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
);
1542 spin_lock_irq(&ctx
->timeout_lock
);
1543 while (!list_empty(&ctx
->timeout_list
)) {
1544 u32 events_needed
, events_got
;
1545 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1546 struct io_kiocb
, timeout
.list
);
1548 if (io_is_timeout_noseq(req
))
1552 * Since seq can easily wrap around over time, subtract
1553 * the last seq at which timeouts were flushed before comparing.
1554 * Assuming not more than 2^31-1 events have happened since,
1555 * these subtractions won't have wrapped, so we can check if
1556 * target is in [last_seq, current_seq] by comparing the two.
1558 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1559 events_got
= seq
- ctx
->cq_last_tm_flush
;
1560 if (events_got
< events_needed
)
1563 list_del_init(&req
->timeout
.list
);
1564 io_kill_timeout(req
, 0);
1566 ctx
->cq_last_tm_flush
= seq
;
1567 spin_unlock_irq(&ctx
->timeout_lock
);
1570 static void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
1572 if (ctx
->off_timeout_used
)
1573 io_flush_timeouts(ctx
);
1574 if (ctx
->drain_active
)
1575 io_queue_deferred(ctx
);
1578 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
1580 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
))
1581 __io_commit_cqring_flush(ctx
);
1582 /* order cqe stores with ring update */
1583 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1586 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1588 struct io_rings
*r
= ctx
->rings
;
1590 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
1593 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1595 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1598 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
1600 struct io_rings
*rings
= ctx
->rings
;
1601 unsigned tail
, mask
= ctx
->cq_entries
- 1;
1604 * writes to the cq entry need to come after reading head; the
1605 * control dependency is enough as we're using WRITE_ONCE to
1608 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
1611 tail
= ctx
->cached_cq_tail
++;
1612 return &rings
->cqes
[tail
& mask
];
1615 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1617 if (likely(!ctx
->cq_ev_fd
))
1619 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1621 return !ctx
->eventfd_async
|| io_wq_current_is_worker();
1625 * This should only get called when at least one event has been posted.
1626 * Some applications rely on the eventfd notification count only changing
1627 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1628 * 1:1 relationship between how many times this function is called (and
1629 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1631 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1634 * wake_up_all() may seem excessive, but io_wake_function() and
1635 * io_should_wake() handle the termination of the loop and only
1636 * wake as many waiters as we need to.
1638 if (wq_has_sleeper(&ctx
->cq_wait
))
1639 wake_up_all(&ctx
->cq_wait
);
1640 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1641 wake_up(&ctx
->sq_data
->wait
);
1642 if (io_should_trigger_evfd(ctx
))
1643 eventfd_signal(ctx
->cq_ev_fd
, 1);
1644 if (waitqueue_active(&ctx
->poll_wait
))
1645 wake_up_interruptible(&ctx
->poll_wait
);
1648 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1650 /* see waitqueue_active() comment */
1653 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1654 if (waitqueue_active(&ctx
->cq_wait
))
1655 wake_up_all(&ctx
->cq_wait
);
1657 if (io_should_trigger_evfd(ctx
))
1658 eventfd_signal(ctx
->cq_ev_fd
, 1);
1659 if (waitqueue_active(&ctx
->poll_wait
))
1660 wake_up_interruptible(&ctx
->poll_wait
);
1663 /* Returns true if there are no backlogged entries after the flush */
1664 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1666 bool all_flushed
, posted
;
1668 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
1672 spin_lock(&ctx
->completion_lock
);
1673 while (!list_empty(&ctx
->cq_overflow_list
)) {
1674 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
1675 struct io_overflow_cqe
*ocqe
;
1679 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
1680 struct io_overflow_cqe
, list
);
1682 memcpy(cqe
, &ocqe
->cqe
, sizeof(*cqe
));
1684 io_account_cq_overflow(ctx
);
1687 list_del(&ocqe
->list
);
1691 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1693 clear_bit(0, &ctx
->check_cq_overflow
);
1694 WRITE_ONCE(ctx
->rings
->sq_flags
,
1695 ctx
->rings
->sq_flags
& ~IORING_SQ_CQ_OVERFLOW
);
1699 io_commit_cqring(ctx
);
1700 spin_unlock(&ctx
->completion_lock
);
1702 io_cqring_ev_posted(ctx
);
1706 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
1710 if (test_bit(0, &ctx
->check_cq_overflow
)) {
1711 /* iopoll syncs against uring_lock, not completion_lock */
1712 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1713 mutex_lock(&ctx
->uring_lock
);
1714 ret
= __io_cqring_overflow_flush(ctx
, false);
1715 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1716 mutex_unlock(&ctx
->uring_lock
);
1722 /* must to be called somewhat shortly after putting a request */
1723 static inline void io_put_task(struct task_struct
*task
, int nr
)
1725 struct io_uring_task
*tctx
= task
->io_uring
;
1727 if (likely(task
== current
)) {
1728 tctx
->cached_refs
+= nr
;
1730 percpu_counter_sub(&tctx
->inflight
, nr
);
1731 if (unlikely(atomic_read(&tctx
->in_idle
)))
1732 wake_up(&tctx
->wait
);
1733 put_task_struct_many(task
, nr
);
1737 static void io_task_refs_refill(struct io_uring_task
*tctx
)
1739 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
1741 percpu_counter_add(&tctx
->inflight
, refill
);
1742 refcount_add(refill
, ¤t
->usage
);
1743 tctx
->cached_refs
+= refill
;
1746 static inline void io_get_task_refs(int nr
)
1748 struct io_uring_task
*tctx
= current
->io_uring
;
1750 tctx
->cached_refs
-= nr
;
1751 if (unlikely(tctx
->cached_refs
< 0))
1752 io_task_refs_refill(tctx
);
1755 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
1757 struct io_uring_task
*tctx
= task
->io_uring
;
1758 unsigned int refs
= tctx
->cached_refs
;
1761 tctx
->cached_refs
= 0;
1762 percpu_counter_sub(&tctx
->inflight
, refs
);
1763 put_task_struct_many(task
, refs
);
1767 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
1768 long res
, unsigned int cflags
)
1770 struct io_overflow_cqe
*ocqe
;
1772 ocqe
= kmalloc(sizeof(*ocqe
), GFP_ATOMIC
| __GFP_ACCOUNT
);
1775 * If we're in ring overflow flush mode, or in task cancel mode,
1776 * or cannot allocate an overflow entry, then we need to drop it
1779 io_account_cq_overflow(ctx
);
1782 if (list_empty(&ctx
->cq_overflow_list
)) {
1783 set_bit(0, &ctx
->check_cq_overflow
);
1784 WRITE_ONCE(ctx
->rings
->sq_flags
,
1785 ctx
->rings
->sq_flags
| IORING_SQ_CQ_OVERFLOW
);
1788 ocqe
->cqe
.user_data
= user_data
;
1789 ocqe
->cqe
.res
= res
;
1790 ocqe
->cqe
.flags
= cflags
;
1791 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
1795 static inline bool __io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1796 long res
, unsigned int cflags
)
1798 struct io_uring_cqe
*cqe
;
1800 trace_io_uring_complete(ctx
, user_data
, res
, cflags
);
1803 * If we can't get a cq entry, userspace overflowed the
1804 * submission (by quite a lot). Increment the overflow count in
1807 cqe
= io_get_cqe(ctx
);
1809 WRITE_ONCE(cqe
->user_data
, user_data
);
1810 WRITE_ONCE(cqe
->res
, res
);
1811 WRITE_ONCE(cqe
->flags
, cflags
);
1814 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
);
1817 /* not as hot to bloat with inlining */
1818 static noinline
bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1819 long res
, unsigned int cflags
)
1821 return __io_cqring_fill_event(ctx
, user_data
, res
, cflags
);
1824 static void io_req_complete_post(struct io_kiocb
*req
, long res
,
1825 unsigned int cflags
)
1827 struct io_ring_ctx
*ctx
= req
->ctx
;
1829 spin_lock(&ctx
->completion_lock
);
1830 __io_cqring_fill_event(ctx
, req
->user_data
, res
, cflags
);
1832 * If we're the last reference to this request, add to our locked
1835 if (req_ref_put_and_test(req
)) {
1836 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1837 if (req
->flags
& IO_DISARM_MASK
)
1838 io_disarm_next(req
);
1840 io_req_task_queue(req
->link
);
1844 io_dismantle_req(req
);
1845 io_put_task(req
->task
, 1);
1846 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
1847 ctx
->locked_free_nr
++;
1849 if (!percpu_ref_tryget(&ctx
->refs
))
1852 io_commit_cqring(ctx
);
1853 spin_unlock(&ctx
->completion_lock
);
1856 io_cqring_ev_posted(ctx
);
1857 percpu_ref_put(&ctx
->refs
);
1861 static inline bool io_req_needs_clean(struct io_kiocb
*req
)
1863 return req
->flags
& IO_REQ_CLEAN_FLAGS
;
1866 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1867 unsigned int cflags
)
1869 if (io_req_needs_clean(req
))
1872 req
->compl.cflags
= cflags
;
1873 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1876 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1877 long res
, unsigned cflags
)
1879 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1880 io_req_complete_state(req
, res
, cflags
);
1882 io_req_complete_post(req
, res
, cflags
);
1885 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1887 __io_req_complete(req
, 0, res
, 0);
1890 static void io_req_complete_failed(struct io_kiocb
*req
, long res
)
1893 io_req_complete_post(req
, res
, 0);
1896 static void io_req_complete_fail_submit(struct io_kiocb
*req
)
1899 * We don't submit, fail them all, for that replace hardlinks with
1900 * normal links. Extra REQ_F_LINK is tolerated.
1902 req
->flags
&= ~REQ_F_HARDLINK
;
1903 req
->flags
|= REQ_F_LINK
;
1904 io_req_complete_failed(req
, req
->result
);
1908 * Don't initialise the fields below on every allocation, but do that in
1909 * advance and keep them valid across allocations.
1911 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1915 req
->async_data
= NULL
;
1916 /* not necessary, but safer to zero */
1920 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1921 struct io_submit_state
*state
)
1923 spin_lock(&ctx
->completion_lock
);
1924 list_splice_init(&ctx
->locked_free_list
, &state
->free_list
);
1925 ctx
->locked_free_nr
= 0;
1926 spin_unlock(&ctx
->completion_lock
);
1929 /* Returns true IFF there are requests in the cache */
1930 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1932 struct io_submit_state
*state
= &ctx
->submit_state
;
1936 * If we have more than a batch's worth of requests in our IRQ side
1937 * locked cache, grab the lock and move them over to our submission
1940 if (READ_ONCE(ctx
->locked_free_nr
) > IO_COMPL_BATCH
)
1941 io_flush_cached_locked_reqs(ctx
, state
);
1943 nr
= state
->free_reqs
;
1944 while (!list_empty(&state
->free_list
)) {
1945 struct io_kiocb
*req
= list_first_entry(&state
->free_list
,
1946 struct io_kiocb
, inflight_entry
);
1948 list_del(&req
->inflight_entry
);
1949 state
->reqs
[nr
++] = req
;
1950 if (nr
== ARRAY_SIZE(state
->reqs
))
1954 state
->free_reqs
= nr
;
1959 * A request might get retired back into the request caches even before opcode
1960 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1961 * Because of that, io_alloc_req() should be called only under ->uring_lock
1962 * and with extra caution to not get a request that is still worked on.
1964 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1965 __must_hold(&ctx
->uring_lock
)
1967 struct io_submit_state
*state
= &ctx
->submit_state
;
1968 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1971 BUILD_BUG_ON(ARRAY_SIZE(state
->reqs
) < IO_REQ_ALLOC_BATCH
);
1973 if (likely(state
->free_reqs
|| io_flush_cached_reqs(ctx
)))
1976 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1980 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1981 * retry single alloc to be on the safe side.
1983 if (unlikely(ret
<= 0)) {
1984 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1985 if (!state
->reqs
[0])
1990 for (i
= 0; i
< ret
; i
++)
1991 io_preinit_req(state
->reqs
[i
], ctx
);
1992 state
->free_reqs
= ret
;
1995 return state
->reqs
[state
->free_reqs
];
1998 static inline void io_put_file(struct file
*file
)
2004 static void io_dismantle_req(struct io_kiocb
*req
)
2006 unsigned int flags
= req
->flags
;
2008 if (io_req_needs_clean(req
))
2010 if (!(flags
& REQ_F_FIXED_FILE
))
2011 io_put_file(req
->file
);
2012 if (req
->fixed_rsrc_refs
)
2013 percpu_ref_put(req
->fixed_rsrc_refs
);
2014 if (req
->async_data
) {
2015 kfree(req
->async_data
);
2016 req
->async_data
= NULL
;
2020 static void __io_free_req(struct io_kiocb
*req
)
2022 struct io_ring_ctx
*ctx
= req
->ctx
;
2024 io_dismantle_req(req
);
2025 io_put_task(req
->task
, 1);
2027 spin_lock(&ctx
->completion_lock
);
2028 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
2029 ctx
->locked_free_nr
++;
2030 spin_unlock(&ctx
->completion_lock
);
2032 percpu_ref_put(&ctx
->refs
);
2035 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2037 struct io_kiocb
*nxt
= req
->link
;
2039 req
->link
= nxt
->link
;
2043 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
2044 __must_hold(&req
->ctx
->completion_lock
)
2045 __must_hold(&req
->ctx
->timeout_lock
)
2047 struct io_kiocb
*link
= req
->link
;
2049 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2050 struct io_timeout_data
*io
= link
->async_data
;
2052 io_remove_next_linked(req
);
2053 link
->timeout
.head
= NULL
;
2054 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
2055 list_del(&link
->timeout
.list
);
2056 io_cqring_fill_event(link
->ctx
, link
->user_data
,
2058 io_put_req_deferred(link
);
2065 static void io_fail_links(struct io_kiocb
*req
)
2066 __must_hold(&req
->ctx
->completion_lock
)
2068 struct io_kiocb
*nxt
, *link
= req
->link
;
2072 long res
= -ECANCELED
;
2074 if (link
->flags
& REQ_F_FAIL
)
2080 trace_io_uring_fail_link(req
, link
);
2081 io_cqring_fill_event(link
->ctx
, link
->user_data
, res
, 0);
2082 io_put_req_deferred(link
);
2087 static bool io_disarm_next(struct io_kiocb
*req
)
2088 __must_hold(&req
->ctx
->completion_lock
)
2090 bool posted
= false;
2092 if (req
->flags
& REQ_F_ARM_LTIMEOUT
) {
2093 struct io_kiocb
*link
= req
->link
;
2095 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
2096 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2097 io_remove_next_linked(req
);
2098 io_cqring_fill_event(link
->ctx
, link
->user_data
,
2100 io_put_req_deferred(link
);
2103 } else if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
2104 struct io_ring_ctx
*ctx
= req
->ctx
;
2106 spin_lock_irq(&ctx
->timeout_lock
);
2107 posted
= io_kill_linked_timeout(req
);
2108 spin_unlock_irq(&ctx
->timeout_lock
);
2110 if (unlikely((req
->flags
& REQ_F_FAIL
) &&
2111 !(req
->flags
& REQ_F_HARDLINK
))) {
2112 posted
|= (req
->link
!= NULL
);
2118 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2120 struct io_kiocb
*nxt
;
2123 * If LINK is set, we have dependent requests in this chain. If we
2124 * didn't fail this request, queue the first one up, moving any other
2125 * dependencies to the next request. In case of failure, fail the rest
2128 if (req
->flags
& IO_DISARM_MASK
) {
2129 struct io_ring_ctx
*ctx
= req
->ctx
;
2132 spin_lock(&ctx
->completion_lock
);
2133 posted
= io_disarm_next(req
);
2135 io_commit_cqring(req
->ctx
);
2136 spin_unlock(&ctx
->completion_lock
);
2138 io_cqring_ev_posted(ctx
);
2145 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2147 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
2149 return __io_req_find_next(req
);
2152 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, bool *locked
)
2157 if (ctx
->submit_state
.compl_nr
)
2158 io_submit_flush_completions(ctx
);
2159 mutex_unlock(&ctx
->uring_lock
);
2162 percpu_ref_put(&ctx
->refs
);
2165 static void tctx_task_work(struct callback_head
*cb
)
2167 bool locked
= false;
2168 struct io_ring_ctx
*ctx
= NULL
;
2169 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
2173 struct io_wq_work_node
*node
;
2175 if (!tctx
->task_list
.first
&& locked
&& ctx
->submit_state
.compl_nr
)
2176 io_submit_flush_completions(ctx
);
2178 spin_lock_irq(&tctx
->task_lock
);
2179 node
= tctx
->task_list
.first
;
2180 INIT_WQ_LIST(&tctx
->task_list
);
2182 tctx
->task_running
= false;
2183 spin_unlock_irq(&tctx
->task_lock
);
2188 struct io_wq_work_node
*next
= node
->next
;
2189 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
2192 if (req
->ctx
!= ctx
) {
2193 ctx_flush_and_put(ctx
, &locked
);
2195 /* if not contended, grab and improve batching */
2196 locked
= mutex_trylock(&ctx
->uring_lock
);
2197 percpu_ref_get(&ctx
->refs
);
2199 req
->io_task_work
.func(req
, &locked
);
2206 ctx_flush_and_put(ctx
, &locked
);
2208 /* relaxed read is enough as only the task itself sets ->in_idle */
2209 if (unlikely(atomic_read(&tctx
->in_idle
)))
2210 io_uring_drop_tctx_refs(current
);
2213 static void io_req_task_work_add(struct io_kiocb
*req
)
2215 struct task_struct
*tsk
= req
->task
;
2216 struct io_uring_task
*tctx
= tsk
->io_uring
;
2217 enum task_work_notify_mode notify
;
2218 struct io_wq_work_node
*node
;
2219 unsigned long flags
;
2222 WARN_ON_ONCE(!tctx
);
2224 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2225 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
2226 running
= tctx
->task_running
;
2228 tctx
->task_running
= true;
2229 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2231 /* task_work already pending, we're done */
2236 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2237 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2238 * processing task_work. There's no reliable way to tell if TWA_RESUME
2241 notify
= (req
->ctx
->flags
& IORING_SETUP_SQPOLL
) ? TWA_NONE
: TWA_SIGNAL
;
2242 if (!task_work_add(tsk
, &tctx
->task_work
, notify
)) {
2243 wake_up_process(tsk
);
2247 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2248 tctx
->task_running
= false;
2249 node
= tctx
->task_list
.first
;
2250 INIT_WQ_LIST(&tctx
->task_list
);
2251 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2254 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
2256 if (llist_add(&req
->io_task_work
.fallback_node
,
2257 &req
->ctx
->fallback_llist
))
2258 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
2262 static void io_req_task_cancel(struct io_kiocb
*req
, bool *locked
)
2264 struct io_ring_ctx
*ctx
= req
->ctx
;
2266 /* not needed for normal modes, but SQPOLL depends on it */
2267 io_tw_lock(ctx
, locked
);
2268 io_req_complete_failed(req
, req
->result
);
2271 static void io_req_task_submit(struct io_kiocb
*req
, bool *locked
)
2273 struct io_ring_ctx
*ctx
= req
->ctx
;
2275 io_tw_lock(ctx
, locked
);
2276 /* req->task == current here, checking PF_EXITING is safe */
2277 if (likely(!(req
->task
->flags
& PF_EXITING
)))
2278 __io_queue_sqe(req
);
2280 io_req_complete_failed(req
, -EFAULT
);
2283 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2286 req
->io_task_work
.func
= io_req_task_cancel
;
2287 io_req_task_work_add(req
);
2290 static void io_req_task_queue(struct io_kiocb
*req
)
2292 req
->io_task_work
.func
= io_req_task_submit
;
2293 io_req_task_work_add(req
);
2296 static void io_req_task_queue_reissue(struct io_kiocb
*req
)
2298 req
->io_task_work
.func
= io_queue_async_work
;
2299 io_req_task_work_add(req
);
2302 static inline void io_queue_next(struct io_kiocb
*req
)
2304 struct io_kiocb
*nxt
= io_req_find_next(req
);
2307 io_req_task_queue(nxt
);
2310 static void io_free_req(struct io_kiocb
*req
)
2316 static void io_free_req_work(struct io_kiocb
*req
, bool *locked
)
2322 struct task_struct
*task
;
2327 static inline void io_init_req_batch(struct req_batch
*rb
)
2334 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2335 struct req_batch
*rb
)
2338 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2340 io_put_task(rb
->task
, rb
->task_refs
);
2343 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2344 struct io_submit_state
*state
)
2347 io_dismantle_req(req
);
2349 if (req
->task
!= rb
->task
) {
2351 io_put_task(rb
->task
, rb
->task_refs
);
2352 rb
->task
= req
->task
;
2358 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2359 state
->reqs
[state
->free_reqs
++] = req
;
2361 list_add(&req
->inflight_entry
, &state
->free_list
);
2364 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
2365 __must_hold(&ctx
->uring_lock
)
2367 struct io_submit_state
*state
= &ctx
->submit_state
;
2368 int i
, nr
= state
->compl_nr
;
2369 struct req_batch rb
;
2371 spin_lock(&ctx
->completion_lock
);
2372 for (i
= 0; i
< nr
; i
++) {
2373 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2375 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2378 io_commit_cqring(ctx
);
2379 spin_unlock(&ctx
->completion_lock
);
2380 io_cqring_ev_posted(ctx
);
2382 io_init_req_batch(&rb
);
2383 for (i
= 0; i
< nr
; i
++) {
2384 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2386 if (req_ref_put_and_test(req
))
2387 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2390 io_req_free_batch_finish(ctx
, &rb
);
2391 state
->compl_nr
= 0;
2395 * Drop reference to request, return next in chain (if there is one) if this
2396 * was the last reference to this request.
2398 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2400 struct io_kiocb
*nxt
= NULL
;
2402 if (req_ref_put_and_test(req
)) {
2403 nxt
= io_req_find_next(req
);
2409 static inline void io_put_req(struct io_kiocb
*req
)
2411 if (req_ref_put_and_test(req
))
2415 static inline void io_put_req_deferred(struct io_kiocb
*req
)
2417 if (req_ref_put_and_test(req
)) {
2418 req
->io_task_work
.func
= io_free_req_work
;
2419 io_req_task_work_add(req
);
2423 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2425 /* See comment at the top of this file */
2427 return __io_cqring_events(ctx
);
2430 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2432 struct io_rings
*rings
= ctx
->rings
;
2434 /* make sure SQ entry isn't read before tail */
2435 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2438 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2440 unsigned int cflags
;
2442 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2443 cflags
|= IORING_CQE_F_BUFFER
;
2444 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2449 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2451 struct io_buffer
*kbuf
;
2453 if (likely(!(req
->flags
& REQ_F_BUFFER_SELECTED
)))
2455 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2456 return io_put_kbuf(req
, kbuf
);
2459 static inline bool io_run_task_work(void)
2461 if (test_thread_flag(TIF_NOTIFY_SIGNAL
) || current
->task_works
) {
2462 __set_current_state(TASK_RUNNING
);
2463 tracehook_notify_signal();
2471 * Find and free completed poll iocbs
2473 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2474 struct list_head
*done
)
2476 struct req_batch rb
;
2477 struct io_kiocb
*req
;
2479 /* order with ->result store in io_complete_rw_iopoll() */
2482 io_init_req_batch(&rb
);
2483 while (!list_empty(done
)) {
2484 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2485 list_del(&req
->inflight_entry
);
2487 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2488 io_put_rw_kbuf(req
));
2491 if (req_ref_put_and_test(req
))
2492 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2495 io_commit_cqring(ctx
);
2496 io_cqring_ev_posted_iopoll(ctx
);
2497 io_req_free_batch_finish(ctx
, &rb
);
2500 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2503 struct io_kiocb
*req
, *tmp
;
2508 * Only spin for completions if we don't have multiple devices hanging
2509 * off our complete list, and we're under the requested amount.
2511 spin
= !ctx
->poll_multi_queue
&& *nr_events
< min
;
2513 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2514 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2518 * Move completed and retryable entries to our local lists.
2519 * If we find a request that requires polling, break out
2520 * and complete those lists first, if we have entries there.
2522 if (READ_ONCE(req
->iopoll_completed
)) {
2523 list_move_tail(&req
->inflight_entry
, &done
);
2526 if (!list_empty(&done
))
2529 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2530 if (unlikely(ret
< 0))
2535 /* iopoll may have completed current req */
2536 if (READ_ONCE(req
->iopoll_completed
))
2537 list_move_tail(&req
->inflight_entry
, &done
);
2540 if (!list_empty(&done
))
2541 io_iopoll_complete(ctx
, nr_events
, &done
);
2547 * We can't just wait for polled events to come to us, we have to actively
2548 * find and complete them.
2550 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2552 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2555 mutex_lock(&ctx
->uring_lock
);
2556 while (!list_empty(&ctx
->iopoll_list
)) {
2557 unsigned int nr_events
= 0;
2559 io_do_iopoll(ctx
, &nr_events
, 0);
2561 /* let it sleep and repeat later if can't complete a request */
2565 * Ensure we allow local-to-the-cpu processing to take place,
2566 * in this case we need to ensure that we reap all events.
2567 * Also let task_work, etc. to progress by releasing the mutex
2569 if (need_resched()) {
2570 mutex_unlock(&ctx
->uring_lock
);
2572 mutex_lock(&ctx
->uring_lock
);
2575 mutex_unlock(&ctx
->uring_lock
);
2578 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2580 unsigned int nr_events
= 0;
2584 * We disallow the app entering submit/complete with polling, but we
2585 * still need to lock the ring to prevent racing with polled issue
2586 * that got punted to a workqueue.
2588 mutex_lock(&ctx
->uring_lock
);
2590 * Don't enter poll loop if we already have events pending.
2591 * If we do, we can potentially be spinning for commands that
2592 * already triggered a CQE (eg in error).
