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
;
626 struct file
*file_in
;
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
;
1456 switch (req
->opcode
) {
1457 case IORING_OP_SPLICE
:
1459 if (!S_ISREG(file_inode(req
->splice
.file_in
)->i_mode
))
1460 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1465 static void io_prep_async_link(struct io_kiocb
*req
)
1467 struct io_kiocb
*cur
;
1469 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
1470 struct io_ring_ctx
*ctx
= req
->ctx
;
1472 spin_lock_irq(&ctx
->timeout_lock
);
1473 io_for_each_link(cur
, req
)
1474 io_prep_async_work(cur
);
1475 spin_unlock_irq(&ctx
->timeout_lock
);
1477 io_for_each_link(cur
, req
)
1478 io_prep_async_work(cur
);
1482 static void io_queue_async_work(struct io_kiocb
*req
, bool *locked
)
1484 struct io_ring_ctx
*ctx
= req
->ctx
;
1485 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1486 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1488 /* must not take the lock, NULL it as a precaution */
1492 BUG_ON(!tctx
->io_wq
);
1494 /* init ->work of the whole link before punting */
1495 io_prep_async_link(req
);
1498 * Not expected to happen, but if we do have a bug where this _can_
1499 * happen, catch it here and ensure the request is marked as
1500 * canceled. That will make io-wq go through the usual work cancel
1501 * procedure rather than attempt to run this request (or create a new
1504 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
1505 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1507 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1508 &req
->work
, req
->flags
);
1509 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
1511 io_queue_linked_timeout(link
);
1514 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1515 __must_hold(&req
->ctx
->completion_lock
)
1516 __must_hold(&req
->ctx
->timeout_lock
)
1518 struct io_timeout_data
*io
= req
->async_data
;
1520 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1523 atomic_set(&req
->ctx
->cq_timeouts
,
1524 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1525 list_del_init(&req
->timeout
.list
);
1526 io_cqring_fill_event(req
->ctx
, req
->user_data
, status
, 0);
1527 io_put_req_deferred(req
);
1531 static void io_queue_deferred(struct io_ring_ctx
*ctx
)
1533 while (!list_empty(&ctx
->defer_list
)) {
1534 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1535 struct io_defer_entry
, list
);
1537 if (req_need_defer(de
->req
, de
->seq
))
1539 list_del_init(&de
->list
);
1540 io_req_task_queue(de
->req
);
1545 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1546 __must_hold(&ctx
->completion_lock
)
1548 u32 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1550 spin_lock_irq(&ctx
->timeout_lock
);
1551 while (!list_empty(&ctx
->timeout_list
)) {
1552 u32 events_needed
, events_got
;
1553 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1554 struct io_kiocb
, timeout
.list
);
1556 if (io_is_timeout_noseq(req
))
1560 * Since seq can easily wrap around over time, subtract
1561 * the last seq at which timeouts were flushed before comparing.
1562 * Assuming not more than 2^31-1 events have happened since,
1563 * these subtractions won't have wrapped, so we can check if
1564 * target is in [last_seq, current_seq] by comparing the two.
1566 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1567 events_got
= seq
- ctx
->cq_last_tm_flush
;
1568 if (events_got
< events_needed
)
1571 list_del_init(&req
->timeout
.list
);
1572 io_kill_timeout(req
, 0);
1574 ctx
->cq_last_tm_flush
= seq
;
1575 spin_unlock_irq(&ctx
->timeout_lock
);
1578 static void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
1580 if (ctx
->off_timeout_used
)
1581 io_flush_timeouts(ctx
);
1582 if (ctx
->drain_active
)
1583 io_queue_deferred(ctx
);
1586 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
1588 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
))
1589 __io_commit_cqring_flush(ctx
);
1590 /* order cqe stores with ring update */
1591 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1594 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1596 struct io_rings
*r
= ctx
->rings
;
1598 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
1601 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1603 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1606 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
1608 struct io_rings
*rings
= ctx
->rings
;
1609 unsigned tail
, mask
= ctx
->cq_entries
- 1;
1612 * writes to the cq entry need to come after reading head; the
1613 * control dependency is enough as we're using WRITE_ONCE to
1616 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
1619 tail
= ctx
->cached_cq_tail
++;
1620 return &rings
->cqes
[tail
& mask
];
1623 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1625 if (likely(!ctx
->cq_ev_fd
))
1627 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1629 return !ctx
->eventfd_async
|| io_wq_current_is_worker();
1633 * This should only get called when at least one event has been posted.
1634 * Some applications rely on the eventfd notification count only changing
1635 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1636 * 1:1 relationship between how many times this function is called (and
1637 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1639 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1642 * wake_up_all() may seem excessive, but io_wake_function() and
1643 * io_should_wake() handle the termination of the loop and only
1644 * wake as many waiters as we need to.
1646 if (wq_has_sleeper(&ctx
->cq_wait
))
1647 wake_up_all(&ctx
->cq_wait
);
1648 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1649 wake_up(&ctx
->sq_data
->wait
);
1650 if (io_should_trigger_evfd(ctx
))
1651 eventfd_signal(ctx
->cq_ev_fd
, 1);
1652 if (waitqueue_active(&ctx
->poll_wait
))
1653 wake_up_interruptible(&ctx
->poll_wait
);
1656 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1658 /* see waitqueue_active() comment */
1661 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1662 if (waitqueue_active(&ctx
->cq_wait
))
1663 wake_up_all(&ctx
->cq_wait
);
1665 if (io_should_trigger_evfd(ctx
))
1666 eventfd_signal(ctx
->cq_ev_fd
, 1);
1667 if (waitqueue_active(&ctx
->poll_wait
))
1668 wake_up_interruptible(&ctx
->poll_wait
);
1671 /* Returns true if there are no backlogged entries after the flush */
1672 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1674 bool all_flushed
, posted
;
1676 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
1680 spin_lock(&ctx
->completion_lock
);
1681 while (!list_empty(&ctx
->cq_overflow_list
)) {
1682 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
1683 struct io_overflow_cqe
*ocqe
;
1687 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
1688 struct io_overflow_cqe
, list
);
1690 memcpy(cqe
, &ocqe
->cqe
, sizeof(*cqe
));
1692 io_account_cq_overflow(ctx
);
1695 list_del(&ocqe
->list
);
1699 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1701 clear_bit(0, &ctx
->check_cq_overflow
);
1702 WRITE_ONCE(ctx
->rings
->sq_flags
,
1703 ctx
->rings
->sq_flags
& ~IORING_SQ_CQ_OVERFLOW
);
1707 io_commit_cqring(ctx
);
1708 spin_unlock(&ctx
->completion_lock
);
1710 io_cqring_ev_posted(ctx
);
1714 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
1718 if (test_bit(0, &ctx
->check_cq_overflow
)) {
1719 /* iopoll syncs against uring_lock, not completion_lock */
1720 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1721 mutex_lock(&ctx
->uring_lock
);
1722 ret
= __io_cqring_overflow_flush(ctx
, false);
1723 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1724 mutex_unlock(&ctx
->uring_lock
);
1730 /* must to be called somewhat shortly after putting a request */
1731 static inline void io_put_task(struct task_struct
*task
, int nr
)
1733 struct io_uring_task
*tctx
= task
->io_uring
;
1735 if (likely(task
== current
)) {
1736 tctx
->cached_refs
+= nr
;
1738 percpu_counter_sub(&tctx
->inflight
, nr
);
1739 if (unlikely(atomic_read(&tctx
->in_idle
)))
1740 wake_up(&tctx
->wait
);
1741 put_task_struct_many(task
, nr
);
1745 static void io_task_refs_refill(struct io_uring_task
*tctx
)
1747 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
1749 percpu_counter_add(&tctx
->inflight
, refill
);
1750 refcount_add(refill
, ¤t
->usage
);
1751 tctx
->cached_refs
+= refill
;
1754 static inline void io_get_task_refs(int nr
)
1756 struct io_uring_task
*tctx
= current
->io_uring
;
1758 tctx
->cached_refs
-= nr
;
1759 if (unlikely(tctx
->cached_refs
< 0))
1760 io_task_refs_refill(tctx
);
1763 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
1764 long res
, unsigned int cflags
)
1766 struct io_overflow_cqe
*ocqe
;
1768 ocqe
= kmalloc(sizeof(*ocqe
), GFP_ATOMIC
| __GFP_ACCOUNT
);
1771 * If we're in ring overflow flush mode, or in task cancel mode,
1772 * or cannot allocate an overflow entry, then we need to drop it
1775 io_account_cq_overflow(ctx
);
1778 if (list_empty(&ctx
->cq_overflow_list
)) {
1779 set_bit(0, &ctx
->check_cq_overflow
);
1780 WRITE_ONCE(ctx
->rings
->sq_flags
,
1781 ctx
->rings
->sq_flags
| IORING_SQ_CQ_OVERFLOW
);
1784 ocqe
->cqe
.user_data
= user_data
;
1785 ocqe
->cqe
.res
= res
;
1786 ocqe
->cqe
.flags
= cflags
;
1787 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
1791 static inline bool __io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1792 long res
, unsigned int cflags
)
1794 struct io_uring_cqe
*cqe
;
1796 trace_io_uring_complete(ctx
, user_data
, res
, cflags
);
1799 * If we can't get a cq entry, userspace overflowed the
1800 * submission (by quite a lot). Increment the overflow count in
1803 cqe
= io_get_cqe(ctx
);
1805 WRITE_ONCE(cqe
->user_data
, user_data
);
1806 WRITE_ONCE(cqe
->res
, res
);
1807 WRITE_ONCE(cqe
->flags
, cflags
);
1810 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
);
1813 /* not as hot to bloat with inlining */
1814 static noinline
bool io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 user_data
,
1815 long res
, unsigned int cflags
)
1817 return __io_cqring_fill_event(ctx
, user_data
, res
, cflags
);
1820 static void io_req_complete_post(struct io_kiocb
*req
, long res
,
1821 unsigned int cflags
)
1823 struct io_ring_ctx
*ctx
= req
->ctx
;
1825 spin_lock(&ctx
->completion_lock
);
1826 __io_cqring_fill_event(ctx
, req
->user_data
, res
, cflags
);
1828 * If we're the last reference to this request, add to our locked
1831 if (req_ref_put_and_test(req
)) {
1832 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1833 if (req
->flags
& IO_DISARM_MASK
)
1834 io_disarm_next(req
);
1836 io_req_task_queue(req
->link
);
1840 io_dismantle_req(req
);
1841 io_put_task(req
->task
, 1);
1842 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
1843 ctx
->locked_free_nr
++;
1845 if (!percpu_ref_tryget(&ctx
->refs
))
1848 io_commit_cqring(ctx
);
1849 spin_unlock(&ctx
->completion_lock
);
1852 io_cqring_ev_posted(ctx
);
1853 percpu_ref_put(&ctx
->refs
);
1857 static inline bool io_req_needs_clean(struct io_kiocb
*req
)
1859 return req
->flags
& IO_REQ_CLEAN_FLAGS
;
1862 static void io_req_complete_state(struct io_kiocb
*req
, long res
,
1863 unsigned int cflags
)
1865 if (io_req_needs_clean(req
))
1868 req
->compl.cflags
= cflags
;
1869 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1872 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1873 long res
, unsigned cflags
)
1875 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1876 io_req_complete_state(req
, res
, cflags
);
1878 io_req_complete_post(req
, res
, cflags
);
1881 static inline void io_req_complete(struct io_kiocb
*req
, long res
)
1883 __io_req_complete(req
, 0, res
, 0);
1886 static void io_req_complete_failed(struct io_kiocb
*req
, long res
)
1889 io_req_complete_post(req
, res
, 0);
1892 static void io_req_complete_fail_submit(struct io_kiocb
*req
)
1895 * We don't submit, fail them all, for that replace hardlinks with
1896 * normal links. Extra REQ_F_LINK is tolerated.
1898 req
->flags
&= ~REQ_F_HARDLINK
;
1899 req
->flags
|= REQ_F_LINK
;
1900 io_req_complete_failed(req
, req
->result
);
1904 * Don't initialise the fields below on every allocation, but do that in
1905 * advance and keep them valid across allocations.
1907 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1911 req
->async_data
= NULL
;
1912 /* not necessary, but safer to zero */
1916 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1917 struct io_submit_state
*state
)
1919 spin_lock(&ctx
->completion_lock
);
1920 list_splice_init(&ctx
->locked_free_list
, &state
->free_list
);
1921 ctx
->locked_free_nr
= 0;
1922 spin_unlock(&ctx
->completion_lock
);
1925 /* Returns true IFF there are requests in the cache */
1926 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1928 struct io_submit_state
*state
= &ctx
->submit_state
;
1932 * If we have more than a batch's worth of requests in our IRQ side
1933 * locked cache, grab the lock and move them over to our submission
1936 if (READ_ONCE(ctx
->locked_free_nr
) > IO_COMPL_BATCH
)
1937 io_flush_cached_locked_reqs(ctx
, state
);
1939 nr
= state
->free_reqs
;
1940 while (!list_empty(&state
->free_list
)) {
1941 struct io_kiocb
*req
= list_first_entry(&state
->free_list
,
1942 struct io_kiocb
, inflight_entry
);
1944 list_del(&req
->inflight_entry
);
1945 state
->reqs
[nr
++] = req
;
1946 if (nr
== ARRAY_SIZE(state
->reqs
))
1950 state
->free_reqs
= nr
;
1955 * A request might get retired back into the request caches even before opcode
1956 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1957 * Because of that, io_alloc_req() should be called only under ->uring_lock
1958 * and with extra caution to not get a request that is still worked on.
1960 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1961 __must_hold(&ctx
->uring_lock
)
1963 struct io_submit_state
*state
= &ctx
->submit_state
;
1964 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1967 BUILD_BUG_ON(ARRAY_SIZE(state
->reqs
) < IO_REQ_ALLOC_BATCH
);
1969 if (likely(state
->free_reqs
|| io_flush_cached_reqs(ctx
)))
1972 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1976 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1977 * retry single alloc to be on the safe side.
