
9412 * lock(&ctx->uring_lock);
9414 * lock(&ctx->uring_lock);
9417 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9418 * pushs them to do the flush.
9420 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->check_cq_overflow
))
9421 mask
|= EPOLLIN
| EPOLLRDNORM
;
9426 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9428 const struct cred
*creds
;
9430 creds
= xa_erase(&ctx
->personalities
, id
);
9439 struct io_tctx_exit
{
9440 struct callback_head task_work
;
9441 struct completion completion
;
9442 struct io_ring_ctx
*ctx
;
9445 static void io_tctx_exit_cb(struct callback_head
*cb
)
9447 struct io_uring_task
*tctx
= current
->io_uring
;
9448 struct io_tctx_exit
*work
;
9450 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
9452 * When @in_idle, we're in cancellation and it's racy to remove the
9453 * node. It'll be removed by the end of cancellation, just ignore it.
9455 if (!atomic_read(&tctx
->in_idle
))
9456 io_uring_del_tctx_node((unsigned long)work
->ctx
);
9457 complete(&work
->completion
);
9460 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
9462 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9464 return req
->ctx
== data
;
9467 static void io_ring_exit_work(struct work_struct
*work
)
9469 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
9470 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
9471 unsigned long interval
= HZ
/ 20;
9472 struct io_tctx_exit exit
;
9473 struct io_tctx_node
*node
;
9477 * If we're doing polled IO and end up having requests being
9478 * submitted async (out-of-line), then completions can come in while
9479 * we're waiting for refs to drop. We need to reap these manually,
9480 * as nobody else will be looking for them.
9483 io_uring_try_cancel_requests(ctx
, NULL
, true);
9485 struct io_sq_data
*sqd
= ctx
->sq_data
;
9486 struct task_struct
*tsk
;
9488 io_sq_thread_park(sqd
);
9490 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
9491 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
9492 io_cancel_ctx_cb
, ctx
, true);
9493 io_sq_thread_unpark(sqd
);
9496 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
9497 /* there is little hope left, don't run it too often */
9500 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
9502 init_completion(&exit
.completion
);
9503 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
9506 * Some may use context even when all refs and requests have been put,
9507 * and they are free to do so while still holding uring_lock or
9508 * completion_lock, see io_req_task_submit(). Apart from other work,
9509 * this lock/unlock section also waits them to finish.
9511 mutex_lock(&ctx
->uring_lock
);
9512 while (!list_empty(&ctx
->tctx_list
)) {
9513 WARN_ON_ONCE(time_after(jiffies
, timeout
));
9515 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
9517 /* don't spin on a single task if cancellation failed */
9518 list_rotate_left(&ctx
->tctx_list
);
9519 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
9520 if (WARN_ON_ONCE(ret
))
9522 wake_up_process(node
->task
);
9524 mutex_unlock(&ctx
->uring_lock
);
9525 wait_for_completion(&exit
.completion
);
9526 mutex_lock(&ctx
->uring_lock
);
9528 mutex_unlock(&ctx
->uring_lock
);
9529 spin_lock(&ctx
->completion_lock
);
9530 spin_unlock(&ctx
->completion_lock
);
9532 io_ring_ctx_free(ctx
);
9535 /* Returns true if we found and killed one or more timeouts */
9536 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
9539 struct io_kiocb
*req
, *tmp
;
9542 spin_lock(&ctx
->completion_lock
);
9543 spin_lock_irq(&ctx
->timeout_lock
);
9544 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
9545 if (io_match_task(req
, tsk
, cancel_all
)) {
9546 io_kill_timeout(req
, -ECANCELED
);
9550 spin_unlock_irq(&ctx
->timeout_lock
);
9552 io_commit_cqring(ctx
);
9553 spin_unlock(&ctx
->completion_lock
);
9555 io_cqring_ev_posted(ctx
);
9556 return canceled
!= 0;
9559 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
9561 unsigned long index
;
9562 struct creds
*creds
;
9564 mutex_lock(&ctx
->uring_lock
);
9565 percpu_ref_kill(&ctx
->refs
);
9567 __io_cqring_overflow_flush(ctx
, true);
9568 xa_for_each(&ctx
->personalities
, index
, creds
)
9569 io_unregister_personality(ctx
, index
);
9570 mutex_unlock(&ctx
->uring_lock
);
9572 io_kill_timeouts(ctx
, NULL
, true);
9573 io_poll_remove_all(ctx
, NULL
, true);
9575 /* if we failed setting up the ctx, we might not have any rings */
9576 io_iopoll_try_reap_events(ctx
);
9578 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
9580 * Use system_unbound_wq to avoid spawning tons of event kworkers
9581 * if we're exiting a ton of rings at the same time. It just adds
9582 * noise and overhead, there's no discernable change in runtime
9583 * over using system_wq.