2594 if (test_bit(0, &ctx
->check_cq_overflow
))
2595 __io_cqring_overflow_flush(ctx
, false);
2596 if (io_cqring_events(ctx
))
2600 * If a submit got punted to a workqueue, we can have the
2601 * application entering polling for a command before it gets
2602 * issued. That app will hold the uring_lock for the duration
2603 * of the poll right here, so we need to take a breather every
2604 * now and then to ensure that the issue has a chance to add
2605 * the poll to the issued list. Otherwise we can spin here
2606 * forever, while the workqueue is stuck trying to acquire the
2609 if (list_empty(&ctx
->iopoll_list
)) {
2610 u32 tail
= ctx
->cached_cq_tail
;
2612 mutex_unlock(&ctx
->uring_lock
);
2614 mutex_lock(&ctx
->uring_lock
);
2616 /* some requests don't go through iopoll_list */
2617 if (tail
!= ctx
->cached_cq_tail
||
2618 list_empty(&ctx
->iopoll_list
))
2621 ret
= io_do_iopoll(ctx
, &nr_events
, min
);
2622 } while (!ret
&& nr_events
< min
&& !need_resched());
2624 mutex_unlock(&ctx
->uring_lock
);
2628 static void kiocb_end_write(struct io_kiocb
*req
)
2631 * Tell lockdep we inherited freeze protection from submission
2634 if (req
->flags
& REQ_F_ISREG
) {
2635 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
2637 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
2643 static bool io_resubmit_prep(struct io_kiocb
*req
)
2645 struct io_async_rw
*rw
= req
->async_data
;
2648 return !io_req_prep_async(req
);
2649 iov_iter_restore(&rw
->iter
, &rw
->iter_state
);
2653 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2655 umode_t mode
= file_inode(req
->file
)->i_mode
;
2656 struct io_ring_ctx
*ctx
= req
->ctx
;
2658 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2660 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2661 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2664 * If ref is dying, we might be running poll reap from the exit work.
2665 * Don't attempt to reissue from that path, just let it fail with
2668 if (percpu_ref_is_dying(&ctx
->refs
))
2671 * Play it safe and assume not safe to re-import and reissue if we're
2672 * not in the original thread group (or in task context).
2674 if (!same_thread_group(req
->task
, current
) || !in_task())
2679 static bool io_resubmit_prep(struct io_kiocb
*req
)
2683 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2689 static bool __io_complete_rw_common(struct io_kiocb
*req
, long res
)
2691 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
)
2692 kiocb_end_write(req
);
2693 if (res
!= req
->result
) {
2694 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
2695 io_rw_should_reissue(req
)) {
2696 req
->flags
|= REQ_F_REISSUE
;
2705 static void io_req_task_complete(struct io_kiocb
*req
, bool *locked
)
2707 unsigned int cflags
= io_put_rw_kbuf(req
);
2708 long res
= req
->result
;
2711 struct io_ring_ctx
*ctx
= req
->ctx
;
2712 struct io_submit_state
*state
= &ctx
->submit_state
;
2714 io_req_complete_state(req
, res
, cflags
);
2715 state
->compl_reqs
[state
->compl_nr
++] = req
;
2716 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
2717 io_submit_flush_completions(ctx
);
2719 io_req_complete_post(req
, res
, cflags
);
2723 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2724 unsigned int issue_flags
)
2726 if (__io_complete_rw_common(req
, res
))
2728 __io_req_complete(req
, issue_flags
, req
->result
, io_put_rw_kbuf(req
));
2731 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2733 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2735 if (__io_complete_rw_common(req
, res
))
2738 req
->io_task_work
.func
= io_req_task_complete
;
2739 io_req_task_work_add(req
);
2742 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2744 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2746 if (kiocb
->ki_flags
& IOCB_WRITE
)
2747 kiocb_end_write(req
);
2748 if (unlikely(res
!= req
->result
)) {
2749 if (res
== -EAGAIN
&& io_rw_should_reissue(req
)) {
2750 req
->flags
|= REQ_F_REISSUE
;
2755 WRITE_ONCE(req
->result
, res
);
2756 /* order with io_iopoll_complete() checking ->result */
2758 WRITE_ONCE(req
->iopoll_completed
, 1);
2762 * After the iocb has been issued, it's safe to be found on the poll list.
2763 * Adding the kiocb to the list AFTER submission ensures that we don't
2764 * find it from a io_do_iopoll() thread before the issuer is done
2765 * accessing the kiocb cookie.
2767 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2769 struct io_ring_ctx
*ctx
= req
->ctx
;
2770 const bool in_async
= io_wq_current_is_worker();
2772 /* workqueue context doesn't hold uring_lock, grab it now */
2773 if (unlikely(in_async
))
2774 mutex_lock(&ctx
->uring_lock
);
2777 * Track whether we have multiple files in our lists. This will impact
2778 * how we do polling eventually, not spinning if we're on potentially
2779 * different devices.
2781 if (list_empty(&ctx
->iopoll_list
)) {
2782 ctx
->poll_multi_queue
= false;
2783 } else if (!ctx
->poll_multi_queue
) {
2784 struct io_kiocb
*list_req
;
2785 unsigned int queue_num0
, queue_num1
;
2787 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2790 if (list_req
->file
!= req
->file
) {
2791 ctx
->poll_multi_queue
= true;
2793 queue_num0
= blk_qc_t_to_queue_num(list_req
->rw
.kiocb
.ki_cookie
);
2794 queue_num1
= blk_qc_t_to_queue_num(req
->rw
.kiocb
.ki_cookie
);
2795 if (queue_num0
!= queue_num1
)
2796 ctx
->poll_multi_queue
= true;
2801 * For fast devices, IO may have already completed. If it has, add
2802 * it to the front so we find it first.
2804 if (READ_ONCE(req
->iopoll_completed
))
2805 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2807 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2809 if (unlikely(in_async
)) {
2811 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2812 * in sq thread task context or in io worker task context. If
2813 * current task context is sq thread, we don't need to check
2814 * whether should wake up sq thread.
2816 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2817 wq_has_sleeper(&ctx
->sq_data
->wait
))
2818 wake_up(&ctx
->sq_data
->wait
);
2820 mutex_unlock(&ctx
->uring_lock
);
2824 static bool io_bdev_nowait(struct block_device
*bdev
)
2826 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2830 * If we tracked the file through the SCM inflight mechanism, we could support
2831 * any file. For now, just ensure that anything potentially problematic is done
2834 static bool __io_file_supports_nowait(struct file
*file
, int rw
)
2836 umode_t mode
= file_inode(file
)->i_mode
;
2838 if (S_ISBLK(mode
)) {
2839 if (IS_ENABLED(CONFIG_BLOCK
) &&
2840 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2846 if (S_ISREG(mode
)) {
2847 if (IS_ENABLED(CONFIG_BLOCK
) &&
2848 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2849 file
->f_op
!= &io_uring_fops
)
2854 /* any ->read/write should understand O_NONBLOCK */
2855 if (file
->f_flags
& O_NONBLOCK
)
2858 if (!(file
->f_mode
& FMODE_NOWAIT
))
2862 return file
->f_op
->read_iter
!= NULL
;
2864 return file
->f_op
->write_iter
!= NULL
;
2867 static bool io_file_supports_nowait(struct io_kiocb
*req
, int rw
)
2869 if (rw
== READ
&& (req
->flags
& REQ_F_NOWAIT_READ
))
2871 else if (rw
== WRITE
&& (req
->flags
& REQ_F_NOWAIT_WRITE
))
2874 return __io_file_supports_nowait(req
->file
, rw
);
2877 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2880 struct io_ring_ctx
*ctx
= req
->ctx
;
2881 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2882 struct file
*file
= req
->file
;
2886 if (!io_req_ffs_set(req
) && S_ISREG(file_inode(file
)->i_mode
))
2887 req
->flags
|= REQ_F_ISREG
;
2889 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2890 if (kiocb
->ki_pos
== -1) {
2891 if (!(file
->f_mode
& FMODE_STREAM
)) {
2892 req
->flags
|= REQ_F_CUR_POS
;
2893 kiocb
->ki_pos
= file
->f_pos
;
2898 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2899 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2900 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2905 * If the file is marked O_NONBLOCK, still allow retry for it if it
2906 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2907 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2909 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) ||
2910 ((file
->f_flags
& O_NONBLOCK
) && !io_file_supports_nowait(req
, rw
)))
2911 req
->flags
|= REQ_F_NOWAIT
;
2913 ioprio
= READ_ONCE(sqe
->ioprio
);
2915 ret
= ioprio_check_cap(ioprio
);
2919 kiocb
->ki_ioprio
= ioprio
;
2921 kiocb
->ki_ioprio
= get_current_ioprio();
2923 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2924 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2925 !kiocb
->ki_filp
->f_op
->iopoll
)
2928 kiocb
->ki_flags
|= IOCB_HIPRI
| IOCB_ALLOC_CACHE
;
2929 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2930 req
->iopoll_completed
= 0;
2932 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2934 kiocb
->ki_complete
= io_complete_rw
;
2937 if (req
->opcode
== IORING_OP_READ_FIXED
||
2938 req
->opcode
== IORING_OP_WRITE_FIXED
) {
2940 io_req_set_rsrc_node(req
);
2943 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2944 req
->rw
.len
= READ_ONCE(sqe
->len
);
2945 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2949 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2955 case -ERESTARTNOINTR
:
2956 case -ERESTARTNOHAND
:
2957 case -ERESTART_RESTARTBLOCK
:
2959 * We can't just restart the syscall, since previously
2960 * submitted sqes may already be in progress. Just fail this
2966 kiocb
->ki_complete(kiocb
, ret
, 0);
2970 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2971 unsigned int issue_flags
)
2973 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2974 struct io_async_rw
*io
= req
->async_data
;
2976 /* add previously done IO, if any */
2977 if (io
&& io
->bytes_done
> 0) {
2979 ret
= io
->bytes_done
;
2981 ret
+= io
->bytes_done
;
2984 if (req
->flags
& REQ_F_CUR_POS
)
2985 req
->file
->f_pos
= kiocb
->ki_pos
;
2986 if (ret
>= 0 && (kiocb
->ki_complete
== io_complete_rw
))
2987 __io_complete_rw(req
, ret
, 0, issue_flags
);
2989 io_rw_done(kiocb
, ret
);
2991 if (req
->flags
& REQ_F_REISSUE
) {
2992 req
->flags
&= ~REQ_F_REISSUE
;
2993 if (io_resubmit_prep(req
)) {
2994 io_req_task_queue_reissue(req
);
2996 unsigned int cflags
= io_put_rw_kbuf(req
);
2997 struct io_ring_ctx
*ctx
= req
->ctx
;
3000 if (!(issue_flags
& IO_URING_F_NONBLOCK
)) {
3001 mutex_lock(&ctx
->uring_lock
);
3002 __io_req_complete(req
, issue_flags
, ret
, cflags
);
3003 mutex_unlock(&ctx
->uring_lock
);
3005 __io_req_complete(req
, issue_flags
, ret
, cflags
);
3011 static int __io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
3012 struct io_mapped_ubuf
*imu
)
3014 size_t len
= req
->rw
.len
;
3015 u64 buf_end
, buf_addr
= req
->rw
.addr
;
3018 if (unlikely(check_add_overflow(buf_addr
, (u64
)len
, &buf_end
)))
3020 /* not inside the mapped region */
3021 if (unlikely(buf_addr
< imu
->ubuf
|| buf_end
> imu
->ubuf_end
))
3025 * May not be a start of buffer, set size appropriately
3026 * and advance us to the beginning.
3028 offset
= buf_addr
- imu
->ubuf
;
3029 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3033 * Don't use iov_iter_advance() here, as it's really slow for
3034 * using the latter parts of a big fixed buffer - it iterates
3035 * over each segment manually. We can cheat a bit here, because
3038 * 1) it's a BVEC iter, we set it up
3039 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3040 * first and last bvec
3042 * So just find our index, and adjust the iterator afterwards.
3043 * If the offset is within the first bvec (or the whole first
3044 * bvec, just use iov_iter_advance(). This makes it easier
3045 * since we can just skip the first segment, which may not
3046 * be PAGE_SIZE aligned.
3048 const struct bio_vec
*bvec
= imu
->bvec
;
3050 if (offset
<= bvec
->bv_len
) {
3051 iov_iter_advance(iter
, offset
);
3053 unsigned long seg_skip
;
3055 /* skip first vec */
3056 offset
-= bvec
->bv_len
;
3057 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3059 iter
->bvec
= bvec
+ seg_skip
;
3060 iter
->nr_segs
-= seg_skip
;
3061 iter
->count
-= bvec
->bv_len
+ offset
;
3062 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3069 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
3071 struct io_ring_ctx
*ctx
= req
->ctx
;
3072 struct io_mapped_ubuf
*imu
= req
->imu
;
3073 u16 index
, buf_index
= req
->buf_index
;
3076 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3078 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3079 imu
= READ_ONCE(ctx
->user_bufs
[index
]);
3082 return __io_import_fixed(req
, rw
, iter
, imu
);
3085 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3088 mutex_unlock(&ctx
->uring_lock
);
3091 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3094 * "Normal" inline submissions always hold the uring_lock, since we
3095 * grab it from the system call. Same is true for the SQPOLL offload.
3096 * The only exception is when we've detached the request and issue it
3097 * from an async worker thread, grab the lock for that case.
3100 mutex_lock(&ctx
->uring_lock
);
3103 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3104 int bgid
, struct io_buffer
*kbuf
,
3107 struct io_buffer
*head
;
3109 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3112 io_ring_submit_lock(req
->ctx
, needs_lock
);
3114 lockdep_assert_held(&req
->ctx
->uring_lock
);
3116 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
3118 if (!list_empty(&head
->list
)) {
3119 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3121 list_del(&kbuf
->list
);
3124 xa_erase(&req
->ctx
->io_buffers
, bgid
);
3126 if (*len
> kbuf
->len
)
3129 kbuf
= ERR_PTR(-ENOBUFS
);
3132 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3137 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3140 struct io_buffer
*kbuf
;
3143 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3144 bgid
= req
->buf_index
;
3145 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3148 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3149 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3150 return u64_to_user_ptr(kbuf
->addr
);
3153 #ifdef CONFIG_COMPAT
3154 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3157 struct compat_iovec __user
*uiov
;
3158 compat_ssize_t clen
;
3162 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3163 if (!access_ok(uiov
, sizeof(*uiov
)))
3165 if (__get_user(clen
, &uiov
->iov_len
))
3171 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3173 return PTR_ERR(buf
);
3174 iov
[0].iov_base
= buf
;
3175 iov
[0].iov_len
= (compat_size_t
) len
;
3180 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3183 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3187 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3190 len
= iov
[0].iov_len
;
3193 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3195 return PTR_ERR(buf
);
3196 iov
[0].iov_base
= buf
;
3197 iov
[0].iov_len
= len
;
3201 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3204 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3205 struct io_buffer
*kbuf
;
3207 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3208 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3209 iov
[0].iov_len
= kbuf
->len
;
3212 if (req
->rw
.len
!= 1)
3215 #ifdef CONFIG_COMPAT
3216 if (req
->ctx
->compat
)
3217 return io_compat_import(req
, iov
, needs_lock
);
3220 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3223 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
3224 struct iov_iter
*iter
, bool needs_lock
)
3226 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3227 size_t sqe_len
= req
->rw
.len
;
3228 u8 opcode
= req
->opcode
;
3231 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3233 return io_import_fixed(req
, rw
, iter
);
3236 /* buffer index only valid with fixed read/write, or buffer select */
3237 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3240 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3241 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3242 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3244 return PTR_ERR(buf
);
3245 req
->rw
.len
= sqe_len
;
3248 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3253 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3254 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3256 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3261 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3265 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3267 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3271 * For files that don't have ->read_iter() and ->write_iter(), handle them
3272 * by looping over ->read() or ->write() manually.
3274 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3276 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3277 struct file
*file
= req
->file
;
3281 * Don't support polled IO through this interface, and we can't
3282 * support non-blocking either. For the latter, this just causes
3283 * the kiocb to be handled from an async context.
3285 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3287 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3290 while (iov_iter_count(iter
)) {
3294 if (!iov_iter_is_bvec(iter
)) {
3295 iovec
= iov_iter_iovec(iter
);
3297 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3298 iovec
.iov_len
= req
->rw
.len
;
3302 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3303 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3305 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3306 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3315 if (!iov_iter_is_bvec(iter
)) {
3316 iov_iter_advance(iter
, nr
);
3323 if (nr
!= iovec
.iov_len
)
3330 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3331 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3333 struct io_async_rw
*rw
= req
->async_data
;
3335 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3336 rw
->free_iovec
= iovec
;
3338 /* can only be fixed buffers, no need to do anything */
3339 if (iov_iter_is_bvec(iter
))
3342 unsigned iov_off
= 0;
3344 rw
->iter
.iov
= rw
->fast_iov
;
3345 if (iter
->iov
!= fast_iov
) {
3346 iov_off
= iter
->iov
- fast_iov
;
3347 rw
->iter
.iov
+= iov_off
;
3349 if (rw
->fast_iov
!= fast_iov
)
3350 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3351 sizeof(struct iovec
) * iter
->nr_segs
);
3353 req
->flags
|= REQ_F_NEED_CLEANUP
;
3357 static inline int io_alloc_async_data(struct io_kiocb
*req
)
3359 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3360 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3361 return req
->async_data
== NULL
;
3364 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3365 const struct iovec
*fast_iov
,
3366 struct iov_iter
*iter
, bool force
)
3368 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_setup
)
3370 if (!req
->async_data
) {
3371 struct io_async_rw
*iorw
;
3373 if (io_alloc_async_data(req
)) {
3378 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3379 iorw
= req
->async_data
;
3380 /* we've copied and mapped the iter, ensure state is saved */
3381 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3386 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3388 struct io_async_rw
*iorw
= req
->async_data
;
3389 struct iovec
*iov
= iorw
->fast_iov
;
3392 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3393 if (unlikely(ret
< 0))
3396 iorw
->bytes_done
= 0;
3397 iorw
->free_iovec
= iov
;
3399 req
->flags
|= REQ_F_NEED_CLEANUP
;
3400 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3404 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3406 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3408 return io_prep_rw(req
, sqe
, READ
);
3412 * This is our waitqueue callback handler, registered through lock_page_async()
3413 * when we initially tried to do the IO with the iocb armed our waitqueue.
3414 * This gets called when the page is unlocked, and we generally expect that to
3415 * happen when the page IO is completed and the page is now uptodate. This will
3416 * queue a task_work based retry of the operation, attempting to copy the data
3417 * again. If the latter fails because the page was NOT uptodate, then we will
3418 * do a thread based blocking retry of the operation. That's the unexpected
3421 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3422 int sync
, void *arg
)
3424 struct wait_page_queue
*wpq
;
3425 struct io_kiocb
*req
= wait
->private;
3426 struct wait_page_key
*key
= arg
;
3428 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3430 if (!wake_page_match(wpq
, key
))
3433 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3434 list_del_init(&wait
->entry
);
3435 io_req_task_queue(req
);
3440 * This controls whether a given IO request should be armed for async page
3441 * based retry. If we return false here, the request is handed to the async
3442 * worker threads for retry. If we're doing buffered reads on a regular file,
3443 * we prepare a private wait_page_queue entry and retry the operation. This
3444 * will either succeed because the page is now uptodate and unlocked, or it
3445 * will register a callback when the page is unlocked at IO completion. Through
3446 * that callback, io_uring uses task_work to setup a retry of the operation.
3447 * That retry will attempt the buffered read again. The retry will generally
3448 * succeed, or in rare cases where it fails, we then fall back to using the
3449 * async worker threads for a blocking retry.
3451 static bool io_rw_should_retry(struct io_kiocb
*req
)
3453 struct io_async_rw
*rw
= req
->async_data
;
3454 struct wait_page_queue
*wait
= &rw
->wpq
;
3455 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3457 /* never retry for NOWAIT, we just complete with -EAGAIN */
3458 if (req
->flags
& REQ_F_NOWAIT
)
3461 /* Only for buffered IO */
3462 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3466 * just use poll if we can, and don't attempt if the fs doesn't
3467 * support callback based unlocks
3469 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3472 wait
->wait
.func
= io_async_buf_func
;
3473 wait
->wait
.private = req
;
3474 wait
->wait
.flags
= 0;
3475 INIT_LIST_HEAD(&wait
->wait
.entry
);
3476 kiocb
->ki_flags
|= IOCB_WAITQ
;
3477 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3478 kiocb
->ki_waitq
= wait
;
3482 static inline int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3484 if (req
->file
->f_op
->read_iter
)
3485 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3486 else if (req
->file
->f_op
->read
)
3487 return loop_rw_iter(READ
, req
, iter
);
3492 static bool need_read_all(struct io_kiocb
*req
)
3494 return req
->flags
& REQ_F_ISREG
||
3495 S_ISBLK(file_inode(req
->file
)->i_mode
);
3498 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3500 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3501 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3502 struct iov_iter __iter
, *iter
= &__iter
;
3503 struct io_async_rw
*rw
= req
->async_data
;
3504 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3505 struct iov_iter_state __state
, *state
;
3510 state
= &rw
->iter_state
;
3512 * We come here from an earlier attempt, restore our state to
3513 * match in case it doesn't. It's cheap enough that we don't
3514 * need to make this conditional.
3516 iov_iter_restore(iter
, state
);
3519 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3523 iov_iter_save_state(iter
, state
);
3525 req
->result
= iov_iter_count(iter
);
3527 /* Ensure we clear previously set non-block flag */
3528 if (!force_nonblock
)
3529 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3531 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3533 /* If the file doesn't support async, just async punt */
3534 if (force_nonblock
&& !io_file_supports_nowait(req
, READ
)) {
3535 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3536 return ret
?: -EAGAIN
;
3539 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3540 if (unlikely(ret
)) {
3545 ret
= io_iter_do_read(req
, iter
);
3547 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
3548 req
->flags
&= ~REQ_F_REISSUE
;
3549 /* IOPOLL retry should happen for io-wq threads */
3550 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3552 /* no retry on NONBLOCK nor RWF_NOWAIT */
3553 if (req
->flags
& REQ_F_NOWAIT
)
3556 } else if (ret
== -EIOCBQUEUED
) {
3558 } else if (ret
<= 0 || ret
== req
->result
|| !force_nonblock
||
3559 (req
->flags
& REQ_F_NOWAIT
) || !need_read_all(req
)) {
3560 /* read all, failed, already did sync or don't want to retry */
3565 * Don't depend on the iter state matching what was consumed, or being
3566 * untouched in case of error. Restore it and we'll advance it
3567 * manually if we need to.
3569 iov_iter_restore(iter
, state
);
3571 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3576 rw
= req
->async_data
;
3578 * Now use our persistent iterator and state, if we aren't already.
3579 * We've restored and mapped the iter to match.
3581 if (iter
!= &rw
->iter
) {
3583 state
= &rw
->iter_state
;
3588 * We end up here because of a partial read, either from
3589 * above or inside this loop. Advance the iter by the bytes
3590 * that were consumed.
3592 iov_iter_advance(iter
, ret
);
3593 if (!iov_iter_count(iter
))
3595 rw
->bytes_done
+= ret
;
3596 iov_iter_save_state(iter
, state
);
3598 /* if we can retry, do so with the callbacks armed */
3599 if (!io_rw_should_retry(req
)) {
3600 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3605 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3606 * we get -EIOCBQUEUED, then we'll get a notification when the
3607 * desired page gets unlocked. We can also get a partial read
3608 * here, and if we do, then just retry at the new offset.
3610 ret
= io_iter_do_read(req
, iter
);
3611 if (ret
== -EIOCBQUEUED
)
3613 /* we got some bytes, but not all. retry. */
3614 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3615 iov_iter_restore(iter
, state
);
3618 kiocb_done(kiocb
, ret
, issue_flags
);
3620 /* it's faster to check here then delegate to kfree */
3626 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3628 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3630 return io_prep_rw(req
, sqe
, WRITE
);
3633 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3635 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3636 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3637 struct iov_iter __iter
, *iter
= &__iter
;
3638 struct io_async_rw
*rw
= req
->async_data
;
3639 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3640 struct iov_iter_state __state
, *state
;
3645 state
= &rw
->iter_state
;
3646 iov_iter_restore(iter
, state
);
3649 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3653 iov_iter_save_state(iter
, state
);
3655 req
->result
= iov_iter_count(iter
);
3657 /* Ensure we clear previously set non-block flag */
3658 if (!force_nonblock
)
3659 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3661 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3663 /* If the file doesn't support async, just async punt */
3664 if (force_nonblock
&& !io_file_supports_nowait(req
, WRITE
))
3667 /* file path doesn't support NOWAIT for non-direct_IO */
3668 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3669 (req
->flags
& REQ_F_ISREG
))
3672 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3677 * Open-code file_start_write here to grab freeze protection,
3678 * which will be released by another thread in
3679 * io_complete_rw(). Fool lockdep by telling it the lock got
3680 * released so that it doesn't complain about the held lock when
3681 * we return to userspace.
3683 if (req
->flags
& REQ_F_ISREG
) {
3684 sb_start_write(file_inode(req
->file
)->i_sb
);
3685 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3688 kiocb
->ki_flags
|= IOCB_WRITE
;
3690 if (req
->file
->f_op
->write_iter
)
3691 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3692 else if (req
->file
->f_op
->write
)
3693 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3697 if (req
->flags
& REQ_F_REISSUE
) {
3698 req
->flags
&= ~REQ_F_REISSUE
;
3703 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3704 * retry them without IOCB_NOWAIT.