1979 if (unlikely(ret
<= 0)) {
1980 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1981 if (!state
->reqs
[0])
1986 for (i
= 0; i
< ret
; i
++)
1987 io_preinit_req(state
->reqs
[i
], ctx
);
1988 state
->free_reqs
= ret
;
1991 return state
->reqs
[state
->free_reqs
];
1994 static inline void io_put_file(struct file
*file
)
2000 static void io_dismantle_req(struct io_kiocb
*req
)
2002 unsigned int flags
= req
->flags
;
2004 if (io_req_needs_clean(req
))
2006 if (!(flags
& REQ_F_FIXED_FILE
))
2007 io_put_file(req
->file
);
2008 if (req
->fixed_rsrc_refs
)
2009 percpu_ref_put(req
->fixed_rsrc_refs
);
2010 if (req
->async_data
) {
2011 kfree(req
->async_data
);
2012 req
->async_data
= NULL
;
2016 static void __io_free_req(struct io_kiocb
*req
)
2018 struct io_ring_ctx
*ctx
= req
->ctx
;
2020 io_dismantle_req(req
);
2021 io_put_task(req
->task
, 1);
2023 spin_lock(&ctx
->completion_lock
);
2024 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
2025 ctx
->locked_free_nr
++;
2026 spin_unlock(&ctx
->completion_lock
);
2028 percpu_ref_put(&ctx
->refs
);
2031 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2033 struct io_kiocb
*nxt
= req
->link
;
2035 req
->link
= nxt
->link
;
2039 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
2040 __must_hold(&req
->ctx
->completion_lock
)
2041 __must_hold(&req
->ctx
->timeout_lock
)
2043 struct io_kiocb
*link
= req
->link
;
2045 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2046 struct io_timeout_data
*io
= link
->async_data
;
2048 io_remove_next_linked(req
);
2049 link
->timeout
.head
= NULL
;
2050 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
2051 list_del(&link
->timeout
.list
);
2052 io_cqring_fill_event(link
->ctx
, link
->user_data
,
2054 io_put_req_deferred(link
);
2061 static void io_fail_links(struct io_kiocb
*req
)
2062 __must_hold(&req
->ctx
->completion_lock
)
2064 struct io_kiocb
*nxt
, *link
= req
->link
;
2068 long res
= -ECANCELED
;
2070 if (link
->flags
& REQ_F_FAIL
)
2076 trace_io_uring_fail_link(req
, link
);
2077 io_cqring_fill_event(link
->ctx
, link
->user_data
, res
, 0);
2078 io_put_req_deferred(link
);
2083 static bool io_disarm_next(struct io_kiocb
*req
)
2084 __must_hold(&req
->ctx
->completion_lock
)
2086 bool posted
= false;
2088 if (req
->flags
& REQ_F_ARM_LTIMEOUT
) {
2089 struct io_kiocb
*link
= req
->link
;
2091 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
2092 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2093 io_remove_next_linked(req
);
2094 io_cqring_fill_event(link
->ctx
, link
->user_data
,
2096 io_put_req_deferred(link
);
2099 } else if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
2100 struct io_ring_ctx
*ctx
= req
->ctx
;
2102 spin_lock_irq(&ctx
->timeout_lock
);
2103 posted
= io_kill_linked_timeout(req
);
2104 spin_unlock_irq(&ctx
->timeout_lock
);
2106 if (unlikely((req
->flags
& REQ_F_FAIL
) &&
2107 !(req
->flags
& REQ_F_HARDLINK
))) {
2108 posted
|= (req
->link
!= NULL
);
2114 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2116 struct io_kiocb
*nxt
;
2119 * If LINK is set, we have dependent requests in this chain. If we
2120 * didn't fail this request, queue the first one up, moving any other
2121 * dependencies to the next request. In case of failure, fail the rest
2124 if (req
->flags
& IO_DISARM_MASK
) {
2125 struct io_ring_ctx
*ctx
= req
->ctx
;
2128 spin_lock(&ctx
->completion_lock
);
2129 posted
= io_disarm_next(req
);
2131 io_commit_cqring(req
->ctx
);
2132 spin_unlock(&ctx
->completion_lock
);
2134 io_cqring_ev_posted(ctx
);
2141 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2143 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
2145 return __io_req_find_next(req
);
2148 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, bool *locked
)
2153 if (ctx
->submit_state
.compl_nr
)
2154 io_submit_flush_completions(ctx
);
2155 mutex_unlock(&ctx
->uring_lock
);
2158 percpu_ref_put(&ctx
->refs
);
2161 static void tctx_task_work(struct callback_head
*cb
)
2163 bool locked
= false;
2164 struct io_ring_ctx
*ctx
= NULL
;
2165 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
2169 struct io_wq_work_node
*node
;
2171 if (!tctx
->task_list
.first
&& locked
&& ctx
->submit_state
.compl_nr
)
2172 io_submit_flush_completions(ctx
);
2174 spin_lock_irq(&tctx
->task_lock
);
2175 node
= tctx
->task_list
.first
;
2176 INIT_WQ_LIST(&tctx
->task_list
);
2178 tctx
->task_running
= false;
2179 spin_unlock_irq(&tctx
->task_lock
);
2184 struct io_wq_work_node
*next
= node
->next
;
2185 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
2188 if (req
->ctx
!= ctx
) {
2189 ctx_flush_and_put(ctx
, &locked
);
2191 /* if not contended, grab and improve batching */
2192 locked
= mutex_trylock(&ctx
->uring_lock
);
2193 percpu_ref_get(&ctx
->refs
);
2195 req
->io_task_work
.func(req
, &locked
);
2202 ctx_flush_and_put(ctx
, &locked
);
2205 static void io_req_task_work_add(struct io_kiocb
*req
)
2207 struct task_struct
*tsk
= req
->task
;
2208 struct io_uring_task
*tctx
= tsk
->io_uring
;
2209 enum task_work_notify_mode notify
;
2210 struct io_wq_work_node
*node
;
2211 unsigned long flags
;
2214 WARN_ON_ONCE(!tctx
);
2216 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2217 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
2218 running
= tctx
->task_running
;
2220 tctx
->task_running
= true;
2221 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2223 /* task_work already pending, we're done */
2228 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2229 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2230 * processing task_work. There's no reliable way to tell if TWA_RESUME
2233 notify
= (req
->ctx
->flags
& IORING_SETUP_SQPOLL
) ? TWA_NONE
: TWA_SIGNAL
;
2234 if (!task_work_add(tsk
, &tctx
->task_work
, notify
)) {
2235 wake_up_process(tsk
);
2239 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2240 tctx
->task_running
= false;
2241 node
= tctx
->task_list
.first
;
2242 INIT_WQ_LIST(&tctx
->task_list
);
2243 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2246 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
2248 if (llist_add(&req
->io_task_work
.fallback_node
,
2249 &req
->ctx
->fallback_llist
))
2250 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
2254 static void io_req_task_cancel(struct io_kiocb
*req
, bool *locked
)
2256 struct io_ring_ctx
*ctx
= req
->ctx
;
2258 /* not needed for normal modes, but SQPOLL depends on it */
2259 io_tw_lock(ctx
, locked
);
2260 io_req_complete_failed(req
, req
->result
);
2263 static void io_req_task_submit(struct io_kiocb
*req
, bool *locked
)
2265 struct io_ring_ctx
*ctx
= req
->ctx
;
2267 io_tw_lock(ctx
, locked
);
2268 /* req->task == current here, checking PF_EXITING is safe */
2269 if (likely(!(req
->task
->flags
& PF_EXITING
)))
2270 __io_queue_sqe(req
);
2272 io_req_complete_failed(req
, -EFAULT
);
2275 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2278 req
->io_task_work
.func
= io_req_task_cancel
;
2279 io_req_task_work_add(req
);
2282 static void io_req_task_queue(struct io_kiocb
*req
)
2284 req
->io_task_work
.func
= io_req_task_submit
;
2285 io_req_task_work_add(req
);
2288 static void io_req_task_queue_reissue(struct io_kiocb
*req
)
2290 req
->io_task_work
.func
= io_queue_async_work
;
2291 io_req_task_work_add(req
);
2294 static inline void io_queue_next(struct io_kiocb
*req
)
2296 struct io_kiocb
*nxt
= io_req_find_next(req
);
2299 io_req_task_queue(nxt
);
2302 static void io_free_req(struct io_kiocb
*req
)
2308 static void io_free_req_work(struct io_kiocb
*req
, bool *locked
)
2314 struct task_struct
*task
;
2319 static inline void io_init_req_batch(struct req_batch
*rb
)
2326 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2327 struct req_batch
*rb
)
2330 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2332 io_put_task(rb
->task
, rb
->task_refs
);
2335 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2336 struct io_submit_state
*state
)
2339 io_dismantle_req(req
);
2341 if (req
->task
!= rb
->task
) {
2343 io_put_task(rb
->task
, rb
->task_refs
);
2344 rb
->task
= req
->task
;
2350 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2351 state
->reqs
[state
->free_reqs
++] = req
;
2353 list_add(&req
->inflight_entry
, &state
->free_list
);
2356 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
2357 __must_hold(&ctx
->uring_lock
)
2359 struct io_submit_state
*state
= &ctx
->submit_state
;
2360 int i
, nr
= state
->compl_nr
;
2361 struct req_batch rb
;
2363 spin_lock(&ctx
->completion_lock
);
2364 for (i
= 0; i
< nr
; i
++) {
2365 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2367 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2370 io_commit_cqring(ctx
);
2371 spin_unlock(&ctx
->completion_lock
);
2372 io_cqring_ev_posted(ctx
);
2374 io_init_req_batch(&rb
);
2375 for (i
= 0; i
< nr
; i
++) {
2376 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2378 if (req_ref_put_and_test(req
))
2379 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2382 io_req_free_batch_finish(ctx
, &rb
);
2383 state
->compl_nr
= 0;
2387 * Drop reference to request, return next in chain (if there is one) if this
2388 * was the last reference to this request.
2390 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2392 struct io_kiocb
*nxt
= NULL
;
2394 if (req_ref_put_and_test(req
)) {
2395 nxt
= io_req_find_next(req
);
2401 static inline void io_put_req(struct io_kiocb
*req
)
2403 if (req_ref_put_and_test(req
))
2407 static inline void io_put_req_deferred(struct io_kiocb
*req
)
2409 if (req_ref_put_and_test(req
)) {
2410 req
->io_task_work
.func
= io_free_req_work
;
2411 io_req_task_work_add(req
);
2415 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2417 /* See comment at the top of this file */
2419 return __io_cqring_events(ctx
);
2422 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2424 struct io_rings
*rings
= ctx
->rings
;
2426 /* make sure SQ entry isn't read before tail */
2427 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2430 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2432 unsigned int cflags
;
2434 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2435 cflags
|= IORING_CQE_F_BUFFER
;
2436 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2441 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2443 struct io_buffer
*kbuf
;
2445 if (likely(!(req
->flags
& REQ_F_BUFFER_SELECTED
)))
2447 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2448 return io_put_kbuf(req
, kbuf
);
2451 static inline bool io_run_task_work(void)
2453 if (test_thread_flag(TIF_NOTIFY_SIGNAL
) || current
->task_works
) {
2454 __set_current_state(TASK_RUNNING
);
2455 tracehook_notify_signal();
2463 * Find and free completed poll iocbs
2465 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2466 struct list_head
*done
)
2468 struct req_batch rb
;
2469 struct io_kiocb
*req
;
2471 /* order with ->result store in io_complete_rw_iopoll() */
2474 io_init_req_batch(&rb
);
2475 while (!list_empty(done
)) {
2476 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2477 list_del(&req
->inflight_entry
);
2479 __io_cqring_fill_event(ctx
, req
->user_data
, req
->result
,
2480 io_put_rw_kbuf(req
));
2483 if (req_ref_put_and_test(req
))
2484 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2487 io_commit_cqring(ctx
);
2488 io_cqring_ev_posted_iopoll(ctx
);
2489 io_req_free_batch_finish(ctx
, &rb
);
2492 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2495 struct io_kiocb
*req
, *tmp
;
2500 * Only spin for completions if we don't have multiple devices hanging
2501 * off our complete list, and we're under the requested amount.
2503 spin
= !ctx
->poll_multi_queue
&& *nr_events
< min
;
2505 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2506 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2510 * Move completed and retryable entries to our local lists.
2511 * If we find a request that requires polling, break out
2512 * and complete those lists first, if we have entries there.
2514 if (READ_ONCE(req
->iopoll_completed
)) {
2515 list_move_tail(&req
->inflight_entry
, &done
);
2518 if (!list_empty(&done
))
2521 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2522 if (unlikely(ret
< 0))
2527 /* iopoll may have completed current req */
2528 if (READ_ONCE(req
->iopoll_completed
))
2529 list_move_tail(&req
->inflight_entry
, &done
);
2532 if (!list_empty(&done
))
2533 io_iopoll_complete(ctx
, nr_events
, &done
);
2539 * We can't just wait for polled events to come to us, we have to actively
2540 * find and complete them.
2542 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2544 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2547 mutex_lock(&ctx
->uring_lock
);
2548 while (!list_empty(&ctx
->iopoll_list
)) {
2549 unsigned int nr_events
= 0;
2551 io_do_iopoll(ctx
, &nr_events
, 0);
2553 /* let it sleep and repeat later if can't complete a request */
2557 * Ensure we allow local-to-the-cpu processing to take place,
2558 * in this case we need to ensure that we reap all events.
2559 * Also let task_work, etc. to progress by releasing the mutex
2561 if (need_resched()) {
2562 mutex_unlock(&ctx
->uring_lock
);
2564 mutex_lock(&ctx
->uring_lock
);
2567 mutex_unlock(&ctx
->uring_lock
);
2570 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2572 unsigned int nr_events
= 0;
2576 * We disallow the app entering submit/complete with polling, but we
2577 * still need to lock the ring to prevent racing with polled issue
2578 * that got punted to a workqueue.
2580 mutex_lock(&ctx
->uring_lock
);
2582 * Don't enter poll loop if we already have events pending.
2583 * If we do, we can potentially be spinning for commands that
2584 * already triggered a CQE (eg in error).
2586 if (test_bit(0, &ctx
->check_cq_overflow
))
2587 __io_cqring_overflow_flush(ctx
, false);
2588 if (io_cqring_events(ctx
))
2592 * If a submit got punted to a workqueue, we can have the
2593 * application entering polling for a command before it gets
2594 * issued. That app will hold the uring_lock for the duration
2595 * of the poll right here, so we need to take a breather every
2596 * now and then to ensure that the issue has a chance to add
2597 * the poll to the issued list. Otherwise we can spin here
2598 * forever, while the workqueue is stuck trying to acquire the
2601 if (list_empty(&ctx
->iopoll_list
)) {
2602 u32 tail
= ctx
->cached_cq_tail
;
2604 mutex_unlock(&ctx
->uring_lock
);
2606 mutex_lock(&ctx
->uring_lock
);
2608 /* some requests don't go through iopoll_list */
2609 if (tail
!= ctx
->cached_cq_tail
||
2610 list_empty(&ctx
->iopoll_list
))
2613 ret
= io_do_iopoll(ctx
, &nr_events
, min
);
2614 } while (!ret
&& nr_events
< min
&& !need_resched());
2616 mutex_unlock(&ctx
->uring_lock
);
2620 static void kiocb_end_write(struct io_kiocb
*req
)
2623 * Tell lockdep we inherited freeze protection from submission
2626 if (req
->flags
& REQ_F_ISREG
) {
2627 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
2629 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
2635 static bool io_resubmit_prep(struct io_kiocb
*req
)
2637 struct io_async_rw
*rw
= req
->async_data
;
2640 return !io_req_prep_async(req
);
2641 iov_iter_restore(&rw
->iter
, &rw
->iter_state
);
2645 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2647 umode_t mode
= file_inode(req
->file
)->i_mode
;
2648 struct io_ring_ctx
*ctx
= req
->ctx
;
2650 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2652 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2653 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2656 * If ref is dying, we might be running poll reap from the exit work.
2657 * Don't attempt to reissue from that path, just let it fail with
2660 if (percpu_ref_is_dying(&ctx
->refs
))
2663 * Play it safe and assume not safe to re-import and reissue if we're
2664 * not in the original thread group (or in task context).
2666 if (!same_thread_group(req
->task
, current
) || !in_task())
2671 static bool io_resubmit_prep(struct io_kiocb
*req
)
2675 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2681 static bool __io_complete_rw_common(struct io_kiocb
*req
, long res
)
2683 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
)
2684 kiocb_end_write(req
);
2685 if (res
!= req
->result
) {
2686 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
2687 io_rw_should_reissue(req
)) {
2688 req
->flags
|= REQ_F_REISSUE
;
2697 static void io_req_task_complete(struct io_kiocb
*req
, bool *locked
)
2699 unsigned int cflags
= io_put_rw_kbuf(req
);
2700 long res
= req
->result
;
2703 struct io_ring_ctx
*ctx
= req
->ctx
;
2704 struct io_submit_state
*state
= &ctx
->submit_state
;
2706 io_req_complete_state(req
, res
, cflags
);
2707 state
->compl_reqs
[state
->compl_nr
++] = req
;
2708 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
2709 io_submit_flush_completions(ctx
);
2711 io_req_complete_post(req
, res
, cflags
);
2715 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2716 unsigned int issue_flags
)
2718 if (__io_complete_rw_common(req
, res
))
2720 __io_req_complete(req
, issue_flags
, req
->result
, io_put_rw_kbuf(req
));
2723 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2725 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2727 if (__io_complete_rw_common(req
, res
))
2730 req
->io_task_work
.func
= io_req_task_complete
;
2731 io_req_task_work_add(req
);
2734 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2736 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2738 if (kiocb
->ki_flags
& IOCB_WRITE
)
2739 kiocb_end_write(req
);
2740 if (unlikely(res
!= req
->result
)) {
2741 if (res
== -EAGAIN
&& io_rw_should_reissue(req
)) {
2742 req
->flags
|= REQ_F_REISSUE
;
2747 WRITE_ONCE(req
->result
, res
);
2748 /* order with io_iopoll_complete() checking ->result */
2750 WRITE_ONCE(req
->iopoll_completed
, 1);
2754 * After the iocb has been issued, it's safe to be found on the poll list.
2755 * Adding the kiocb to the list AFTER submission ensures that we don't
2756 * find it from a io_do_iopoll() thread before the issuer is done
2757 * accessing the kiocb cookie.
2759 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2761 struct io_ring_ctx
*ctx
= req
->ctx
;
2762 const bool in_async
= io_wq_current_is_worker();
2764 /* workqueue context doesn't hold uring_lock, grab it now */
2765 if (unlikely(in_async
))
2766 mutex_lock(&ctx
->uring_lock
);
2769 * Track whether we have multiple files in our lists. This will impact
2770 * how we do polling eventually, not spinning if we're on potentially
2771 * different devices.
2773 if (list_empty(&ctx
->iopoll_list
)) {
2774 ctx
->poll_multi_queue
= false;
2775 } else if (!ctx
->poll_multi_queue
) {
2776 struct io_kiocb
*list_req
;
2777 unsigned int queue_num0
, queue_num1
;
2779 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2782 if (list_req
->file
!= req
->file
) {
2783 ctx
->poll_multi_queue
= true;
2785 queue_num0
= blk_qc_t_to_queue_num(list_req
->rw
.kiocb
.ki_cookie
);
2786 queue_num1
= blk_qc_t_to_queue_num(req
->rw
.kiocb
.ki_cookie
);
2787 if (queue_num0
!= queue_num1
)
2788 ctx
->poll_multi_queue
= true;
2793 * For fast devices, IO may have already completed. If it has, add
2794 * it to the front so we find it first.
2796 if (READ_ONCE(req
->iopoll_completed
))
2797 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2799 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2801 if (unlikely(in_async
)) {
2803 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2804 * in sq thread task context or in io worker task context. If
2805 * current task context is sq thread, we don't need to check
2806 * whether should wake up sq thread.