9585 queue_work(system_unbound_wq
, &ctx
->exit_work
);
9588 static int io_uring_release(struct inode
*inode
, struct file
*file
)
9590 struct io_ring_ctx
*ctx
= file
->private_data
;
9592 file
->private_data
= NULL
;
9593 io_ring_ctx_wait_and_kill(ctx
);
9597 struct io_task_cancel
{
9598 struct task_struct
*task
;
9602 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
9604 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9605 struct io_task_cancel
*cancel
= data
;
9607 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
9610 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
9611 struct task_struct
*task
, bool cancel_all
)
9613 struct io_defer_entry
*de
;
9616 spin_lock(&ctx
->completion_lock
);
9617 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
9618 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
9619 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
9623 spin_unlock(&ctx
->completion_lock
);
9624 if (list_empty(&list
))
9627 while (!list_empty(&list
)) {
9628 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
9629 list_del_init(&de
->list
);
9630 io_req_complete_failed(de
->req
, -ECANCELED
);
9636 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
9638 struct io_tctx_node
*node
;
9639 enum io_wq_cancel cret
;
9642 mutex_lock(&ctx
->uring_lock
);
9643 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
9644 struct io_uring_task
*tctx
= node
->task
->io_uring
;
9647 * io_wq will stay alive while we hold uring_lock, because it's
9648 * killed after ctx nodes, which requires to take the lock.
9650 if (!tctx
|| !tctx
->io_wq
)
9652 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
9653 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9655 mutex_unlock(&ctx
->uring_lock
);
9660 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
9661 struct task_struct
*task
,
9664 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
9665 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
9668 enum io_wq_cancel cret
;
9672 ret
|= io_uring_try_cancel_iowq(ctx
);
9673 } else if (tctx
&& tctx
->io_wq
) {
9675 * Cancels requests of all rings, not only @ctx, but
9676 * it's fine as the task is in exit/exec.
9678 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
9680 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9683 /* SQPOLL thread does its own polling */
9684 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
9685 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
9686 while (!list_empty_careful(&ctx
->iopoll_list
)) {
9687 io_iopoll_try_reap_events(ctx
);
9692 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
9693 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
9694 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
9696 ret
|= io_run_task_work();
9703 static int __io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9705 struct io_uring_task
*tctx
= current
->io_uring
;
9706 struct io_tctx_node
*node
;
9709 if (unlikely(!tctx
)) {
9710 ret
= io_uring_alloc_task_context(current
, ctx
);
9714 tctx
= current
->io_uring
;
9715 if (ctx
->iowq_limits_set
) {
9716 unsigned int limits
[2] = { ctx
->iowq_limits
[0],
9717 ctx
->iowq_limits
[1], };
9719 ret
= io_wq_max_workers(tctx
->io_wq
, limits
);
9724 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
9725 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
9729 node
->task
= current
;
9731 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
9738 mutex_lock(&ctx
->uring_lock
);
9739 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
9740 mutex_unlock(&ctx
->uring_lock
);
9747 * Note that this task has used io_uring. We use it for cancelation purposes.
9749 static inline int io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9751 struct io_uring_task
*tctx
= current
->io_uring
;
9753 if (likely(tctx
&& tctx
->last
== ctx
))
9755 return __io_uring_add_tctx_node(ctx
);
9759 * Remove this io_uring_file -> task mapping.