3706 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3708 /* no retry on NONBLOCK nor RWF_NOWAIT */
3709 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3711 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3712 /* IOPOLL retry should happen for io-wq threads */
3713 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3716 kiocb_done(kiocb
, ret2
, issue_flags
);
3719 iov_iter_restore(iter
, state
);
3720 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3721 return ret
?: -EAGAIN
;
3724 /* it's reportedly faster than delegating the null check to kfree() */
3730 static int io_renameat_prep(struct io_kiocb
*req
,
3731 const struct io_uring_sqe
*sqe
)
3733 struct io_rename
*ren
= &req
->rename
;
3734 const char __user
*oldf
, *newf
;
3736 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3738 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3740 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3743 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3744 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3745 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3746 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3747 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3749 ren
->oldpath
= getname(oldf
);
3750 if (IS_ERR(ren
->oldpath
))
3751 return PTR_ERR(ren
->oldpath
);
3753 ren
->newpath
= getname(newf
);
3754 if (IS_ERR(ren
->newpath
)) {
3755 putname(ren
->oldpath
);
3756 return PTR_ERR(ren
->newpath
);
3759 req
->flags
|= REQ_F_NEED_CLEANUP
;
3763 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3765 struct io_rename
*ren
= &req
->rename
;
3768 if (issue_flags
& IO_URING_F_NONBLOCK
)
3771 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3772 ren
->newpath
, ren
->flags
);
3774 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3777 io_req_complete(req
, ret
);
3781 static int io_unlinkat_prep(struct io_kiocb
*req
,
3782 const struct io_uring_sqe
*sqe
)
3784 struct io_unlink
*un
= &req
->unlink
;
3785 const char __user
*fname
;
3787 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3789 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
3792 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3795 un
->dfd
= READ_ONCE(sqe
->fd
);
3797 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3798 if (un
->flags
& ~AT_REMOVEDIR
)
3801 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3802 un
->filename
= getname(fname
);
3803 if (IS_ERR(un
->filename
))
3804 return PTR_ERR(un
->filename
);
3806 req
->flags
|= REQ_F_NEED_CLEANUP
;
3810 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3812 struct io_unlink
*un
= &req
->unlink
;
3815 if (issue_flags
& IO_URING_F_NONBLOCK
)
3818 if (un
->flags
& AT_REMOVEDIR
)
3819 ret
= do_rmdir(un
->dfd
, un
->filename
);
3821 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3823 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3826 io_req_complete(req
, ret
);
3830 static int io_mkdirat_prep(struct io_kiocb
*req
,
3831 const struct io_uring_sqe
*sqe
)
3833 struct io_mkdir
*mkd
= &req
->mkdir
;
3834 const char __user
*fname
;
3836 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3838 if (sqe
->ioprio
|| sqe
->off
|| sqe
->rw_flags
|| sqe
->buf_index
||
3841 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3844 mkd
->dfd
= READ_ONCE(sqe
->fd
);
3845 mkd
->mode
= READ_ONCE(sqe
->len
);
3847 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3848 mkd
->filename
= getname(fname
);
3849 if (IS_ERR(mkd
->filename
))
3850 return PTR_ERR(mkd
->filename
);
3852 req
->flags
|= REQ_F_NEED_CLEANUP
;
3856 static int io_mkdirat(struct io_kiocb
*req
, int issue_flags
)
3858 struct io_mkdir
*mkd
= &req
->mkdir
;
3861 if (issue_flags
& IO_URING_F_NONBLOCK
)
3864 ret
= do_mkdirat(mkd
->dfd
, mkd
->filename
, mkd
->mode
);
3866 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3869 io_req_complete(req
, ret
);
3873 static int io_symlinkat_prep(struct io_kiocb
*req
,
3874 const struct io_uring_sqe
*sqe
)
3876 struct io_symlink
*sl
= &req
->symlink
;
3877 const char __user
*oldpath
, *newpath
;
3879 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3881 if (sqe
->ioprio
|| sqe
->len
|| sqe
->rw_flags
|| sqe
->buf_index
||
3884 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3887 sl
->new_dfd
= READ_ONCE(sqe
->fd
);
3888 oldpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3889 newpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3891 sl
->oldpath
= getname(oldpath
);
3892 if (IS_ERR(sl
->oldpath
))
3893 return PTR_ERR(sl
->oldpath
);
3895 sl
->newpath
= getname(newpath
);
3896 if (IS_ERR(sl
->newpath
)) {
3897 putname(sl
->oldpath
);
3898 return PTR_ERR(sl
->newpath
);
3901 req
->flags
|= REQ_F_NEED_CLEANUP
;
3905 static int io_symlinkat(struct io_kiocb
*req
, int issue_flags
)
3907 struct io_symlink
*sl
= &req
->symlink
;
3910 if (issue_flags
& IO_URING_F_NONBLOCK
)
3913 ret
= do_symlinkat(sl
->oldpath
, sl
->new_dfd
, sl
->newpath
);
3915 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3918 io_req_complete(req
, ret
);
3922 static int io_linkat_prep(struct io_kiocb
*req
,
3923 const struct io_uring_sqe
*sqe
)
3925 struct io_hardlink
*lnk
= &req
->hardlink
;
3926 const char __user
*oldf
, *newf
;
3928 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3930 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3932 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3935 lnk
->old_dfd
= READ_ONCE(sqe
->fd
);
3936 lnk
->new_dfd
= READ_ONCE(sqe
->len
);
3937 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3938 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3939 lnk
->flags
= READ_ONCE(sqe
->hardlink_flags
);
3941 lnk
->oldpath
= getname(oldf
);
3942 if (IS_ERR(lnk
->oldpath
))
3943 return PTR_ERR(lnk
->oldpath
);
3945 lnk
->newpath
= getname(newf
);
3946 if (IS_ERR(lnk
->newpath
)) {
3947 putname(lnk
->oldpath
);
3948 return PTR_ERR(lnk
->newpath
);
3951 req
->flags
|= REQ_F_NEED_CLEANUP
;
3955 static int io_linkat(struct io_kiocb
*req
, int issue_flags
)
3957 struct io_hardlink
*lnk
= &req
->hardlink
;
3960 if (issue_flags
& IO_URING_F_NONBLOCK
)
3963 ret
= do_linkat(lnk
->old_dfd
, lnk
->oldpath
, lnk
->new_dfd
,
3964 lnk
->newpath
, lnk
->flags
);
3966 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3969 io_req_complete(req
, ret
);
3973 static int io_shutdown_prep(struct io_kiocb
*req
,
3974 const struct io_uring_sqe
*sqe
)
3976 #if defined(CONFIG_NET)
3977 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3979 if (unlikely(sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3980 sqe
->buf_index
|| sqe
->splice_fd_in
))
3983 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3990 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3992 #if defined(CONFIG_NET)
3993 struct socket
*sock
;
3996 if (issue_flags
& IO_URING_F_NONBLOCK
)
3999 sock
= sock_from_file(req
->file
);
4000 if (unlikely(!sock
))
4003 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
4006 io_req_complete(req
, ret
);
4013 static int __io_splice_prep(struct io_kiocb
*req
,
4014 const struct io_uring_sqe
*sqe
)
4016 struct io_splice
*sp
= &req
->splice
;
4017 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
4019 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4022 sp
->len
= READ_ONCE(sqe
->len
);
4023 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
4024 if (unlikely(sp
->flags
& ~valid_flags
))
4026 sp
->splice_fd_in
= READ_ONCE(sqe
->splice_fd_in
);
4030 static int io_tee_prep(struct io_kiocb
*req
,
4031 const struct io_uring_sqe
*sqe
)
4033 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
4035 return __io_splice_prep(req
, sqe
);
4038 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
4040 struct io_splice
*sp
= &req
->splice
;
4041 struct file
*out
= sp
->file_out
;
4042 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4046 if (issue_flags
& IO_URING_F_NONBLOCK
)
4049 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4050 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4057 ret
= do_tee(in
, out
, sp
->len
, flags
);
4059 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4064 io_req_complete(req
, ret
);
4068 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4070 struct io_splice
*sp
= &req
->splice
;
4072 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
4073 sp
->off_out
= READ_ONCE(sqe
->off
);
4074 return __io_splice_prep(req
, sqe
);
4077 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
4079 struct io_splice
*sp
= &req
->splice
;
4080 struct file
*out
= sp
->file_out
;
4081 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4082 loff_t
*poff_in
, *poff_out
;
4086 if (issue_flags
& IO_URING_F_NONBLOCK
)
4089 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4090 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4096 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
4097 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
4100 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
4102 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4107 io_req_complete(req
, ret
);
4112 * IORING_OP_NOP just posts a completion event, nothing else.
4114 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
4116 struct io_ring_ctx
*ctx
= req
->ctx
;
4118 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4121 __io_req_complete(req
, issue_flags
, 0, 0);
4125 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4127 struct io_ring_ctx
*ctx
= req
->ctx
;
4129 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4131 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4135 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
4136 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
4139 req
->sync
.off
= READ_ONCE(sqe
->off
);
4140 req
->sync
.len
= READ_ONCE(sqe
->len
);
4144 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
4146 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
4149 /* fsync always requires a blocking context */
4150 if (issue_flags
& IO_URING_F_NONBLOCK
)
4153 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
4154 end
> 0 ? end
: LLONG_MAX
,
4155 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
4158 io_req_complete(req
, ret
);
4162 static int io_fallocate_prep(struct io_kiocb
*req
,
4163 const struct io_uring_sqe
*sqe
)
4165 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
||
4168 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4171 req
->sync
.off
= READ_ONCE(sqe
->off
);
4172 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4173 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4177 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
4181 /* fallocate always requiring blocking context */
4182 if (issue_flags
& IO_URING_F_NONBLOCK
)
4184 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4188 io_req_complete(req
, ret
);
4192 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4194 const char __user
*fname
;
4197 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4199 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4201 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4204 /* open.how should be already initialised */
4205 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4206 req
->open
.how
.flags
|= O_LARGEFILE
;
4208 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4209 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4210 req
->open
.filename
= getname(fname
);
4211 if (IS_ERR(req
->open
.filename
)) {
4212 ret
= PTR_ERR(req
->open
.filename
);
4213 req
->open
.filename
= NULL
;
4217 req
->open
.file_slot
= READ_ONCE(sqe
->file_index
);
4218 if (req
->open
.file_slot
&& (req
->open
.how
.flags
& O_CLOEXEC
))
4221 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4222 req
->flags
|= REQ_F_NEED_CLEANUP
;
4226 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4228 u64 mode
= READ_ONCE(sqe
->len
);
4229 u64 flags
= READ_ONCE(sqe
->open_flags
);
4231 req
->open
.how
= build_open_how(flags
, mode
);
4232 return __io_openat_prep(req
, sqe
);
4235 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4237 struct open_how __user
*how
;
4241 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4242 len
= READ_ONCE(sqe
->len
);
4243 if (len
< OPEN_HOW_SIZE_VER0
)
4246 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4251 return __io_openat_prep(req
, sqe
);
4254 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
4256 struct open_flags op
;
4258 bool resolve_nonblock
, nonblock_set
;
4259 bool fixed
= !!req
->open
.file_slot
;
4262 ret
= build_open_flags(&req
->open
.how
, &op
);
4265 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
4266 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
4267 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4269 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4270 * it'll always -EAGAIN
4272 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
4274 op
.lookup_flags
|= LOOKUP_CACHED
;
4275 op
.open_flag
|= O_NONBLOCK
;
4279 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4284 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4287 * We could hang on to this 'fd' on retrying, but seems like
4288 * marginal gain for something that is now known to be a slower
4289 * path. So just put it, and we'll get a new one when we retry.
4294 ret
= PTR_ERR(file
);
4295 /* only retry if RESOLVE_CACHED wasn't already set by application */
4296 if (ret
== -EAGAIN
&&
4297 (!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)))
4302 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
4303 file
->f_flags
&= ~O_NONBLOCK
;
4304 fsnotify_open(file
);
4307 fd_install(ret
, file
);
4309 ret
= io_install_fixed_file(req
, file
, issue_flags
,
4310 req
->open
.file_slot
- 1);
4312 putname(req
->open
.filename
);
4313 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4316 __io_req_complete(req
, issue_flags
, ret
, 0);
4320 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
4322 return io_openat2(req
, issue_flags
);
4325 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4326 const struct io_uring_sqe
*sqe
)
4328 struct io_provide_buf
*p
= &req
->pbuf
;
4331 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
4335 tmp
= READ_ONCE(sqe
->fd
);
4336 if (!tmp
|| tmp
> USHRT_MAX
)
4339 memset(p
, 0, sizeof(*p
));
4341 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4345 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4346 int bgid
, unsigned nbufs
)
4350 /* shouldn't happen */
4354 /* the head kbuf is the list itself */
4355 while (!list_empty(&buf
->list
)) {
4356 struct io_buffer
*nxt
;
4358 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4359 list_del(&nxt
->list
);
4367 xa_erase(&ctx
->io_buffers
, bgid
);
4372 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4374 struct io_provide_buf
*p
= &req
->pbuf
;
4375 struct io_ring_ctx
*ctx
= req
->ctx
;
4376 struct io_buffer
*head
;
4378 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4380 io_ring_submit_lock(ctx
, !force_nonblock
);
4382 lockdep_assert_held(&ctx
->uring_lock
);
4385 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4387 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4391 /* complete before unlock, IOPOLL may need the lock */
4392 __io_req_complete(req
, issue_flags
, ret
, 0);
4393 io_ring_submit_unlock(ctx
, !force_nonblock
);
4397 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4398 const struct io_uring_sqe
*sqe
)
4400 unsigned long size
, tmp_check
;
4401 struct io_provide_buf
*p
= &req
->pbuf
;
4404 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
4407 tmp
= READ_ONCE(sqe
->fd
);
4408 if (!tmp
|| tmp
> USHRT_MAX
)
4411 p
->addr
= READ_ONCE(sqe
->addr
);
4412 p
->len
= READ_ONCE(sqe
->len
);
4414 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
4417 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
4420 size
= (unsigned long)p
->len
* p
->nbufs
;
4421 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4424 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4425 tmp
= READ_ONCE(sqe
->off
);
4426 if (tmp
> USHRT_MAX
)
4432 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4434 struct io_buffer
*buf
;
4435 u64 addr
= pbuf
->addr
;
4436 int i
, bid
= pbuf
->bid
;
4438 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4439 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL_ACCOUNT
);
4444 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
4449 INIT_LIST_HEAD(&buf
->list
);
4452 list_add_tail(&buf
->list
, &(*head
)->list
);
4457 return i
? i
: -ENOMEM
;
4460 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4462 struct io_provide_buf
*p
= &req
->pbuf
;
4463 struct io_ring_ctx
*ctx
= req
->ctx
;
4464 struct io_buffer
*head
, *list
;
4466 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4468 io_ring_submit_lock(ctx
, !force_nonblock
);
4470 lockdep_assert_held(&ctx
->uring_lock
);
4472 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4474 ret
= io_add_buffers(p
, &head
);
4475 if (ret
>= 0 && !list
) {
4476 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4478 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4482 /* complete before unlock, IOPOLL may need the lock */
4483 __io_req_complete(req
, issue_flags
, ret
, 0);
4484 io_ring_submit_unlock(ctx
, !force_nonblock
);
4488 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4489 const struct io_uring_sqe
*sqe
)
4491 #if defined(CONFIG_EPOLL)
4492 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4494 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4497 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4498 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4499 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4501 if (ep_op_has_event(req
->epoll
.op
)) {
4502 struct epoll_event __user
*ev
;
4504 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4505 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4515 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4517 #if defined(CONFIG_EPOLL)
4518 struct io_epoll
*ie
= &req
->epoll
;
4520 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4522 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4523 if (force_nonblock
&& ret
== -EAGAIN
)
4528 __io_req_complete(req
, issue_flags
, ret
, 0);
4535 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4537 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4538 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->splice_fd_in
)
4540 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4543 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4544 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4545 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4552 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4554 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4555 struct io_madvise
*ma
= &req
->madvise
;
4558 if (issue_flags
& IO_URING_F_NONBLOCK
)
4561 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4564 io_req_complete(req
, ret
);
4571 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4573 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
|| sqe
->splice_fd_in
)
4575 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4578 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4579 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4580 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4584 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4586 struct io_fadvise
*fa
= &req
->fadvise
;
4589 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4590 switch (fa
->advice
) {
4591 case POSIX_FADV_NORMAL
:
4592 case POSIX_FADV_RANDOM
:
4593 case POSIX_FADV_SEQUENTIAL
:
4600 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4603 __io_req_complete(req
, issue_flags
, ret
, 0);
4607 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4609 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4611 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4613 if (req
->flags
& REQ_F_FIXED_FILE
)
4616 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4617 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4618 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4619 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4620 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4625 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4627 struct io_statx
*ctx
= &req
->statx
;
4630 if (issue_flags
& IO_URING_F_NONBLOCK
)
4633 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4638 io_req_complete(req
, ret
);
4642 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4644 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4646 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4647 sqe
->rw_flags
|| sqe
->buf_index
)
4649 if (req
->flags
& REQ_F_FIXED_FILE
)
4652 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4653 req
->close
.file_slot
= READ_ONCE(sqe
->file_index
);
4654 if (req
->close
.file_slot
&& req
->close
.fd
)
4660 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4662 struct files_struct
*files
= current
->files
;
4663 struct io_close
*close
= &req
->close
;
4664 struct fdtable
*fdt
;
4665 struct file
*file
= NULL
;
4668 if (req
->close
.file_slot
) {
4669 ret
= io_close_fixed(req
, issue_flags
);
4673 spin_lock(&files
->file_lock
);
4674 fdt
= files_fdtable(files
);
4675 if (close
->fd
>= fdt
->max_fds
) {
4676 spin_unlock(&files
->file_lock
);
4679 file
= fdt
->fd
[close
->fd
];
4680 if (!file
|| file
->f_op
== &io_uring_fops
) {
4681 spin_unlock(&files
->file_lock
);
4686 /* if the file has a flush method, be safe and punt to async */
4687 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4688 spin_unlock(&files
->file_lock
);
4692 ret
= __close_fd_get_file(close
->fd
, &file
);
4693 spin_unlock(&files
->file_lock
);
4700 /* No ->flush() or already async, safely close from here */
4701 ret
= filp_close(file
, current
->files
);
4707 __io_req_complete(req
, issue_flags
, ret
, 0);
4711 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4713 struct io_ring_ctx
*ctx
= req
->ctx
;
4715 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4717 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4721 req
->sync
.off
= READ_ONCE(sqe
->off
);
4722 req
->sync
.len
= READ_ONCE(sqe
->len
);
4723 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4727 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4731 /* sync_file_range always requires a blocking context */
4732 if (issue_flags
& IO_URING_F_NONBLOCK
)
4735 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4739 io_req_complete(req
, ret
);
4743 #if defined(CONFIG_NET)
4744 static int io_setup_async_msg(struct io_kiocb
*req
,
4745 struct io_async_msghdr
*kmsg
)
4747 struct io_async_msghdr
*async_msg
= req
->async_data
;
4751 if (io_alloc_async_data(req
)) {
4752 kfree(kmsg
->free_iov
);
4755 async_msg
= req
->async_data
;
4756 req
->flags
|= REQ_F_NEED_CLEANUP
;
4757 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4758 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4759 /* if were using fast_iov, set it to the new one */
4760 if (!async_msg
->free_iov
)
4761 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4766 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4767 struct io_async_msghdr
*iomsg
)
4769 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4770 iomsg
->free_iov
= iomsg
->fast_iov
;
4771 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4772 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4775 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4779 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4781 req
->flags
|= REQ_F_NEED_CLEANUP
;
4785 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4787 struct io_sr_msg
*sr
= &req
->sr_msg
;
4789 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
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
))
5013 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5014 sr
->len
= READ_ONCE(sqe
->len
);
5015 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
5016 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
5017 if (sr
->msg_flags
& MSG_DONTWAIT
)
5018 req
->flags
|= REQ_F_NOWAIT
;
5020 #ifdef CONFIG_COMPAT
5021 if (req
->ctx
->compat
)
5022 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
5027 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
5029 struct io_async_msghdr iomsg
, *kmsg
;
5030 struct socket
*sock
;
5031 struct io_buffer
*kbuf
;
5034 int ret
, cflags
= 0;
5035 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5037 sock
= sock_from_file(req
->file
);
5038 if (unlikely(!sock
))
5041 kmsg
= req
->async_data
;
5043 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
5049 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5050 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5052 return PTR_ERR(kbuf
);
5053 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
5054 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
5055 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
5056 1, req
->sr_msg
.len
);
5059 flags
= req
->sr_msg
.msg_flags
;
5061 flags
|= MSG_DONTWAIT
;
5062 if (flags
& MSG_WAITALL
)
5063 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
5065 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
5066 kmsg
->uaddr
, flags
);
5067 if (force_nonblock
&& ret
== -EAGAIN
)
5068 return io_setup_async_msg(req
, kmsg
);
5069 if (ret
== -ERESTARTSYS
)
5072 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5073 cflags
= io_put_recv_kbuf(req
);
5074 /* fast path, check for non-NULL to avoid function call */
5076 kfree(kmsg
->free_iov
);
5077 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5078 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5080 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5084 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
5086 struct io_buffer
*kbuf
;
5087 struct io_sr_msg
*sr
= &req
->sr_msg
;
5089 void __user
*buf
= sr
->buf
;
5090 struct socket
*sock
;
5094 int ret
, cflags
= 0;
5095 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5097 sock
= sock_from_file(req
->file
);
5098 if (unlikely(!sock
))
5101 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5102 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5104 return PTR_ERR(kbuf
);
5105 buf
= u64_to_user_ptr(kbuf
->addr
);
5108 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
5112 msg
.msg_name
= NULL
;
5113 msg
.msg_control
= NULL
;
5114 msg
.msg_controllen
= 0;
5115 msg
.msg_namelen
= 0;
5116 msg
.msg_iocb
= NULL
;
5119 flags
= req
->sr_msg
.msg_flags
;
5121 flags
|= MSG_DONTWAIT
;
5122 if (flags
& MSG_WAITALL
)
5123 min_ret
= iov_iter_count(&msg
.msg_iter
);
5125 ret
= sock_recvmsg(sock
, &msg
, flags
);
5126 if (force_nonblock
&& ret
== -EAGAIN
)
5128 if (ret
== -ERESTARTSYS
)
5131 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5132 cflags
= io_put_recv_kbuf(req
);
5133 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5135 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5139 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5141 struct io_accept
*accept
= &req
->accept
;
5143 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5145 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
5148 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5149 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5150 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
5151 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
5153 accept
->file_slot
= READ_ONCE(sqe
->file_index
);
5154 if (accept
->file_slot
&& (accept
->flags
& SOCK_CLOEXEC
))
5156 if (accept
->flags
& ~(SOCK_CLOEXEC
| SOCK_NONBLOCK
))
5158 if (SOCK_NONBLOCK
!= O_NONBLOCK
&& (accept
->flags
& SOCK_NONBLOCK
))
5159 accept
->flags
= (accept
->flags
& ~SOCK_NONBLOCK
) | O_NONBLOCK
;
5163 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
5165 struct io_accept
*accept
= &req
->accept
;
5166 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5167 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5168 bool fixed
= !!accept
->file_slot
;
5172 if (req
->file
->f_flags
& O_NONBLOCK
)
5173 req
->flags
|= REQ_F_NOWAIT
;
5176 fd
= __get_unused_fd_flags(accept
->flags
, accept
->nofile
);
5177 if (unlikely(fd
< 0))
5180 file
= do_accept(req
->file
, file_flags
, accept
->addr
, accept
->addr_len
,
5185 ret
= PTR_ERR(file
);
5186 if (ret
== -EAGAIN
&& force_nonblock
)
5188 if (ret
== -ERESTARTSYS
)
5191 } else if (!fixed
) {
5192 fd_install(fd
, file
);
5195 ret
= io_install_fixed_file(req
, file
, issue_flags
,
5196 accept
->file_slot
- 1);
5198 __io_req_complete(req
, issue_flags
, ret
, 0);
5202 static int io_connect_prep_async(struct io_kiocb
*req
)
5204 struct io_async_connect
*io
= req
->async_data
;
5205 struct io_connect
*conn
= &req
->connect
;
5207 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
5210 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5212 struct io_connect
*conn
= &req
->connect
;
5214 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5216 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
||
5220 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5221 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5225 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
5227 struct io_async_connect __io
, *io
;
5228 unsigned file_flags
;
5230 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5232 if (req
->async_data
) {
5233 io
= req
->async_data
;
5235 ret
= move_addr_to_kernel(req
->connect
.addr
,
5236 req
->connect
.addr_len
,
5243 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5245 ret
= __sys_connect_file(req
->file
, &io
->address
,
5246 req
->connect
.addr_len
, file_flags
);
5247 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5248 if (req
->async_data
)
5250 if (io_alloc_async_data(req
)) {
5254 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5257 if (ret
== -ERESTARTSYS
)
5262 __io_req_complete(req
, issue_flags
, ret
, 0);
5265 #else /* !CONFIG_NET */
5266 #define IO_NETOP_FN(op) \
5267 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5269 return -EOPNOTSUPP; \
5272 #define IO_NETOP_PREP(op) \
5274 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5276 return -EOPNOTSUPP; \
5279 #define IO_NETOP_PREP_ASYNC(op) \
5281 static int io_##op##_prep_async(struct io_kiocb *req) \
5283 return -EOPNOTSUPP; \
5286 IO_NETOP_PREP_ASYNC(sendmsg
);
5287 IO_NETOP_PREP_ASYNC(recvmsg
);
5288 IO_NETOP_PREP_ASYNC(connect
);
5289 IO_NETOP_PREP(accept
);
5292 #endif /* CONFIG_NET */
5294 struct io_poll_table
{
5295 struct poll_table_struct pt
;
5296 struct io_kiocb
*req
;
5301 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5302 __poll_t mask
, io_req_tw_func_t func
)
5304 /* for instances that support it check for an event match first: */
5305 if (mask
&& !(mask
& poll
->events
))
5308 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5310 list_del_init(&poll
->wait
.entry
);
5313 req
->io_task_work
.func
= func
;
5316 * If this fails, then the task is exiting. When a task exits, the
5317 * work gets canceled, so just cancel this request as well instead
5318 * of executing it. We can't safely execute it anyway, as we may not
5319 * have the needed state needed for it anyway.