2808 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2809 wq_has_sleeper(&ctx
->sq_data
->wait
))
2810 wake_up(&ctx
->sq_data
->wait
);
2812 mutex_unlock(&ctx
->uring_lock
);
2816 static bool io_bdev_nowait(struct block_device
*bdev
)
2818 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2822 * If we tracked the file through the SCM inflight mechanism, we could support
2823 * any file. For now, just ensure that anything potentially problematic is done
2826 static bool __io_file_supports_nowait(struct file
*file
, int rw
)
2828 umode_t mode
= file_inode(file
)->i_mode
;
2830 if (S_ISBLK(mode
)) {
2831 if (IS_ENABLED(CONFIG_BLOCK
) &&
2832 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2838 if (S_ISREG(mode
)) {
2839 if (IS_ENABLED(CONFIG_BLOCK
) &&
2840 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2841 file
->f_op
!= &io_uring_fops
)
2846 /* any ->read/write should understand O_NONBLOCK */
2847 if (file
->f_flags
& O_NONBLOCK
)
2850 if (!(file
->f_mode
& FMODE_NOWAIT
))
2854 return file
->f_op
->read_iter
!= NULL
;
2856 return file
->f_op
->write_iter
!= NULL
;
2859 static bool io_file_supports_nowait(struct io_kiocb
*req
, int rw
)
2861 if (rw
== READ
&& (req
->flags
& REQ_F_NOWAIT_READ
))
2863 else if (rw
== WRITE
&& (req
->flags
& REQ_F_NOWAIT_WRITE
))
2866 return __io_file_supports_nowait(req
->file
, rw
);
2869 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2872 struct io_ring_ctx
*ctx
= req
->ctx
;
2873 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2874 struct file
*file
= req
->file
;
2878 if (!io_req_ffs_set(req
) && S_ISREG(file_inode(file
)->i_mode
))
2879 req
->flags
|= REQ_F_ISREG
;
2881 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2882 if (kiocb
->ki_pos
== -1 && !(file
->f_mode
& FMODE_STREAM
)) {
2883 req
->flags
|= REQ_F_CUR_POS
;
2884 kiocb
->ki_pos
= file
->f_pos
;
2886 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2887 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2888 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2893 * If the file is marked O_NONBLOCK, still allow retry for it if it
2894 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2895 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2897 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) ||
2898 ((file
->f_flags
& O_NONBLOCK
) && !io_file_supports_nowait(req
, rw
)))
2899 req
->flags
|= REQ_F_NOWAIT
;
2901 ioprio
= READ_ONCE(sqe
->ioprio
);
2903 ret
= ioprio_check_cap(ioprio
);
2907 kiocb
->ki_ioprio
= ioprio
;
2909 kiocb
->ki_ioprio
= get_current_ioprio();
2911 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2912 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2913 !kiocb
->ki_filp
->f_op
->iopoll
)
2916 kiocb
->ki_flags
|= IOCB_HIPRI
| IOCB_ALLOC_CACHE
;
2917 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2918 req
->iopoll_completed
= 0;
2920 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2922 kiocb
->ki_complete
= io_complete_rw
;
2925 if (req
->opcode
== IORING_OP_READ_FIXED
||
2926 req
->opcode
== IORING_OP_WRITE_FIXED
) {
2928 io_req_set_rsrc_node(req
);
2931 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2932 req
->rw
.len
= READ_ONCE(sqe
->len
);
2933 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2937 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2943 case -ERESTARTNOINTR
:
2944 case -ERESTARTNOHAND
:
2945 case -ERESTART_RESTARTBLOCK
:
2947 * We can't just restart the syscall, since previously
2948 * submitted sqes may already be in progress. Just fail this
2954 kiocb
->ki_complete(kiocb
, ret
, 0);
2958 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2959 unsigned int issue_flags
)
2961 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2962 struct io_async_rw
*io
= req
->async_data
;
2964 /* add previously done IO, if any */
2965 if (io
&& io
->bytes_done
> 0) {
2967 ret
= io
->bytes_done
;
2969 ret
+= io
->bytes_done
;
2972 if (req
->flags
& REQ_F_CUR_POS
)
2973 req
->file
->f_pos
= kiocb
->ki_pos
;
2974 if (ret
>= 0 && (kiocb
->ki_complete
== io_complete_rw
))
2975 __io_complete_rw(req
, ret
, 0, issue_flags
);
2977 io_rw_done(kiocb
, ret
);
2979 if (req
->flags
& REQ_F_REISSUE
) {
2980 req
->flags
&= ~REQ_F_REISSUE
;
2981 if (io_resubmit_prep(req
)) {
2982 io_req_task_queue_reissue(req
);
2984 unsigned int cflags
= io_put_rw_kbuf(req
);
2985 struct io_ring_ctx
*ctx
= req
->ctx
;
2988 if (!(issue_flags
& IO_URING_F_NONBLOCK
)) {
2989 mutex_lock(&ctx
->uring_lock
);
2990 __io_req_complete(req
, issue_flags
, ret
, cflags
);
2991 mutex_unlock(&ctx
->uring_lock
);
2993 __io_req_complete(req
, issue_flags
, ret
, cflags
);
2999 static int __io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
3000 struct io_mapped_ubuf
*imu
)
3002 size_t len
= req
->rw
.len
;
3003 u64 buf_end
, buf_addr
= req
->rw
.addr
;
3006 if (unlikely(check_add_overflow(buf_addr
, (u64
)len
, &buf_end
)))
3008 /* not inside the mapped region */
3009 if (unlikely(buf_addr
< imu
->ubuf
|| buf_end
> imu
->ubuf_end
))
3013 * May not be a start of buffer, set size appropriately
3014 * and advance us to the beginning.
3016 offset
= buf_addr
- imu
->ubuf
;
3017 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3021 * Don't use iov_iter_advance() here, as it's really slow for
3022 * using the latter parts of a big fixed buffer - it iterates
3023 * over each segment manually. We can cheat a bit here, because
3026 * 1) it's a BVEC iter, we set it up
3027 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3028 * first and last bvec
3030 * So just find our index, and adjust the iterator afterwards.
3031 * If the offset is within the first bvec (or the whole first
3032 * bvec, just use iov_iter_advance(). This makes it easier
3033 * since we can just skip the first segment, which may not
3034 * be PAGE_SIZE aligned.
3036 const struct bio_vec
*bvec
= imu
->bvec
;
3038 if (offset
<= bvec
->bv_len
) {
3039 iov_iter_advance(iter
, offset
);
3041 unsigned long seg_skip
;
3043 /* skip first vec */
3044 offset
-= bvec
->bv_len
;
3045 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3047 iter
->bvec
= bvec
+ seg_skip
;
3048 iter
->nr_segs
-= seg_skip
;
3049 iter
->count
-= bvec
->bv_len
+ offset
;
3050 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3057 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
3059 struct io_ring_ctx
*ctx
= req
->ctx
;
3060 struct io_mapped_ubuf
*imu
= req
->imu
;
3061 u16 index
, buf_index
= req
->buf_index
;
3064 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
3066 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
3067 imu
= READ_ONCE(ctx
->user_bufs
[index
]);
3070 return __io_import_fixed(req
, rw
, iter
, imu
);
3073 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3076 mutex_unlock(&ctx
->uring_lock
);
3079 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3082 * "Normal" inline submissions always hold the uring_lock, since we
3083 * grab it from the system call. Same is true for the SQPOLL offload.
3084 * The only exception is when we've detached the request and issue it
3085 * from an async worker thread, grab the lock for that case.
3088 mutex_lock(&ctx
->uring_lock
);
3091 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3092 int bgid
, struct io_buffer
*kbuf
,
3095 struct io_buffer
*head
;
3097 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3100 io_ring_submit_lock(req
->ctx
, needs_lock
);
3102 lockdep_assert_held(&req
->ctx
->uring_lock
);
3104 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
3106 if (!list_empty(&head
->list
)) {
3107 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3109 list_del(&kbuf
->list
);
3112 xa_erase(&req
->ctx
->io_buffers
, bgid
);
3114 if (*len
> kbuf
->len
)
3117 kbuf
= ERR_PTR(-ENOBUFS
);
3120 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3125 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3128 struct io_buffer
*kbuf
;
3131 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3132 bgid
= req
->buf_index
;
3133 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3136 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3137 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3138 return u64_to_user_ptr(kbuf
->addr
);
3141 #ifdef CONFIG_COMPAT
3142 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3145 struct compat_iovec __user
*uiov
;
3146 compat_ssize_t clen
;
3150 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3151 if (!access_ok(uiov
, sizeof(*uiov
)))
3153 if (__get_user(clen
, &uiov
->iov_len
))
3159 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3161 return PTR_ERR(buf
);
3162 iov
[0].iov_base
= buf
;
3163 iov
[0].iov_len
= (compat_size_t
) len
;
3168 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3171 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3175 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3178 len
= iov
[0].iov_len
;
3181 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3183 return PTR_ERR(buf
);
3184 iov
[0].iov_base
= buf
;
3185 iov
[0].iov_len
= len
;
3189 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3192 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3193 struct io_buffer
*kbuf
;
3195 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3196 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3197 iov
[0].iov_len
= kbuf
->len
;
3200 if (req
->rw
.len
!= 1)
3203 #ifdef CONFIG_COMPAT
3204 if (req
->ctx
->compat
)
3205 return io_compat_import(req
, iov
, needs_lock
);
3208 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3211 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
3212 struct iov_iter
*iter
, bool needs_lock
)
3214 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3215 size_t sqe_len
= req
->rw
.len
;
3216 u8 opcode
= req
->opcode
;
3219 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3221 return io_import_fixed(req
, rw
, iter
);
3224 /* buffer index only valid with fixed read/write, or buffer select */
3225 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3228 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3229 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3230 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3232 return PTR_ERR(buf
);
3233 req
->rw
.len
= sqe_len
;
3236 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3241 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3242 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3244 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3249 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3253 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3255 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3259 * For files that don't have ->read_iter() and ->write_iter(), handle them
3260 * by looping over ->read() or ->write() manually.
3262 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3264 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3265 struct file
*file
= req
->file
;
3269 * Don't support polled IO through this interface, and we can't
3270 * support non-blocking either. For the latter, this just causes
3271 * the kiocb to be handled from an async context.
3273 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3275 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3278 while (iov_iter_count(iter
)) {
3282 if (!iov_iter_is_bvec(iter
)) {
3283 iovec
= iov_iter_iovec(iter
);
3285 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3286 iovec
.iov_len
= req
->rw
.len
;
3290 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3291 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3293 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3294 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3302 if (!iov_iter_is_bvec(iter
)) {
3303 iov_iter_advance(iter
, nr
);
3309 if (nr
!= iovec
.iov_len
)
3316 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3317 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3319 struct io_async_rw
*rw
= req
->async_data
;
3321 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3322 rw
->free_iovec
= iovec
;
3324 /* can only be fixed buffers, no need to do anything */
3325 if (iov_iter_is_bvec(iter
))
3328 unsigned iov_off
= 0;
3330 rw
->iter
.iov
= rw
->fast_iov
;
3331 if (iter
->iov
!= fast_iov
) {
3332 iov_off
= iter
->iov
- fast_iov
;
3333 rw
->iter
.iov
+= iov_off
;
3335 if (rw
->fast_iov
!= fast_iov
)
3336 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3337 sizeof(struct iovec
) * iter
->nr_segs
);
3339 req
->flags
|= REQ_F_NEED_CLEANUP
;
3343 static inline int io_alloc_async_data(struct io_kiocb
*req
)
3345 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3346 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3347 return req
->async_data
== NULL
;
3350 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3351 const struct iovec
*fast_iov
,
3352 struct iov_iter
*iter
, bool force
)
3354 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_setup
)
3356 if (!req
->async_data
) {
3357 struct io_async_rw
*iorw
;
3359 if (io_alloc_async_data(req
)) {
3364 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3365 iorw
= req
->async_data
;
3366 /* we've copied and mapped the iter, ensure state is saved */
3367 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3372 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3374 struct io_async_rw
*iorw
= req
->async_data
;
3375 struct iovec
*iov
= iorw
->fast_iov
;
3378 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3379 if (unlikely(ret
< 0))
3382 iorw
->bytes_done
= 0;
3383 iorw
->free_iovec
= iov
;
3385 req
->flags
|= REQ_F_NEED_CLEANUP
;
3386 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3390 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3392 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3394 return io_prep_rw(req
, sqe
, READ
);
3398 * This is our waitqueue callback handler, registered through lock_page_async()
3399 * when we initially tried to do the IO with the iocb armed our waitqueue.
3400 * This gets called when the page is unlocked, and we generally expect that to
3401 * happen when the page IO is completed and the page is now uptodate. This will
3402 * queue a task_work based retry of the operation, attempting to copy the data
3403 * again. If the latter fails because the page was NOT uptodate, then we will
3404 * do a thread based blocking retry of the operation. That's the unexpected
3407 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3408 int sync
, void *arg
)
3410 struct wait_page_queue
*wpq
;
3411 struct io_kiocb
*req
= wait
->private;
3412 struct wait_page_key
*key
= arg
;
3414 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3416 if (!wake_page_match(wpq
, key
))
3419 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3420 list_del_init(&wait
->entry
);
3421 io_req_task_queue(req
);
3426 * This controls whether a given IO request should be armed for async page
3427 * based retry. If we return false here, the request is handed to the async
3428 * worker threads for retry. If we're doing buffered reads on a regular file,
3429 * we prepare a private wait_page_queue entry and retry the operation. This
3430 * will either succeed because the page is now uptodate and unlocked, or it
3431 * will register a callback when the page is unlocked at IO completion. Through
3432 * that callback, io_uring uses task_work to setup a retry of the operation.
3433 * That retry will attempt the buffered read again. The retry will generally
3434 * succeed, or in rare cases where it fails, we then fall back to using the
3435 * async worker threads for a blocking retry.
3437 static bool io_rw_should_retry(struct io_kiocb
*req
)
3439 struct io_async_rw
*rw
= req
->async_data
;
3440 struct wait_page_queue
*wait
= &rw
->wpq
;
3441 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3443 /* never retry for NOWAIT, we just complete with -EAGAIN */
3444 if (req
->flags
& REQ_F_NOWAIT
)
3447 /* Only for buffered IO */
3448 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3452 * just use poll if we can, and don't attempt if the fs doesn't
3453 * support callback based unlocks
3455 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3458 wait
->wait
.func
= io_async_buf_func
;
3459 wait
->wait
.private = req
;
3460 wait
->wait
.flags
= 0;
3461 INIT_LIST_HEAD(&wait
->wait
.entry
);
3462 kiocb
->ki_flags
|= IOCB_WAITQ
;
3463 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3464 kiocb
->ki_waitq
= wait
;
3468 static inline int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3470 if (req
->file
->f_op
->read_iter
)
3471 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3472 else if (req
->file
->f_op
->read
)
3473 return loop_rw_iter(READ
, req
, iter
);
3478 static bool need_read_all(struct io_kiocb
*req
)
3480 return req
->flags
& REQ_F_ISREG
||
3481 S_ISBLK(file_inode(req
->file
)->i_mode
);
3484 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3486 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3487 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3488 struct iov_iter __iter
, *iter
= &__iter
;
3489 struct io_async_rw
*rw
= req
->async_data
;
3490 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3491 struct iov_iter_state __state
, *state
;
3496 state
= &rw
->iter_state
;
3498 * We come here from an earlier attempt, restore our state to
3499 * match in case it doesn't. It's cheap enough that we don't
3500 * need to make this conditional.
3502 iov_iter_restore(iter
, state
);
3505 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3509 iov_iter_save_state(iter
, state
);
3511 req
->result
= iov_iter_count(iter
);
3513 /* Ensure we clear previously set non-block flag */
3514 if (!force_nonblock
)
3515 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3517 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3519 /* If the file doesn't support async, just async punt */
3520 if (force_nonblock
&& !io_file_supports_nowait(req
, READ
)) {
3521 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3522 return ret
?: -EAGAIN
;
3525 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3526 if (unlikely(ret
)) {
3531 ret
= io_iter_do_read(req
, iter
);
3533 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
3534 req
->flags
&= ~REQ_F_REISSUE
;
3535 /* IOPOLL retry should happen for io-wq threads */
3536 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3538 /* no retry on NONBLOCK nor RWF_NOWAIT */
3539 if (req
->flags
& REQ_F_NOWAIT
)
3542 } else if (ret
== -EIOCBQUEUED
) {
3544 } else if (ret
<= 0 || ret
== req
->result
|| !force_nonblock
||
3545 (req
->flags
& REQ_F_NOWAIT
) || !need_read_all(req
)) {
3546 /* read all, failed, already did sync or don't want to retry */
3551 * Don't depend on the iter state matching what was consumed, or being
3552 * untouched in case of error. Restore it and we'll advance it
3553 * manually if we need to.
3555 iov_iter_restore(iter
, state
);
3557 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3562 rw
= req
->async_data
;
3564 * Now use our persistent iterator and state, if we aren't already.
3565 * We've restored and mapped the iter to match.
3567 if (iter
!= &rw
->iter
) {
3569 state
= &rw
->iter_state
;
3574 * We end up here because of a partial read, either from
3575 * above or inside this loop. Advance the iter by the bytes
3576 * that were consumed.
3578 iov_iter_advance(iter
, ret
);
3579 if (!iov_iter_count(iter
))
3581 rw
->bytes_done
+= ret
;
3582 iov_iter_save_state(iter
, state
);
3584 /* if we can retry, do so with the callbacks armed */
3585 if (!io_rw_should_retry(req
)) {
3586 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3591 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3592 * we get -EIOCBQUEUED, then we'll get a notification when the
3593 * desired page gets unlocked. We can also get a partial read
3594 * here, and if we do, then just retry at the new offset.