9761 static void io_uring_del_tctx_node(unsigned long index
)
9763 struct io_uring_task
*tctx
= current
->io_uring
;
9764 struct io_tctx_node
*node
;
9768 node
= xa_erase(&tctx
->xa
, index
);
9772 WARN_ON_ONCE(current
!= node
->task
);
9773 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
9775 mutex_lock(&node
->ctx
->uring_lock
);
9776 list_del(&node
->ctx_node
);
9777 mutex_unlock(&node
->ctx
->uring_lock
);
9779 if (tctx
->last
== node
->ctx
)
9784 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
9786 struct io_wq
*wq
= tctx
->io_wq
;
9787 struct io_tctx_node
*node
;
9788 unsigned long index
;
9790 xa_for_each(&tctx
->xa
, index
, node
) {
9791 io_uring_del_tctx_node(index
);
9796 * Must be after io_uring_del_task_file() (removes nodes under
9797 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9799 io_wq_put_and_exit(wq
);
9804 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
9807 return atomic_read(&tctx
->inflight_tracked
);
9808 return percpu_counter_sum(&tctx
->inflight
);
9812 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9813 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9815 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
9817 struct io_uring_task
*tctx
= current
->io_uring
;
9818 struct io_ring_ctx
*ctx
;
9822 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
9824 if (!current
->io_uring
)
9827 io_wq_exit_start(tctx
->io_wq
);
9829 atomic_inc(&tctx
->in_idle
);
9831 io_uring_drop_tctx_refs(current
);
9832 /* read completions before cancelations */
9833 inflight
= tctx_inflight(tctx
, !cancel_all
);
9838 struct io_tctx_node
*node
;
9839 unsigned long index
;
9841 xa_for_each(&tctx
->xa
, index
, node
) {
9842 /* sqpoll task will cancel all its requests */
9843 if (node
->ctx
->sq_data
)
9845 io_uring_try_cancel_requests(node
->ctx
, current
,
9849 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9850 io_uring_try_cancel_requests(ctx
, current
,
9854 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
9856 io_uring_drop_tctx_refs(current
);
9859 * If we've seen completions, retry without waiting. This
9860 * avoids a race where a completion comes in before we did
9861 * prepare_to_wait().
9863 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
9865 finish_wait(&tctx
->wait
, &wait
);
9868 io_uring_clean_tctx(tctx
);
9871 * We shouldn't run task_works after cancel, so just leave
9872 * ->in_idle set for normal exit.
9874 atomic_dec(&tctx
->in_idle
);
9875 /* for exec all current's requests should be gone, kill tctx */
9876 __io_uring_free(current
);
9880 void __io_uring_cancel(bool cancel_all
)
9882 io_uring_cancel_generic(cancel_all
, NULL
);
9885 static void *io_uring_validate_mmap_request(struct file
*file
,
9886 loff_t pgoff
, size_t sz
)
9888 struct io_ring_ctx
*ctx
= file
->private_data
;
9889 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9894 case IORING_OFF_SQ_RING
:
9895 case IORING_OFF_CQ_RING
:
9898 case IORING_OFF_SQES
:
9902 return ERR_PTR(-EINVAL
);
9905 page
= virt_to_head_page(ptr
);
9906 if (sz
> page_size(page
))
9907 return ERR_PTR(-EINVAL
);
9914 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9916 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9920 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9922 return PTR_ERR(ptr
);
9924 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9925 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9928 #else /* !CONFIG_MMU */
9930 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9932 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9935 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9937 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9940 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9941 unsigned long addr
, unsigned long len
,
9942 unsigned long pgoff
, unsigned long flags
)
9946 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9948 return PTR_ERR(ptr
);
9950 return (unsigned long) ptr
;
9953 #endif /* !CONFIG_MMU */
9955 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9960 if (!io_sqring_full(ctx
))
9962 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9964 if (!io_sqring_full(ctx
))
9967 } while (!signal_pending(current
));
9969 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9973 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9974 struct __kernel_timespec __user
**ts
,
9975 const sigset_t __user
**sig
)
9977 struct io_uring_getevents_arg arg
;
9980 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9981 * is just a pointer to the sigset_t.
9983 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9984 *sig
= (const sigset_t __user
*) argp
;
9990 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9991 * timespec and sigset_t pointers if good.
9993 if (*argsz
!= sizeof(arg
))
9995 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9999 *sig
= u64_to_user_ptr(arg
.sigmask
);
10000 *argsz
= arg
.sigmask_sz
;
10001 *ts
= u64_to_user_ptr(arg
.ts
);
10005 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
10006 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
10009 struct io_ring_ctx
*ctx
;
10014 io_run_task_work();
10016 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
10017 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
)))
10021 if (unlikely(!f
.file
))
10025 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
10029 ctx
= f
.file
->private_data
;
10030 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
10034 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10038 * For SQ polling, the thread will do all submissions and completions.
10039 * Just return the requested submit count, and wake the thread if
10040 * we were asked to.