5321 io_req_task_work_add(req
);
5325 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5326 __acquires(&req
->ctx
->completion_lock
)
5328 struct io_ring_ctx
*ctx
= req
->ctx
;
5330 /* req->task == current here, checking PF_EXITING is safe */
5331 if (unlikely(req
->task
->flags
& PF_EXITING
))
5332 WRITE_ONCE(poll
->canceled
, true);
5334 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5335 struct poll_table_struct pt
= { ._key
= poll
->events
};
5337 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5340 spin_lock(&ctx
->completion_lock
);
5341 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5342 add_wait_queue(poll
->head
, &poll
->wait
);
5349 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5351 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5352 if (req
->opcode
== IORING_OP_POLL_ADD
)
5353 return req
->async_data
;
5354 return req
->apoll
->double_poll
;
5357 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5359 if (req
->opcode
== IORING_OP_POLL_ADD
)
5361 return &req
->apoll
->poll
;
5364 static void io_poll_remove_double(struct io_kiocb
*req
)
5365 __must_hold(&req
->ctx
->completion_lock
)
5367 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5369 lockdep_assert_held(&req
->ctx
->completion_lock
);
5371 if (poll
&& poll
->head
) {
5372 struct wait_queue_head
*head
= poll
->head
;
5374 spin_lock_irq(&head
->lock
);
5375 list_del_init(&poll
->wait
.entry
);
5376 if (poll
->wait
.private)
5379 spin_unlock_irq(&head
->lock
);
5383 static bool __io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
5384 __must_hold(&req
->ctx
->completion_lock
)
5386 struct io_ring_ctx
*ctx
= req
->ctx
;
5387 unsigned flags
= IORING_CQE_F_MORE
;
5390 if (READ_ONCE(req
->poll
.canceled
)) {
5392 req
->poll
.events
|= EPOLLONESHOT
;
5394 error
= mangle_poll(mask
);
5396 if (req
->poll
.events
& EPOLLONESHOT
)
5398 if (!io_cqring_fill_event(ctx
, req
->user_data
, error
, flags
)) {
5399 req
->poll
.events
|= EPOLLONESHOT
;
5402 if (flags
& IORING_CQE_F_MORE
)
5405 return !(flags
& IORING_CQE_F_MORE
);
5408 static inline bool io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
5409 __must_hold(&req
->ctx
->completion_lock
)
5413 done
= __io_poll_complete(req
, mask
);
5414 io_commit_cqring(req
->ctx
);
5418 static void io_poll_task_func(struct io_kiocb
*req
, bool *locked
)
5420 struct io_ring_ctx
*ctx
= req
->ctx
;
5421 struct io_kiocb
*nxt
;
5423 if (io_poll_rewait(req
, &req
->poll
)) {
5424 spin_unlock(&ctx
->completion_lock
);
5428 if (req
->poll
.done
) {
5429 spin_unlock(&ctx
->completion_lock
);
5432 done
= __io_poll_complete(req
, req
->result
);
5434 io_poll_remove_double(req
);
5435 hash_del(&req
->hash_node
);
5436 req
->poll
.done
= true;
5439 add_wait_queue(req
->poll
.head
, &req
->poll
.wait
);
5441 io_commit_cqring(ctx
);
5442 spin_unlock(&ctx
->completion_lock
);
5443 io_cqring_ev_posted(ctx
);
5446 nxt
= io_put_req_find_next(req
);
5448 io_req_task_submit(nxt
, locked
);
5453 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5454 int sync
, void *key
)
5456 struct io_kiocb
*req
= wait
->private;
5457 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5458 __poll_t mask
= key_to_poll(key
);
5459 unsigned long flags
;
5461 /* for instances that support it check for an event match first: */
5462 if (mask
&& !(mask
& poll
->events
))
5464 if (!(poll
->events
& EPOLLONESHOT
))
5465 return poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5467 list_del_init(&wait
->entry
);
5472 spin_lock_irqsave(&poll
->head
->lock
, flags
);
5473 done
= list_empty(&poll
->wait
.entry
);
5475 list_del_init(&poll
->wait
.entry
);
5476 /* make sure double remove sees this as being gone */
5477 wait
->private = NULL
;
5478 spin_unlock_irqrestore(&poll
->head
->lock
, flags
);
5480 /* use wait func handler, so it matches the rq type */
5481 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5488 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5489 wait_queue_func_t wake_func
)
5493 poll
->canceled
= false;
5494 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5495 /* mask in events that we always want/need */
5496 poll
->events
= events
| IO_POLL_UNMASK
;
5497 INIT_LIST_HEAD(&poll
->wait
.entry
);
5498 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5501 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5502 struct wait_queue_head
*head
,
5503 struct io_poll_iocb
**poll_ptr
)
5505 struct io_kiocb
*req
= pt
->req
;
5508 * The file being polled uses multiple waitqueues for poll handling
5509 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5512 if (unlikely(pt
->nr_entries
)) {
5513 struct io_poll_iocb
*poll_one
= poll
;
5515 /* double add on the same waitqueue head, ignore */
5516 if (poll_one
->head
== head
)
5518 /* already have a 2nd entry, fail a third attempt */
5520 if ((*poll_ptr
)->head
== head
)
5522 pt
->error
= -EINVAL
;
5526 * Can't handle multishot for double wait for now, turn it
5527 * into one-shot mode.
5529 if (!(poll_one
->events
& EPOLLONESHOT
))
5530 poll_one
->events
|= EPOLLONESHOT
;
5531 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5533 pt
->error
= -ENOMEM
;
5536 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5538 poll
->wait
.private = req
;
5545 if (poll
->events
& EPOLLEXCLUSIVE
)
5546 add_wait_queue_exclusive(head
, &poll
->wait
);
5548 add_wait_queue(head
, &poll
->wait
);
5551 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5552 struct poll_table_struct
*p
)
5554 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5555 struct async_poll
*apoll
= pt
->req
->apoll
;
5557 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5560 static void io_async_task_func(struct io_kiocb
*req
, bool *locked
)
5562 struct async_poll
*apoll
= req
->apoll
;
5563 struct io_ring_ctx
*ctx
= req
->ctx
;
5565 trace_io_uring_task_run(req
->ctx
, req
, req
->opcode
, req
->user_data
);
5567 if (io_poll_rewait(req
, &apoll
->poll
)) {
5568 spin_unlock(&ctx
->completion_lock
);
5572 hash_del(&req
->hash_node
);
5573 io_poll_remove_double(req
);
5574 apoll
->poll
.done
= true;
5575 spin_unlock(&ctx
->completion_lock
);
5577 if (!READ_ONCE(apoll
->poll
.canceled
))
5578 io_req_task_submit(req
, locked
);
5580 io_req_complete_failed(req
, -ECANCELED
);
5583 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5586 struct io_kiocb
*req
= wait
->private;
5587 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5589 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5592 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5595 static void io_poll_req_insert(struct io_kiocb
*req
)
5597 struct io_ring_ctx
*ctx
= req
->ctx
;
5598 struct hlist_head
*list
;
5600 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5601 hlist_add_head(&req
->hash_node
, list
);
5604 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5605 struct io_poll_iocb
*poll
,
5606 struct io_poll_table
*ipt
, __poll_t mask
,
5607 wait_queue_func_t wake_func
)
5608 __acquires(&ctx
->completion_lock
)
5610 struct io_ring_ctx
*ctx
= req
->ctx
;
5611 bool cancel
= false;
5613 INIT_HLIST_NODE(&req
->hash_node
);
5614 io_init_poll_iocb(poll
, mask
, wake_func
);
5615 poll
->file
= req
->file
;
5616 poll
->wait
.private = req
;
5618 ipt
->pt
._key
= mask
;
5621 ipt
->nr_entries
= 0;
5623 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5624 if (unlikely(!ipt
->nr_entries
) && !ipt
->error
)
5625 ipt
->error
= -EINVAL
;
5627 spin_lock(&ctx
->completion_lock
);
5628 if (ipt
->error
|| (mask
&& (poll
->events
& EPOLLONESHOT
)))
5629 io_poll_remove_double(req
);
5630 if (likely(poll
->head
)) {
5631 spin_lock_irq(&poll
->head
->lock
);
5632 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5638 if ((mask
&& (poll
->events
& EPOLLONESHOT
)) || ipt
->error
)
5639 list_del_init(&poll
->wait
.entry
);
5641 WRITE_ONCE(poll
->canceled
, true);
5642 else if (!poll
->done
) /* actually waiting for an event */
5643 io_poll_req_insert(req
);
5644 spin_unlock_irq(&poll
->head
->lock
);
5656 static int io_arm_poll_handler(struct io_kiocb
*req
)
5658 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5659 struct io_ring_ctx
*ctx
= req
->ctx
;
5660 struct async_poll
*apoll
;
5661 struct io_poll_table ipt
;
5662 __poll_t ret
, mask
= EPOLLONESHOT
| POLLERR
| POLLPRI
;
5665 if (!req
->file
|| !file_can_poll(req
->file
))
5666 return IO_APOLL_ABORTED
;
5667 if (req
->flags
& REQ_F_POLLED
)
5668 return IO_APOLL_ABORTED
;
5669 if (!def
->pollin
&& !def
->pollout
)
5670 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
))
5682 mask
|= POLLOUT
| POLLWRNORM
;
5685 /* if we can't nonblock try, then no point in arming a poll handler */
5686 if (!io_file_supports_nowait(req
, rw
))
5687 return IO_APOLL_ABORTED
;
5689 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5690 if (unlikely(!apoll
))
5691 return IO_APOLL_ABORTED
;
5692 apoll
->double_poll
= NULL
;
5694 req
->flags
|= REQ_F_POLLED
;
5695 ipt
.pt
._qproc
= io_async_queue_proc
;
5696 io_req_set_refcount(req
);
5698 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5700 spin_unlock(&ctx
->completion_lock
);
5701 if (ret
|| ipt
.error
)
5702 return ret
? IO_APOLL_READY
: IO_APOLL_ABORTED
;
5704 trace_io_uring_poll_arm(ctx
, req
, req
->opcode
, req
->user_data
,
5705 mask
, apoll
->poll
.events
);
5709 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5710 struct io_poll_iocb
*poll
, bool do_cancel
)
5711 __must_hold(&req
->ctx
->completion_lock
)
5713 bool do_complete
= false;
5717 spin_lock_irq(&poll
->head
->lock
);
5719 WRITE_ONCE(poll
->canceled
, true);
5720 if (!list_empty(&poll
->wait
.entry
)) {
5721 list_del_init(&poll
->wait
.entry
);
5724 spin_unlock_irq(&poll
->head
->lock
);
5725 hash_del(&req
->hash_node
);
5729 static bool io_poll_remove_one(struct io_kiocb
*req
)
5730 __must_hold(&req
->ctx
->completion_lock
)
5734 io_poll_remove_double(req
);
5735 do_complete
= __io_poll_remove_one(req
, io_poll_get_single(req
), true);
5738 io_cqring_fill_event(req
->ctx
, req
->user_data
, -ECANCELED
, 0);
5739 io_commit_cqring(req
->ctx
);
5741 io_put_req_deferred(req
);
5747 * Returns true if we found and killed one or more poll requests
5749 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5752 struct hlist_node
*tmp
;
5753 struct io_kiocb
*req
;
5756 spin_lock(&ctx
->completion_lock
);
5757 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5758 struct hlist_head
*list
;
5760 list
= &ctx
->cancel_hash
[i
];
5761 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5762 if (io_match_task_safe(req
, tsk
, cancel_all
))
5763 posted
+= io_poll_remove_one(req
);
5766 spin_unlock(&ctx
->completion_lock
);
5769 io_cqring_ev_posted(ctx
);
5774 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5776 __must_hold(&ctx
->completion_lock
)
5778 struct hlist_head
*list
;
5779 struct io_kiocb
*req
;
5781 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5782 hlist_for_each_entry(req
, list
, hash_node
) {
5783 if (sqe_addr
!= req
->user_data
)
5785 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
5792 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5794 __must_hold(&ctx
->completion_lock
)
5796 struct io_kiocb
*req
;
5798 req
= io_poll_find(ctx
, sqe_addr
, poll_only
);
5801 if (io_poll_remove_one(req
))
5807 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
5812 events
= READ_ONCE(sqe
->poll32_events
);
5814 events
= swahw32(events
);
5816 if (!(flags
& IORING_POLL_ADD_MULTI
))
5817 events
|= EPOLLONESHOT
;
5818 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
5821 static int io_poll_update_prep(struct io_kiocb
*req
,
5822 const struct io_uring_sqe
*sqe
)
5824 struct io_poll_update
*upd
= &req
->poll_update
;
5827 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5829 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
5831 flags
= READ_ONCE(sqe
->len
);
5832 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
5833 IORING_POLL_ADD_MULTI
))
5835 /* meaningless without update */
5836 if (flags
== IORING_POLL_ADD_MULTI
)
5839 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
5840 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
5841 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
5843 upd
->new_user_data
= READ_ONCE(sqe
->off
);
5844 if (!upd
->update_user_data
&& upd
->new_user_data
)
5846 if (upd
->update_events
)
5847 upd
->events
= io_poll_parse_events(sqe
, flags
);
5848 else if (sqe
->poll32_events
)
5854 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5857 struct io_kiocb
*req
= wait
->private;
5858 struct io_poll_iocb
*poll
= &req
->poll
;
5860 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5863 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5864 struct poll_table_struct
*p
)
5866 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5868 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5871 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5873 struct io_poll_iocb
*poll
= &req
->poll
;
5876 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5878 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
5880 flags
= READ_ONCE(sqe
->len
);
5881 if (flags
& ~IORING_POLL_ADD_MULTI
)
5884 io_req_set_refcount(req
);
5885 poll
->events
= io_poll_parse_events(sqe
, flags
);
5889 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5891 struct io_poll_iocb
*poll
= &req
->poll
;
5892 struct io_ring_ctx
*ctx
= req
->ctx
;
5893 struct io_poll_table ipt
;
5897 ipt
.pt
._qproc
= io_poll_queue_proc
;
5899 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5902 if (mask
) { /* no async, we'd stolen it */
5904 done
= io_poll_complete(req
, mask
);
5906 spin_unlock(&ctx
->completion_lock
);
5909 io_cqring_ev_posted(ctx
);
5916 static int io_poll_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5918 struct io_ring_ctx
*ctx
= req
->ctx
;
5919 struct io_kiocb
*preq
;
5923 spin_lock(&ctx
->completion_lock
);
5924 preq
= io_poll_find(ctx
, req
->poll_update
.old_user_data
, true);
5930 if (!req
->poll_update
.update_events
&& !req
->poll_update
.update_user_data
) {
5932 ret
= io_poll_remove_one(preq
) ? 0 : -EALREADY
;
5937 * Don't allow racy completion with singleshot, as we cannot safely
5938 * update those. For multishot, if we're racing with completion, just
5939 * let completion re-add it.
5941 io_poll_remove_double(preq
);
5942 completing
= !__io_poll_remove_one(preq
, &preq
->poll
, false);
5943 if (completing
&& (preq
->poll
.events
& EPOLLONESHOT
)) {
5947 /* we now have a detached poll request. reissue. */
5951 spin_unlock(&ctx
->completion_lock
);
5953 io_req_complete(req
, ret
);
5956 /* only mask one event flags, keep behavior flags */
5957 if (req
->poll_update
.update_events
) {
5958 preq
->poll
.events
&= ~0xffff;
5959 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
5960 preq
->poll
.events
|= IO_POLL_UNMASK
;
5962 if (req
->poll_update
.update_user_data
)
5963 preq
->user_data
= req
->poll_update
.new_user_data
;
5964 spin_unlock(&ctx
->completion_lock
);
5966 /* complete update request, we're done with it */
5967 io_req_complete(req
, ret
);
5970 ret
= io_poll_add(preq
, issue_flags
);
5973 io_req_complete(preq
, ret
);
5979 static void io_req_task_timeout(struct io_kiocb
*req
, bool *locked
)
5982 io_req_complete_post(req
, -ETIME
, 0);
5985 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5987 struct io_timeout_data
*data
= container_of(timer
,
5988 struct io_timeout_data
, timer
);
5989 struct io_kiocb
*req
= data
->req
;
5990 struct io_ring_ctx
*ctx
= req
->ctx
;
5991 unsigned long flags
;
5993 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
5994 list_del_init(&req
->timeout
.list
);
5995 atomic_set(&req
->ctx
->cq_timeouts
,
5996 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5997 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
5999 req
->io_task_work
.func
= io_req_task_timeout
;
6000 io_req_task_work_add(req
);
6001 return HRTIMER_NORESTART
;
6004 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
6006 __must_hold(&ctx
->timeout_lock
)
6008 struct io_timeout_data
*io
;
6009 struct io_kiocb
*req
;
6012 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
6013 found
= user_data
== req
->user_data
;
6018 return ERR_PTR(-ENOENT
);
6020 io
= req
->async_data
;
6021 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6022 return ERR_PTR(-EALREADY
);
6023 list_del_init(&req
->timeout
.list
);
6027 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
6028 __must_hold(&ctx
->completion_lock
)
6029 __must_hold(&ctx
->timeout_lock
)
6031 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6034 return PTR_ERR(req
);
6037 io_cqring_fill_event(ctx
, req
->user_data
, -ECANCELED
, 0);
6038 io_put_req_deferred(req
);
6042 static clockid_t
io_timeout_get_clock(struct io_timeout_data
*data
)
6044 switch (data
->flags
& IORING_TIMEOUT_CLOCK_MASK
) {
6045 case IORING_TIMEOUT_BOOTTIME
:
6046 return CLOCK_BOOTTIME
;
6047 case IORING_TIMEOUT_REALTIME
:
6048 return CLOCK_REALTIME
;
6050 /* can't happen, vetted at prep time */
6054 return CLOCK_MONOTONIC
;
6058 static int io_linked_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6059 struct timespec64
*ts
, enum hrtimer_mode mode
)
6060 __must_hold(&ctx
->timeout_lock
)
6062 struct io_timeout_data
*io
;
6063 struct io_kiocb
*req
;
6066 list_for_each_entry(req
, &ctx
->ltimeout_list
, timeout
.list
) {
6067 found
= user_data
== req
->user_data
;
6074 io
= req
->async_data
;
6075 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6077 hrtimer_init(&io
->timer
, io_timeout_get_clock(io
), mode
);
6078 io
->timer
.function
= io_link_timeout_fn
;
6079 hrtimer_start(&io
->timer
, timespec64_to_ktime(*ts
), mode
);
6083 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6084 struct timespec64
*ts
, enum hrtimer_mode mode
)
6085 __must_hold(&ctx
->timeout_lock
)
6087 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6088 struct io_timeout_data
*data
;
6091 return PTR_ERR(req
);
6093 req
->timeout
.off
= 0; /* noseq */
6094 data
= req
->async_data
;
6095 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
6096 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), mode
);
6097 data
->timer
.function
= io_timeout_fn
;
6098 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
6102 static int io_timeout_remove_prep(struct io_kiocb
*req
,
6103 const struct io_uring_sqe
*sqe
)
6105 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6107 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6109 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6111 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->splice_fd_in
)
6114 tr
->ltimeout
= false;
6115 tr
->addr
= READ_ONCE(sqe
->addr
);
6116 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
6117 if (tr
->flags
& IORING_TIMEOUT_UPDATE_MASK
) {
6118 if (hweight32(tr
->flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6120 if (tr
->flags
& IORING_LINK_TIMEOUT_UPDATE
)
6121 tr
->ltimeout
= true;
6122 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE_MASK
|IORING_TIMEOUT_ABS
))
6124 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
6126 } else if (tr
->flags
) {
6127 /* timeout removal doesn't support flags */
6134 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
6136 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
6141 * Remove or update an existing timeout command
6143 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
6145 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6146 struct io_ring_ctx
*ctx
= req
->ctx
;
6149 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
)) {
6150 spin_lock(&ctx
->completion_lock
);
6151 spin_lock_irq(&ctx
->timeout_lock
);
6152 ret
= io_timeout_cancel(ctx
, tr
->addr
);
6153 spin_unlock_irq(&ctx
->timeout_lock
);
6154 spin_unlock(&ctx
->completion_lock
);
6156 enum hrtimer_mode mode
= io_translate_timeout_mode(tr
->flags
);
6158 spin_lock_irq(&ctx
->timeout_lock
);
6160 ret
= io_linked_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6162 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6163 spin_unlock_irq(&ctx
->timeout_lock
);
6168 io_req_complete_post(req
, ret
, 0);
6172 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6173 bool is_timeout_link
)
6175 struct io_timeout_data
*data
;
6177 u32 off
= READ_ONCE(sqe
->off
);
6179 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6181 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1 ||
6184 if (off
&& is_timeout_link
)
6186 flags
= READ_ONCE(sqe
->timeout_flags
);
6187 if (flags
& ~(IORING_TIMEOUT_ABS
| IORING_TIMEOUT_CLOCK_MASK
))
6189 /* more than one clock specified is invalid, obviously */
6190 if (hweight32(flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6193 INIT_LIST_HEAD(&req
->timeout
.list
);
6194 req
->timeout
.off
= off
;
6195 if (unlikely(off
&& !req
->ctx
->off_timeout_used
))
6196 req
->ctx
->off_timeout_used
= true;
6198 if (!req
->async_data
&& io_alloc_async_data(req
))
6201 data
= req
->async_data
;
6203 data
->flags
= flags
;
6205 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
6208 data
->mode
= io_translate_timeout_mode(flags
);
6209 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), data
->mode
);
6211 if (is_timeout_link
) {
6212 struct io_submit_link
*link
= &req
->ctx
->submit_state
.link
;
6216 if (link
->last
->opcode
== IORING_OP_LINK_TIMEOUT
)
6218 req
->timeout
.head
= link
->last
;
6219 link
->last
->flags
|= REQ_F_ARM_LTIMEOUT
;
6224 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
6226 struct io_ring_ctx
*ctx
= req
->ctx
;
6227 struct io_timeout_data
*data
= req
->async_data
;
6228 struct list_head
*entry
;
6229 u32 tail
, off
= req
->timeout
.off
;
6231 spin_lock_irq(&ctx
->timeout_lock
);
6234 * sqe->off holds how many events that need to occur for this
6235 * timeout event to be satisfied. If it isn't set, then this is
6236 * a pure timeout request, sequence isn't used.
6238 if (io_is_timeout_noseq(req
)) {
6239 entry
= ctx
->timeout_list
.prev
;
6243 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
6244 req
->timeout
.target_seq
= tail
+ off
;
6246 /* Update the last seq here in case io_flush_timeouts() hasn't.
6247 * This is safe because ->completion_lock is held, and submissions
6248 * and completions are never mixed in the same ->completion_lock section.
6250 ctx
->cq_last_tm_flush
= tail
;
6253 * Insertion sort, ensuring the first entry in the list is always
6254 * the one we need first.