3596 ret
= io_iter_do_read(req
, iter
);
3597 if (ret
== -EIOCBQUEUED
)
3599 /* we got some bytes, but not all. retry. */
3600 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3601 iov_iter_restore(iter
, state
);
3604 kiocb_done(kiocb
, ret
, issue_flags
);
3606 /* it's faster to check here then delegate to kfree */
3612 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3614 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3616 return io_prep_rw(req
, sqe
, WRITE
);
3619 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3621 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3622 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3623 struct iov_iter __iter
, *iter
= &__iter
;
3624 struct io_async_rw
*rw
= req
->async_data
;
3625 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3626 struct iov_iter_state __state
, *state
;
3631 state
= &rw
->iter_state
;
3632 iov_iter_restore(iter
, state
);
3635 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3639 iov_iter_save_state(iter
, state
);
3641 req
->result
= iov_iter_count(iter
);
3643 /* Ensure we clear previously set non-block flag */
3644 if (!force_nonblock
)
3645 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3647 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3649 /* If the file doesn't support async, just async punt */
3650 if (force_nonblock
&& !io_file_supports_nowait(req
, WRITE
))
3653 /* file path doesn't support NOWAIT for non-direct_IO */
3654 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3655 (req
->flags
& REQ_F_ISREG
))
3658 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3663 * Open-code file_start_write here to grab freeze protection,
3664 * which will be released by another thread in
3665 * io_complete_rw(). Fool lockdep by telling it the lock got
3666 * released so that it doesn't complain about the held lock when
3667 * we return to userspace.
3669 if (req
->flags
& REQ_F_ISREG
) {
3670 sb_start_write(file_inode(req
->file
)->i_sb
);
3671 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3674 kiocb
->ki_flags
|= IOCB_WRITE
;
3676 if (req
->file
->f_op
->write_iter
)
3677 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3678 else if (req
->file
->f_op
->write
)
3679 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3683 if (req
->flags
& REQ_F_REISSUE
) {
3684 req
->flags
&= ~REQ_F_REISSUE
;
3689 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3690 * retry them without IOCB_NOWAIT.
3692 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3694 /* no retry on NONBLOCK nor RWF_NOWAIT */
3695 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3697 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3698 /* IOPOLL retry should happen for io-wq threads */
3699 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3702 kiocb_done(kiocb
, ret2
, issue_flags
);
3705 iov_iter_restore(iter
, state
);
3706 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3707 return ret
?: -EAGAIN
;
3710 /* it's reportedly faster than delegating the null check to kfree() */
3716 static int io_renameat_prep(struct io_kiocb
*req
,
3717 const struct io_uring_sqe
*sqe
)
3719 struct io_rename
*ren
= &req
->rename
;
3720 const char __user
*oldf
, *newf
;
3722 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3724 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3726 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3729 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3730 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3731 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3732 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3733 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3735 ren
->oldpath
= getname(oldf
);
3736 if (IS_ERR(ren
->oldpath
))
3737 return PTR_ERR(ren
->oldpath
);
3739 ren
->newpath
= getname(newf
);
3740 if (IS_ERR(ren
->newpath
)) {
3741 putname(ren
->oldpath
);
3742 return PTR_ERR(ren
->newpath
);
3745 req
->flags
|= REQ_F_NEED_CLEANUP
;
3749 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3751 struct io_rename
*ren
= &req
->rename
;
3754 if (issue_flags
& IO_URING_F_NONBLOCK
)
3757 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3758 ren
->newpath
, ren
->flags
);
3760 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3763 io_req_complete(req
, ret
);
3767 static int io_unlinkat_prep(struct io_kiocb
*req
,
3768 const struct io_uring_sqe
*sqe
)
3770 struct io_unlink
*un
= &req
->unlink
;
3771 const char __user
*fname
;
3773 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3775 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
3778 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3781 un
->dfd
= READ_ONCE(sqe
->fd
);
3783 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3784 if (un
->flags
& ~AT_REMOVEDIR
)
3787 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3788 un
->filename
= getname(fname
);
3789 if (IS_ERR(un
->filename
))
3790 return PTR_ERR(un
->filename
);
3792 req
->flags
|= REQ_F_NEED_CLEANUP
;
3796 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3798 struct io_unlink
*un
= &req
->unlink
;
3801 if (issue_flags
& IO_URING_F_NONBLOCK
)
3804 if (un
->flags
& AT_REMOVEDIR
)
3805 ret
= do_rmdir(un
->dfd
, un
->filename
);
3807 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3809 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3812 io_req_complete(req
, ret
);
3816 static int io_mkdirat_prep(struct io_kiocb
*req
,
3817 const struct io_uring_sqe
*sqe
)
3819 struct io_mkdir
*mkd
= &req
->mkdir
;
3820 const char __user
*fname
;
3822 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3824 if (sqe
->ioprio
|| sqe
->off
|| sqe
->rw_flags
|| sqe
->buf_index
||
3827 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3830 mkd
->dfd
= READ_ONCE(sqe
->fd
);
3831 mkd
->mode
= READ_ONCE(sqe
->len
);
3833 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3834 mkd
->filename
= getname(fname
);
3835 if (IS_ERR(mkd
->filename
))
3836 return PTR_ERR(mkd
->filename
);
3838 req
->flags
|= REQ_F_NEED_CLEANUP
;
3842 static int io_mkdirat(struct io_kiocb
*req
, int issue_flags
)
3844 struct io_mkdir
*mkd
= &req
->mkdir
;
3847 if (issue_flags
& IO_URING_F_NONBLOCK
)
3850 ret
= do_mkdirat(mkd
->dfd
, mkd
->filename
, mkd
->mode
);
3852 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3855 io_req_complete(req
, ret
);
3859 static int io_symlinkat_prep(struct io_kiocb
*req
,
3860 const struct io_uring_sqe
*sqe
)
3862 struct io_symlink
*sl
= &req
->symlink
;
3863 const char __user
*oldpath
, *newpath
;
3865 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3867 if (sqe
->ioprio
|| sqe
->len
|| sqe
->rw_flags
|| sqe
->buf_index
||
3870 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3873 sl
->new_dfd
= READ_ONCE(sqe
->fd
);
3874 oldpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3875 newpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3877 sl
->oldpath
= getname(oldpath
);
3878 if (IS_ERR(sl
->oldpath
))
3879 return PTR_ERR(sl
->oldpath
);
3881 sl
->newpath
= getname(newpath
);
3882 if (IS_ERR(sl
->newpath
)) {
3883 putname(sl
->oldpath
);
3884 return PTR_ERR(sl
->newpath
);
3887 req
->flags
|= REQ_F_NEED_CLEANUP
;
3891 static int io_symlinkat(struct io_kiocb
*req
, int issue_flags
)
3893 struct io_symlink
*sl
= &req
->symlink
;
3896 if (issue_flags
& IO_URING_F_NONBLOCK
)
3899 ret
= do_symlinkat(sl
->oldpath
, sl
->new_dfd
, sl
->newpath
);
3901 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3904 io_req_complete(req
, ret
);
3908 static int io_linkat_prep(struct io_kiocb
*req
,
3909 const struct io_uring_sqe
*sqe
)
3911 struct io_hardlink
*lnk
= &req
->hardlink
;
3912 const char __user
*oldf
, *newf
;
3914 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3916 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3918 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3921 lnk
->old_dfd
= READ_ONCE(sqe
->fd
);
3922 lnk
->new_dfd
= READ_ONCE(sqe
->len
);
3923 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3924 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3925 lnk
->flags
= READ_ONCE(sqe
->hardlink_flags
);
3927 lnk
->oldpath
= getname(oldf
);
3928 if (IS_ERR(lnk
->oldpath
))
3929 return PTR_ERR(lnk
->oldpath
);
3931 lnk
->newpath
= getname(newf
);
3932 if (IS_ERR(lnk
->newpath
)) {
3933 putname(lnk
->oldpath
);
3934 return PTR_ERR(lnk
->newpath
);
3937 req
->flags
|= REQ_F_NEED_CLEANUP
;
3941 static int io_linkat(struct io_kiocb
*req
, int issue_flags
)
3943 struct io_hardlink
*lnk
= &req
->hardlink
;
3946 if (issue_flags
& IO_URING_F_NONBLOCK
)
3949 ret
= do_linkat(lnk
->old_dfd
, lnk
->oldpath
, lnk
->new_dfd
,
3950 lnk
->newpath
, lnk
->flags
);
3952 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3955 io_req_complete(req
, ret
);
3959 static int io_shutdown_prep(struct io_kiocb
*req
,
3960 const struct io_uring_sqe
*sqe
)
3962 #if defined(CONFIG_NET)
3963 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3965 if (unlikely(sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3966 sqe
->buf_index
|| sqe
->splice_fd_in
))
3969 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3976 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3978 #if defined(CONFIG_NET)
3979 struct socket
*sock
;
3982 if (issue_flags
& IO_URING_F_NONBLOCK
)
3985 sock
= sock_from_file(req
->file
);
3986 if (unlikely(!sock
))
3989 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
3992 io_req_complete(req
, ret
);
3999 static int __io_splice_prep(struct io_kiocb
*req
,
4000 const struct io_uring_sqe
*sqe
)
4002 struct io_splice
*sp
= &req
->splice
;
4003 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
4005 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4009 sp
->len
= READ_ONCE(sqe
->len
);
4010 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
4012 if (unlikely(sp
->flags
& ~valid_flags
))
4015 sp
->file_in
= io_file_get(req
->ctx
, req
, READ_ONCE(sqe
->splice_fd_in
),
4016 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4019 req
->flags
|= REQ_F_NEED_CLEANUP
;
4023 static int io_tee_prep(struct io_kiocb
*req
,
4024 const struct io_uring_sqe
*sqe
)
4026 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
4028 return __io_splice_prep(req
, sqe
);
4031 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
4033 struct io_splice
*sp
= &req
->splice
;
4034 struct file
*in
= sp
->file_in
;
4035 struct file
*out
= sp
->file_out
;
4036 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4039 if (issue_flags
& IO_URING_F_NONBLOCK
)
4042 ret
= do_tee(in
, out
, sp
->len
, flags
);
4044 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4046 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4050 io_req_complete(req
, ret
);
4054 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4056 struct io_splice
*sp
= &req
->splice
;
4058 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
4059 sp
->off_out
= READ_ONCE(sqe
->off
);
4060 return __io_splice_prep(req
, sqe
);
4063 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
4065 struct io_splice
*sp
= &req
->splice
;
4066 struct file
*in
= sp
->file_in
;
4067 struct file
*out
= sp
->file_out
;
4068 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4069 loff_t
*poff_in
, *poff_out
;
4072 if (issue_flags
& IO_URING_F_NONBLOCK
)
4075 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
4076 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
4079 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
4081 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4083 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4087 io_req_complete(req
, ret
);
4092 * IORING_OP_NOP just posts a completion event, nothing else.
4094 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
4096 struct io_ring_ctx
*ctx
= req
->ctx
;
4098 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4101 __io_req_complete(req
, issue_flags
, 0, 0);
4105 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4107 struct io_ring_ctx
*ctx
= req
->ctx
;
4112 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4114 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4118 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
4119 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
4122 req
->sync
.off
= READ_ONCE(sqe
->off
);
4123 req
->sync
.len
= READ_ONCE(sqe
->len
);
4127 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
4129 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
4132 /* fsync always requires a blocking context */
4133 if (issue_flags
& IO_URING_F_NONBLOCK
)
4136 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
4137 end
> 0 ? end
: LLONG_MAX
,
4138 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
4141 io_req_complete(req
, ret
);
4145 static int io_fallocate_prep(struct io_kiocb
*req
,
4146 const struct io_uring_sqe
*sqe
)
4148 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
||
4151 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4154 req
->sync
.off
= READ_ONCE(sqe
->off
);
4155 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4156 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4160 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
4164 /* fallocate always requiring blocking context */
4165 if (issue_flags
& IO_URING_F_NONBLOCK
)
4167 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4171 io_req_complete(req
, ret
);
4175 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4177 const char __user
*fname
;
4180 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4182 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4184 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4187 /* open.how should be already initialised */
4188 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4189 req
->open
.how
.flags
|= O_LARGEFILE
;
4191 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4192 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4193 req
->open
.filename
= getname(fname
);
4194 if (IS_ERR(req
->open
.filename
)) {
4195 ret
= PTR_ERR(req
->open
.filename
);
4196 req
->open
.filename
= NULL
;
4200 req
->open
.file_slot
= READ_ONCE(sqe
->file_index
);
4201 if (req
->open
.file_slot
&& (req
->open
.how
.flags
& O_CLOEXEC
))
4204 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4205 req
->flags
|= REQ_F_NEED_CLEANUP
;
4209 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4211 u64 mode
= READ_ONCE(sqe
->len
);
4212 u64 flags
= READ_ONCE(sqe
->open_flags
);
4214 req
->open
.how
= build_open_how(flags
, mode
);
4215 return __io_openat_prep(req
, sqe
);
4218 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4220 struct open_how __user
*how
;
4224 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4225 len
= READ_ONCE(sqe
->len
);
4226 if (len
< OPEN_HOW_SIZE_VER0
)
4229 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4234 return __io_openat_prep(req
, sqe
);
4237 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
4239 struct open_flags op
;
4241 bool resolve_nonblock
, nonblock_set
;
4242 bool fixed
= !!req
->open
.file_slot
;
4245 ret
= build_open_flags(&req
->open
.how
, &op
);
4248 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
4249 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
4250 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4252 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4253 * it'll always -EAGAIN
4255 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
4257 op
.lookup_flags
|= LOOKUP_CACHED
;
4258 op
.open_flag
|= O_NONBLOCK
;
4262 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4267 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4270 * We could hang on to this 'fd' on retrying, but seems like
4271 * marginal gain for something that is now known to be a slower
4272 * path. So just put it, and we'll get a new one when we retry.