10043 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10044 io_cqring_overflow_flush(ctx
);
10046 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
10050 if (flags
& IORING_ENTER_SQ_WAKEUP
)
10051 wake_up(&ctx
->sq_data
->wait
);
10052 if (flags
& IORING_ENTER_SQ_WAIT
) {
10053 ret
= io_sqpoll_wait_sq(ctx
);
10057 submitted
= to_submit
;
10058 } else if (to_submit
) {
10059 ret
= io_uring_add_tctx_node(ctx
);
10062 mutex_lock(&ctx
->uring_lock
);
10063 submitted
= io_submit_sqes(ctx
, to_submit
);
10064 mutex_unlock(&ctx
->uring_lock
);
10066 if (submitted
!= to_submit
)
10069 if (flags
& IORING_ENTER_GETEVENTS
) {
10070 const sigset_t __user
*sig
;
10071 struct __kernel_timespec __user
*ts
;
10073 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
10077 min_complete
= min(min_complete
, ctx
->cq_entries
);
10080 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10081 * space applications don't need to do io completion events
10082 * polling again, they can rely on io_sq_thread to do polling
10083 * work, which can reduce cpu usage and uring_lock contention.
10085 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
10086 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10087 ret
= io_iopoll_check(ctx
, min_complete
);
10089 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
10094 percpu_ref_put(&ctx
->refs
);
10097 return submitted
? submitted
: ret
;
10100 #ifdef CONFIG_PROC_FS
10101 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
10102 const struct cred
*cred
)
10104 struct user_namespace
*uns
= seq_user_ns(m
);
10105 struct group_info
*gi
;
10110 seq_printf(m
, "%5d\n", id
);
10111 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
10112 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
10113 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
10114 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
10115 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
10116 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
10117 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
10118 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
10119 seq_puts(m
, "\n\tGroups:\t");
10120 gi
= cred
->group_info
;
10121 for (g
= 0; g
< gi
->ngroups
; g
++) {
10122 seq_put_decimal_ull(m
, g
? " " : "",
10123 from_kgid_munged(uns
, gi
->gid
[g
]));
10125 seq_puts(m
, "\n\tCapEff:\t");
10126 cap
= cred
->cap_effective
;
10127 CAP_FOR_EACH_U32(__capi
)
10128 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
10133 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
10135 struct io_sq_data
*sq
= NULL
;
10140 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10141 * since fdinfo case grabs it in the opposite direction of normal use
10142 * cases. If we fail to get the lock, we just don't iterate any
10143 * structures that could be going away outside the io_uring mutex.
10145 has_lock
= mutex_trylock(&ctx
->uring_lock
);
10147 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10153 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
10154 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
10155 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
10156 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
10157 struct file
*f
= io_file_from_index(ctx
, i
);
10160 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
10162 seq_printf(m
, "%5u: <none>\n", i
);
10164 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
10165 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
10166 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
10167 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
10169 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
10171 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
10172 unsigned long index
;
10173 const struct cred
*cred
;
10175 seq_printf(m
, "Personalities:\n");
10176 xa_for_each(&ctx
->personalities
, index
, cred
)
10177 io_uring_show_cred(m
, index
, cred
);
10179 seq_printf(m
, "PollList:\n");
10180 spin_lock(&ctx
->completion_lock
);
10181 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
10182 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
10183 struct io_kiocb
*req
;
10185 hlist_for_each_entry(req
, list
, hash_node
)
10186 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
10187 req
->task
->task_works
!= NULL
);
10189 spin_unlock(&ctx
->completion_lock
);
10191 mutex_unlock(&ctx
->uring_lock
);
10194 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
10196 struct io_ring_ctx
*ctx
= f
->private_data
;
10198 if (percpu_ref_tryget(&ctx
->refs
)) {
10199 __io_uring_show_fdinfo(ctx
, m
);
10200 percpu_ref_put(&ctx
->refs
);
10205 static const struct file_operations io_uring_fops
= {
10206 .release
= io_uring_release
,
10207 .mmap
= io_uring_mmap
,
10209 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
10210 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
10212 .poll
= io_uring_poll
,
10213 #ifdef CONFIG_PROC_FS
10214 .show_fdinfo
= io_uring_show_fdinfo
,
10218 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
10219 struct io_uring_params
*p
)
10221 struct io_rings
*rings
;
10222 size_t size
, sq_array_offset
;
10224 /* make sure these are sane, as we already accounted them */
10225 ctx
->sq_entries
= p
->sq_entries
;
10226 ctx
->cq_entries
= p
->cq_entries
;
10228 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
10229 if (size
== SIZE_MAX
)
10232 rings
= io_mem_alloc(size
);
10236 ctx
->rings
= rings
;
10237 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
10238 rings
->sq_ring_mask
= p
->sq_entries
- 1;
10239 rings
->cq_ring_mask
= p
->cq_entries
- 1;
10240 rings
->sq_ring_entries
= p
->sq_entries
;
10241 rings
->cq_ring_entries
= p
->cq_entries
;
10243 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
10244 if (size
== SIZE_MAX
) {
10245 io_mem_free(ctx
->rings
);
10250 ctx
->sq_sqes
= io_mem_alloc(size
);
10251 if (!ctx
->sq_sqes
) {
10252 io_mem_free(ctx
->rings
);
10260 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
10264 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
10268 ret
= io_uring_add_tctx_node(ctx
);
10273 fd_install(fd
, file
);
10278 * Allocate an anonymous fd, this is what constitutes the application
10279 * visible backing of an io_uring instance. The application mmaps this
10280 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10281 * we have to tie this fd to a socket for file garbage collection purposes.