6256 list_for_each_prev(entry
, &ctx
->timeout_list
) {
6257 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
6260 if (io_is_timeout_noseq(nxt
))
6262 /* nxt.seq is behind @tail, otherwise would've been completed */
6263 if (off
>= nxt
->timeout
.target_seq
- tail
)
6267 list_add(&req
->timeout
.list
, entry
);
6268 data
->timer
.function
= io_timeout_fn
;
6269 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
6270 spin_unlock_irq(&ctx
->timeout_lock
);
6274 struct io_cancel_data
{
6275 struct io_ring_ctx
*ctx
;
6279 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
6281 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6282 struct io_cancel_data
*cd
= data
;
6284 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
6287 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
6288 struct io_ring_ctx
*ctx
)
6290 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
6291 enum io_wq_cancel cancel_ret
;
6294 if (!tctx
|| !tctx
->io_wq
)
6297 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
6298 switch (cancel_ret
) {
6299 case IO_WQ_CANCEL_OK
:
6302 case IO_WQ_CANCEL_RUNNING
:
6305 case IO_WQ_CANCEL_NOTFOUND
:
6313 static int io_try_cancel_userdata(struct io_kiocb
*req
, u64 sqe_addr
)
6315 struct io_ring_ctx
*ctx
= req
->ctx
;
6318 WARN_ON_ONCE(!io_wq_current_is_worker() && req
->task
!= current
);
6320 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
6324 spin_lock(&ctx
->completion_lock
);
6325 spin_lock_irq(&ctx
->timeout_lock
);
6326 ret
= io_timeout_cancel(ctx
, sqe_addr
);
6327 spin_unlock_irq(&ctx
->timeout_lock
);
6330 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
6332 spin_unlock(&ctx
->completion_lock
);
6336 static int io_async_cancel_prep(struct io_kiocb
*req
,
6337 const struct io_uring_sqe
*sqe
)
6339 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6341 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6343 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
||
6347 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
6351 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
6353 struct io_ring_ctx
*ctx
= req
->ctx
;
6354 u64 sqe_addr
= req
->cancel
.addr
;
6355 struct io_tctx_node
*node
;
6358 ret
= io_try_cancel_userdata(req
, sqe_addr
);
6362 /* slow path, try all io-wq's */
6363 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6365 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
6366 struct io_uring_task
*tctx
= node
->task
->io_uring
;
6368 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
6372 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6376 io_req_complete_post(req
, ret
, 0);
6380 static int io_rsrc_update_prep(struct io_kiocb
*req
,
6381 const struct io_uring_sqe
*sqe
)
6383 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6385 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
6388 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
6389 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
6390 if (!req
->rsrc_update
.nr_args
)
6392 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
6396 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
6398 struct io_ring_ctx
*ctx
= req
->ctx
;
6399 struct io_uring_rsrc_update2 up
;
6402 up
.offset
= req
->rsrc_update
.offset
;
6403 up
.data
= req
->rsrc_update
.arg
;
6408 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6409 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
6410 &up
, req
->rsrc_update
.nr_args
);
6411 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6415 __io_req_complete(req
, issue_flags
, ret
, 0);
6419 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6421 switch (req
->opcode
) {
6424 case IORING_OP_READV
:
6425 case IORING_OP_READ_FIXED
:
6426 case IORING_OP_READ
:
6427 return io_read_prep(req
, sqe
);
6428 case IORING_OP_WRITEV
:
6429 case IORING_OP_WRITE_FIXED
:
6430 case IORING_OP_WRITE
:
6431 return io_write_prep(req
, sqe
);
6432 case IORING_OP_POLL_ADD
:
6433 return io_poll_add_prep(req
, sqe
);
6434 case IORING_OP_POLL_REMOVE
:
6435 return io_poll_update_prep(req
, sqe
);
6436 case IORING_OP_FSYNC
:
6437 return io_fsync_prep(req
, sqe
);
6438 case IORING_OP_SYNC_FILE_RANGE
:
6439 return io_sfr_prep(req
, sqe
);
6440 case IORING_OP_SENDMSG
:
6441 case IORING_OP_SEND
:
6442 return io_sendmsg_prep(req
, sqe
);
6443 case IORING_OP_RECVMSG
:
6444 case IORING_OP_RECV
:
6445 return io_recvmsg_prep(req
, sqe
);
6446 case IORING_OP_CONNECT
:
6447 return io_connect_prep(req
, sqe
);
6448 case IORING_OP_TIMEOUT
:
6449 return io_timeout_prep(req
, sqe
, false);
6450 case IORING_OP_TIMEOUT_REMOVE
:
6451 return io_timeout_remove_prep(req
, sqe
);
6452 case IORING_OP_ASYNC_CANCEL
:
6453 return io_async_cancel_prep(req
, sqe
);
6454 case IORING_OP_LINK_TIMEOUT
:
6455 return io_timeout_prep(req
, sqe
, true);
6456 case IORING_OP_ACCEPT
:
6457 return io_accept_prep(req
, sqe
);
6458 case IORING_OP_FALLOCATE
:
6459 return io_fallocate_prep(req
, sqe
);
6460 case IORING_OP_OPENAT
:
6461 return io_openat_prep(req
, sqe
);
6462 case IORING_OP_CLOSE
:
6463 return io_close_prep(req
, sqe
);
6464 case IORING_OP_FILES_UPDATE
:
6465 return io_rsrc_update_prep(req
, sqe
);
6466 case IORING_OP_STATX
:
6467 return io_statx_prep(req
, sqe
);
6468 case IORING_OP_FADVISE
:
6469 return io_fadvise_prep(req
, sqe
);
6470 case IORING_OP_MADVISE
:
6471 return io_madvise_prep(req
, sqe
);
6472 case IORING_OP_OPENAT2
:
6473 return io_openat2_prep(req
, sqe
);
6474 case IORING_OP_EPOLL_CTL
:
6475 return io_epoll_ctl_prep(req
, sqe
);
6476 case IORING_OP_SPLICE
:
6477 return io_splice_prep(req
, sqe
);
6478 case IORING_OP_PROVIDE_BUFFERS
:
6479 return io_provide_buffers_prep(req
, sqe
);
6480 case IORING_OP_REMOVE_BUFFERS
:
6481 return io_remove_buffers_prep(req
, sqe
);
6483 return io_tee_prep(req
, sqe
);
6484 case IORING_OP_SHUTDOWN
:
6485 return io_shutdown_prep(req
, sqe
);
6486 case IORING_OP_RENAMEAT
:
6487 return io_renameat_prep(req
, sqe
);
6488 case IORING_OP_UNLINKAT
:
6489 return io_unlinkat_prep(req
, sqe
);
6490 case IORING_OP_MKDIRAT
:
6491 return io_mkdirat_prep(req
, sqe
);
6492 case IORING_OP_SYMLINKAT
:
6493 return io_symlinkat_prep(req
, sqe
);
6494 case IORING_OP_LINKAT
:
6495 return io_linkat_prep(req
, sqe
);
6498 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6503 static int io_req_prep_async(struct io_kiocb
*req
)
6505 if (!io_op_defs
[req
->opcode
].needs_async_setup
)
6507 if (WARN_ON_ONCE(req
->async_data
))
6509 if (io_alloc_async_data(req
))
6512 switch (req
->opcode
) {
6513 case IORING_OP_READV
:
6514 return io_rw_prep_async(req
, READ
);
6515 case IORING_OP_WRITEV
:
6516 return io_rw_prep_async(req
, WRITE
);
6517 case IORING_OP_SENDMSG
:
6518 return io_sendmsg_prep_async(req
);
6519 case IORING_OP_RECVMSG
:
6520 return io_recvmsg_prep_async(req
);
6521 case IORING_OP_CONNECT
:
6522 return io_connect_prep_async(req
);
6524 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
6529 static u32
io_get_sequence(struct io_kiocb
*req
)
6531 u32 seq
= req
->ctx
->cached_sq_head
;
6533 /* need original cached_sq_head, but it was increased for each req */
6534 io_for_each_link(req
, req
)
6539 static bool io_drain_req(struct io_kiocb
*req
)
6541 struct io_kiocb
*pos
;
6542 struct io_ring_ctx
*ctx
= req
->ctx
;
6543 struct io_defer_entry
*de
;
6547 if (req
->flags
& REQ_F_FAIL
) {
6548 io_req_complete_fail_submit(req
);
6553 * If we need to drain a request in the middle of a link, drain the
6554 * head request and the next request/link after the current link.
6555 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6556 * maintained for every request of our link.
6558 if (ctx
->drain_next
) {
6559 req
->flags
|= REQ_F_IO_DRAIN
;
6560 ctx
->drain_next
= false;
6562 /* not interested in head, start from the first linked */
6563 io_for_each_link(pos
, req
->link
) {
6564 if (pos
->flags
& REQ_F_IO_DRAIN
) {
6565 ctx
->drain_next
= true;
6566 req
->flags
|= REQ_F_IO_DRAIN
;
6571 /* Still need defer if there is pending req in defer list. */
6572 spin_lock(&ctx
->completion_lock
);
6573 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6574 !(req
->flags
& REQ_F_IO_DRAIN
))) {
6575 spin_unlock(&ctx
->completion_lock
);
6576 ctx
->drain_active
= false;
6579 spin_unlock(&ctx
->completion_lock
);
6581 seq
= io_get_sequence(req
);
6582 /* Still a chance to pass the sequence check */
6583 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6586 ret
= io_req_prep_async(req
);
6589 io_prep_async_link(req
);
6590 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6594 io_req_complete_failed(req
, ret
);
6598 spin_lock(&ctx
->completion_lock
);
6599 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6600 spin_unlock(&ctx
->completion_lock
);
6602 io_queue_async_work(req
, NULL
);
6606 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6609 list_add_tail(&de
->list
, &ctx
->defer_list
);
6610 spin_unlock(&ctx
->completion_lock
);
6614 static void io_clean_op(struct io_kiocb
*req
)
6616 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6617 switch (req
->opcode
) {
6618 case IORING_OP_READV
:
6619 case IORING_OP_READ_FIXED
:
6620 case IORING_OP_READ
:
6621 kfree((void *)(unsigned long)req
->rw
.addr
);
6623 case IORING_OP_RECVMSG
:
6624 case IORING_OP_RECV
:
6625 kfree(req
->sr_msg
.kbuf
);
6630 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6631 switch (req
->opcode
) {
6632 case IORING_OP_READV
:
6633 case IORING_OP_READ_FIXED
:
6634 case IORING_OP_READ
:
6635 case IORING_OP_WRITEV
:
6636 case IORING_OP_WRITE_FIXED
:
6637 case IORING_OP_WRITE
: {
6638 struct io_async_rw
*io
= req
->async_data
;
6640 kfree(io
->free_iovec
);
6643 case IORING_OP_RECVMSG
:
6644 case IORING_OP_SENDMSG
: {
6645 struct io_async_msghdr
*io
= req
->async_data
;
6647 kfree(io
->free_iov
);
6650 case IORING_OP_OPENAT
:
6651 case IORING_OP_OPENAT2
:
6652 if (req
->open
.filename
)
6653 putname(req
->open
.filename
);
6655 case IORING_OP_RENAMEAT
:
6656 putname(req
->rename
.oldpath
);
6657 putname(req
->rename
.newpath
);
6659 case IORING_OP_UNLINKAT
:
6660 putname(req
->unlink
.filename
);
6662 case IORING_OP_MKDIRAT
:
6663 putname(req
->mkdir
.filename
);
6665 case IORING_OP_SYMLINKAT
:
6666 putname(req
->symlink
.oldpath
);
6667 putname(req
->symlink
.newpath
);
6669 case IORING_OP_LINKAT
:
6670 putname(req
->hardlink
.oldpath
);
6671 putname(req
->hardlink
.newpath
);
6675 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
6676 kfree(req
->apoll
->double_poll
);
6680 if (req
->flags
& REQ_F_INFLIGHT
) {
6681 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6683 atomic_dec(&tctx
->inflight_tracked
);
6685 if (req
->flags
& REQ_F_CREDS
)
6686 put_cred(req
->creds
);
6688 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
6691 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6693 struct io_ring_ctx
*ctx
= req
->ctx
;
6694 const struct cred
*creds
= NULL
;
6697 if ((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred())
6698 creds
= override_creds(req
->creds
);
6700 switch (req
->opcode
) {
6702 ret
= io_nop(req
, issue_flags
);
6704 case IORING_OP_READV
:
6705 case IORING_OP_READ_FIXED
:
6706 case IORING_OP_READ
:
6707 ret
= io_read(req
, issue_flags
);
6709 case IORING_OP_WRITEV
:
6710 case IORING_OP_WRITE_FIXED
:
6711 case IORING_OP_WRITE
:
6712 ret
= io_write(req
, issue_flags
);
6714 case IORING_OP_FSYNC
:
6715 ret
= io_fsync(req
, issue_flags
);
6717 case IORING_OP_POLL_ADD
:
6718 ret
= io_poll_add(req
, issue_flags
);
6720 case IORING_OP_POLL_REMOVE
:
6721 ret
= io_poll_update(req
, issue_flags
);
6723 case IORING_OP_SYNC_FILE_RANGE
:
6724 ret
= io_sync_file_range(req
, issue_flags
);
6726 case IORING_OP_SENDMSG
:
6727 ret
= io_sendmsg(req
, issue_flags
);
6729 case IORING_OP_SEND
:
6730 ret
= io_send(req
, issue_flags
);
6732 case IORING_OP_RECVMSG
:
6733 ret
= io_recvmsg(req
, issue_flags
);
6735 case IORING_OP_RECV
:
6736 ret
= io_recv(req
, issue_flags
);
6738 case IORING_OP_TIMEOUT
:
6739 ret
= io_timeout(req
, issue_flags
);
6741 case IORING_OP_TIMEOUT_REMOVE
:
6742 ret
= io_timeout_remove(req
, issue_flags
);
6744 case IORING_OP_ACCEPT
:
6745 ret
= io_accept(req
, issue_flags
);
6747 case IORING_OP_CONNECT
:
6748 ret
= io_connect(req
, issue_flags
);
6750 case IORING_OP_ASYNC_CANCEL
:
6751 ret
= io_async_cancel(req
, issue_flags
);
6753 case IORING_OP_FALLOCATE
:
6754 ret
= io_fallocate(req
, issue_flags
);
6756 case IORING_OP_OPENAT
:
6757 ret
= io_openat(req
, issue_flags
);
6759 case IORING_OP_CLOSE
:
6760 ret
= io_close(req
, issue_flags
);
6762 case IORING_OP_FILES_UPDATE
:
6763 ret
= io_files_update(req
, issue_flags
);
6765 case IORING_OP_STATX
:
6766 ret
= io_statx(req
, issue_flags
);
6768 case IORING_OP_FADVISE
:
6769 ret
= io_fadvise(req
, issue_flags
);
6771 case IORING_OP_MADVISE
:
6772 ret
= io_madvise(req
, issue_flags
);
6774 case IORING_OP_OPENAT2
:
6775 ret
= io_openat2(req
, issue_flags
);
6777 case IORING_OP_EPOLL_CTL
:
6778 ret
= io_epoll_ctl(req
, issue_flags
);
6780 case IORING_OP_SPLICE
:
6781 ret
= io_splice(req
, issue_flags
);
6783 case IORING_OP_PROVIDE_BUFFERS
:
6784 ret
= io_provide_buffers(req
, issue_flags
);
6786 case IORING_OP_REMOVE_BUFFERS
:
6787 ret
= io_remove_buffers(req
, issue_flags
);
6790 ret
= io_tee(req
, issue_flags
);
6792 case IORING_OP_SHUTDOWN
:
6793 ret
= io_shutdown(req
, issue_flags
);
6795 case IORING_OP_RENAMEAT
:
6796 ret
= io_renameat(req
, issue_flags
);
6798 case IORING_OP_UNLINKAT
:
6799 ret
= io_unlinkat(req
, issue_flags
);
6801 case IORING_OP_MKDIRAT
:
6802 ret
= io_mkdirat(req
, issue_flags
);
6804 case IORING_OP_SYMLINKAT
:
6805 ret
= io_symlinkat(req
, issue_flags
);
6807 case IORING_OP_LINKAT
:
6808 ret
= io_linkat(req
, issue_flags
);
6816 revert_creds(creds
);
6819 /* If the op doesn't have a file, we're not polling for it */
6820 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
6821 io_iopoll_req_issued(req
);
6826 static struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
6828 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6830 req
= io_put_req_find_next(req
);
6831 return req
? &req
->work
: NULL
;
6834 static void io_wq_submit_work(struct io_wq_work
*work
)
6836 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6837 struct io_kiocb
*timeout
;
6840 /* one will be dropped by ->io_free_work() after returning to io-wq */
6841 if (!(req
->flags
& REQ_F_REFCOUNT
))
6842 __io_req_set_refcount(req
, 2);
6846 timeout
= io_prep_linked_timeout(req
);
6848 io_queue_linked_timeout(timeout
);
6850 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6851 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6856 ret
= io_issue_sqe(req
, 0);
6858 * We can get EAGAIN for polled IO even though we're
6859 * forcing a sync submission from here, since we can't
6860 * wait for request slots on the block side.
6868 /* avoid locking problems by failing it from a clean context */
6870 io_req_task_queue_fail(req
, ret
);
6873 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
6876 return &table
->files
[i
];
6879 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6882 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
6884 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
6887 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
6889 unsigned long file_ptr
= (unsigned long) file
;
6891 if (__io_file_supports_nowait(file
, READ
))
6892 file_ptr
|= FFS_ASYNC_READ
;
6893 if (__io_file_supports_nowait(file
, WRITE
))
6894 file_ptr
|= FFS_ASYNC_WRITE
;
6895 if (S_ISREG(file_inode(file
)->i_mode
))
6896 file_ptr
|= FFS_ISREG
;
6897 file_slot
->file_ptr
= file_ptr
;
6900 static inline struct file
*io_file_get_fixed(struct io_ring_ctx
*ctx
,
6901 struct io_kiocb
*req
, int fd
)
6904 unsigned long file_ptr
;
6906 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6908 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6909 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
6910 file
= (struct file
*) (file_ptr
& FFS_MASK
);
6911 file_ptr
&= ~FFS_MASK
;
6912 /* mask in overlapping REQ_F and FFS bits */
6913 req
->flags
|= (file_ptr
<< REQ_F_NOWAIT_READ_BIT
);
6914 io_req_set_rsrc_node(req
);
6918 static struct file
*io_file_get_normal(struct io_ring_ctx
*ctx
,
6919 struct io_kiocb
*req
, int fd
)
6921 struct file
*file
= fget(fd
);
6923 trace_io_uring_file_get(ctx
, fd
);
6925 /* we don't allow fixed io_uring files */
6926 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6927 io_req_track_inflight(req
);
6931 static inline struct file
*io_file_get(struct io_ring_ctx
*ctx
,
6932 struct io_kiocb
*req
, int fd
, bool fixed
)
6935 return io_file_get_fixed(ctx
, req
, fd
);
6937 return io_file_get_normal(ctx
, req
, fd
);
6940 static void io_req_task_link_timeout(struct io_kiocb
*req
, bool *locked
)
6942 struct io_kiocb
*prev
= req
->timeout
.prev
;
6946 if (!(req
->task
->flags
& PF_EXITING
))
6947 ret
= io_try_cancel_userdata(req
, prev
->user_data
);
6948 io_req_complete_post(req
, ret
?: -ETIME
, 0);
6951 io_req_complete_post(req
, -ETIME
, 0);
6955 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6957 struct io_timeout_data
*data
= container_of(timer
,
6958 struct io_timeout_data
, timer
);
6959 struct io_kiocb
*prev
, *req
= data
->req
;
6960 struct io_ring_ctx
*ctx
= req
->ctx
;
6961 unsigned long flags
;
6963 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
6964 prev
= req
->timeout
.head
;
6965 req
->timeout
.head
= NULL
;
6968 * We don't expect the list to be empty, that will only happen if we
6969 * race with the completion of the linked work.
6972 io_remove_next_linked(prev
);
6973 if (!req_ref_inc_not_zero(prev
))
6976 list_del(&req
->timeout
.list
);
6977 req
->timeout
.prev
= prev
;
6978 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
6980 req
->io_task_work
.func
= io_req_task_link_timeout
;
6981 io_req_task_work_add(req
);
6982 return HRTIMER_NORESTART
;
6985 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6987 struct io_ring_ctx
*ctx
= req
->ctx
;
6989 spin_lock_irq(&ctx
->timeout_lock
);
6991 * If the back reference is NULL, then our linked request finished
6992 * before we got a chance to setup the timer
6994 if (req
->timeout
.head
) {
6995 struct io_timeout_data
*data
= req
->async_data
;
6997 data
->timer
.function
= io_link_timeout_fn
;
6998 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
7000 list_add_tail(&req
->timeout
.list
, &ctx
->ltimeout_list
);
7002 spin_unlock_irq(&ctx
->timeout_lock
);
7003 /* drop submission reference */
7007 static void __io_queue_sqe(struct io_kiocb
*req
)
7008 __must_hold(&req
->ctx
->uring_lock
)
7010 struct io_kiocb
*linked_timeout
;
7014 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
7017 * We async punt it if the file wasn't marked NOWAIT, or if the file
7018 * doesn't support non-blocking read/write attempts
7021 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
7022 struct io_ring_ctx
*ctx
= req
->ctx
;
7023 struct io_submit_state
*state
= &ctx
->submit_state
;
7025 state
->compl_reqs
[state
->compl_nr
++] = req
;
7026 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
7027 io_submit_flush_completions(ctx
);
7031 linked_timeout
= io_prep_linked_timeout(req
);
7033 io_queue_linked_timeout(linked_timeout
);
7034 } else if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
7035 linked_timeout
= io_prep_linked_timeout(req
);
7037 switch (io_arm_poll_handler(req
)) {
7038 case IO_APOLL_READY
:
7040 io_queue_linked_timeout(linked_timeout
);
7042 case IO_APOLL_ABORTED
:
7044 * Queued up for async execution, worker will release
7045 * submit reference when the iocb is actually submitted.
7047 io_queue_async_work(req
, NULL
);
7052 io_queue_linked_timeout(linked_timeout
);
7054 io_req_complete_failed(req
, ret
);
7058 static inline void io_queue_sqe(struct io_kiocb
*req
)
7059 __must_hold(&req
->ctx
->uring_lock
)
7061 if (unlikely(req
->ctx
->drain_active
) && io_drain_req(req
))
7064 if (likely(!(req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
)))) {
7065 __io_queue_sqe(req
);
7066 } else if (req
->flags
& REQ_F_FAIL
) {
7067 io_req_complete_fail_submit(req
);
7069 int ret
= io_req_prep_async(req
);
7072 io_req_complete_failed(req
, ret
);
7074 io_queue_async_work(req
, NULL
);
7079 * Check SQE restrictions (opcode and flags).
7081 * Returns 'true' if SQE is allowed, 'false' otherwise.
7083 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
7084 struct io_kiocb
*req
,
7085 unsigned int sqe_flags
)
7087 if (likely(!ctx
->restricted
))
7090 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
7093 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
7094 ctx
->restrictions
.sqe_flags_required
)
7097 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
7098 ctx
->restrictions
.sqe_flags_required
))
7104 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7105 const struct io_uring_sqe
*sqe
)
7106 __must_hold(&ctx
->uring_lock
)
7108 struct io_submit_state
*state
;
7109 unsigned int sqe_flags
;
7110 int personality
, ret
= 0;
7112 /* req is partially pre-initialised, see io_preinit_req() */
7113 req
->opcode
= READ_ONCE(sqe
->opcode
);
7114 /* same numerical values with corresponding REQ_F_*, safe to copy */
7115 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
7116 req
->user_data
= READ_ONCE(sqe
->user_data
);
7118 req
->fixed_rsrc_refs
= NULL
;
7119 req
->task
= current
;
7121 /* enforce forwards compatibility on users */
7122 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
7124 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
7126 if (!io_check_restriction(ctx
, req
, sqe_flags
))
7129 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
7130 !io_op_defs
[req
->opcode
].buffer_select
)
7132 if (unlikely(sqe_flags
& IOSQE_IO_DRAIN
))
7133 ctx
->drain_active
= true;
7135 personality
= READ_ONCE(sqe
->personality
);
7137 req
->creds
= xa_load(&ctx
->personalities
, personality
);
7140 get_cred(req
->creds
);
7141 req
->flags
|= REQ_F_CREDS
;
7143 state
= &ctx
->submit_state
;
7146 * Plug now if we have more than 1 IO left after this, and the target
7147 * is potentially a read/write to block based storage.
7149 if (!state
->plug_started
&& state
->ios_left
> 1 &&
7150 io_op_defs
[req
->opcode
].plug
) {
7151 blk_start_plug(&state
->plug
);
7152 state
->plug_started
= true;
7155 if (io_op_defs
[req
->opcode
].needs_file
) {
7156 req
->file
= io_file_get(ctx
, req
, READ_ONCE(sqe
->fd
),
7157 (sqe_flags
& IOSQE_FIXED_FILE
));
7158 if (unlikely(!req
->file
))
7166 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7167 const struct io_uring_sqe
*sqe
)
7168 __must_hold(&ctx
->uring_lock
)
7170 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
7173 ret
= io_init_req(ctx
, req
, sqe
);
7174 if (unlikely(ret
)) {
7176 /* fail even hard links since we don't submit */
7179 * we can judge a link req is failed or cancelled by if
7180 * REQ_F_FAIL is set, but the head is an exception since
7181 * it may be set REQ_F_FAIL because of other req's failure
7182 * so let's leverage req->result to distinguish if a head
7183 * is set REQ_F_FAIL because of its failure or other req's
7184 * failure so that we can set the correct ret code for it.
7185 * init result here to avoid affecting the normal path.
7187 if (!(link
->head
->flags
& REQ_F_FAIL
))
7188 req_fail_link_node(link
->head
, -ECANCELED
);
7189 } else if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7191 * the current req is a normal req, we should return
7192 * error and thus break the submittion loop.
7194 io_req_complete_failed(req
, ret
);
7197 req_fail_link_node(req
, ret
);
7199 ret
= io_req_prep(req
, sqe
);
7204 /* don't need @sqe from now on */
7205 trace_io_uring_submit_sqe(ctx
, req
, req
->opcode
, req
->user_data
,
7207 ctx
->flags
& IORING_SETUP_SQPOLL
);
7210 * If we already have a head request, queue this one for async
7211 * submittal once the head completes. If we don't have a head but
7212 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7213 * submitted sync once the chain is complete. If none of those
7214 * conditions are true (normal request), then just queue it.
7217 struct io_kiocb
*head
= link
->head
;
7219 if (!(req
->flags
& REQ_F_FAIL
)) {
7220 ret
= io_req_prep_async(req
);
7221 if (unlikely(ret
)) {
7222 req_fail_link_node(req
, ret
);
7223 if (!(head
->flags
& REQ_F_FAIL
))
7224 req_fail_link_node(head
, -ECANCELED
);
7227 trace_io_uring_link(ctx
, req
, head
);
7228 link
->last
->link
= req
;
7231 /* last request of a link, enqueue the link */
7232 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7237 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
7249 * Batched submission is done, ensure local IO is flushed out.
7251 static void io_submit_state_end(struct io_submit_state
*state
,
7252 struct io_ring_ctx
*ctx
)
7254 if (state
->link
.head
)
7255 io_queue_sqe(state
->link
.head
);
7256 if (state
->compl_nr
)
7257 io_submit_flush_completions(ctx
);
7258 if (state
->plug_started
)
7259 blk_finish_plug(&state
->plug
);
7263 * Start submission side cache.
7265 static void io_submit_state_start(struct io_submit_state
*state
,
7266 unsigned int max_ios
)
7268 state
->plug_started
= false;
7269 state
->ios_left
= max_ios
;
7270 /* set only head, no need to init link_last in advance */
7271 state
->link
.head
= NULL
;
7274 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
7276 struct io_rings
*rings
= ctx
->rings
;
7279 * Ensure any loads from the SQEs are done at this point,
7280 * since once we write the new head, the application could
7281 * write new data to them.
7283 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
7287 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7288 * that is mapped by userspace. This means that care needs to be taken to
7289 * ensure that reads are stable, as we cannot rely on userspace always
7290 * being a good citizen. If members of the sqe are validated and then later
7291 * used, it's important that those reads are done through READ_ONCE() to
7292 * prevent a re-load down the line.
7294 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
7296 unsigned head
, mask
= ctx
->sq_entries
- 1;
7297 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
7300 * The cached sq head (or cq tail) serves two purposes:
7302 * 1) allows us to batch the cost of updating the user visible
7304 * 2) allows the kernel side to track the head on its own, even
7305 * though the application is the one updating it.