4277 ret
= PTR_ERR(file
);
4278 /* only retry if RESOLVE_CACHED wasn't already set by application */
4279 if (ret
== -EAGAIN
&&
4280 (!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)))
4285 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
4286 file
->f_flags
&= ~O_NONBLOCK
;
4287 fsnotify_open(file
);
4290 fd_install(ret
, file
);
4292 ret
= io_install_fixed_file(req
, file
, issue_flags
,
4293 req
->open
.file_slot
- 1);
4295 putname(req
->open
.filename
);
4296 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4299 __io_req_complete(req
, issue_flags
, ret
, 0);
4303 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
4305 return io_openat2(req
, issue_flags
);
4308 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4309 const struct io_uring_sqe
*sqe
)
4311 struct io_provide_buf
*p
= &req
->pbuf
;
4314 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
4318 tmp
= READ_ONCE(sqe
->fd
);
4319 if (!tmp
|| tmp
> USHRT_MAX
)
4322 memset(p
, 0, sizeof(*p
));
4324 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4328 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4329 int bgid
, unsigned nbufs
)
4333 /* shouldn't happen */
4337 /* the head kbuf is the list itself */
4338 while (!list_empty(&buf
->list
)) {
4339 struct io_buffer
*nxt
;
4341 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4342 list_del(&nxt
->list
);
4349 xa_erase(&ctx
->io_buffers
, bgid
);
4354 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4356 struct io_provide_buf
*p
= &req
->pbuf
;
4357 struct io_ring_ctx
*ctx
= req
->ctx
;
4358 struct io_buffer
*head
;
4360 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4362 io_ring_submit_lock(ctx
, !force_nonblock
);
4364 lockdep_assert_held(&ctx
->uring_lock
);
4367 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4369 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4373 /* complete before unlock, IOPOLL may need the lock */
4374 __io_req_complete(req
, issue_flags
, ret
, 0);
4375 io_ring_submit_unlock(ctx
, !force_nonblock
);
4379 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4380 const struct io_uring_sqe
*sqe
)
4382 unsigned long size
, tmp_check
;
4383 struct io_provide_buf
*p
= &req
->pbuf
;
4386 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
4389 tmp
= READ_ONCE(sqe
->fd
);
4390 if (!tmp
|| tmp
> USHRT_MAX
)
4393 p
->addr
= READ_ONCE(sqe
->addr
);
4394 p
->len
= READ_ONCE(sqe
->len
);
4396 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
4399 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
4402 size
= (unsigned long)p
->len
* p
->nbufs
;
4403 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4406 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4407 tmp
= READ_ONCE(sqe
->off
);
4408 if (tmp
> USHRT_MAX
)
4414 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4416 struct io_buffer
*buf
;
4417 u64 addr
= pbuf
->addr
;
4418 int i
, bid
= pbuf
->bid
;
4420 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4421 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL_ACCOUNT
);
4426 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
4431 INIT_LIST_HEAD(&buf
->list
);
4434 list_add_tail(&buf
->list
, &(*head
)->list
);
4438 return i
? i
: -ENOMEM
;
4441 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4443 struct io_provide_buf
*p
= &req
->pbuf
;
4444 struct io_ring_ctx
*ctx
= req
->ctx
;
4445 struct io_buffer
*head
, *list
;
4447 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4449 io_ring_submit_lock(ctx
, !force_nonblock
);
4451 lockdep_assert_held(&ctx
->uring_lock
);
4453 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4455 ret
= io_add_buffers(p
, &head
);
4456 if (ret
>= 0 && !list
) {
4457 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4459 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4463 /* complete before unlock, IOPOLL may need the lock */
4464 __io_req_complete(req
, issue_flags
, ret
, 0);
4465 io_ring_submit_unlock(ctx
, !force_nonblock
);
4469 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4470 const struct io_uring_sqe
*sqe
)
4472 #if defined(CONFIG_EPOLL)
4473 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4475 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4478 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4479 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4480 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4482 if (ep_op_has_event(req
->epoll
.op
)) {
4483 struct epoll_event __user
*ev
;
4485 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4486 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4496 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4498 #if defined(CONFIG_EPOLL)
4499 struct io_epoll
*ie
= &req
->epoll
;
4501 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4503 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4504 if (force_nonblock
&& ret
== -EAGAIN
)
4509 __io_req_complete(req
, issue_flags
, ret
, 0);
4516 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4518 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4519 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->splice_fd_in
)
4521 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4524 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4525 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4526 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4533 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4535 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4536 struct io_madvise
*ma
= &req
->madvise
;
4539 if (issue_flags
& IO_URING_F_NONBLOCK
)
4542 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4545 io_req_complete(req
, ret
);
4552 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4554 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
|| sqe
->splice_fd_in
)
4556 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4559 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4560 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4561 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4565 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4567 struct io_fadvise
*fa
= &req
->fadvise
;
4570 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4571 switch (fa
->advice
) {
4572 case POSIX_FADV_NORMAL
:
4573 case POSIX_FADV_RANDOM
:
4574 case POSIX_FADV_SEQUENTIAL
:
4581 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4584 __io_req_complete(req
, issue_flags
, ret
, 0);
4588 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4590 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4592 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4594 if (req
->flags
& REQ_F_FIXED_FILE
)
4597 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4598 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4599 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4600 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4601 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4606 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4608 struct io_statx
*ctx
= &req
->statx
;
4611 if (issue_flags
& IO_URING_F_NONBLOCK
)
4614 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4619 io_req_complete(req
, ret
);
4623 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4625 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4627 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4628 sqe
->rw_flags
|| sqe
->buf_index
)
4630 if (req
->flags
& REQ_F_FIXED_FILE
)
4633 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4634 req
->close
.file_slot
= READ_ONCE(sqe
->file_index
);
4635 if (req
->close
.file_slot
&& req
->close
.fd
)
4641 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4643 struct files_struct
*files
= current
->files
;
4644 struct io_close
*close
= &req
->close
;
4645 struct fdtable
*fdt
;
4646 struct file
*file
= NULL
;
4649 if (req
->close
.file_slot
) {
4650 ret
= io_close_fixed(req
, issue_flags
);
4654 spin_lock(&files
->file_lock
);
4655 fdt
= files_fdtable(files
);
4656 if (close
->fd
>= fdt
->max_fds
) {
4657 spin_unlock(&files
->file_lock
);
4660 file
= fdt
->fd
[close
->fd
];
4661 if (!file
|| file
->f_op
== &io_uring_fops
) {
4662 spin_unlock(&files
->file_lock
);
4667 /* if the file has a flush method, be safe and punt to async */
4668 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4669 spin_unlock(&files
->file_lock
);
4673 ret
= __close_fd_get_file(close
->fd
, &file
);
4674 spin_unlock(&files
->file_lock
);
4681 /* No ->flush() or already async, safely close from here */
4682 ret
= filp_close(file
, current
->files
);
4688 __io_req_complete(req
, issue_flags
, ret
, 0);
4692 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4694 struct io_ring_ctx
*ctx
= req
->ctx
;
4696 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4698 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4702 req
->sync
.off
= READ_ONCE(sqe
->off
);
4703 req
->sync
.len
= READ_ONCE(sqe
->len
);
4704 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4708 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4712 /* sync_file_range always requires a blocking context */
4713 if (issue_flags
& IO_URING_F_NONBLOCK
)
4716 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4720 io_req_complete(req
, ret
);
4724 #if defined(CONFIG_NET)
4725 static int io_setup_async_msg(struct io_kiocb
*req
,
4726 struct io_async_msghdr
*kmsg
)
4728 struct io_async_msghdr
*async_msg
= req
->async_data
;
4732 if (io_alloc_async_data(req
)) {
4733 kfree(kmsg
->free_iov
);
4736 async_msg
= req
->async_data
;
4737 req
->flags
|= REQ_F_NEED_CLEANUP
;
4738 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4739 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4740 /* if were using fast_iov, set it to the new one */
4741 if (!async_msg
->free_iov
)
4742 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4747 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4748 struct io_async_msghdr
*iomsg
)
4750 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4751 iomsg
->free_iov
= iomsg
->fast_iov
;
4752 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4753 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4756 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4760 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4762 req
->flags
|= REQ_F_NEED_CLEANUP
;
4766 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4768 struct io_sr_msg
*sr
= &req
->sr_msg
;
4770 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4773 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4774 sr
->len
= READ_ONCE(sqe
->len
);
4775 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4776 if (sr
->msg_flags
& MSG_DONTWAIT
)
4777 req
->flags
|= REQ_F_NOWAIT
;
4779 #ifdef CONFIG_COMPAT
4780 if (req
->ctx
->compat
)
4781 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4786 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4788 struct io_async_msghdr iomsg
, *kmsg
;
4789 struct socket
*sock
;
4794 sock
= sock_from_file(req
->file
);
4795 if (unlikely(!sock
))
4798 kmsg
= req
->async_data
;
4800 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4806 flags
= req
->sr_msg
.msg_flags
;
4807 if (issue_flags
& IO_URING_F_NONBLOCK
)
4808 flags
|= MSG_DONTWAIT
;
4809 if (flags
& MSG_WAITALL
)
4810 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4812 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4813 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4814 return io_setup_async_msg(req
, kmsg
);
4815 if (ret
== -ERESTARTSYS
)
4818 /* fast path, check for non-NULL to avoid function call */
4820 kfree(kmsg
->free_iov
);
4821 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4824 __io_req_complete(req
, issue_flags
, ret
, 0);
4828 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4830 struct io_sr_msg
*sr
= &req
->sr_msg
;
4833 struct socket
*sock
;
4838 sock
= sock_from_file(req
->file
);
4839 if (unlikely(!sock
))
4842 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4846 msg
.msg_name
= NULL
;
4847 msg
.msg_control
= NULL
;
4848 msg
.msg_controllen
= 0;
4849 msg
.msg_namelen
= 0;
4851 flags
= req
->sr_msg
.msg_flags
;
4852 if (issue_flags
& IO_URING_F_NONBLOCK
)
4853 flags
|= MSG_DONTWAIT
;
4854 if (flags
& MSG_WAITALL
)
4855 min_ret
= iov_iter_count(&msg
.msg_iter
);
4857 msg
.msg_flags
= flags
;
4858 ret
= sock_sendmsg(sock
, &msg
);
4859 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4861 if (ret
== -ERESTARTSYS
)
4866 __io_req_complete(req
, issue_flags
, ret
, 0);
4870 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4871 struct io_async_msghdr
*iomsg
)
4873 struct io_sr_msg
*sr
= &req
->sr_msg
;
4874 struct iovec __user
*uiov
;
4878 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4879 &iomsg
->uaddr
, &uiov
, &iov_len
);
4883 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4886 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4888 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4889 iomsg
->free_iov
= NULL
;
4891 iomsg
->free_iov
= iomsg
->fast_iov
;
4892 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4893 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4902 #ifdef CONFIG_COMPAT
4903 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4904 struct io_async_msghdr
*iomsg
)
4906 struct io_sr_msg
*sr
= &req
->sr_msg
;
4907 struct compat_iovec __user
*uiov
;
4912 ret
= __get_compat_msghdr(&iomsg
->msg
, sr
->umsg_compat
, &iomsg
->uaddr
,
4917 uiov
= compat_ptr(ptr
);
4918 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4919 compat_ssize_t clen
;
4923 if (!access_ok(uiov
, sizeof(*uiov
)))
4925 if (__get_user(clen
, &uiov
->iov_len
))
4930 iomsg
->free_iov
= NULL
;
4932 iomsg
->free_iov
= iomsg
->fast_iov
;
4933 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4934 UIO_FASTIOV
, &iomsg
->free_iov
,
4935 &iomsg
->msg
.msg_iter
, true);
4944 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4945 struct io_async_msghdr
*iomsg
)
4947 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4949 #ifdef CONFIG_COMPAT
4950 if (req
->ctx
->compat
)
4951 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4954 return __io_recvmsg_copy_hdr(req
, iomsg
);
4957 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4960 struct io_sr_msg
*sr
= &req
->sr_msg
;
4961 struct io_buffer
*kbuf
;
4963 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4968 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4972 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
4974 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
4977 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
4981 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
4983 req
->flags
|= REQ_F_NEED_CLEANUP
;
4987 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4989 struct io_sr_msg
*sr
= &req
->sr_msg
;
4991 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4994 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4995 sr
->len
= READ_ONCE(sqe
->len
);
4996 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4997 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4998 if (sr
->msg_flags
& MSG_DONTWAIT
)
4999 req
->flags
|= REQ_F_NOWAIT
;
5001 #ifdef CONFIG_COMPAT
5002 if (req
->ctx
->compat
)
5003 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
5008 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
5010 struct io_async_msghdr iomsg
, *kmsg
;
5011 struct socket
*sock
;
5012 struct io_buffer
*kbuf
;
5015 int ret
, cflags
= 0;
5016 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5018 sock
= sock_from_file(req
->file
);
5019 if (unlikely(!sock
))
5022 kmsg
= req
->async_data
;
5024 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
5030 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5031 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5033 return PTR_ERR(kbuf
);
5034 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
5035 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
5036 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
5037 1, req
->sr_msg
.len
);
5040 flags
= req
->sr_msg
.msg_flags
;
5042 flags
|= MSG_DONTWAIT
;
5043 if (flags
& MSG_WAITALL
)
5044 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
5046 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
5047 kmsg
->uaddr
, flags
);
5048 if (force_nonblock
&& ret
== -EAGAIN
)
5049 return io_setup_async_msg(req
, kmsg
);
5050 if (ret
== -ERESTARTSYS
)
5053 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5054 cflags
= io_put_recv_kbuf(req
);
5055 /* fast path, check for non-NULL to avoid function call */
5057 kfree(kmsg
->free_iov
);
5058 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5059 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5061 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5065 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
5067 struct io_buffer
*kbuf
;
5068 struct io_sr_msg
*sr
= &req
->sr_msg
;
5070 void __user
*buf
= sr
->buf
;
5071 struct socket
*sock
;
5075 int ret
, cflags
= 0;
5076 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5078 sock
= sock_from_file(req
->file
);
5079 if (unlikely(!sock
))
5082 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5083 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5085 return PTR_ERR(kbuf
);
5086 buf
= u64_to_user_ptr(kbuf
->addr
);
5089 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
5093 msg
.msg_name
= NULL
;
5094 msg
.msg_control
= NULL
;
5095 msg
.msg_controllen
= 0;
5096 msg
.msg_namelen
= 0;
5097 msg
.msg_iocb
= NULL
;
5100 flags
= req
->sr_msg
.msg_flags
;
5102 flags
|= MSG_DONTWAIT
;
5103 if (flags
& MSG_WAITALL
)
5104 min_ret
= iov_iter_count(&msg
.msg_iter
);
5106 ret
= sock_recvmsg(sock
, &msg
, flags
);
5107 if (force_nonblock
&& ret
== -EAGAIN
)
5109 if (ret
== -ERESTARTSYS
)
5112 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5113 cflags
= io_put_recv_kbuf(req
);
5114 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5116 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5120 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5122 struct io_accept
*accept
= &req
->accept
;
5124 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5126 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
5129 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5130 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5131 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
5132 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
5134 accept
->file_slot
= READ_ONCE(sqe
->file_index
);
5135 if (accept
->file_slot
&& ((req
->open
.how
.flags
& O_CLOEXEC
) ||
5136 (accept
->flags
& SOCK_CLOEXEC
)))
5138 if (accept
->flags
& ~(SOCK_CLOEXEC
| SOCK_NONBLOCK
))
5140 if (SOCK_NONBLOCK
!= O_NONBLOCK
&& (accept
->flags
& SOCK_NONBLOCK
))
5141 accept
->flags
= (accept
->flags
& ~SOCK_NONBLOCK
) | O_NONBLOCK
;
5145 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
5147 struct io_accept
*accept
= &req
->accept
;
5148 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5149 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5150 bool fixed
= !!accept
->file_slot
;
5154 if (req
->file
->f_flags
& O_NONBLOCK
)
5155 req
->flags
|= REQ_F_NOWAIT
;
5158 fd
= __get_unused_fd_flags(accept
->flags
, accept
->nofile
);
5159 if (unlikely(fd
< 0))
5162 file
= do_accept(req
->file
, file_flags
, accept
->addr
, accept
->addr_len
,
5167 ret
= PTR_ERR(file
);
5168 if (ret
== -EAGAIN
&& force_nonblock
)
5170 if (ret
== -ERESTARTSYS
)
5173 } else if (!fixed
) {
5174 fd_install(fd
, file
);
5177 ret
= io_install_fixed_file(req
, file
, issue_flags
,
5178 accept
->file_slot
- 1);
5180 __io_req_complete(req
, issue_flags
, ret
, 0);
5184 static int io_connect_prep_async(struct io_kiocb
*req
)
5186 struct io_async_connect
*io
= req
->async_data
;
5187 struct io_connect
*conn
= &req
->connect
;
5189 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
5192 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5194 struct io_connect
*conn
= &req
->connect
;
5196 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5198 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
||
5202 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5203 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5207 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
5209 struct io_async_connect __io
, *io
;
5210 unsigned file_flags
;
5212 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5214 if (req
->async_data
) {
5215 io
= req
->async_data
;
5217 ret
= move_addr_to_kernel(req
->connect
.addr
,
5218 req
->connect
.addr_len
,
5225 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5227 ret
= __sys_connect_file(req
->file
, &io
->address
,
5228 req
->connect
.addr_len
, file_flags
);
5229 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5230 if (req
->async_data
)
5232 if (io_alloc_async_data(req
)) {
5236 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5239 if (ret
== -ERESTARTSYS
)
5244 __io_req_complete(req
, issue_flags
, ret
, 0);
5247 #else /* !CONFIG_NET */
5248 #define IO_NETOP_FN(op) \
5249 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5251 return -EOPNOTSUPP; \
5254 #define IO_NETOP_PREP(op) \
5256 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5258 return -EOPNOTSUPP; \
5261 #define IO_NETOP_PREP_ASYNC(op) \
5263 static int io_##op##_prep_async(struct io_kiocb *req) \
5265 return -EOPNOTSUPP; \
5268 IO_NETOP_PREP_ASYNC(sendmsg
);
5269 IO_NETOP_PREP_ASYNC(recvmsg
);
5270 IO_NETOP_PREP_ASYNC(connect
);
5271 IO_NETOP_PREP(accept
);
5274 #endif /* CONFIG_NET */
5276 struct io_poll_table
{
5277 struct poll_table_struct pt
;
5278 struct io_kiocb
*req
;
5283 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
5284 __poll_t mask
, io_req_tw_func_t func
)
5286 /* for instances that support it check for an event match first: */
5287 if (mask
&& !(mask
& poll
->events
))
5290 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5292 list_del_init(&poll
->wait
.entry
);
5295 req
->io_task_work
.func
= func
;
5298 * If this fails, then the task is exiting. When a task exits, the
5299 * work gets canceled, so just cancel this request as well instead
5300 * of executing it. We can't safely execute it anyway, as we may not
5301 * have the needed state needed for it anyway.