10283 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
10286 #if defined(CONFIG_UNIX)
10289 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
10292 return ERR_PTR(ret
);
10295 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
10296 O_RDWR
| O_CLOEXEC
);
10297 #if defined(CONFIG_UNIX)
10298 if (IS_ERR(file
)) {
10299 sock_release(ctx
->ring_sock
);
10300 ctx
->ring_sock
= NULL
;
10302 ctx
->ring_sock
->file
= file
;
10308 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
10309 struct io_uring_params __user
*params
)
10311 struct io_ring_ctx
*ctx
;
10317 if (entries
> IORING_MAX_ENTRIES
) {
10318 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10320 entries
= IORING_MAX_ENTRIES
;
10324 * Use twice as many entries for the CQ ring. It's possible for the
10325 * application to drive a higher depth than the size of the SQ ring,
10326 * since the sqes are only used at submission time. This allows for
10327 * some flexibility in overcommitting a bit. If the application has
10328 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10329 * of CQ ring entries manually.
10331 p
->sq_entries
= roundup_pow_of_two(entries
);
10332 if (p
->flags
& IORING_SETUP_CQSIZE
) {
10334 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10335 * to a power-of-two, if it isn't already. We do NOT impose
10336 * any cq vs sq ring sizing.
10338 if (!p
->cq_entries
)
10340 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
10341 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10343 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
10345 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
10346 if (p
->cq_entries
< p
->sq_entries
)
10349 p
->cq_entries
= 2 * p
->sq_entries
;
10352 ctx
= io_ring_ctx_alloc(p
);
10355 ctx
->compat
= in_compat_syscall();
10356 if (!capable(CAP_IPC_LOCK
))
10357 ctx
->user
= get_uid(current_user());
10360 * This is just grabbed for accounting purposes. When a process exits,
10361 * the mm is exited and dropped before the files, hence we need to hang
10362 * on to this mm purely for the purposes of being able to unaccount
10363 * memory (locked/pinned vm). It's not used for anything else.
10365 mmgrab(current
->mm
);
10366 ctx
->mm_account
= current
->mm
;
10368 ret
= io_allocate_scq_urings(ctx
, p
);
10372 ret
= io_sq_offload_create(ctx
, p
);
10375 /* always set a rsrc node */
10376 ret
= io_rsrc_node_switch_start(ctx
);
10379 io_rsrc_node_switch(ctx
, NULL
);
10381 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
10382 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
10383 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
10384 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
10385 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
10386 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
10387 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
10388 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
10390 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
10391 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
10392 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
10393 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
10394 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
10395 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
10396 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
10397 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
10399 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
10400 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
10401 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
10402 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
10403 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
10404 IORING_FEAT_RSRC_TAGS
;
10406 if (copy_to_user(params
, p
, sizeof(*p
))) {
10411 file
= io_uring_get_file(ctx
);
10412 if (IS_ERR(file
)) {
10413 ret
= PTR_ERR(file
);
10418 * Install ring fd as the very last thing, so we don't risk someone
10419 * having closed it before we finish setup
10421 ret
= io_uring_install_fd(ctx
, file
);
10423 /* fput will clean it up */
10428 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
10431 io_ring_ctx_wait_and_kill(ctx
);
10436 * Sets up an aio uring context, and returns the fd. Applications asks for a
10437 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10438 * params structure passed in.