7307 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
7308 if (likely(head
< ctx
->sq_entries
))
7309 return &ctx
->sq_sqes
[head
];
7311 /* drop invalid entries */
7313 WRITE_ONCE(ctx
->rings
->sq_dropped
,
7314 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
7318 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
7319 __must_hold(&ctx
->uring_lock
)
7323 /* make sure SQ entry isn't read before tail */
7324 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
7325 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
7327 io_get_task_refs(nr
);
7329 io_submit_state_start(&ctx
->submit_state
, nr
);
7330 while (submitted
< nr
) {
7331 const struct io_uring_sqe
*sqe
;
7332 struct io_kiocb
*req
;
7334 req
= io_alloc_req(ctx
);
7335 if (unlikely(!req
)) {
7337 submitted
= -EAGAIN
;
7340 sqe
= io_get_sqe(ctx
);
7341 if (unlikely(!sqe
)) {
7342 list_add(&req
->inflight_entry
, &ctx
->submit_state
.free_list
);
7345 /* will complete beyond this point, count as submitted */
7347 if (io_submit_sqe(ctx
, req
, sqe
))
7351 if (unlikely(submitted
!= nr
)) {
7352 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
7353 int unused
= nr
- ref_used
;
7355 current
->io_uring
->cached_refs
+= unused
;
7356 percpu_ref_put_many(&ctx
->refs
, unused
);
7359 io_submit_state_end(&ctx
->submit_state
, ctx
);
7360 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7361 io_commit_sqring(ctx
);
7366 static inline bool io_sqd_events_pending(struct io_sq_data
*sqd
)
7368 return READ_ONCE(sqd
->state
);
7371 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
7373 /* Tell userspace we may need a wakeup call */
7374 spin_lock(&ctx
->completion_lock
);
7375 WRITE_ONCE(ctx
->rings
->sq_flags
,
7376 ctx
->rings
->sq_flags
| IORING_SQ_NEED_WAKEUP
);
7377 spin_unlock(&ctx
->completion_lock
);
7380 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
7382 spin_lock(&ctx
->completion_lock
);
7383 WRITE_ONCE(ctx
->rings
->sq_flags
,
7384 ctx
->rings
->sq_flags
& ~IORING_SQ_NEED_WAKEUP
);
7385 spin_unlock(&ctx
->completion_lock
);
7388 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
7390 unsigned int to_submit
;
7393 to_submit
= io_sqring_entries(ctx
);
7394 /* if we're handling multiple rings, cap submit size for fairness */
7395 if (cap_entries
&& to_submit
> IORING_SQPOLL_CAP_ENTRIES_VALUE
)
7396 to_submit
= IORING_SQPOLL_CAP_ENTRIES_VALUE
;
7398 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
7399 unsigned nr_events
= 0;
7400 const struct cred
*creds
= NULL
;
7402 if (ctx
->sq_creds
!= current_cred())
7403 creds
= override_creds(ctx
->sq_creds
);
7405 mutex_lock(&ctx
->uring_lock
);
7406 if (!list_empty(&ctx
->iopoll_list
))
7407 io_do_iopoll(ctx
, &nr_events
, 0);
7410 * Don't submit if refs are dying, good for io_uring_register(),
7411 * but also it is relied upon by io_ring_exit_work()
7413 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
7414 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
7415 ret
= io_submit_sqes(ctx
, to_submit
);
7416 mutex_unlock(&ctx
->uring_lock
);
7418 if (to_submit
&& wq_has_sleeper(&ctx
->sqo_sq_wait
))
7419 wake_up(&ctx
->sqo_sq_wait
);
7421 revert_creds(creds
);
7427 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7429 struct io_ring_ctx
*ctx
;
7430 unsigned sq_thread_idle
= 0;
7432 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7433 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
7434 sqd
->sq_thread_idle
= sq_thread_idle
;
7437 static bool io_sqd_handle_event(struct io_sq_data
*sqd
)
7439 bool did_sig
= false;
7440 struct ksignal ksig
;
7442 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
7443 signal_pending(current
)) {
7444 mutex_unlock(&sqd
->lock
);
7445 if (signal_pending(current
))
7446 did_sig
= get_signal(&ksig
);
7448 mutex_lock(&sqd
->lock
);
7450 return did_sig
|| test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7453 static int io_sq_thread(void *data
)
7455 struct io_sq_data
*sqd
= data
;
7456 struct io_ring_ctx
*ctx
;
7457 unsigned long timeout
= 0;
7458 char buf
[TASK_COMM_LEN
];
7461 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
7462 set_task_comm(current
, buf
);
7464 if (sqd
->sq_cpu
!= -1)
7465 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
7467 set_cpus_allowed_ptr(current
, cpu_online_mask
);
7468 current
->flags
|= PF_NO_SETAFFINITY
;
7470 mutex_lock(&sqd
->lock
);
7472 bool cap_entries
, sqt_spin
= false;
7474 if (io_sqd_events_pending(sqd
) || signal_pending(current
)) {
7475 if (io_sqd_handle_event(sqd
))
7477 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7480 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7481 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7482 int ret
= __io_sq_thread(ctx
, cap_entries
);
7484 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7487 if (io_run_task_work())
7490 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7493 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7497 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7498 if (!io_sqd_events_pending(sqd
) && !current
->task_works
) {
7499 bool needs_sched
= true;
7501 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7502 io_ring_set_wakeup_flag(ctx
);
7504 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7505 !list_empty_careful(&ctx
->iopoll_list
)) {
7506 needs_sched
= false;
7509 if (io_sqring_entries(ctx
)) {
7510 needs_sched
= false;
7516 mutex_unlock(&sqd
->lock
);
7518 mutex_lock(&sqd
->lock
);
7520 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7521 io_ring_clear_wakeup_flag(ctx
);
7524 finish_wait(&sqd
->wait
, &wait
);
7525 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7528 io_uring_cancel_generic(true, sqd
);
7530 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7531 io_ring_set_wakeup_flag(ctx
);
7533 mutex_unlock(&sqd
->lock
);
7535 complete(&sqd
->exited
);
7539 struct io_wait_queue
{
7540 struct wait_queue_entry wq
;
7541 struct io_ring_ctx
*ctx
;
7543 unsigned nr_timeouts
;
7546 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7548 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7549 int dist
= ctx
->cached_cq_tail
- (int) iowq
->cq_tail
;
7552 * Wake up if we have enough events, or if a timeout occurred since we
7553 * started waiting. For timeouts, we always want to return to userspace,
7554 * regardless of event count.
7556 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7559 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7560 int wake_flags
, void *key
)
7562 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7566 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7567 * the task, and the next invocation will do it.
7569 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->check_cq_overflow
))
7570 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7574 static int io_run_task_work_sig(void)
7576 if (io_run_task_work())
7578 if (!signal_pending(current
))
7580 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
7581 return -ERESTARTSYS
;
7585 /* when returns >0, the caller should retry */
7586 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7587 struct io_wait_queue
*iowq
,
7592 /* make sure we run task_work before checking for signals */
7593 ret
= io_run_task_work_sig();
7594 if (ret
|| io_should_wake(iowq
))
7596 /* let the caller flush overflows, retry */
7597 if (test_bit(0, &ctx
->check_cq_overflow
))
7600 if (!schedule_hrtimeout(&timeout
, HRTIMER_MODE_ABS
))
7606 * Wait until events become available, if we don't already have some. The
7607 * application must reap them itself, as they reside on the shared cq ring.
7609 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7610 const sigset_t __user
*sig
, size_t sigsz
,
7611 struct __kernel_timespec __user
*uts
)
7613 struct io_wait_queue iowq
;
7614 struct io_rings
*rings
= ctx
->rings
;
7615 ktime_t timeout
= KTIME_MAX
;
7619 io_cqring_overflow_flush(ctx
);
7620 if (io_cqring_events(ctx
) >= min_events
)
7622 if (!io_run_task_work())
7627 struct timespec64 ts
;
7629 if (get_timespec64(&ts
, uts
))
7631 timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
7635 #ifdef CONFIG_COMPAT
7636 if (in_compat_syscall())
7637 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7641 ret
= set_user_sigmask(sig
, sigsz
);
7647 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
7648 iowq
.wq
.private = current
;
7649 INIT_LIST_HEAD(&iowq
.wq
.entry
);
7651 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7652 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
7654 trace_io_uring_cqring_wait(ctx
, min_events
);
7656 /* if we can't even flush overflow, don't wait for more */
7657 if (!io_cqring_overflow_flush(ctx
)) {
7661 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
7662 TASK_INTERRUPTIBLE
);
7663 ret
= io_cqring_wait_schedule(ctx
, &iowq
, timeout
);
7664 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
7668 restore_saved_sigmask_unless(ret
== -EINTR
);
7670 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7673 static void io_free_page_table(void **table
, size_t size
)
7675 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7677 for (i
= 0; i
< nr_tables
; i
++)
7682 static void **io_alloc_page_table(size_t size
)
7684 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7685 size_t init_size
= size
;
7688 table
= kcalloc(nr_tables
, sizeof(*table
), GFP_KERNEL_ACCOUNT
);
7692 for (i
= 0; i
< nr_tables
; i
++) {
7693 unsigned int this_size
= min_t(size_t, size
, PAGE_SIZE
);
7695 table
[i
] = kzalloc(this_size
, GFP_KERNEL_ACCOUNT
);
7697 io_free_page_table(table
, init_size
);
7705 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
7707 percpu_ref_exit(&ref_node
->refs
);
7711 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7713 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
7714 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
7715 unsigned long flags
;
7716 bool first_add
= false;
7717 unsigned long delay
= HZ
;
7719 spin_lock_irqsave(&ctx
->rsrc_ref_lock
, flags
);
7722 /* if we are mid-quiesce then do not delay */
7723 if (node
->rsrc_data
->quiesce
)
7726 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7727 node
= list_first_entry(&ctx
->rsrc_ref_list
,
7728 struct io_rsrc_node
, node
);
7729 /* recycle ref nodes in order */
7732 list_del(&node
->node
);
7733 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
7735 spin_unlock_irqrestore(&ctx
->rsrc_ref_lock
, flags
);
7738 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
7741 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
)
7743 struct io_rsrc_node
*ref_node
;
7745 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7749 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7754 INIT_LIST_HEAD(&ref_node
->node
);
7755 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7756 ref_node
->done
= false;
7760 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
7761 struct io_rsrc_data
*data_to_kill
)
7763 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
7764 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
7767 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
7769 rsrc_node
->rsrc_data
= data_to_kill
;
7770 spin_lock_irq(&ctx
->rsrc_ref_lock
);
7771 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
7772 spin_unlock_irq(&ctx
->rsrc_ref_lock
);
7774 atomic_inc(&data_to_kill
->refs
);
7775 percpu_ref_kill(&rsrc_node
->refs
);
7776 ctx
->rsrc_node
= NULL
;
7779 if (!ctx
->rsrc_node
) {
7780 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
7781 ctx
->rsrc_backup_node
= NULL
;
7785 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
7787 if (ctx
->rsrc_backup_node
)
7789 ctx
->rsrc_backup_node
= io_rsrc_node_alloc(ctx
);
7790 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
7793 static int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
, struct io_ring_ctx
*ctx
)
7797 /* As we may drop ->uring_lock, other task may have started quiesce */
7801 data
->quiesce
= true;
7803 ret
= io_rsrc_node_switch_start(ctx
);
7806 io_rsrc_node_switch(ctx
, data
);
7808 /* kill initial ref, already quiesced if zero */
7809 if (atomic_dec_and_test(&data
->refs
))
7811 mutex_unlock(&ctx
->uring_lock
);
7812 flush_delayed_work(&ctx
->rsrc_put_work
);
7813 ret
= wait_for_completion_interruptible(&data
->done
);
7815 mutex_lock(&ctx
->uring_lock
);
7816 if (atomic_read(&data
->refs
) > 0) {
7818 * it has been revived by another thread while
7821 mutex_unlock(&ctx
->uring_lock
);
7827 atomic_inc(&data
->refs
);
7828 /* wait for all works potentially completing data->done */
7829 flush_delayed_work(&ctx
->rsrc_put_work
);
7830 reinit_completion(&data
->done
);
7832 ret
= io_run_task_work_sig();
7833 mutex_lock(&ctx
->uring_lock
);
7835 data
->quiesce
= false;
7840 static u64
*io_get_tag_slot(struct io_rsrc_data
*data
, unsigned int idx
)
7842 unsigned int off
= idx
& IO_RSRC_TAG_TABLE_MASK
;
7843 unsigned int table_idx
= idx
>> IO_RSRC_TAG_TABLE_SHIFT
;
7845 return &data
->tags
[table_idx
][off
];
7848 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
7850 size_t size
= data
->nr
* sizeof(data
->tags
[0][0]);
7853 io_free_page_table((void **)data
->tags
, size
);
7857 static int io_rsrc_data_alloc(struct io_ring_ctx
*ctx
, rsrc_put_fn
*do_put
,
7858 u64 __user
*utags
, unsigned nr
,
7859 struct io_rsrc_data
**pdata
)
7861 struct io_rsrc_data
*data
;
7865 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7868 data
->tags
= (u64
**)io_alloc_page_table(nr
* sizeof(data
->tags
[0][0]));
7876 data
->do_put
= do_put
;
7879 for (i
= 0; i
< nr
; i
++) {
7880 u64
*tag_slot
= io_get_tag_slot(data
, i
);
7882 if (copy_from_user(tag_slot
, &utags
[i
],
7888 atomic_set(&data
->refs
, 1);
7889 init_completion(&data
->done
);
7893 io_rsrc_data_free(data
);
7897 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7899 table
->files
= kvcalloc(nr_files
, sizeof(table
->files
[0]),
7900 GFP_KERNEL_ACCOUNT
);
7901 return !!table
->files
;
7904 static void io_free_file_tables(struct io_file_table
*table
)
7906 kvfree(table
->files
);
7907 table
->files
= NULL
;
7910 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7912 #if defined(CONFIG_UNIX)
7913 if (ctx
->ring_sock
) {
7914 struct sock
*sock
= ctx
->ring_sock
->sk
;
7915 struct sk_buff
*skb
;
7917 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7923 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7926 file
= io_file_from_index(ctx
, i
);
7931 io_free_file_tables(&ctx
->file_table
);
7932 io_rsrc_data_free(ctx
->file_data
);
7933 ctx
->file_data
= NULL
;
7934 ctx
->nr_user_files
= 0;
7937 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7941 if (!ctx
->file_data
)
7943 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
7945 __io_sqe_files_unregister(ctx
);
7949 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7950 __releases(&sqd
->lock
)
7952 WARN_ON_ONCE(sqd
->thread
== current
);
7955 * Do the dance but not conditional clear_bit() because it'd race with
7956 * other threads incrementing park_pending and setting the bit.
7958 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7959 if (atomic_dec_return(&sqd
->park_pending
))
7960 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7961 mutex_unlock(&sqd
->lock
);
7964 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7965 __acquires(&sqd
->lock
)
7967 WARN_ON_ONCE(sqd
->thread
== current
);
7969 atomic_inc(&sqd
->park_pending
);
7970 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7971 mutex_lock(&sqd
->lock
);
7973 wake_up_process(sqd
->thread
);
7976 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7978 WARN_ON_ONCE(sqd
->thread
== current
);
7979 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
7981 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7982 mutex_lock(&sqd
->lock
);
7984 wake_up_process(sqd
->thread
);
7985 mutex_unlock(&sqd
->lock
);
7986 wait_for_completion(&sqd
->exited
);
7989 static void io_put_sq_data(struct io_sq_data
*sqd
)
7991 if (refcount_dec_and_test(&sqd
->refs
)) {
7992 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
7994 io_sq_thread_stop(sqd
);
7999 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
8001 struct io_sq_data
*sqd
= ctx
->sq_data
;
8004 io_sq_thread_park(sqd
);
8005 list_del_init(&ctx
->sqd_list
);
8006 io_sqd_update_thread_idle(sqd
);
8007 io_sq_thread_unpark(sqd
);
8009 io_put_sq_data(sqd
);
8010 ctx
->sq_data
= NULL
;
8014 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
8016 struct io_ring_ctx
*ctx_attach
;
8017 struct io_sq_data
*sqd
;
8020 f
= fdget(p
->wq_fd
);
8022 return ERR_PTR(-ENXIO
);
8023 if (f
.file
->f_op
!= &io_uring_fops
) {
8025 return ERR_PTR(-EINVAL
);
8028 ctx_attach
= f
.file
->private_data
;
8029 sqd
= ctx_attach
->sq_data
;
8032 return ERR_PTR(-EINVAL
);
8034 if (sqd
->task_tgid
!= current
->tgid
) {
8036 return ERR_PTR(-EPERM
);
8039 refcount_inc(&sqd
->refs
);
8044 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
8047 struct io_sq_data
*sqd
;
8050 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
8051 sqd
= io_attach_sq_data(p
);
8056 /* fall through for EPERM case, setup new sqd/task */
8057 if (PTR_ERR(sqd
) != -EPERM
)
8061 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
8063 return ERR_PTR(-ENOMEM
);
8065 atomic_set(&sqd
->park_pending
, 0);
8066 refcount_set(&sqd
->refs
, 1);
8067 INIT_LIST_HEAD(&sqd
->ctx_list
);
8068 mutex_init(&sqd
->lock
);
8069 init_waitqueue_head(&sqd
->wait
);
8070 init_completion(&sqd
->exited
);
8074 #if defined(CONFIG_UNIX)
8076 * Ensure the UNIX gc is aware of our file set, so we are certain that
8077 * the io_uring can be safely unregistered on process exit, even if we have
8078 * loops in the file referencing.
8080 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
8082 struct sock
*sk
= ctx
->ring_sock
->sk
;
8083 struct scm_fp_list
*fpl
;
8084 struct sk_buff
*skb
;
8087 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
8091 skb
= alloc_skb(0, GFP_KERNEL
);
8100 fpl
->user
= get_uid(current_user());
8101 for (i
= 0; i
< nr
; i
++) {
8102 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8106 fpl
->fp
[nr_files
] = get_file(file
);
8107 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
8112 fpl
->max
= SCM_MAX_FD
;
8113 fpl
->count
= nr_files
;
8114 UNIXCB(skb
).fp
= fpl
;
8115 skb
->destructor
= unix_destruct_scm
;
8116 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
8117 skb_queue_head(&sk
->sk_receive_queue
, skb
);
8119 for (i
= 0; i
< nr
; i
++) {
8120 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8127 free_uid(fpl
->user
);
8135 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8136 * causes regular reference counting to break down. We rely on the UNIX
8137 * garbage collection to take care of this problem for us.
8139 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8141 unsigned left
, total
;
8145 left
= ctx
->nr_user_files
;
8147 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
8149 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
8153 total
+= this_files
;
8159 while (total
< ctx
->nr_user_files
) {
8160 struct file
*file
= io_file_from_index(ctx
, total
);
8170 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8176 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8178 struct file
*file
= prsrc
->file
;
8179 #if defined(CONFIG_UNIX)
8180 struct sock
*sock
= ctx
->ring_sock
->sk
;
8181 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
8182 struct sk_buff
*skb
;
8185 __skb_queue_head_init(&list
);
8188 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8189 * remove this entry and rearrange the file array.
8191 skb
= skb_dequeue(head
);
8193 struct scm_fp_list
*fp
;
8195 fp
= UNIXCB(skb
).fp
;
8196 for (i
= 0; i
< fp
->count
; i
++) {
8199 if (fp
->fp
[i
] != file
)
8202 unix_notinflight(fp
->user
, fp
->fp
[i
]);
8203 left
= fp
->count
- 1 - i
;
8205 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
8206 left
* sizeof(struct file
*));
8213 __skb_queue_tail(&list
, skb
);
8223 __skb_queue_tail(&list
, skb
);
8225 skb
= skb_dequeue(head
);
8228 if (skb_peek(&list
)) {
8229 spin_lock_irq(&head
->lock
);
8230 while ((skb
= __skb_dequeue(&list
)) != NULL
)
8231 __skb_queue_tail(head
, skb
);
8232 spin_unlock_irq(&head
->lock
);
8239 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
8241 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
8242 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
8243 struct io_rsrc_put
*prsrc
, *tmp
;
8245 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
8246 list_del(&prsrc
->list
);
8249 bool lock_ring
= ctx
->flags
& IORING_SETUP_IOPOLL
;
8251 io_ring_submit_lock(ctx
, lock_ring
);
8252 spin_lock(&ctx
->completion_lock
);
8253 io_cqring_fill_event(ctx
, prsrc
->tag
, 0, 0);
8255 io_commit_cqring(ctx
);
8256 spin_unlock(&ctx
->completion_lock
);
8257 io_cqring_ev_posted(ctx
);
8258 io_ring_submit_unlock(ctx
, lock_ring
);
8261 rsrc_data
->do_put(ctx
, prsrc
);
8265 io_rsrc_node_destroy(ref_node
);
8266 if (atomic_dec_and_test(&rsrc_data
->refs
))
8267 complete(&rsrc_data
->done
);
8270 static void io_rsrc_put_work(struct work_struct
*work
)
8272 struct io_ring_ctx
*ctx
;
8273 struct llist_node
*node
;
8275 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
8276 node
= llist_del_all(&ctx
->rsrc_put_llist
);
8279 struct io_rsrc_node
*ref_node
;
8280 struct llist_node
*next
= node
->next
;
8282 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
8283 __io_rsrc_put_work(ref_node
);
8288 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8289 unsigned nr_args
, u64 __user
*tags
)
8291 __s32 __user
*fds
= (__s32 __user
*) arg
;
8300 if (nr_args
> IORING_MAX_FIXED_FILES
)
8302 if (nr_args
> rlimit(RLIMIT_NOFILE
))
8304 ret
= io_rsrc_node_switch_start(ctx
);
8307 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, tags
, nr_args
,
8313 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
))
8316 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
8317 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
8321 /* allow sparse sets */
8324 if (unlikely(*io_get_tag_slot(ctx
->file_data
, i
)))
8331 if (unlikely(!file
))
8335 * Don't allow io_uring instances to be registered. If UNIX
8336 * isn't enabled, then this causes a reference cycle and this
8337 * instance can never get freed. If UNIX is enabled we'll
8338 * handle it just fine, but there's still no point in allowing
8339 * a ring fd as it doesn't support regular read/write anyway.
8341 if (file
->f_op
== &io_uring_fops
) {
8345 io_fixed_file_set(io_fixed_file_slot(&ctx
->file_table
, i
), file
);
8348 ret
= io_sqe_files_scm(ctx
);
8350 __io_sqe_files_unregister(ctx
);
8354 io_rsrc_node_switch(ctx
, NULL
);
8357 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
8358 file
= io_file_from_index(ctx
, i
);
8362 io_free_file_tables(&ctx
->file_table
);
8363 ctx
->nr_user_files
= 0;
8365 io_rsrc_data_free(ctx
->file_data
);
8366 ctx
->file_data
= NULL
;
8370 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
8373 #if defined(CONFIG_UNIX)
8374 struct sock
*sock
= ctx
->ring_sock
->sk
;
8375 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
8376 struct sk_buff
*skb
;
8379 * See if we can merge this file into an existing skb SCM_RIGHTS
8380 * file set. If there's no room, fall back to allocating a new skb
8381 * and filling it in.
8383 spin_lock_irq(&head
->lock
);
8384 skb
= skb_peek(head
);
8386 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
8388 if (fpl
->count
< SCM_MAX_FD
) {
8389 __skb_unlink(skb
, head
);
8390 spin_unlock_irq(&head
->lock
);
8391 fpl
->fp
[fpl
->count
] = get_file(file
);
8392 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
8394 spin_lock_irq(&head
->lock
);
8395 __skb_queue_head(head
, skb
);
8400 spin_unlock_irq(&head
->lock
);
8407 return __io_sqe_files_scm(ctx
, 1, index
);
8413 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
8414 struct io_rsrc_node
*node
, void *rsrc
)
8416 struct io_rsrc_put
*prsrc
;
8418 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
8422 prsrc
->tag
= *io_get_tag_slot(data
, idx
);
8424 list_add(&prsrc
->list
, &node
->rsrc_list
);
8428 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
8429 unsigned int issue_flags
, u32 slot_index
)
8431 struct io_ring_ctx
*ctx
= req
->ctx
;
8432 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
8433 bool needs_switch
= false;
8434 struct io_fixed_file
*file_slot
;
8437 io_ring_submit_lock(ctx
, !force_nonblock
);
8438 if (file
->f_op
== &io_uring_fops
)
8441 if (!ctx
->file_data
)
8444 if (slot_index
>= ctx
->nr_user_files
)
8447 slot_index
= array_index_nospec(slot_index
, ctx
->nr_user_files
);
8448 file_slot
= io_fixed_file_slot(&ctx
->file_table
, slot_index
);
8450 if (file_slot
->file_ptr
) {
8451 struct file
*old_file
;
8453 ret
= io_rsrc_node_switch_start(ctx
);
8457 old_file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8458 ret
= io_queue_rsrc_removal(ctx
->file_data
, slot_index
,
8459 ctx
->rsrc_node
, old_file
);
8462 file_slot
->file_ptr
= 0;
8463 needs_switch
= true;
8466 *io_get_tag_slot(ctx
->file_data
, slot_index
) = 0;
8467 io_fixed_file_set(file_slot
, file
);
8468 ret
= io_sqe_file_register(ctx
, file
, slot_index
);
8470 file_slot
->file_ptr
= 0;
8477 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8478 io_ring_submit_unlock(ctx
, !force_nonblock
);
8484 static int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
)
8486 unsigned int offset
= req
->close
.file_slot
- 1;
8487 struct io_ring_ctx
*ctx
= req
->ctx
;
8488 struct io_fixed_file
*file_slot
;
8492 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8494 if (unlikely(!ctx
->file_data
))
8497 if (offset
>= ctx
->nr_user_files
)
8499 ret
= io_rsrc_node_switch_start(ctx
);
8503 i
= array_index_nospec(offset
, ctx
->nr_user_files
);
8504 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
8506 if (!file_slot
->file_ptr
)
8509 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8510 ret
= io_queue_rsrc_removal(ctx
->file_data
, offset
, ctx
->rsrc_node
, file
);
8514 file_slot
->file_ptr
= 0;
8515 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8518 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8522 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
8523 struct io_uring_rsrc_update2
*up
,
8526 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
8527 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
8528 struct io_rsrc_data
*data
= ctx
->file_data
;
8529 struct io_fixed_file
*file_slot
;
8533 bool needs_switch
= false;
8535 if (!ctx
->file_data
)
8537 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
8540 for (done
= 0; done
< nr_args
; done
++) {
8543 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
8544 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8548 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
8552 if (fd
== IORING_REGISTER_FILES_SKIP
)
8555 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
8556 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
8558 if (file_slot
->file_ptr
) {
8559 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8560 err
= io_queue_rsrc_removal(data
, up
->offset
+ done
,
8561 ctx
->rsrc_node
, file
);
8564 file_slot
->file_ptr
= 0;
8565 needs_switch
= true;
8574 * Don't allow io_uring instances to be registered. If
8575 * UNIX isn't enabled, then this causes a reference
8576 * cycle and this instance can never get freed. If UNIX
8577 * is enabled we'll handle it just fine, but there's
8578 * still no point in allowing a ring fd as it doesn't
8579 * support regular read/write anyway.