5303 io_req_task_work_add(req
);
5307 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
5308 __acquires(&req
->ctx
->completion_lock
)
5310 struct io_ring_ctx
*ctx
= req
->ctx
;
5312 /* req->task == current here, checking PF_EXITING is safe */
5313 if (unlikely(req
->task
->flags
& PF_EXITING
))
5314 WRITE_ONCE(poll
->canceled
, true);
5316 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5317 struct poll_table_struct pt
= { ._key
= poll
->events
};
5319 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5322 spin_lock(&ctx
->completion_lock
);
5323 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
5324 add_wait_queue(poll
->head
, &poll
->wait
);
5331 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5333 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5334 if (req
->opcode
== IORING_OP_POLL_ADD
)
5335 return req
->async_data
;
5336 return req
->apoll
->double_poll
;
5339 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5341 if (req
->opcode
== IORING_OP_POLL_ADD
)
5343 return &req
->apoll
->poll
;
5346 static void io_poll_remove_double(struct io_kiocb
*req
)
5347 __must_hold(&req
->ctx
->completion_lock
)
5349 struct io_poll_iocb
*poll
= io_poll_get_double(req
);
5351 lockdep_assert_held(&req
->ctx
->completion_lock
);
5353 if (poll
&& poll
->head
) {
5354 struct wait_queue_head
*head
= poll
->head
;
5356 spin_lock_irq(&head
->lock
);
5357 list_del_init(&poll
->wait
.entry
);
5358 if (poll
->wait
.private)
5361 spin_unlock_irq(&head
->lock
);
5365 static bool __io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
5366 __must_hold(&req
->ctx
->completion_lock
)
5368 struct io_ring_ctx
*ctx
= req
->ctx
;
5369 unsigned flags
= IORING_CQE_F_MORE
;
5372 if (READ_ONCE(req
->poll
.canceled
)) {
5374 req
->poll
.events
|= EPOLLONESHOT
;
5376 error
= mangle_poll(mask
);
5378 if (req
->poll
.events
& EPOLLONESHOT
)
5380 if (!io_cqring_fill_event(ctx
, req
->user_data
, error
, flags
)) {
5381 req
->poll
.events
|= EPOLLONESHOT
;
5384 if (flags
& IORING_CQE_F_MORE
)
5387 return !(flags
& IORING_CQE_F_MORE
);
5390 static inline bool io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
5391 __must_hold(&req
->ctx
->completion_lock
)
5395 done
= __io_poll_complete(req
, mask
);
5396 io_commit_cqring(req
->ctx
);
5400 static void io_poll_task_func(struct io_kiocb
*req
, bool *locked
)
5402 struct io_ring_ctx
*ctx
= req
->ctx
;
5403 struct io_kiocb
*nxt
;
5405 if (io_poll_rewait(req
, &req
->poll
)) {
5406 spin_unlock(&ctx
->completion_lock
);
5410 if (req
->poll
.done
) {
5411 spin_unlock(&ctx
->completion_lock
);
5414 done
= __io_poll_complete(req
, req
->result
);
5416 io_poll_remove_double(req
);
5417 hash_del(&req
->hash_node
);
5418 req
->poll
.done
= true;
5421 add_wait_queue(req
->poll
.head
, &req
->poll
.wait
);
5423 io_commit_cqring(ctx
);
5424 spin_unlock(&ctx
->completion_lock
);
5425 io_cqring_ev_posted(ctx
);
5428 nxt
= io_put_req_find_next(req
);
5430 io_req_task_submit(nxt
, locked
);
5435 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
5436 int sync
, void *key
)
5438 struct io_kiocb
*req
= wait
->private;
5439 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5440 __poll_t mask
= key_to_poll(key
);
5441 unsigned long flags
;
5443 /* for instances that support it check for an event match first: */
5444 if (mask
&& !(mask
& poll
->events
))
5446 if (!(poll
->events
& EPOLLONESHOT
))
5447 return poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5449 list_del_init(&wait
->entry
);
5454 spin_lock_irqsave(&poll
->head
->lock
, flags
);
5455 done
= list_empty(&poll
->wait
.entry
);
5457 list_del_init(&poll
->wait
.entry
);
5458 /* make sure double remove sees this as being gone */
5459 wait
->private = NULL
;
5460 spin_unlock_irqrestore(&poll
->head
->lock
, flags
);
5462 /* use wait func handler, so it matches the rq type */
5463 poll
->wait
.func(&poll
->wait
, mode
, sync
, key
);
5470 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5471 wait_queue_func_t wake_func
)
5475 poll
->canceled
= false;
5476 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5477 /* mask in events that we always want/need */
5478 poll
->events
= events
| IO_POLL_UNMASK
;
5479 INIT_LIST_HEAD(&poll
->wait
.entry
);
5480 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5483 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5484 struct wait_queue_head
*head
,
5485 struct io_poll_iocb
**poll_ptr
)
5487 struct io_kiocb
*req
= pt
->req
;
5490 * The file being polled uses multiple waitqueues for poll handling
5491 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5494 if (unlikely(pt
->nr_entries
)) {
5495 struct io_poll_iocb
*poll_one
= poll
;
5497 /* double add on the same waitqueue head, ignore */
5498 if (poll_one
->head
== head
)
5500 /* already have a 2nd entry, fail a third attempt */
5502 if ((*poll_ptr
)->head
== head
)
5504 pt
->error
= -EINVAL
;
5508 * Can't handle multishot for double wait for now, turn it
5509 * into one-shot mode.
5511 if (!(poll_one
->events
& EPOLLONESHOT
))
5512 poll_one
->events
|= EPOLLONESHOT
;
5513 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5515 pt
->error
= -ENOMEM
;
5518 io_init_poll_iocb(poll
, poll_one
->events
, io_poll_double_wake
);
5520 poll
->wait
.private = req
;
5527 if (poll
->events
& EPOLLEXCLUSIVE
)
5528 add_wait_queue_exclusive(head
, &poll
->wait
);
5530 add_wait_queue(head
, &poll
->wait
);
5533 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5534 struct poll_table_struct
*p
)
5536 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5537 struct async_poll
*apoll
= pt
->req
->apoll
;
5539 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5542 static void io_async_task_func(struct io_kiocb
*req
, bool *locked
)
5544 struct async_poll
*apoll
= req
->apoll
;
5545 struct io_ring_ctx
*ctx
= req
->ctx
;
5547 trace_io_uring_task_run(req
->ctx
, req
, req
->opcode
, req
->user_data
);
5549 if (io_poll_rewait(req
, &apoll
->poll
)) {
5550 spin_unlock(&ctx
->completion_lock
);
5554 hash_del(&req
->hash_node
);
5555 io_poll_remove_double(req
);
5556 apoll
->poll
.done
= true;
5557 spin_unlock(&ctx
->completion_lock
);
5559 if (!READ_ONCE(apoll
->poll
.canceled
))
5560 io_req_task_submit(req
, locked
);
5562 io_req_complete_failed(req
, -ECANCELED
);
5565 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5568 struct io_kiocb
*req
= wait
->private;
5569 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
5571 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
5574 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
5577 static void io_poll_req_insert(struct io_kiocb
*req
)
5579 struct io_ring_ctx
*ctx
= req
->ctx
;
5580 struct hlist_head
*list
;
5582 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5583 hlist_add_head(&req
->hash_node
, list
);
5586 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
5587 struct io_poll_iocb
*poll
,
5588 struct io_poll_table
*ipt
, __poll_t mask
,
5589 wait_queue_func_t wake_func
)
5590 __acquires(&ctx
->completion_lock
)
5592 struct io_ring_ctx
*ctx
= req
->ctx
;
5593 bool cancel
= false;
5595 INIT_HLIST_NODE(&req
->hash_node
);
5596 io_init_poll_iocb(poll
, mask
, wake_func
);
5597 poll
->file
= req
->file
;
5598 poll
->wait
.private = req
;
5600 ipt
->pt
._key
= mask
;
5603 ipt
->nr_entries
= 0;
5605 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5606 if (unlikely(!ipt
->nr_entries
) && !ipt
->error
)
5607 ipt
->error
= -EINVAL
;
5609 spin_lock(&ctx
->completion_lock
);
5610 if (ipt
->error
|| (mask
&& (poll
->events
& EPOLLONESHOT
)))
5611 io_poll_remove_double(req
);
5612 if (likely(poll
->head
)) {
5613 spin_lock_irq(&poll
->head
->lock
);
5614 if (unlikely(list_empty(&poll
->wait
.entry
))) {
5620 if ((mask
&& (poll
->events
& EPOLLONESHOT
)) || ipt
->error
)
5621 list_del_init(&poll
->wait
.entry
);
5623 WRITE_ONCE(poll
->canceled
, true);
5624 else if (!poll
->done
) /* actually waiting for an event */
5625 io_poll_req_insert(req
);
5626 spin_unlock_irq(&poll
->head
->lock
);
5638 static int io_arm_poll_handler(struct io_kiocb
*req
)
5640 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5641 struct io_ring_ctx
*ctx
= req
->ctx
;
5642 struct async_poll
*apoll
;
5643 struct io_poll_table ipt
;
5644 __poll_t ret
, mask
= EPOLLONESHOT
| POLLERR
| POLLPRI
;
5647 if (!req
->file
|| !file_can_poll(req
->file
))
5648 return IO_APOLL_ABORTED
;
5649 if (req
->flags
& REQ_F_POLLED
)
5650 return IO_APOLL_ABORTED
;
5651 if (!def
->pollin
&& !def
->pollout
)
5652 return IO_APOLL_ABORTED
;
5656 mask
|= POLLIN
| POLLRDNORM
;
5658 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5659 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5660 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5664 mask
|= POLLOUT
| POLLWRNORM
;
5667 /* if we can't nonblock try, then no point in arming a poll handler */
5668 if (!io_file_supports_nowait(req
, rw
))
5669 return IO_APOLL_ABORTED
;
5671 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5672 if (unlikely(!apoll
))
5673 return IO_APOLL_ABORTED
;
5674 apoll
->double_poll
= NULL
;
5676 req
->flags
|= REQ_F_POLLED
;
5677 ipt
.pt
._qproc
= io_async_queue_proc
;
5678 io_req_set_refcount(req
);
5680 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
5682 spin_unlock(&ctx
->completion_lock
);
5683 if (ret
|| ipt
.error
)
5684 return ret
? IO_APOLL_READY
: IO_APOLL_ABORTED
;
5686 trace_io_uring_poll_arm(ctx
, req
, req
->opcode
, req
->user_data
,
5687 mask
, apoll
->poll
.events
);
5691 static bool __io_poll_remove_one(struct io_kiocb
*req
,
5692 struct io_poll_iocb
*poll
, bool do_cancel
)
5693 __must_hold(&req
->ctx
->completion_lock
)
5695 bool do_complete
= false;
5699 spin_lock_irq(&poll
->head
->lock
);
5701 WRITE_ONCE(poll
->canceled
, true);
5702 if (!list_empty(&poll
->wait
.entry
)) {
5703 list_del_init(&poll
->wait
.entry
);
5706 spin_unlock_irq(&poll
->head
->lock
);
5707 hash_del(&req
->hash_node
);
5711 static bool io_poll_remove_one(struct io_kiocb
*req
)
5712 __must_hold(&req
->ctx
->completion_lock
)
5716 io_poll_remove_double(req
);
5717 do_complete
= __io_poll_remove_one(req
, io_poll_get_single(req
), true);
5720 io_cqring_fill_event(req
->ctx
, req
->user_data
, -ECANCELED
, 0);
5721 io_commit_cqring(req
->ctx
);
5723 io_put_req_deferred(req
);
5729 * Returns true if we found and killed one or more poll requests
5731 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5734 struct hlist_node
*tmp
;
5735 struct io_kiocb
*req
;
5738 spin_lock(&ctx
->completion_lock
);
5739 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5740 struct hlist_head
*list
;
5742 list
= &ctx
->cancel_hash
[i
];
5743 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5744 if (io_match_task_safe(req
, tsk
, cancel_all
))
5745 posted
+= io_poll_remove_one(req
);
5748 spin_unlock(&ctx
->completion_lock
);
5751 io_cqring_ev_posted(ctx
);
5756 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5758 __must_hold(&ctx
->completion_lock
)
5760 struct hlist_head
*list
;
5761 struct io_kiocb
*req
;
5763 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5764 hlist_for_each_entry(req
, list
, hash_node
) {
5765 if (sqe_addr
!= req
->user_data
)
5767 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
5774 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5776 __must_hold(&ctx
->completion_lock
)
5778 struct io_kiocb
*req
;
5780 req
= io_poll_find(ctx
, sqe_addr
, poll_only
);
5783 if (io_poll_remove_one(req
))
5789 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
5794 events
= READ_ONCE(sqe
->poll32_events
);
5796 events
= swahw32(events
);
5798 if (!(flags
& IORING_POLL_ADD_MULTI
))
5799 events
|= EPOLLONESHOT
;
5800 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
5803 static int io_poll_update_prep(struct io_kiocb
*req
,
5804 const struct io_uring_sqe
*sqe
)
5806 struct io_poll_update
*upd
= &req
->poll_update
;
5809 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5811 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
5813 flags
= READ_ONCE(sqe
->len
);
5814 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
5815 IORING_POLL_ADD_MULTI
))
5817 /* meaningless without update */
5818 if (flags
== IORING_POLL_ADD_MULTI
)
5821 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
5822 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
5823 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
5825 upd
->new_user_data
= READ_ONCE(sqe
->off
);
5826 if (!upd
->update_user_data
&& upd
->new_user_data
)
5828 if (upd
->update_events
)
5829 upd
->events
= io_poll_parse_events(sqe
, flags
);
5830 else if (sqe
->poll32_events
)
5836 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5839 struct io_kiocb
*req
= wait
->private;
5840 struct io_poll_iocb
*poll
= &req
->poll
;
5842 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
5845 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5846 struct poll_table_struct
*p
)
5848 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5850 __io_queue_proc(&pt
->req
->poll
, pt
, head
, (struct io_poll_iocb
**) &pt
->req
->async_data
);
5853 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5855 struct io_poll_iocb
*poll
= &req
->poll
;
5858 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5860 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
5862 flags
= READ_ONCE(sqe
->len
);
5863 if (flags
& ~IORING_POLL_ADD_MULTI
)
5866 io_req_set_refcount(req
);
5867 poll
->events
= io_poll_parse_events(sqe
, flags
);
5871 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5873 struct io_poll_iocb
*poll
= &req
->poll
;
5874 struct io_ring_ctx
*ctx
= req
->ctx
;
5875 struct io_poll_table ipt
;
5879 ipt
.pt
._qproc
= io_poll_queue_proc
;
5881 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
5884 if (mask
) { /* no async, we'd stolen it */
5886 done
= io_poll_complete(req
, mask
);
5888 spin_unlock(&ctx
->completion_lock
);
5891 io_cqring_ev_posted(ctx
);
5898 static int io_poll_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5900 struct io_ring_ctx
*ctx
= req
->ctx
;
5901 struct io_kiocb
*preq
;
5905 spin_lock(&ctx
->completion_lock
);
5906 preq
= io_poll_find(ctx
, req
->poll_update
.old_user_data
, true);
5912 if (!req
->poll_update
.update_events
&& !req
->poll_update
.update_user_data
) {
5914 ret
= io_poll_remove_one(preq
) ? 0 : -EALREADY
;
5919 * Don't allow racy completion with singleshot, as we cannot safely
5920 * update those. For multishot, if we're racing with completion, just
5921 * let completion re-add it.