10440 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
10442 struct io_uring_params p
;
10445 if (copy_from_user(&p
, params
, sizeof(p
)))
10447 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
10452 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
10453 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
10454 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
10455 IORING_SETUP_R_DISABLED
))
10458 return io_uring_create(entries
, &p
, params
);
10461 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
10462 struct io_uring_params __user
*, params
)
10464 return io_uring_setup(entries
, params
);
10467 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
10469 struct io_uring_probe
*p
;
10473 size
= struct_size(p
, ops
, nr_args
);
10474 if (size
== SIZE_MAX
)
10476 p
= kzalloc(size
, GFP_KERNEL
);
10481 if (copy_from_user(p
, arg
, size
))
10484 if (memchr_inv(p
, 0, size
))
10487 p
->last_op
= IORING_OP_LAST
- 1;
10488 if (nr_args
> IORING_OP_LAST
)
10489 nr_args
= IORING_OP_LAST
;
10491 for (i
= 0; i
< nr_args
; i
++) {
10493 if (!io_op_defs
[i
].not_supported
)
10494 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
10499 if (copy_to_user(arg
, p
, size
))
10506 static int io_register_personality(struct io_ring_ctx
*ctx
)
10508 const struct cred
*creds
;
10512 creds
= get_current_cred();
10514 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
10515 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
10523 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
10524 unsigned int nr_args
)
10526 struct io_uring_restriction
*res
;
10530 /* Restrictions allowed only if rings started disabled */
10531 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10534 /* We allow only a single restrictions registration */
10535 if (ctx
->restrictions
.registered
)
10538 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
10541 size
= array_size(nr_args
, sizeof(*res
));
10542 if (size
== SIZE_MAX
)
10545 res
= memdup_user(arg
, size
);
10547 return PTR_ERR(res
);
10551 for (i
= 0; i
< nr_args
; i
++) {
10552 switch (res
[i
].opcode
) {
10553 case IORING_RESTRICTION_REGISTER_OP
:
10554 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
10559 __set_bit(res
[i
].register_op
,
10560 ctx
->restrictions
.register_op
);
10562 case IORING_RESTRICTION_SQE_OP
:
10563 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
10568 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
10570 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
10571 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
10573 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
10574 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
10583 /* Reset all restrictions if an error happened */
10585 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
10587 ctx
->restrictions
.registered
= true;
10593 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10595 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10598 if (ctx
->restrictions
.registered
)
10599 ctx
->restricted
= 1;
10601 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
10602 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
10603 wake_up(&ctx
->sq_data
->wait
);
10607 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
10608 struct io_uring_rsrc_update2
*up
,
10614 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
10616 err
= io_rsrc_node_switch_start(ctx
);
10621 case IORING_RSRC_FILE
:
10622 return __io_sqe_files_update(ctx
, up
, nr_args
);
10623 case IORING_RSRC_BUFFER
:
10624 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
10629 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10632 struct io_uring_rsrc_update2 up
;
10636 memset(&up
, 0, sizeof(up
));
10637 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
10639 if (up
.resv
|| up
.resv2
)
10641 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
10644 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10645 unsigned size
, unsigned type
)
10647 struct io_uring_rsrc_update2 up
;
10649 if (size
!= sizeof(up
))
10651 if (copy_from_user(&up
, arg
, sizeof(up
)))
10653 if (!up
.nr
|| up
.resv
|| up
.resv2
)
10655 return __io_register_rsrc_update(ctx
, type
, &up
, up
.nr
);
10658 static int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
10659 unsigned int size
, unsigned int type
)
10661 struct io_uring_rsrc_register rr
;
10663 /* keep it extendible */
10664 if (size
!= sizeof(rr
))
10667 memset(&rr
, 0, sizeof(rr
));
10668 if (copy_from_user(&rr
, arg
, size
))
10670 if (!rr
.nr
|| rr
.resv
|| rr
.resv2
)
10674 case IORING_RSRC_FILE
:
10675 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
10676 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10677 case IORING_RSRC_BUFFER
:
10678 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
10679 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10684 static int io_register_iowq_aff(struct io_ring_ctx
*ctx
, void __user
*arg
,
10687 struct io_uring_task
*tctx
= current
->io_uring
;
10688 cpumask_var_t new_mask
;
10691 if (!tctx
|| !tctx
->io_wq
)
10694 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
10697 cpumask_clear(new_mask
);
10698 if (len
> cpumask_size())
10699 len
= cpumask_size();
10701 if (in_compat_syscall()) {
10702 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
10703 (const compat_ulong_t __user
*)arg
,
10704 len
* 8 /* CHAR_BIT */);
10706 ret
= copy_from_user(new_mask
, arg
, len
);
10710 free_cpumask_var(new_mask
);
10714 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
10715 free_cpumask_var(new_mask
);
10719 static int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
10721 struct io_uring_task
*tctx
= current
->io_uring
;
10723 if (!tctx
|| !tctx
->io_wq
)
10726 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
10729 static int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
10731 __must_hold(&ctx
->uring_lock
)
10733 struct io_tctx_node
*node
;
10734 struct io_uring_task
*tctx
= NULL
;
10735 struct io_sq_data
*sqd
= NULL
;
10736 __u32 new_count
[2];
10739 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
10741 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10742 if (new_count
[i
] > INT_MAX
)
10745 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10746 sqd
= ctx
->sq_data
;
10749 * Observe the correct sqd->lock -> ctx->uring_lock
10750 * ordering. Fine to drop uring_lock here, we hold
10751 * a ref to the ctx.