8581 if (file
->f_op
== &io_uring_fops
) {
8586 *io_get_tag_slot(data
, i
) = tag
;
8587 io_fixed_file_set(file_slot
, file
);
8588 err
= io_sqe_file_register(ctx
, file
, i
);
8590 file_slot
->file_ptr
= 0;
8598 io_rsrc_node_switch(ctx
, data
);
8599 return done
? done
: err
;
8602 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
8603 struct task_struct
*task
)
8605 struct io_wq_hash
*hash
;
8606 struct io_wq_data data
;
8607 unsigned int concurrency
;
8609 mutex_lock(&ctx
->uring_lock
);
8610 hash
= ctx
->hash_map
;
8612 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
8614 mutex_unlock(&ctx
->uring_lock
);
8615 return ERR_PTR(-ENOMEM
);
8617 refcount_set(&hash
->refs
, 1);
8618 init_waitqueue_head(&hash
->wait
);
8619 ctx
->hash_map
= hash
;
8621 mutex_unlock(&ctx
->uring_lock
);
8625 data
.free_work
= io_wq_free_work
;
8626 data
.do_work
= io_wq_submit_work
;
8628 /* Do QD, or 4 * CPUS, whatever is smallest */
8629 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8631 return io_wq_create(concurrency
, &data
);
8634 static int io_uring_alloc_task_context(struct task_struct
*task
,
8635 struct io_ring_ctx
*ctx
)
8637 struct io_uring_task
*tctx
;
8640 tctx
= kzalloc(sizeof(*tctx
), GFP_KERNEL
);
8641 if (unlikely(!tctx
))
8644 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8645 if (unlikely(ret
)) {
8650 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
8651 if (IS_ERR(tctx
->io_wq
)) {
8652 ret
= PTR_ERR(tctx
->io_wq
);
8653 percpu_counter_destroy(&tctx
->inflight
);
8659 init_waitqueue_head(&tctx
->wait
);
8660 atomic_set(&tctx
->in_idle
, 0);
8661 atomic_set(&tctx
->inflight_tracked
, 0);
8662 task
->io_uring
= tctx
;
8663 spin_lock_init(&tctx
->task_lock
);
8664 INIT_WQ_LIST(&tctx
->task_list
);
8665 init_task_work(&tctx
->task_work
, tctx_task_work
);
8669 void __io_uring_free(struct task_struct
*tsk
)
8671 struct io_uring_task
*tctx
= tsk
->io_uring
;
8673 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8674 WARN_ON_ONCE(tctx
->io_wq
);
8675 WARN_ON_ONCE(tctx
->cached_refs
);
8677 percpu_counter_destroy(&tctx
->inflight
);
8679 tsk
->io_uring
= NULL
;
8682 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8683 struct io_uring_params
*p
)
8687 /* Retain compatibility with failing for an invalid attach attempt */
8688 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
8689 IORING_SETUP_ATTACH_WQ
) {
8692 f
= fdget(p
->wq_fd
);
8695 if (f
.file
->f_op
!= &io_uring_fops
) {
8701 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8702 struct task_struct
*tsk
;
8703 struct io_sq_data
*sqd
;
8706 sqd
= io_get_sq_data(p
, &attached
);
8712 ctx
->sq_creds
= get_current_cred();
8714 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8715 if (!ctx
->sq_thread_idle
)
8716 ctx
->sq_thread_idle
= HZ
;
8718 io_sq_thread_park(sqd
);
8719 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
8720 io_sqd_update_thread_idle(sqd
);
8721 /* don't attach to a dying SQPOLL thread, would be racy */
8722 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
8723 io_sq_thread_unpark(sqd
);
8730 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8731 int cpu
= p
->sq_thread_cpu
;
8734 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
8741 sqd
->task_pid
= current
->pid
;
8742 sqd
->task_tgid
= current
->tgid
;
8743 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
8750 ret
= io_uring_alloc_task_context(tsk
, ctx
);
8751 wake_up_new_task(tsk
);
8754 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8755 /* Can't have SQ_AFF without SQPOLL */
8762 complete(&ctx
->sq_data
->exited
);
8764 io_sq_thread_finish(ctx
);
8768 static inline void __io_unaccount_mem(struct user_struct
*user
,
8769 unsigned long nr_pages
)
8771 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8774 static inline int __io_account_mem(struct user_struct
*user
,
8775 unsigned long nr_pages
)
8777 unsigned long page_limit
, cur_pages
, new_pages
;
8779 /* Don't allow more pages than we can safely lock */
8780 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8783 cur_pages
= atomic_long_read(&user
->locked_vm
);
8784 new_pages
= cur_pages
+ nr_pages
;
8785 if (new_pages
> page_limit
)
8787 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8788 new_pages
) != cur_pages
);
8793 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8796 __io_unaccount_mem(ctx
->user
, nr_pages
);
8798 if (ctx
->mm_account
)
8799 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8802 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8807 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8812 if (ctx
->mm_account
)
8813 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8818 static void io_mem_free(void *ptr
)
8825 page
= virt_to_head_page(ptr
);
8826 if (put_page_testzero(page
))
8827 free_compound_page(page
);
8830 static void *io_mem_alloc(size_t size
)
8832 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
8834 return (void *) __get_free_pages(gfp
, get_order(size
));
8837 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8840 struct io_rings
*rings
;
8841 size_t off
, sq_array_size
;
8843 off
= struct_size(rings
, cqes
, cq_entries
);
8844 if (off
== SIZE_MAX
)
8848 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8856 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8857 if (sq_array_size
== SIZE_MAX
)
8860 if (check_add_overflow(off
, sq_array_size
, &off
))
8866 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
8868 struct io_mapped_ubuf
*imu
= *slot
;
8871 if (imu
!= ctx
->dummy_ubuf
) {
8872 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
8873 unpin_user_page(imu
->bvec
[i
].bv_page
);
8874 if (imu
->acct_pages
)
8875 io_unaccount_mem(ctx
, imu
->acct_pages
);
8881 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8883 io_buffer_unmap(ctx
, &prsrc
->buf
);
8887 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8891 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
8892 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
8893 kfree(ctx
->user_bufs
);
8894 io_rsrc_data_free(ctx
->buf_data
);
8895 ctx
->user_bufs
= NULL
;
8896 ctx
->buf_data
= NULL
;
8897 ctx
->nr_user_bufs
= 0;
8900 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8907 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
8909 __io_sqe_buffers_unregister(ctx
);
8913 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8914 void __user
*arg
, unsigned index
)
8916 struct iovec __user
*src
;
8918 #ifdef CONFIG_COMPAT
8920 struct compat_iovec __user
*ciovs
;
8921 struct compat_iovec ciov
;
8923 ciovs
= (struct compat_iovec __user
*) arg
;
8924 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8927 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8928 dst
->iov_len
= ciov
.iov_len
;
8932 src
= (struct iovec __user
*) arg
;
8933 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8939 * Not super efficient, but this is just a registration time. And we do cache
8940 * the last compound head, so generally we'll only do a full search if we don't
8943 * We check if the given compound head page has already been accounted, to
8944 * avoid double accounting it. This allows us to account the full size of the
8945 * page, not just the constituent pages of a huge page.
8947 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8948 int nr_pages
, struct page
*hpage
)
8952 /* check current page array */
8953 for (i
= 0; i
< nr_pages
; i
++) {
8954 if (!PageCompound(pages
[i
]))
8956 if (compound_head(pages
[i
]) == hpage
)
8960 /* check previously registered pages */
8961 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8962 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
8964 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8965 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8967 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8975 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8976 int nr_pages
, struct io_mapped_ubuf
*imu
,
8977 struct page
**last_hpage
)
8981 imu
->acct_pages
= 0;
8982 for (i
= 0; i
< nr_pages
; i
++) {
8983 if (!PageCompound(pages
[i
])) {
8988 hpage
= compound_head(pages
[i
]);
8989 if (hpage
== *last_hpage
)
8991 *last_hpage
= hpage
;
8992 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8994 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8998 if (!imu
->acct_pages
)
9001 ret
= io_account_mem(ctx
, imu
->acct_pages
);
9003 imu
->acct_pages
= 0;
9007 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
9008 struct io_mapped_ubuf
**pimu
,
9009 struct page
**last_hpage
)
9011 struct io_mapped_ubuf
*imu
= NULL
;
9012 struct vm_area_struct
**vmas
= NULL
;
9013 struct page
**pages
= NULL
;
9014 unsigned long off
, start
, end
, ubuf
;
9016 int ret
, pret
, nr_pages
, i
;
9018 if (!iov
->iov_base
) {
9019 *pimu
= ctx
->dummy_ubuf
;
9023 ubuf
= (unsigned long) iov
->iov_base
;
9024 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
9025 start
= ubuf
>> PAGE_SHIFT
;
9026 nr_pages
= end
- start
;
9031 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
9035 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
9040 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
9045 mmap_read_lock(current
->mm
);
9046 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
9048 if (pret
== nr_pages
) {
9049 /* don't support file backed memory */
9050 for (i
= 0; i
< nr_pages
; i
++) {
9051 struct vm_area_struct
*vma
= vmas
[i
];
9053 if (vma_is_shmem(vma
))
9056 !is_file_hugepages(vma
->vm_file
)) {
9062 ret
= pret
< 0 ? pret
: -EFAULT
;
9064 mmap_read_unlock(current
->mm
);
9067 * if we did partial map, or found file backed vmas,
9068 * release any pages we did get
9071 unpin_user_pages(pages
, pret
);
9075 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
9077 unpin_user_pages(pages
, pret
);
9081 off
= ubuf
& ~PAGE_MASK
;
9082 size
= iov
->iov_len
;
9083 for (i
= 0; i
< nr_pages
; i
++) {
9086 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
9087 imu
->bvec
[i
].bv_page
= pages
[i
];
9088 imu
->bvec
[i
].bv_len
= vec_len
;
9089 imu
->bvec
[i
].bv_offset
= off
;
9093 /* store original address for later verification */
9095 imu
->ubuf_end
= ubuf
+ iov
->iov_len
;
9096 imu
->nr_bvecs
= nr_pages
;
9107 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
9109 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
9110 return ctx
->user_bufs
? 0 : -ENOMEM
;
9113 static int io_buffer_validate(struct iovec
*iov
)
9115 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
9118 * Don't impose further limits on the size and buffer
9119 * constraints here, we'll -EINVAL later when IO is
9120 * submitted if they are wrong.
9123 return iov
->iov_len
? -EFAULT
: 0;
9127 /* arbitrary limit, but we need something */
9128 if (iov
->iov_len
> SZ_1G
)
9131 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
9137 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
9138 unsigned int nr_args
, u64 __user
*tags
)
9140 struct page
*last_hpage
= NULL
;
9141 struct io_rsrc_data
*data
;
9147 if (!nr_args
|| nr_args
> IORING_MAX_REG_BUFFERS
)
9149 ret
= io_rsrc_node_switch_start(ctx
);
9152 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, tags
, nr_args
, &data
);
9155 ret
= io_buffers_map_alloc(ctx
, nr_args
);
9157 io_rsrc_data_free(data
);
9161 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
9162 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
9165 ret
= io_buffer_validate(&iov
);
9168 if (!iov
.iov_base
&& *io_get_tag_slot(data
, i
)) {
9173 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
9179 WARN_ON_ONCE(ctx
->buf_data
);
9181 ctx
->buf_data
= data
;
9183 __io_sqe_buffers_unregister(ctx
);
9185 io_rsrc_node_switch(ctx
, NULL
);
9189 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
9190 struct io_uring_rsrc_update2
*up
,
9191 unsigned int nr_args
)
9193 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
9194 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
9195 struct page
*last_hpage
= NULL
;
9196 bool needs_switch
= false;
9202 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
9205 for (done
= 0; done
< nr_args
; done
++) {
9206 struct io_mapped_ubuf
*imu
;
9207 int offset
= up
->offset
+ done
;
9210 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
9213 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
9217 err
= io_buffer_validate(&iov
);
9220 if (!iov
.iov_base
&& tag
) {
9224 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
9228 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
9229 if (ctx
->user_bufs
[i
] != ctx
->dummy_ubuf
) {
9230 err
= io_queue_rsrc_removal(ctx
->buf_data
, offset
,
9231 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
9232 if (unlikely(err
)) {
9233 io_buffer_unmap(ctx
, &imu
);
9236 ctx
->user_bufs
[i
] = NULL
;
9237 needs_switch
= true;
9240 ctx
->user_bufs
[i
] = imu
;
9241 *io_get_tag_slot(ctx
->buf_data
, offset
) = tag
;
9245 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
9246 return done
? done
: err
;
9249 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
9251 __s32 __user
*fds
= arg
;
9257 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
9260 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
9261 if (IS_ERR(ctx
->cq_ev_fd
)) {
9262 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
9264 ctx
->cq_ev_fd
= NULL
;
9271 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
9273 if (ctx
->cq_ev_fd
) {
9274 eventfd_ctx_put(ctx
->cq_ev_fd
);
9275 ctx
->cq_ev_fd
= NULL
;
9282 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
9284 struct io_buffer
*buf
;
9285 unsigned long index
;
9287 xa_for_each(&ctx
->io_buffers
, index
, buf
)
9288 __io_remove_buffers(ctx
, buf
, index
, -1U);
9291 static void io_req_cache_free(struct list_head
*list
)
9293 struct io_kiocb
*req
, *nxt
;
9295 list_for_each_entry_safe(req
, nxt
, list
, inflight_entry
) {
9296 list_del(&req
->inflight_entry
);
9297 kmem_cache_free(req_cachep
, req
);
9301 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
9303 struct io_submit_state
*state
= &ctx
->submit_state
;
9305 mutex_lock(&ctx
->uring_lock
);
9307 if (state
->free_reqs
) {
9308 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
9309 state
->free_reqs
= 0;
9312 io_flush_cached_locked_reqs(ctx
, state
);
9313 io_req_cache_free(&state
->free_list
);
9314 mutex_unlock(&ctx
->uring_lock
);
9317 static void io_wait_rsrc_data(struct io_rsrc_data
*data
)
9319 if (data
&& !atomic_dec_and_test(&data
->refs
))
9320 wait_for_completion(&data
->done
);
9323 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
9325 io_sq_thread_finish(ctx
);
9327 if (ctx
->mm_account
) {
9328 mmdrop(ctx
->mm_account
);
9329 ctx
->mm_account
= NULL
;
9332 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9333 io_wait_rsrc_data(ctx
->buf_data
);
9334 io_wait_rsrc_data(ctx
->file_data
);
9336 mutex_lock(&ctx
->uring_lock
);
9338 __io_sqe_buffers_unregister(ctx
);
9340 __io_sqe_files_unregister(ctx
);
9342 __io_cqring_overflow_flush(ctx
, true);
9343 mutex_unlock(&ctx
->uring_lock
);
9344 io_eventfd_unregister(ctx
);
9345 io_destroy_buffers(ctx
);
9347 put_cred(ctx
->sq_creds
);
9349 /* there are no registered resources left, nobody uses it */
9351 io_rsrc_node_destroy(ctx
->rsrc_node
);
9352 if (ctx
->rsrc_backup_node
)
9353 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
9354 flush_delayed_work(&ctx
->rsrc_put_work
);
9356 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
9357 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
9359 #if defined(CONFIG_UNIX)
9360 if (ctx
->ring_sock
) {
9361 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
9362 sock_release(ctx
->ring_sock
);
9365 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
9367 io_mem_free(ctx
->rings
);
9368 io_mem_free(ctx
->sq_sqes
);
9370 percpu_ref_exit(&ctx
->refs
);
9371 free_uid(ctx
->user
);
9372 io_req_caches_free(ctx
);
9374 io_wq_put_hash(ctx
->hash_map
);
9375 kfree(ctx
->cancel_hash
);
9376 kfree(ctx
->dummy_ubuf
);
9380 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
9382 struct io_ring_ctx
*ctx
= file
->private_data
;
9385 poll_wait(file
, &ctx
->poll_wait
, wait
);
9387 * synchronizes with barrier from wq_has_sleeper call in
9391 if (!io_sqring_full(ctx
))
9392 mask
|= EPOLLOUT
| EPOLLWRNORM
;
9395 * Don't flush cqring overflow list here, just do a simple check.
9396 * Otherwise there could possible be ABBA deadlock:
9399 * lock(&ctx->uring_lock);
9401 * lock(&ctx->uring_lock);
9404 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9405 * pushs them to do the flush.
9407 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->check_cq_overflow
))
9408 mask
|= EPOLLIN
| EPOLLRDNORM
;
9413 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9415 const struct cred
*creds
;
9417 creds
= xa_erase(&ctx
->personalities
, id
);
9426 struct io_tctx_exit
{
9427 struct callback_head task_work
;
9428 struct completion completion
;
9429 struct io_ring_ctx
*ctx
;
9432 static void io_tctx_exit_cb(struct callback_head
*cb
)
9434 struct io_uring_task
*tctx
= current
->io_uring
;
9435 struct io_tctx_exit
*work
;
9437 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
9439 * When @in_idle, we're in cancellation and it's racy to remove the
9440 * node. It'll be removed by the end of cancellation, just ignore it.
9442 if (!atomic_read(&tctx
->in_idle
))
9443 io_uring_del_tctx_node((unsigned long)work
->ctx
);
9444 complete(&work
->completion
);
9447 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
9449 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9451 return req
->ctx
== data
;
9454 static void io_ring_exit_work(struct work_struct
*work
)
9456 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
9457 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
9458 unsigned long interval
= HZ
/ 20;
9459 struct io_tctx_exit exit
;
9460 struct io_tctx_node
*node
;
9464 * If we're doing polled IO and end up having requests being
9465 * submitted async (out-of-line), then completions can come in while
9466 * we're waiting for refs to drop. We need to reap these manually,
9467 * as nobody else will be looking for them.
9470 io_uring_try_cancel_requests(ctx
, NULL
, true);
9472 struct io_sq_data
*sqd
= ctx
->sq_data
;
9473 struct task_struct
*tsk
;
9475 io_sq_thread_park(sqd
);
9477 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
9478 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
9479 io_cancel_ctx_cb
, ctx
, true);
9480 io_sq_thread_unpark(sqd
);
9483 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
9484 /* there is little hope left, don't run it too often */
9487 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
9489 init_completion(&exit
.completion
);
9490 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
9493 * Some may use context even when all refs and requests have been put,
9494 * and they are free to do so while still holding uring_lock or
9495 * completion_lock, see io_req_task_submit(). Apart from other work,
9496 * this lock/unlock section also waits them to finish.
9498 mutex_lock(&ctx
->uring_lock
);
9499 while (!list_empty(&ctx
->tctx_list
)) {
9500 WARN_ON_ONCE(time_after(jiffies
, timeout
));
9502 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
9504 /* don't spin on a single task if cancellation failed */
9505 list_rotate_left(&ctx
->tctx_list
);
9506 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
9507 if (WARN_ON_ONCE(ret
))
9509 wake_up_process(node
->task
);
9511 mutex_unlock(&ctx
->uring_lock
);
9512 wait_for_completion(&exit
.completion
);
9513 mutex_lock(&ctx
->uring_lock
);
9515 mutex_unlock(&ctx
->uring_lock
);
9516 spin_lock(&ctx
->completion_lock
);
9517 spin_unlock(&ctx
->completion_lock
);
9519 io_ring_ctx_free(ctx
);
9522 /* Returns true if we found and killed one or more timeouts */
9523 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
9526 struct io_kiocb
*req
, *tmp
;
9529 spin_lock(&ctx
->completion_lock
);
9530 spin_lock_irq(&ctx
->timeout_lock
);
9531 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
9532 if (io_match_task(req
, tsk
, cancel_all
)) {
9533 io_kill_timeout(req
, -ECANCELED
);
9537 spin_unlock_irq(&ctx
->timeout_lock
);
9539 io_commit_cqring(ctx
);
9540 spin_unlock(&ctx
->completion_lock
);
9542 io_cqring_ev_posted(ctx
);
9543 return canceled
!= 0;
9546 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
9548 unsigned long index
;
9549 struct creds
*creds
;
9551 mutex_lock(&ctx
->uring_lock
);
9552 percpu_ref_kill(&ctx
->refs
);
9554 __io_cqring_overflow_flush(ctx
, true);
9555 xa_for_each(&ctx
->personalities
, index
, creds
)
9556 io_unregister_personality(ctx
, index
);
9557 mutex_unlock(&ctx
->uring_lock
);
9559 io_kill_timeouts(ctx
, NULL
, true);
9560 io_poll_remove_all(ctx
, NULL
, true);
9562 /* if we failed setting up the ctx, we might not have any rings */
9563 io_iopoll_try_reap_events(ctx
);
9565 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
9567 * Use system_unbound_wq to avoid spawning tons of event kworkers
9568 * if we're exiting a ton of rings at the same time. It just adds
9569 * noise and overhead, there's no discernable change in runtime
9570 * over using system_wq.
9572 queue_work(system_unbound_wq
, &ctx
->exit_work
);
9575 static int io_uring_release(struct inode
*inode
, struct file
*file
)
9577 struct io_ring_ctx
*ctx
= file
->private_data
;
9579 file
->private_data
= NULL
;
9580 io_ring_ctx_wait_and_kill(ctx
);
9584 struct io_task_cancel
{
9585 struct task_struct
*task
;
9589 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
9591 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9592 struct io_task_cancel
*cancel
= data
;
9594 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
9597 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
9598 struct task_struct
*task
, bool cancel_all
)
9600 struct io_defer_entry
*de
;
9603 spin_lock(&ctx
->completion_lock
);
9604 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
9605 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
9606 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
9610 spin_unlock(&ctx
->completion_lock
);
9611 if (list_empty(&list
))
9614 while (!list_empty(&list
)) {
9615 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
9616 list_del_init(&de
->list
);
9617 io_req_complete_failed(de
->req
, -ECANCELED
);
9623 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
9625 struct io_tctx_node
*node
;
9626 enum io_wq_cancel cret
;
9629 mutex_lock(&ctx
->uring_lock
);
9630 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
9631 struct io_uring_task
*tctx
= node
->task
->io_uring
;
9634 * io_wq will stay alive while we hold uring_lock, because it's
9635 * killed after ctx nodes, which requires to take the lock.
9637 if (!tctx
|| !tctx
->io_wq
)
9639 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
9640 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9642 mutex_unlock(&ctx
->uring_lock
);
9647 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
9648 struct task_struct
*task
,
9651 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
9652 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
9655 enum io_wq_cancel cret
;
9659 ret
|= io_uring_try_cancel_iowq(ctx
);
9660 } else if (tctx
&& tctx
->io_wq
) {
9662 * Cancels requests of all rings, not only @ctx, but
9663 * it's fine as the task is in exit/exec.
9665 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
9667 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9670 /* SQPOLL thread does its own polling */
9671 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
9672 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
9673 while (!list_empty_careful(&ctx
->iopoll_list
)) {
9674 io_iopoll_try_reap_events(ctx
);
9679 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
9680 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
9681 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
9683 ret
|= io_run_task_work();
9690 static int __io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9692 struct io_uring_task
*tctx
= current
->io_uring
;
9693 struct io_tctx_node
*node
;
9696 if (unlikely(!tctx
)) {
9697 ret
= io_uring_alloc_task_context(current
, ctx
);
9701 tctx
= current
->io_uring
;
9702 if (ctx
->iowq_limits_set
) {
9703 unsigned int limits
[2] = { ctx
->iowq_limits
[0],
9704 ctx
->iowq_limits
[1], };
9706 ret
= io_wq_max_workers(tctx
->io_wq
, limits
);
9711 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
9712 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
9716 node
->task
= current
;
9718 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
9725 mutex_lock(&ctx
->uring_lock
);
9726 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
9727 mutex_unlock(&ctx
->uring_lock
);
9734 * Note that this task has used io_uring. We use it for cancelation purposes.
9736 static inline int io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9738 struct io_uring_task
*tctx
= current
->io_uring
;
9740 if (likely(tctx
&& tctx
->last
== ctx
))
9742 return __io_uring_add_tctx_node(ctx
);
9746 * Remove this io_uring_file -> task mapping.