5923 completing
= !__io_poll_remove_one(preq
, &preq
->poll
, false);
5924 if (completing
&& (preq
->poll
.events
& EPOLLONESHOT
)) {
5928 /* we now have a detached poll request. reissue. */
5932 spin_unlock(&ctx
->completion_lock
);
5934 io_req_complete(req
, ret
);
5937 /* only mask one event flags, keep behavior flags */
5938 if (req
->poll_update
.update_events
) {
5939 preq
->poll
.events
&= ~0xffff;
5940 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
5941 preq
->poll
.events
|= IO_POLL_UNMASK
;
5943 if (req
->poll_update
.update_user_data
)
5944 preq
->user_data
= req
->poll_update
.new_user_data
;
5945 spin_unlock(&ctx
->completion_lock
);
5947 /* complete update request, we're done with it */
5948 io_req_complete(req
, ret
);
5951 ret
= io_poll_add(preq
, issue_flags
);
5954 io_req_complete(preq
, ret
);
5960 static void io_req_task_timeout(struct io_kiocb
*req
, bool *locked
)
5963 io_req_complete_post(req
, -ETIME
, 0);
5966 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5968 struct io_timeout_data
*data
= container_of(timer
,
5969 struct io_timeout_data
, timer
);
5970 struct io_kiocb
*req
= data
->req
;
5971 struct io_ring_ctx
*ctx
= req
->ctx
;
5972 unsigned long flags
;
5974 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
5975 list_del_init(&req
->timeout
.list
);
5976 atomic_set(&req
->ctx
->cq_timeouts
,
5977 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5978 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
5980 req
->io_task_work
.func
= io_req_task_timeout
;
5981 io_req_task_work_add(req
);
5982 return HRTIMER_NORESTART
;
5985 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5987 __must_hold(&ctx
->timeout_lock
)
5989 struct io_timeout_data
*io
;
5990 struct io_kiocb
*req
;
5993 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5994 found
= user_data
== req
->user_data
;
5999 return ERR_PTR(-ENOENT
);
6001 io
= req
->async_data
;
6002 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6003 return ERR_PTR(-EALREADY
);
6004 list_del_init(&req
->timeout
.list
);
6008 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
6009 __must_hold(&ctx
->completion_lock
)
6010 __must_hold(&ctx
->timeout_lock
)
6012 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6015 return PTR_ERR(req
);
6018 io_cqring_fill_event(ctx
, req
->user_data
, -ECANCELED
, 0);
6019 io_put_req_deferred(req
);
6023 static clockid_t
io_timeout_get_clock(struct io_timeout_data
*data
)
6025 switch (data
->flags
& IORING_TIMEOUT_CLOCK_MASK
) {
6026 case IORING_TIMEOUT_BOOTTIME
:
6027 return CLOCK_BOOTTIME
;
6028 case IORING_TIMEOUT_REALTIME
:
6029 return CLOCK_REALTIME
;
6031 /* can't happen, vetted at prep time */
6035 return CLOCK_MONOTONIC
;
6039 static int io_linked_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6040 struct timespec64
*ts
, enum hrtimer_mode mode
)
6041 __must_hold(&ctx
->timeout_lock
)
6043 struct io_timeout_data
*io
;
6044 struct io_kiocb
*req
;
6047 list_for_each_entry(req
, &ctx
->ltimeout_list
, timeout
.list
) {
6048 found
= user_data
== req
->user_data
;
6055 io
= req
->async_data
;
6056 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6058 hrtimer_init(&io
->timer
, io_timeout_get_clock(io
), mode
);
6059 io
->timer
.function
= io_link_timeout_fn
;
6060 hrtimer_start(&io
->timer
, timespec64_to_ktime(*ts
), mode
);
6064 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6065 struct timespec64
*ts
, enum hrtimer_mode mode
)
6066 __must_hold(&ctx
->timeout_lock
)
6068 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6069 struct io_timeout_data
*data
;
6072 return PTR_ERR(req
);
6074 req
->timeout
.off
= 0; /* noseq */
6075 data
= req
->async_data
;
6076 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
6077 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), mode
);
6078 data
->timer
.function
= io_timeout_fn
;
6079 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
6083 static int io_timeout_remove_prep(struct io_kiocb
*req
,
6084 const struct io_uring_sqe
*sqe
)
6086 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6088 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6090 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6092 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->splice_fd_in
)
6095 tr
->ltimeout
= false;
6096 tr
->addr
= READ_ONCE(sqe
->addr
);
6097 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
6098 if (tr
->flags
& IORING_TIMEOUT_UPDATE_MASK
) {
6099 if (hweight32(tr
->flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6101 if (tr
->flags
& IORING_LINK_TIMEOUT_UPDATE
)
6102 tr
->ltimeout
= true;
6103 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE_MASK
|IORING_TIMEOUT_ABS
))
6105 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
6107 } else if (tr
->flags
) {
6108 /* timeout removal doesn't support flags */
6115 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
6117 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
6122 * Remove or update an existing timeout command
6124 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
6126 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6127 struct io_ring_ctx
*ctx
= req
->ctx
;
6130 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
)) {
6131 spin_lock(&ctx
->completion_lock
);
6132 spin_lock_irq(&ctx
->timeout_lock
);
6133 ret
= io_timeout_cancel(ctx
, tr
->addr
);
6134 spin_unlock_irq(&ctx
->timeout_lock
);
6135 spin_unlock(&ctx
->completion_lock
);
6137 enum hrtimer_mode mode
= io_translate_timeout_mode(tr
->flags
);
6139 spin_lock_irq(&ctx
->timeout_lock
);
6141 ret
= io_linked_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6143 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6144 spin_unlock_irq(&ctx
->timeout_lock
);
6149 io_req_complete_post(req
, ret
, 0);
6153 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6154 bool is_timeout_link
)
6156 struct io_timeout_data
*data
;
6158 u32 off
= READ_ONCE(sqe
->off
);
6160 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6162 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1 ||
6165 if (off
&& is_timeout_link
)
6167 flags
= READ_ONCE(sqe
->timeout_flags
);
6168 if (flags
& ~(IORING_TIMEOUT_ABS
| IORING_TIMEOUT_CLOCK_MASK
))
6170 /* more than one clock specified is invalid, obviously */
6171 if (hweight32(flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6174 INIT_LIST_HEAD(&req
->timeout
.list
);
6175 req
->timeout
.off
= off
;
6176 if (unlikely(off
&& !req
->ctx
->off_timeout_used
))
6177 req
->ctx
->off_timeout_used
= true;
6179 if (!req
->async_data
&& io_alloc_async_data(req
))
6182 data
= req
->async_data
;
6184 data
->flags
= flags
;
6186 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
6189 data
->mode
= io_translate_timeout_mode(flags
);
6190 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), data
->mode
);
6192 if (is_timeout_link
) {
6193 struct io_submit_link
*link
= &req
->ctx
->submit_state
.link
;
6197 if (link
->last
->opcode
== IORING_OP_LINK_TIMEOUT
)
6199 req
->timeout
.head
= link
->last
;
6200 link
->last
->flags
|= REQ_F_ARM_LTIMEOUT
;
6205 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
6207 struct io_ring_ctx
*ctx
= req
->ctx
;
6208 struct io_timeout_data
*data
= req
->async_data
;
6209 struct list_head
*entry
;
6210 u32 tail
, off
= req
->timeout
.off
;
6212 spin_lock_irq(&ctx
->timeout_lock
);
6215 * sqe->off holds how many events that need to occur for this
6216 * timeout event to be satisfied. If it isn't set, then this is
6217 * a pure timeout request, sequence isn't used.
6219 if (io_is_timeout_noseq(req
)) {
6220 entry
= ctx
->timeout_list
.prev
;
6224 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
6225 req
->timeout
.target_seq
= tail
+ off
;
6227 /* Update the last seq here in case io_flush_timeouts() hasn't.
6228 * This is safe because ->completion_lock is held, and submissions
6229 * and completions are never mixed in the same ->completion_lock section.
6231 ctx
->cq_last_tm_flush
= tail
;
6234 * Insertion sort, ensuring the first entry in the list is always
6235 * the one we need first.
6237 list_for_each_prev(entry
, &ctx
->timeout_list
) {
6238 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
6241 if (io_is_timeout_noseq(nxt
))
6243 /* nxt.seq is behind @tail, otherwise would've been completed */
6244 if (off
>= nxt
->timeout
.target_seq
- tail
)
6248 list_add(&req
->timeout
.list
, entry
);
6249 data
->timer
.function
= io_timeout_fn
;
6250 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
6251 spin_unlock_irq(&ctx
->timeout_lock
);
6255 struct io_cancel_data
{
6256 struct io_ring_ctx
*ctx
;
6260 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
6262 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6263 struct io_cancel_data
*cd
= data
;
6265 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
6268 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
6269 struct io_ring_ctx
*ctx
)
6271 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
6272 enum io_wq_cancel cancel_ret
;
6275 if (!tctx
|| !tctx
->io_wq
)
6278 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
6279 switch (cancel_ret
) {
6280 case IO_WQ_CANCEL_OK
:
6283 case IO_WQ_CANCEL_RUNNING
:
6286 case IO_WQ_CANCEL_NOTFOUND
:
6294 static int io_try_cancel_userdata(struct io_kiocb
*req
, u64 sqe_addr
)
6296 struct io_ring_ctx
*ctx
= req
->ctx
;
6299 WARN_ON_ONCE(!io_wq_current_is_worker() && req
->task
!= current
);
6301 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
6305 spin_lock(&ctx
->completion_lock
);
6306 spin_lock_irq(&ctx
->timeout_lock
);
6307 ret
= io_timeout_cancel(ctx
, sqe_addr
);
6308 spin_unlock_irq(&ctx
->timeout_lock
);
6311 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
6313 spin_unlock(&ctx
->completion_lock
);
6317 static int io_async_cancel_prep(struct io_kiocb
*req
,
6318 const struct io_uring_sqe
*sqe
)
6320 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6322 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6324 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
||
6328 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
6332 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
6334 struct io_ring_ctx
*ctx
= req
->ctx
;
6335 u64 sqe_addr
= req
->cancel
.addr
;
6336 struct io_tctx_node
*node
;
6339 ret
= io_try_cancel_userdata(req
, sqe_addr
);
6343 /* slow path, try all io-wq's */
6344 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6346 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
6347 struct io_uring_task
*tctx
= node
->task
->io_uring
;
6349 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
6353 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6357 io_req_complete_post(req
, ret
, 0);
6361 static int io_rsrc_update_prep(struct io_kiocb
*req
,
6362 const struct io_uring_sqe
*sqe
)
6364 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6366 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
6369 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
6370 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
6371 if (!req
->rsrc_update
.nr_args
)
6373 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
6377 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
6379 struct io_ring_ctx
*ctx
= req
->ctx
;
6380 struct io_uring_rsrc_update2 up
;
6383 up
.offset
= req
->rsrc_update
.offset
;
6384 up
.data
= req
->rsrc_update
.arg
;
6389 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6390 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
6391 &up
, req
->rsrc_update
.nr_args
);
6392 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6396 __io_req_complete(req
, issue_flags
, ret
, 0);
6400 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6402 switch (req
->opcode
) {
6405 case IORING_OP_READV
:
6406 case IORING_OP_READ_FIXED
:
6407 case IORING_OP_READ
:
6408 return io_read_prep(req
, sqe
);
6409 case IORING_OP_WRITEV
:
6410 case IORING_OP_WRITE_FIXED
:
6411 case IORING_OP_WRITE
:
6412 return io_write_prep(req
, sqe
);
6413 case IORING_OP_POLL_ADD
:
6414 return io_poll_add_prep(req
, sqe
);
6415 case IORING_OP_POLL_REMOVE
:
6416 return io_poll_update_prep(req
, sqe
);
6417 case IORING_OP_FSYNC
:
6418 return io_fsync_prep(req
, sqe
);
6419 case IORING_OP_SYNC_FILE_RANGE
:
6420 return io_sfr_prep(req
, sqe
);
6421 case IORING_OP_SENDMSG
:
6422 case IORING_OP_SEND
:
6423 return io_sendmsg_prep(req
, sqe
);
6424 case IORING_OP_RECVMSG
:
6425 case IORING_OP_RECV
:
6426 return io_recvmsg_prep(req
, sqe
);
6427 case IORING_OP_CONNECT
:
6428 return io_connect_prep(req
, sqe
);
6429 case IORING_OP_TIMEOUT
:
6430 return io_timeout_prep(req
, sqe
, false);
6431 case IORING_OP_TIMEOUT_REMOVE
:
6432 return io_timeout_remove_prep(req
, sqe
);
6433 case IORING_OP_ASYNC_CANCEL
:
6434 return io_async_cancel_prep(req
, sqe
);
6435 case IORING_OP_LINK_TIMEOUT
:
6436 return io_timeout_prep(req
, sqe
, true);
6437 case IORING_OP_ACCEPT
:
6438 return io_accept_prep(req
, sqe
);
6439 case IORING_OP_FALLOCATE
:
6440 return io_fallocate_prep(req
, sqe
);
6441 case IORING_OP_OPENAT
:
6442 return io_openat_prep(req
, sqe
);
6443 case IORING_OP_CLOSE
:
6444 return io_close_prep(req
, sqe
);
6445 case IORING_OP_FILES_UPDATE
:
6446 return io_rsrc_update_prep(req
, sqe
);
6447 case IORING_OP_STATX
:
6448 return io_statx_prep(req
, sqe
);
6449 case IORING_OP_FADVISE
:
6450 return io_fadvise_prep(req
, sqe
);
6451 case IORING_OP_MADVISE
:
6452 return io_madvise_prep(req
, sqe
);
6453 case IORING_OP_OPENAT2
:
6454 return io_openat2_prep(req
, sqe
);
6455 case IORING_OP_EPOLL_CTL
:
6456 return io_epoll_ctl_prep(req
, sqe
);
6457 case IORING_OP_SPLICE
:
6458 return io_splice_prep(req
, sqe
);
6459 case IORING_OP_PROVIDE_BUFFERS
:
6460 return io_provide_buffers_prep(req
, sqe
);
6461 case IORING_OP_REMOVE_BUFFERS
:
6462 return io_remove_buffers_prep(req
, sqe
);
6464 return io_tee_prep(req
, sqe
);
6465 case IORING_OP_SHUTDOWN
:
6466 return io_shutdown_prep(req
, sqe
);
6467 case IORING_OP_RENAMEAT
:
6468 return io_renameat_prep(req
, sqe
);
6469 case IORING_OP_UNLINKAT
:
6470 return io_unlinkat_prep(req
, sqe
);
6471 case IORING_OP_MKDIRAT
:
6472 return io_mkdirat_prep(req
, sqe
);
6473 case IORING_OP_SYMLINKAT
:
6474 return io_symlinkat_prep(req
, sqe
);
6475 case IORING_OP_LINKAT
:
6476 return io_linkat_prep(req
, sqe
);
6479 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6484 static int io_req_prep_async(struct io_kiocb
*req
)
6486 if (!io_op_defs
[req
->opcode
].needs_async_setup
)
6488 if (WARN_ON_ONCE(req
->async_data
))
6490 if (io_alloc_async_data(req
))
6493 switch (req
->opcode
) {
6494 case IORING_OP_READV
:
6495 return io_rw_prep_async(req
, READ
);
6496 case IORING_OP_WRITEV
:
6497 return io_rw_prep_async(req
, WRITE
);
6498 case IORING_OP_SENDMSG
:
6499 return io_sendmsg_prep_async(req
);
6500 case IORING_OP_RECVMSG
:
6501 return io_recvmsg_prep_async(req
);
6502 case IORING_OP_CONNECT
:
6503 return io_connect_prep_async(req
);
6505 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
6510 static u32
io_get_sequence(struct io_kiocb
*req
)
6512 u32 seq
= req
->ctx
->cached_sq_head
;
6514 /* need original cached_sq_head, but it was increased for each req */
6515 io_for_each_link(req
, req
)
6520 static bool io_drain_req(struct io_kiocb
*req
)
6522 struct io_kiocb
*pos
;
6523 struct io_ring_ctx
*ctx
= req
->ctx
;
6524 struct io_defer_entry
*de
;
6528 if (req
->flags
& REQ_F_FAIL
) {
6529 io_req_complete_fail_submit(req
);
6534 * If we need to drain a request in the middle of a link, drain the
6535 * head request and the next request/link after the current link.
6536 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6537 * maintained for every request of our link.