10753 refcount_inc(&sqd
->refs
);
10754 mutex_unlock(&ctx
->uring_lock
);
10755 mutex_lock(&sqd
->lock
);
10756 mutex_lock(&ctx
->uring_lock
);
10758 tctx
= sqd
->thread
->io_uring
;
10761 tctx
= current
->io_uring
;
10764 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
10766 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10768 ctx
->iowq_limits
[i
] = new_count
[i
];
10769 ctx
->iowq_limits_set
= true;
10772 if (tctx
&& tctx
->io_wq
) {
10773 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
10777 memset(new_count
, 0, sizeof(new_count
));
10781 mutex_unlock(&sqd
->lock
);
10782 io_put_sq_data(sqd
);
10785 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
10788 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10792 /* now propagate the restriction to all registered users */
10793 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
10794 struct io_uring_task
*tctx
= node
->task
->io_uring
;
10796 if (WARN_ON_ONCE(!tctx
->io_wq
))
10799 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10800 new_count
[i
] = ctx
->iowq_limits
[i
];
10801 /* ignore errors, it always returns zero anyway */
10802 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
10807 mutex_unlock(&sqd
->lock
);
10808 io_put_sq_data(sqd
);
10813 static bool io_register_op_must_quiesce(int op
)
10816 case IORING_REGISTER_BUFFERS
:
10817 case IORING_UNREGISTER_BUFFERS
:
10818 case IORING_REGISTER_FILES
:
10819 case IORING_UNREGISTER_FILES
:
10820 case IORING_REGISTER_FILES_UPDATE
:
10821 case IORING_REGISTER_PROBE
:
10822 case IORING_REGISTER_PERSONALITY
:
10823 case IORING_UNREGISTER_PERSONALITY
:
10824 case IORING_REGISTER_FILES2
:
10825 case IORING_REGISTER_FILES_UPDATE2
:
10826 case IORING_REGISTER_BUFFERS2
:
10827 case IORING_REGISTER_BUFFERS_UPDATE
:
10828 case IORING_REGISTER_IOWQ_AFF
:
10829 case IORING_UNREGISTER_IOWQ_AFF
:
10830 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10837 static int io_ctx_quiesce(struct io_ring_ctx
*ctx
)
10841 percpu_ref_kill(&ctx
->refs
);
10844 * Drop uring mutex before waiting for references to exit. If another
10845 * thread is currently inside io_uring_enter() it might need to grab the
10846 * uring_lock to make progress. If we hold it here across the drain
10847 * wait, then we can deadlock. It's safe to drop the mutex here, since
10848 * no new references will come in after we've killed the percpu ref.
10850 mutex_unlock(&ctx
->uring_lock
);
10852 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10855 ret
= io_run_task_work_sig();
10856 } while (ret
>= 0);
10857 mutex_lock(&ctx
->uring_lock
);
10860 io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
);
10864 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10865 void __user
*arg
, unsigned nr_args
)
10866 __releases(ctx
->uring_lock
)
10867 __acquires(ctx
->uring_lock
)
10872 * We're inside the ring mutex, if the ref is already dying, then
10873 * someone else killed the ctx or is already going through
10874 * io_uring_register().