9748 static void io_uring_del_tctx_node(unsigned long index
)
9750 struct io_uring_task
*tctx
= current
->io_uring
;
9751 struct io_tctx_node
*node
;
9755 node
= xa_erase(&tctx
->xa
, index
);
9759 WARN_ON_ONCE(current
!= node
->task
);
9760 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
9762 mutex_lock(&node
->ctx
->uring_lock
);
9763 list_del(&node
->ctx_node
);
9764 mutex_unlock(&node
->ctx
->uring_lock
);
9766 if (tctx
->last
== node
->ctx
)
9771 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
9773 struct io_wq
*wq
= tctx
->io_wq
;
9774 struct io_tctx_node
*node
;
9775 unsigned long index
;
9777 xa_for_each(&tctx
->xa
, index
, node
) {
9778 io_uring_del_tctx_node(index
);
9783 * Must be after io_uring_del_task_file() (removes nodes under
9784 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9786 io_wq_put_and_exit(wq
);
9791 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
9794 return atomic_read(&tctx
->inflight_tracked
);
9795 return percpu_counter_sum(&tctx
->inflight
);
9799 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9800 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9802 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
9804 struct io_uring_task
*tctx
= current
->io_uring
;
9805 struct io_ring_ctx
*ctx
;
9809 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
9811 if (!current
->io_uring
)
9814 io_wq_exit_start(tctx
->io_wq
);
9816 atomic_inc(&tctx
->in_idle
);
9818 io_uring_drop_tctx_refs(current
);
9819 /* read completions before cancelations */
9820 inflight
= tctx_inflight(tctx
, !cancel_all
);
9825 struct io_tctx_node
*node
;
9826 unsigned long index
;
9828 xa_for_each(&tctx
->xa
, index
, node
) {
9829 /* sqpoll task will cancel all its requests */
9830 if (node
->ctx
->sq_data
)
9832 io_uring_try_cancel_requests(node
->ctx
, current
,
9836 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9837 io_uring_try_cancel_requests(ctx
, current
,
9841 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
9843 io_uring_drop_tctx_refs(current
);
9846 * If we've seen completions, retry without waiting. This
9847 * avoids a race where a completion comes in before we did
9848 * prepare_to_wait().
9850 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
9852 finish_wait(&tctx
->wait
, &wait
);
9855 io_uring_clean_tctx(tctx
);
9858 * We shouldn't run task_works after cancel, so just leave
9859 * ->in_idle set for normal exit.
9861 atomic_dec(&tctx
->in_idle
);
9862 /* for exec all current's requests should be gone, kill tctx */
9863 __io_uring_free(current
);
9867 void __io_uring_cancel(bool cancel_all
)
9869 io_uring_cancel_generic(cancel_all
, NULL
);
9872 static void *io_uring_validate_mmap_request(struct file
*file
,
9873 loff_t pgoff
, size_t sz
)
9875 struct io_ring_ctx
*ctx
= file
->private_data
;
9876 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9881 case IORING_OFF_SQ_RING
:
9882 case IORING_OFF_CQ_RING
:
9885 case IORING_OFF_SQES
:
9889 return ERR_PTR(-EINVAL
);
9892 page
= virt_to_head_page(ptr
);
9893 if (sz
> page_size(page
))
9894 return ERR_PTR(-EINVAL
);
9901 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9903 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9907 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9909 return PTR_ERR(ptr
);
9911 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9912 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9915 #else /* !CONFIG_MMU */
9917 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9919 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9922 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9924 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9927 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9928 unsigned long addr
, unsigned long len
,
9929 unsigned long pgoff
, unsigned long flags
)
9933 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9935 return PTR_ERR(ptr
);
9937 return (unsigned long) ptr
;
9940 #endif /* !CONFIG_MMU */
9942 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9947 if (!io_sqring_full(ctx
))
9949 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9951 if (!io_sqring_full(ctx
))
9954 } while (!signal_pending(current
));
9956 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9960 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9961 struct __kernel_timespec __user
**ts
,
9962 const sigset_t __user
**sig
)
9964 struct io_uring_getevents_arg arg
;
9967 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9968 * is just a pointer to the sigset_t.
9970 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9971 *sig
= (const sigset_t __user
*) argp
;
9977 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9978 * timespec and sigset_t pointers if good.
9980 if (*argsz
!= sizeof(arg
))
9982 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9984 *sig
= u64_to_user_ptr(arg
.sigmask
);
9985 *argsz
= arg
.sigmask_sz
;
9986 *ts
= u64_to_user_ptr(arg
.ts
);
9990 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9991 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9994 struct io_ring_ctx
*ctx
;
10001 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
10002 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
)))
10006 if (unlikely(!f
.file
))
10010 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
10014 ctx
= f
.file
->private_data
;
10015 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
10019 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10023 * For SQ polling, the thread will do all submissions and completions.
10024 * Just return the requested submit count, and wake the thread if
10025 * we were asked to.
10028 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10029 io_cqring_overflow_flush(ctx
);
10031 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
10035 if (flags
& IORING_ENTER_SQ_WAKEUP
)
10036 wake_up(&ctx
->sq_data
->wait
);
10037 if (flags
& IORING_ENTER_SQ_WAIT
) {
10038 ret
= io_sqpoll_wait_sq(ctx
);
10042 submitted
= to_submit
;
10043 } else if (to_submit
) {
10044 ret
= io_uring_add_tctx_node(ctx
);
10047 mutex_lock(&ctx
->uring_lock
);
10048 submitted
= io_submit_sqes(ctx
, to_submit
);
10049 mutex_unlock(&ctx
->uring_lock
);
10051 if (submitted
!= to_submit
)
10054 if (flags
& IORING_ENTER_GETEVENTS
) {
10055 const sigset_t __user
*sig
;
10056 struct __kernel_timespec __user
*ts
;
10058 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
10062 min_complete
= min(min_complete
, ctx
->cq_entries
);
10065 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10066 * space applications don't need to do io completion events
10067 * polling again, they can rely on io_sq_thread to do polling
10068 * work, which can reduce cpu usage and uring_lock contention.
10070 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
10071 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10072 ret
= io_iopoll_check(ctx
, min_complete
);
10074 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
10079 percpu_ref_put(&ctx
->refs
);
10082 return submitted
? submitted
: ret
;
10085 #ifdef CONFIG_PROC_FS
10086 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
10087 const struct cred
*cred
)
10089 struct user_namespace
*uns
= seq_user_ns(m
);
10090 struct group_info
*gi
;
10095 seq_printf(m
, "%5d\n", id
);
10096 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
10097 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
10098 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
10099 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
10100 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
10101 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
10102 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
10103 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
10104 seq_puts(m
, "\n\tGroups:\t");
10105 gi
= cred
->group_info
;
10106 for (g
= 0; g
< gi
->ngroups
; g
++) {
10107 seq_put_decimal_ull(m
, g
? " " : "",
10108 from_kgid_munged(uns
, gi
->gid
[g
]));
10110 seq_puts(m
, "\n\tCapEff:\t");
10111 cap
= cred
->cap_effective
;
10112 CAP_FOR_EACH_U32(__capi
)
10113 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
10118 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
10120 struct io_sq_data
*sq
= NULL
;
10125 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10126 * since fdinfo case grabs it in the opposite direction of normal use
10127 * cases. If we fail to get the lock, we just don't iterate any
10128 * structures that could be going away outside the io_uring mutex.
10130 has_lock
= mutex_trylock(&ctx
->uring_lock
);
10132 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10138 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
10139 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
10140 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
10141 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
10142 struct file
*f
= io_file_from_index(ctx
, i
);
10145 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
10147 seq_printf(m
, "%5u: <none>\n", i
);
10149 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
10150 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
10151 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
10152 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
10154 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
10156 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
10157 unsigned long index
;
10158 const struct cred
*cred
;
10160 seq_printf(m
, "Personalities:\n");
10161 xa_for_each(&ctx
->personalities
, index
, cred
)
10162 io_uring_show_cred(m
, index
, cred
);
10164 seq_printf(m
, "PollList:\n");
10165 spin_lock(&ctx
->completion_lock
);
10166 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
10167 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
10168 struct io_kiocb
*req
;
10170 hlist_for_each_entry(req
, list
, hash_node
)
10171 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
10172 req
->task
->task_works
!= NULL
);
10174 spin_unlock(&ctx
->completion_lock
);
10176 mutex_unlock(&ctx
->uring_lock
);
10179 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
10181 struct io_ring_ctx
*ctx
= f
->private_data
;
10183 if (percpu_ref_tryget(&ctx
->refs
)) {
10184 __io_uring_show_fdinfo(ctx
, m
);
10185 percpu_ref_put(&ctx
->refs
);
10190 static const struct file_operations io_uring_fops
= {
10191 .release
= io_uring_release
,
10192 .mmap
= io_uring_mmap
,
10194 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
10195 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
10197 .poll
= io_uring_poll
,
10198 #ifdef CONFIG_PROC_FS
10199 .show_fdinfo
= io_uring_show_fdinfo
,
10203 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
10204 struct io_uring_params
*p
)
10206 struct io_rings
*rings
;
10207 size_t size
, sq_array_offset
;
10209 /* make sure these are sane, as we already accounted them */
10210 ctx
->sq_entries
= p
->sq_entries
;
10211 ctx
->cq_entries
= p
->cq_entries
;
10213 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
10214 if (size
== SIZE_MAX
)
10217 rings
= io_mem_alloc(size
);
10221 ctx
->rings
= rings
;
10222 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
10223 rings
->sq_ring_mask
= p
->sq_entries
- 1;
10224 rings
->cq_ring_mask
= p
->cq_entries
- 1;
10225 rings
->sq_ring_entries
= p
->sq_entries
;
10226 rings
->cq_ring_entries
= p
->cq_entries
;
10228 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
10229 if (size
== SIZE_MAX
) {
10230 io_mem_free(ctx
->rings
);
10235 ctx
->sq_sqes
= io_mem_alloc(size
);
10236 if (!ctx
->sq_sqes
) {
10237 io_mem_free(ctx
->rings
);
10245 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
10249 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
10253 ret
= io_uring_add_tctx_node(ctx
);
10258 fd_install(fd
, file
);
10263 * Allocate an anonymous fd, this is what constitutes the application
10264 * visible backing of an io_uring instance. The application mmaps this
10265 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10266 * we have to tie this fd to a socket for file garbage collection purposes.
10268 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
10271 #if defined(CONFIG_UNIX)
10274 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
10277 return ERR_PTR(ret
);
10280 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
10281 O_RDWR
| O_CLOEXEC
);
10282 #if defined(CONFIG_UNIX)
10283 if (IS_ERR(file
)) {
10284 sock_release(ctx
->ring_sock
);
10285 ctx
->ring_sock
= NULL
;
10287 ctx
->ring_sock
->file
= file
;
10293 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
10294 struct io_uring_params __user
*params
)
10296 struct io_ring_ctx
*ctx
;
10302 if (entries
> IORING_MAX_ENTRIES
) {
10303 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10305 entries
= IORING_MAX_ENTRIES
;
10309 * Use twice as many entries for the CQ ring. It's possible for the
10310 * application to drive a higher depth than the size of the SQ ring,
10311 * since the sqes are only used at submission time. This allows for
10312 * some flexibility in overcommitting a bit. If the application has
10313 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10314 * of CQ ring entries manually.
10316 p
->sq_entries
= roundup_pow_of_two(entries
);
10317 if (p
->flags
& IORING_SETUP_CQSIZE
) {
10319 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10320 * to a power-of-two, if it isn't already. We do NOT impose
10321 * any cq vs sq ring sizing.
10323 if (!p
->cq_entries
)
10325 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
10326 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10328 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
10330 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
10331 if (p
->cq_entries
< p
->sq_entries
)
10334 p
->cq_entries
= 2 * p
->sq_entries
;
10337 ctx
= io_ring_ctx_alloc(p
);
10340 ctx
->compat
= in_compat_syscall();
10341 if (!capable(CAP_IPC_LOCK
))
10342 ctx
->user
= get_uid(current_user());
10345 * This is just grabbed for accounting purposes. When a process exits,
10346 * the mm is exited and dropped before the files, hence we need to hang
10347 * on to this mm purely for the purposes of being able to unaccount
10348 * memory (locked/pinned vm). It's not used for anything else.
10350 mmgrab(current
->mm
);
10351 ctx
->mm_account
= current
->mm
;
10353 ret
= io_allocate_scq_urings(ctx
, p
);
10357 ret
= io_sq_offload_create(ctx
, p
);
10360 /* always set a rsrc node */
10361 ret
= io_rsrc_node_switch_start(ctx
);
10364 io_rsrc_node_switch(ctx
, NULL
);
10366 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
10367 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
10368 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
10369 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
10370 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
10371 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
10372 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
10373 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
10375 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
10376 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
10377 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
10378 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
10379 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
10380 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
10381 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
10382 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
10384 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
10385 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
10386 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
10387 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
10388 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
10389 IORING_FEAT_RSRC_TAGS
;
10391 if (copy_to_user(params
, p
, sizeof(*p
))) {
10396 file
= io_uring_get_file(ctx
);
10397 if (IS_ERR(file
)) {
10398 ret
= PTR_ERR(file
);
10403 * Install ring fd as the very last thing, so we don't risk someone
10404 * having closed it before we finish setup
10406 ret
= io_uring_install_fd(ctx
, file
);
10408 /* fput will clean it up */
10413 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
10416 io_ring_ctx_wait_and_kill(ctx
);
10421 * Sets up an aio uring context, and returns the fd. Applications asks for a
10422 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10423 * params structure passed in.
10425 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
10427 struct io_uring_params p
;
10430 if (copy_from_user(&p
, params
, sizeof(p
)))
10432 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
10437 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
10438 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
10439 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
10440 IORING_SETUP_R_DISABLED
))
10443 return io_uring_create(entries
, &p
, params
);
10446 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
10447 struct io_uring_params __user
*, params
)
10449 return io_uring_setup(entries
, params
);
10452 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
10454 struct io_uring_probe
*p
;
10458 size
= struct_size(p
, ops
, nr_args
);
10459 if (size
== SIZE_MAX
)
10461 p
= kzalloc(size
, GFP_KERNEL
);
10466 if (copy_from_user(p
, arg
, size
))
10469 if (memchr_inv(p
, 0, size
))
10472 p
->last_op
= IORING_OP_LAST
- 1;
10473 if (nr_args
> IORING_OP_LAST
)
10474 nr_args
= IORING_OP_LAST
;
10476 for (i
= 0; i
< nr_args
; i
++) {
10478 if (!io_op_defs
[i
].not_supported
)
10479 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
10484 if (copy_to_user(arg
, p
, size
))
10491 static int io_register_personality(struct io_ring_ctx
*ctx
)
10493 const struct cred
*creds
;
10497 creds
= get_current_cred();
10499 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
10500 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
10508 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
10509 unsigned int nr_args
)
10511 struct io_uring_restriction
*res
;
10515 /* Restrictions allowed only if rings started disabled */
10516 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10519 /* We allow only a single restrictions registration */
10520 if (ctx
->restrictions
.registered
)
10523 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
10526 size
= array_size(nr_args
, sizeof(*res
));
10527 if (size
== SIZE_MAX
)
10530 res
= memdup_user(arg
, size
);
10532 return PTR_ERR(res
);
10536 for (i
= 0; i
< nr_args
; i
++) {
10537 switch (res
[i
].opcode
) {
10538 case IORING_RESTRICTION_REGISTER_OP
:
10539 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
10544 __set_bit(res
[i
].register_op
,
10545 ctx
->restrictions
.register_op
);
10547 case IORING_RESTRICTION_SQE_OP
:
10548 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
10553 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
10555 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
10556 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
10558 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
10559 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
10568 /* Reset all restrictions if an error happened */
10570 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
10572 ctx
->restrictions
.registered
= true;
10578 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10580 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10583 if (ctx
->restrictions
.registered
)
10584 ctx
->restricted
= 1;
10586 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
10587 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
10588 wake_up(&ctx
->sq_data
->wait
);
10592 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
10593 struct io_uring_rsrc_update2
*up
,
10601 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
10603 err
= io_rsrc_node_switch_start(ctx
);
10608 case IORING_RSRC_FILE
:
10609 return __io_sqe_files_update(ctx
, up
, nr_args
);
10610 case IORING_RSRC_BUFFER
:
10611 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
10616 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10619 struct io_uring_rsrc_update2 up
;
10623 memset(&up
, 0, sizeof(up
));
10624 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
10626 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
10629 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10630 unsigned size
, unsigned type
)
10632 struct io_uring_rsrc_update2 up
;
10634 if (size
!= sizeof(up
))
10636 if (copy_from_user(&up
, arg
, sizeof(up
)))
10638 if (!up
.nr
|| up
.resv
)
10640 return __io_register_rsrc_update(ctx
, type
, &up
, up
.nr
);
10643 static int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
10644 unsigned int size
, unsigned int type
)
10646 struct io_uring_rsrc_register rr
;
10648 /* keep it extendible */
10649 if (size
!= sizeof(rr
))
10652 memset(&rr
, 0, sizeof(rr
));
10653 if (copy_from_user(&rr
, arg
, size
))
10655 if (!rr
.nr
|| rr
.resv
|| rr
.resv2
)
10659 case IORING_RSRC_FILE
:
10660 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
10661 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10662 case IORING_RSRC_BUFFER
:
10663 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
10664 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10669 static int io_register_iowq_aff(struct io_ring_ctx
*ctx
, void __user
*arg
,
10672 struct io_uring_task
*tctx
= current
->io_uring
;
10673 cpumask_var_t new_mask
;
10676 if (!tctx
|| !tctx
->io_wq
)
10679 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
10682 cpumask_clear(new_mask
);
10683 if (len
> cpumask_size())
10684 len
= cpumask_size();
10686 if (copy_from_user(new_mask
, arg
, len
)) {
10687 free_cpumask_var(new_mask
);
10691 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
10692 free_cpumask_var(new_mask
);
10696 static int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
10698 struct io_uring_task
*tctx
= current
->io_uring
;
10700 if (!tctx
|| !tctx
->io_wq
)
10703 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
10706 static int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
10708 __must_hold(&ctx
->uring_lock
)
10710 struct io_tctx_node
*node
;
10711 struct io_uring_task
*tctx
= NULL
;
10712 struct io_sq_data
*sqd
= NULL
;
10713 __u32 new_count
[2];
10716 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
10718 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10719 if (new_count
[i
] > INT_MAX
)
10722 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10723 sqd
= ctx
->sq_data
;
10726 * Observe the correct sqd->lock -> ctx->uring_lock
10727 * ordering. Fine to drop uring_lock here, we hold
10728 * a ref to the ctx.
10730 refcount_inc(&sqd
->refs
);
10731 mutex_unlock(&ctx
->uring_lock
);
10732 mutex_lock(&sqd
->lock
);
10733 mutex_lock(&ctx
->uring_lock
);
10735 tctx
= sqd
->thread
->io_uring
;
10738 tctx
= current
->io_uring
;
10741 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
10743 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10745 ctx
->iowq_limits
[i
] = new_count
[i
];
10746 ctx
->iowq_limits_set
= true;
10749 if (tctx
&& tctx
->io_wq
) {
10750 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
10754 memset(new_count
, 0, sizeof(new_count
));
10758 mutex_unlock(&sqd
->lock
);
10759 io_put_sq_data(sqd
);
10762 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
10765 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10769 /* now propagate the restriction to all registered users */
10770 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
10771 struct io_uring_task
*tctx
= node
->task
->io_uring
;
10773 if (WARN_ON_ONCE(!tctx
->io_wq
))
10776 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10777 new_count
[i
] = ctx
->iowq_limits
[i
];
10778 /* ignore errors, it always returns zero anyway */
10779 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
10784 mutex_unlock(&sqd
->lock
);
10785 io_put_sq_data(sqd
);
10790 static bool io_register_op_must_quiesce(int op
)
10793 case IORING_REGISTER_BUFFERS
:
10794 case IORING_UNREGISTER_BUFFERS
:
10795 case IORING_REGISTER_FILES
:
10796 case IORING_UNREGISTER_FILES
:
10797 case IORING_REGISTER_FILES_UPDATE
:
10798 case IORING_REGISTER_PROBE
:
10799 case IORING_REGISTER_PERSONALITY
:
10800 case IORING_UNREGISTER_PERSONALITY
:
10801 case IORING_REGISTER_FILES2
:
10802 case IORING_REGISTER_FILES_UPDATE2
:
10803 case IORING_REGISTER_BUFFERS2
:
10804 case IORING_REGISTER_BUFFERS_UPDATE
:
10805 case IORING_REGISTER_IOWQ_AFF
:
10806 case IORING_UNREGISTER_IOWQ_AFF
:
10807 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10814 static int io_ctx_quiesce(struct io_ring_ctx
*ctx
)
10818 percpu_ref_kill(&ctx
->refs
);
10821 * Drop uring mutex before waiting for references to exit. If another
10822 * thread is currently inside io_uring_enter() it might need to grab the
10823 * uring_lock to make progress. If we hold it here across the drain
10824 * wait, then we can deadlock. It's safe to drop the mutex here, since
10825 * no new references will come in after we've killed the percpu ref.
10827 mutex_unlock(&ctx
->uring_lock
);
10829 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10832 ret
= io_run_task_work_sig();
10833 } while (ret
>= 0);
10834 mutex_lock(&ctx
->uring_lock
);
10837 io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
);
10841 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10842 void __user
*arg
, unsigned nr_args
)
10843 __releases(ctx
->uring_lock
)
10844 __acquires(ctx
->uring_lock
)
10849 * We're inside the ring mutex, if the ref is already dying, then
10850 * someone else killed the ctx or is already going through
10851 * io_uring_register().
10853 if (percpu_ref_is_dying(&ctx
->refs
))
10856 if (ctx
->restricted
) {
10857 if (opcode
>= IORING_REGISTER_LAST
)
10859 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
10860 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
10864 if (io_register_op_must_quiesce(opcode
)) {
10865 ret
= io_ctx_quiesce(ctx
);
10871 case IORING_REGISTER_BUFFERS
:
10872 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
10874 case IORING_UNREGISTER_BUFFERS
:
10876 if (arg
|| nr_args
)
10878 ret
= io_sqe_buffers_unregister(ctx
);
10880 case IORING_REGISTER_FILES
:
10881 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
10883 case IORING_UNREGISTER_FILES
:
10885 if (arg
|| nr_args
)
10887 ret
= io_sqe_files_unregister(ctx
);
10889 case IORING_REGISTER_FILES_UPDATE
:
10890 ret
= io_register_files_update(ctx
, arg
, nr_args
);
10892 case IORING_REGISTER_EVENTFD
:
10893 case IORING_REGISTER_EVENTFD_ASYNC
:
10897 ret
= io_eventfd_register(ctx
, arg
);
10900 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10901 ctx
->eventfd_async
= 1;
10903 ctx
->eventfd_async
= 0;
10905 case IORING_UNREGISTER_EVENTFD
:
10907 if (arg
|| nr_args
)
10909 ret
= io_eventfd_unregister(ctx
);
10911 case IORING_REGISTER_PROBE
:
10913 if (!arg
|| nr_args
> 256)
10915 ret
= io_probe(ctx
, arg
, nr_args
);
10917 case IORING_REGISTER_PERSONALITY
:
10919 if (arg
|| nr_args
)
10921 ret
= io_register_personality(ctx
);
10923 case IORING_UNREGISTER_PERSONALITY
:
10927 ret
= io_unregister_personality(ctx
, nr_args
);
10929 case IORING_REGISTER_ENABLE_RINGS
:
10931 if (arg
|| nr_args
)
10933 ret
= io_register_enable_rings(ctx
);
10935 case IORING_REGISTER_RESTRICTIONS
:
10936 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10938 case IORING_REGISTER_FILES2
:
10939 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
10941 case IORING_REGISTER_FILES_UPDATE2
:
10942 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10945 case IORING_REGISTER_BUFFERS2
:
10946 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
10948 case IORING_REGISTER_BUFFERS_UPDATE
:
10949 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10950 IORING_RSRC_BUFFER
);
10952 case IORING_REGISTER_IOWQ_AFF
:
10954 if (!arg
|| !nr_args
)
10956 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
10958 case IORING_UNREGISTER_IOWQ_AFF
:
10960 if (arg
|| nr_args
)
10962 ret
= io_unregister_iowq_aff(ctx
);
10964 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10966 if (!arg
|| nr_args
!= 2)
10968 ret
= io_register_iowq_max_workers(ctx
, arg
);
10975 if (io_register_op_must_quiesce(opcode
)) {
10976 /* bring the ctx back to life */
10977 percpu_ref_reinit(&ctx
->refs
);
10978 reinit_completion(&ctx
->ref_comp
);
10983 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10984 void __user
*, arg
, unsigned int, nr_args
)
10986 struct io_ring_ctx
*ctx
;
10995 if (f
.file
->f_op
!= &io_uring_fops
)
10998 ctx
= f
.file
->private_data
;
11000 io_run_task_work();
11002 mutex_lock(&ctx
->uring_lock
);
11003 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
11004 mutex_unlock(&ctx
->uring_lock
);
11005 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
11006 ctx
->cq_ev_fd
!= NULL
, ret
);
11012 static int __init
io_uring_init(void)
11014 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11015 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11016 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11019 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11020 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11021 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
11022 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
11023 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
11024 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
11025 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
11026 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
11027 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
11028 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
11029 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
11030 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
11031 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
11032 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
11033 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
11034 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
11035 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
11036 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
11037 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
11038 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
11039 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
11040 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
11041 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
11042 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
11043 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
11044 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
11045 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
11046 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
11047 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
11048 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
11049 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
11050 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
11051 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
11053 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
11054 sizeof(struct io_uring_rsrc_update
));
11055 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
11056 sizeof(struct io_uring_rsrc_update2
));
11058 /* ->buf_index is u16 */
11059 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS
>= (1u << 16));
11061 /* should fit into one byte */
11062 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
11064 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
11065 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
11067 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
11071 __initcall(io_uring_init
);