10876 if (percpu_ref_is_dying(&ctx
->refs
))
10879 if (ctx
->restricted
) {
10880 if (opcode
>= IORING_REGISTER_LAST
)
10882 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
10883 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
10887 if (io_register_op_must_quiesce(opcode
)) {
10888 ret
= io_ctx_quiesce(ctx
);
10894 case IORING_REGISTER_BUFFERS
:
10895 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
10897 case IORING_UNREGISTER_BUFFERS
:
10899 if (arg
|| nr_args
)
10901 ret
= io_sqe_buffers_unregister(ctx
);
10903 case IORING_REGISTER_FILES
:
10904 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
10906 case IORING_UNREGISTER_FILES
:
10908 if (arg
|| nr_args
)
10910 ret
= io_sqe_files_unregister(ctx
);
10912 case IORING_REGISTER_FILES_UPDATE
:
10913 ret
= io_register_files_update(ctx
, arg
, nr_args
);
10915 case IORING_REGISTER_EVENTFD
:
10916 case IORING_REGISTER_EVENTFD_ASYNC
:
10920 ret
= io_eventfd_register(ctx
, arg
);
10923 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10924 ctx
->eventfd_async
= 1;
10926 ctx
->eventfd_async
= 0;
10928 case IORING_UNREGISTER_EVENTFD
:
10930 if (arg
|| nr_args
)
10932 ret
= io_eventfd_unregister(ctx
);
10934 case IORING_REGISTER_PROBE
:
10936 if (!arg
|| nr_args
> 256)
10938 ret
= io_probe(ctx
, arg
, nr_args
);
10940 case IORING_REGISTER_PERSONALITY
:
10942 if (arg
|| nr_args
)
10944 ret
= io_register_personality(ctx
);
10946 case IORING_UNREGISTER_PERSONALITY
:
10950 ret
= io_unregister_personality(ctx
, nr_args
);
10952 case IORING_REGISTER_ENABLE_RINGS
:
10954 if (arg
|| nr_args
)
10956 ret
= io_register_enable_rings(ctx
);
10958 case IORING_REGISTER_RESTRICTIONS
:
10959 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10961 case IORING_REGISTER_FILES2
:
10962 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
10964 case IORING_REGISTER_FILES_UPDATE2
:
10965 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10968 case IORING_REGISTER_BUFFERS2
:
10969 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
10971 case IORING_REGISTER_BUFFERS_UPDATE
:
10972 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10973 IORING_RSRC_BUFFER
);
10975 case IORING_REGISTER_IOWQ_AFF
:
10977 if (!arg
|| !nr_args
)
10979 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
10981 case IORING_UNREGISTER_IOWQ_AFF
:
10983 if (arg
|| nr_args
)
10985 ret
= io_unregister_iowq_aff(ctx
);
10987 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10989 if (!arg
|| nr_args
!= 2)
10991 ret
= io_register_iowq_max_workers(ctx
, arg
);
10998 if (io_register_op_must_quiesce(opcode
)) {
10999 /* bring the ctx back to life */
11000 percpu_ref_reinit(&ctx
->refs
);
11001 reinit_completion(&ctx
->ref_comp
);
11006 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
11007 void __user
*, arg
, unsigned int, nr_args
)
11009 struct io_ring_ctx
*ctx
;
11018 if (f
.file
->f_op
!= &io_uring_fops
)
11021 ctx
= f
.file
->private_data
;
11023 io_run_task_work();
11025 mutex_lock(&ctx
->uring_lock
);
11026 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
11027 mutex_unlock(&ctx
->uring_lock
);
11028 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
11029 ctx
->cq_ev_fd
!= NULL
, ret
);
11035 static int __init
io_uring_init(void)
11037 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11038 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11039 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11042 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11043 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11044 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
11045 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
11046 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
11047 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
11048 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
11049 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
11050 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
11051 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
11052 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
11053 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
11054 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
11055 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
11056 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
11057 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
11058 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
11059 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
11060 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
11061 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
11062 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
11063 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
11064 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
11065 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
11066 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
11067 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
11068 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
11069 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
11070 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
11071 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
11072 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
11073 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
11074 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
11076 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
11077 sizeof(struct io_uring_rsrc_update
));
11078 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
11079 sizeof(struct io_uring_rsrc_update2
));
11081 /* ->buf_index is u16 */
11082 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS
>= (1u << 16));
11084 /* should fit into one byte */
11085 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
11087 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
11088 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
11090 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
11094 __initcall(io_uring_init
);