6539 if (ctx
->drain_next
) {
6540 req
->flags
|= REQ_F_IO_DRAIN
;
6541 ctx
->drain_next
= false;
6543 /* not interested in head, start from the first linked */
6544 io_for_each_link(pos
, req
->link
) {
6545 if (pos
->flags
& REQ_F_IO_DRAIN
) {
6546 ctx
->drain_next
= true;
6547 req
->flags
|= REQ_F_IO_DRAIN
;
6552 /* Still need defer if there is pending req in defer list. */
6553 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6554 !(req
->flags
& REQ_F_IO_DRAIN
))) {
6555 ctx
->drain_active
= false;
6559 seq
= io_get_sequence(req
);
6560 /* Still a chance to pass the sequence check */
6561 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6564 ret
= io_req_prep_async(req
);
6567 io_prep_async_link(req
);
6568 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6572 io_req_complete_failed(req
, ret
);
6576 spin_lock(&ctx
->completion_lock
);
6577 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6578 spin_unlock(&ctx
->completion_lock
);
6580 io_queue_async_work(req
, NULL
);
6584 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6587 list_add_tail(&de
->list
, &ctx
->defer_list
);
6588 spin_unlock(&ctx
->completion_lock
);
6592 static void io_clean_op(struct io_kiocb
*req
)
6594 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6595 switch (req
->opcode
) {
6596 case IORING_OP_READV
:
6597 case IORING_OP_READ_FIXED
:
6598 case IORING_OP_READ
:
6599 kfree((void *)(unsigned long)req
->rw
.addr
);
6601 case IORING_OP_RECVMSG
:
6602 case IORING_OP_RECV
:
6603 kfree(req
->sr_msg
.kbuf
);
6608 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6609 switch (req
->opcode
) {
6610 case IORING_OP_READV
:
6611 case IORING_OP_READ_FIXED
:
6612 case IORING_OP_READ
:
6613 case IORING_OP_WRITEV
:
6614 case IORING_OP_WRITE_FIXED
:
6615 case IORING_OP_WRITE
: {
6616 struct io_async_rw
*io
= req
->async_data
;
6618 kfree(io
->free_iovec
);
6621 case IORING_OP_RECVMSG
:
6622 case IORING_OP_SENDMSG
: {
6623 struct io_async_msghdr
*io
= req
->async_data
;
6625 kfree(io
->free_iov
);
6628 case IORING_OP_SPLICE
:
6630 if (!(req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
))
6631 io_put_file(req
->splice
.file_in
);
6633 case IORING_OP_OPENAT
:
6634 case IORING_OP_OPENAT2
:
6635 if (req
->open
.filename
)
6636 putname(req
->open
.filename
);
6638 case IORING_OP_RENAMEAT
:
6639 putname(req
->rename
.oldpath
);
6640 putname(req
->rename
.newpath
);
6642 case IORING_OP_UNLINKAT
:
6643 putname(req
->unlink
.filename
);
6645 case IORING_OP_MKDIRAT
:
6646 putname(req
->mkdir
.filename
);
6648 case IORING_OP_SYMLINKAT
:
6649 putname(req
->symlink
.oldpath
);
6650 putname(req
->symlink
.newpath
);
6652 case IORING_OP_LINKAT
:
6653 putname(req
->hardlink
.oldpath
);
6654 putname(req
->hardlink
.newpath
);
6658 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
6659 kfree(req
->apoll
->double_poll
);
6663 if (req
->flags
& REQ_F_INFLIGHT
) {
6664 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6666 atomic_dec(&tctx
->inflight_tracked
);
6668 if (req
->flags
& REQ_F_CREDS
)
6669 put_cred(req
->creds
);
6671 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
6674 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6676 struct io_ring_ctx
*ctx
= req
->ctx
;
6677 const struct cred
*creds
= NULL
;
6680 if ((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred())
6681 creds
= override_creds(req
->creds
);
6683 switch (req
->opcode
) {
6685 ret
= io_nop(req
, issue_flags
);
6687 case IORING_OP_READV
:
6688 case IORING_OP_READ_FIXED
:
6689 case IORING_OP_READ
:
6690 ret
= io_read(req
, issue_flags
);
6692 case IORING_OP_WRITEV
:
6693 case IORING_OP_WRITE_FIXED
:
6694 case IORING_OP_WRITE
:
6695 ret
= io_write(req
, issue_flags
);
6697 case IORING_OP_FSYNC
:
6698 ret
= io_fsync(req
, issue_flags
);
6700 case IORING_OP_POLL_ADD
:
6701 ret
= io_poll_add(req
, issue_flags
);
6703 case IORING_OP_POLL_REMOVE
:
6704 ret
= io_poll_update(req
, issue_flags
);
6706 case IORING_OP_SYNC_FILE_RANGE
:
6707 ret
= io_sync_file_range(req
, issue_flags
);
6709 case IORING_OP_SENDMSG
:
6710 ret
= io_sendmsg(req
, issue_flags
);
6712 case IORING_OP_SEND
:
6713 ret
= io_send(req
, issue_flags
);
6715 case IORING_OP_RECVMSG
:
6716 ret
= io_recvmsg(req
, issue_flags
);
6718 case IORING_OP_RECV
:
6719 ret
= io_recv(req
, issue_flags
);
6721 case IORING_OP_TIMEOUT
:
6722 ret
= io_timeout(req
, issue_flags
);
6724 case IORING_OP_TIMEOUT_REMOVE
:
6725 ret
= io_timeout_remove(req
, issue_flags
);
6727 case IORING_OP_ACCEPT
:
6728 ret
= io_accept(req
, issue_flags
);
6730 case IORING_OP_CONNECT
:
6731 ret
= io_connect(req
, issue_flags
);
6733 case IORING_OP_ASYNC_CANCEL
:
6734 ret
= io_async_cancel(req
, issue_flags
);
6736 case IORING_OP_FALLOCATE
:
6737 ret
= io_fallocate(req
, issue_flags
);
6739 case IORING_OP_OPENAT
:
6740 ret
= io_openat(req
, issue_flags
);
6742 case IORING_OP_CLOSE
:
6743 ret
= io_close(req
, issue_flags
);
6745 case IORING_OP_FILES_UPDATE
:
6746 ret
= io_files_update(req
, issue_flags
);
6748 case IORING_OP_STATX
:
6749 ret
= io_statx(req
, issue_flags
);
6751 case IORING_OP_FADVISE
:
6752 ret
= io_fadvise(req
, issue_flags
);
6754 case IORING_OP_MADVISE
:
6755 ret
= io_madvise(req
, issue_flags
);
6757 case IORING_OP_OPENAT2
:
6758 ret
= io_openat2(req
, issue_flags
);
6760 case IORING_OP_EPOLL_CTL
:
6761 ret
= io_epoll_ctl(req
, issue_flags
);
6763 case IORING_OP_SPLICE
:
6764 ret
= io_splice(req
, issue_flags
);
6766 case IORING_OP_PROVIDE_BUFFERS
:
6767 ret
= io_provide_buffers(req
, issue_flags
);
6769 case IORING_OP_REMOVE_BUFFERS
:
6770 ret
= io_remove_buffers(req
, issue_flags
);
6773 ret
= io_tee(req
, issue_flags
);
6775 case IORING_OP_SHUTDOWN
:
6776 ret
= io_shutdown(req
, issue_flags
);
6778 case IORING_OP_RENAMEAT
:
6779 ret
= io_renameat(req
, issue_flags
);
6781 case IORING_OP_UNLINKAT
:
6782 ret
= io_unlinkat(req
, issue_flags
);
6784 case IORING_OP_MKDIRAT
:
6785 ret
= io_mkdirat(req
, issue_flags
);
6787 case IORING_OP_SYMLINKAT
:
6788 ret
= io_symlinkat(req
, issue_flags
);
6790 case IORING_OP_LINKAT
:
6791 ret
= io_linkat(req
, issue_flags
);
6799 revert_creds(creds
);
6802 /* If the op doesn't have a file, we're not polling for it */
6803 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
6804 io_iopoll_req_issued(req
);
6809 static struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
6811 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6813 req
= io_put_req_find_next(req
);
6814 return req
? &req
->work
: NULL
;
6817 static void io_wq_submit_work(struct io_wq_work
*work
)
6819 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6820 struct io_kiocb
*timeout
;
6823 /* one will be dropped by ->io_free_work() after returning to io-wq */
6824 if (!(req
->flags
& REQ_F_REFCOUNT
))
6825 __io_req_set_refcount(req
, 2);
6829 timeout
= io_prep_linked_timeout(req
);
6831 io_queue_linked_timeout(timeout
);
6833 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6834 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6839 ret
= io_issue_sqe(req
, 0);
6841 * We can get EAGAIN for polled IO even though we're
6842 * forcing a sync submission from here, since we can't
6843 * wait for request slots on the block side.
6851 /* avoid locking problems by failing it from a clean context */
6853 io_req_task_queue_fail(req
, ret
);
6856 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
6859 return &table
->files
[i
];
6862 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6865 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
6867 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
6870 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
6872 unsigned long file_ptr
= (unsigned long) file
;
6874 if (__io_file_supports_nowait(file
, READ
))
6875 file_ptr
|= FFS_ASYNC_READ
;
6876 if (__io_file_supports_nowait(file
, WRITE
))
6877 file_ptr
|= FFS_ASYNC_WRITE
;
6878 if (S_ISREG(file_inode(file
)->i_mode
))
6879 file_ptr
|= FFS_ISREG
;
6880 file_slot
->file_ptr
= file_ptr
;
6883 static inline struct file
*io_file_get_fixed(struct io_ring_ctx
*ctx
,
6884 struct io_kiocb
*req
, int fd
)
6887 unsigned long file_ptr
;
6889 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6891 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6892 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
6893 file
= (struct file
*) (file_ptr
& FFS_MASK
);
6894 file_ptr
&= ~FFS_MASK
;
6895 /* mask in overlapping REQ_F and FFS bits */
6896 req
->flags
|= (file_ptr
<< REQ_F_NOWAIT_READ_BIT
);
6897 io_req_set_rsrc_node(req
);
6901 static struct file
*io_file_get_normal(struct io_ring_ctx
*ctx
,
6902 struct io_kiocb
*req
, int fd
)
6904 struct file
*file
= fget(fd
);
6906 trace_io_uring_file_get(ctx
, fd
);
6908 /* we don't allow fixed io_uring files */
6909 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6910 io_req_track_inflight(req
);
6914 static inline struct file
*io_file_get(struct io_ring_ctx
*ctx
,
6915 struct io_kiocb
*req
, int fd
, bool fixed
)
6918 return io_file_get_fixed(ctx
, req
, fd
);
6920 return io_file_get_normal(ctx
, req
, fd
);
6923 static void io_req_task_link_timeout(struct io_kiocb
*req
, bool *locked
)
6925 struct io_kiocb
*prev
= req
->timeout
.prev
;
6929 if (!(req
->task
->flags
& PF_EXITING
))
6930 ret
= io_try_cancel_userdata(req
, prev
->user_data
);
6931 io_req_complete_post(req
, ret
?: -ETIME
, 0);
6934 io_req_complete_post(req
, -ETIME
, 0);
6938 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6940 struct io_timeout_data
*data
= container_of(timer
,
6941 struct io_timeout_data
, timer
);
6942 struct io_kiocb
*prev
, *req
= data
->req
;
6943 struct io_ring_ctx
*ctx
= req
->ctx
;
6944 unsigned long flags
;
6946 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
6947 prev
= req
->timeout
.head
;
6948 req
->timeout
.head
= NULL
;
6951 * We don't expect the list to be empty, that will only happen if we
6952 * race with the completion of the linked work.
6955 io_remove_next_linked(prev
);
6956 if (!req_ref_inc_not_zero(prev
))
6959 list_del(&req
->timeout
.list
);
6960 req
->timeout
.prev
= prev
;
6961 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
6963 req
->io_task_work
.func
= io_req_task_link_timeout
;
6964 io_req_task_work_add(req
);
6965 return HRTIMER_NORESTART
;
6968 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6970 struct io_ring_ctx
*ctx
= req
->ctx
;
6972 spin_lock_irq(&ctx
->timeout_lock
);
6974 * If the back reference is NULL, then our linked request finished
6975 * before we got a chance to setup the timer
6977 if (req
->timeout
.head
) {
6978 struct io_timeout_data
*data
= req
->async_data
;
6980 data
->timer
.function
= io_link_timeout_fn
;
6981 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6983 list_add_tail(&req
->timeout
.list
, &ctx
->ltimeout_list
);
6985 spin_unlock_irq(&ctx
->timeout_lock
);
6986 /* drop submission reference */
6990 static void __io_queue_sqe(struct io_kiocb
*req
)
6991 __must_hold(&req
->ctx
->uring_lock
)
6993 struct io_kiocb
*linked_timeout
;
6997 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
7000 * We async punt it if the file wasn't marked NOWAIT, or if the file
7001 * doesn't support non-blocking read/write attempts
7004 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
7005 struct io_ring_ctx
*ctx
= req
->ctx
;
7006 struct io_submit_state
*state
= &ctx
->submit_state
;
7008 state
->compl_reqs
[state
->compl_nr
++] = req
;
7009 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
7010 io_submit_flush_completions(ctx
);
7014 linked_timeout
= io_prep_linked_timeout(req
);
7016 io_queue_linked_timeout(linked_timeout
);
7017 } else if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
7018 linked_timeout
= io_prep_linked_timeout(req
);
7020 switch (io_arm_poll_handler(req
)) {
7021 case IO_APOLL_READY
:
7023 io_queue_linked_timeout(linked_timeout
);
7025 case IO_APOLL_ABORTED
:
7027 * Queued up for async execution, worker will release
7028 * submit reference when the iocb is actually submitted.
7030 io_queue_async_work(req
, NULL
);
7035 io_queue_linked_timeout(linked_timeout
);
7037 io_req_complete_failed(req
, ret
);
7041 static inline void io_queue_sqe(struct io_kiocb
*req
)
7042 __must_hold(&req
->ctx
->uring_lock
)
7044 if (unlikely(req
->ctx
->drain_active
) && io_drain_req(req
))
7047 if (likely(!(req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
)))) {
7048 __io_queue_sqe(req
);
7049 } else if (req
->flags
& REQ_F_FAIL
) {
7050 io_req_complete_fail_submit(req
);
7052 int ret
= io_req_prep_async(req
);
7055 io_req_complete_failed(req
, ret
);
7057 io_queue_async_work(req
, NULL
);
7062 * Check SQE restrictions (opcode and flags).
7064 * Returns 'true' if SQE is allowed, 'false' otherwise.
7066 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
7067 struct io_kiocb
*req
,
7068 unsigned int sqe_flags
)
7070 if (likely(!ctx
->restricted
))
7073 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
7076 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
7077 ctx
->restrictions
.sqe_flags_required
)
7080 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
7081 ctx
->restrictions
.sqe_flags_required
))
7087 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7088 const struct io_uring_sqe
*sqe
)
7089 __must_hold(&ctx
->uring_lock
)
7091 struct io_submit_state
*state
;
7092 unsigned int sqe_flags
;
7093 int personality
, ret
= 0;
7095 /* req is partially pre-initialised, see io_preinit_req() */
7096 req
->opcode
= READ_ONCE(sqe
->opcode
);
7097 /* same numerical values with corresponding REQ_F_*, safe to copy */
7098 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
7099 req
->user_data
= READ_ONCE(sqe
->user_data
);
7101 req
->fixed_rsrc_refs
= NULL
;
7102 req
->task
= current
;
7104 /* enforce forwards compatibility on users */
7105 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
7107 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
7109 if (!io_check_restriction(ctx
, req
, sqe_flags
))
7112 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
7113 !io_op_defs
[req
->opcode
].buffer_select
)
7115 if (unlikely(sqe_flags
& IOSQE_IO_DRAIN
))
7116 ctx
->drain_active
= true;
7118 personality
= READ_ONCE(sqe
->personality
);
7120 req
->creds
= xa_load(&ctx
->personalities
, personality
);
7123 get_cred(req
->creds
);
7124 req
->flags
|= REQ_F_CREDS
;
7126 state
= &ctx
->submit_state
;
7129 * Plug now if we have more than 1 IO left after this, and the target
7130 * is potentially a read/write to block based storage.
7132 if (!state
->plug_started
&& state
->ios_left
> 1 &&
7133 io_op_defs
[req
->opcode
].plug
) {
7134 blk_start_plug(&state
->plug
);
7135 state
->plug_started
= true;
7138 if (io_op_defs
[req
->opcode
].needs_file
) {
7139 req
->file
= io_file_get(ctx
, req
, READ_ONCE(sqe
->fd
),
7140 (sqe_flags
& IOSQE_FIXED_FILE
));
7141 if (unlikely(!req
->file
))
7149 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7150 const struct io_uring_sqe
*sqe
)
7151 __must_hold(&ctx
->uring_lock
)
7153 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
7156 ret
= io_init_req(ctx
, req
, sqe
);
7157 if (unlikely(ret
)) {
7159 /* fail even hard links since we don't submit */
7162 * we can judge a link req is failed or cancelled by if
7163 * REQ_F_FAIL is set, but the head is an exception since
7164 * it may be set REQ_F_FAIL because of other req's failure
7165 * so let's leverage req->result to distinguish if a head
7166 * is set REQ_F_FAIL because of its failure or other req's
7167 * failure so that we can set the correct ret code for it.
7168 * init result here to avoid affecting the normal path.
7170 if (!(link
->head
->flags
& REQ_F_FAIL
))
7171 req_fail_link_node(link
->head
, -ECANCELED
);
7172 } else if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7174 * the current req is a normal req, we should return
7175 * error and thus break the submittion loop.
7177 io_req_complete_failed(req
, ret
);
7180 req_fail_link_node(req
, ret
);
7182 ret
= io_req_prep(req
, sqe
);
7187 /* don't need @sqe from now on */
7188 trace_io_uring_submit_sqe(ctx
, req
, req
->opcode
, req
->user_data
,
7190 ctx
->flags
& IORING_SETUP_SQPOLL
);
7193 * If we already have a head request, queue this one for async
7194 * submittal once the head completes. If we don't have a head but
7195 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7196 * submitted sync once the chain is complete. If none of those
7197 * conditions are true (normal request), then just queue it.
7200 struct io_kiocb
*head
= link
->head
;
7202 if (!(req
->flags
& REQ_F_FAIL
)) {
7203 ret
= io_req_prep_async(req
);
7204 if (unlikely(ret
)) {
7205 req_fail_link_node(req
, ret
);
7206 if (!(head
->flags
& REQ_F_FAIL
))
7207 req_fail_link_node(head
, -ECANCELED
);
7210 trace_io_uring_link(ctx
, req
, head
);
7211 link
->last
->link
= req
;
7214 /* last request of a link, enqueue the link */
7215 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7220 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
7232 * Batched submission is done, ensure local IO is flushed out.
7234 static void io_submit_state_end(struct io_submit_state
*state
,
7235 struct io_ring_ctx
*ctx
)
7237 if (state
->link
.head
)
7238 io_queue_sqe(state
->link
.head
);
7239 if (state
->compl_nr
)
7240 io_submit_flush_completions(ctx
);
7241 if (state
->plug_started
)
7242 blk_finish_plug(&state
->plug
);
7246 * Start submission side cache.
7248 static void io_submit_state_start(struct io_submit_state
*state
,
7249 unsigned int max_ios
)
7251 state
->plug_started
= false;
7252 state
->ios_left
= max_ios
;
7253 /* set only head, no need to init link_last in advance */
7254 state
->link
.head
= NULL
;
7257 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
7259 struct io_rings
*rings
= ctx
->rings
;
7262 * Ensure any loads from the SQEs are done at this point,
7263 * since once we write the new head, the application could
7264 * write new data to them.
7266 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
7270 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7271 * that is mapped by userspace. This means that care needs to be taken to
7272 * ensure that reads are stable, as we cannot rely on userspace always
7273 * being a good citizen. If members of the sqe are validated and then later
7274 * used, it's important that those reads are done through READ_ONCE() to
7275 * prevent a re-load down the line.
7277 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
7279 unsigned head
, mask
= ctx
->sq_entries
- 1;
7280 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
7283 * The cached sq head (or cq tail) serves two purposes:
7285 * 1) allows us to batch the cost of updating the user visible
7287 * 2) allows the kernel side to track the head on its own, even
7288 * though the application is the one updating it.
7290 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
7291 if (likely(head
< ctx
->sq_entries
))
7292 return &ctx
->sq_sqes
[head
];
7294 /* drop invalid entries */
7296 WRITE_ONCE(ctx
->rings
->sq_dropped
,
7297 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
7301 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
7302 __must_hold(&ctx
->uring_lock
)
7306 /* make sure SQ entry isn't read before tail */
7307 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
7308 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
7310 io_get_task_refs(nr
);
7312 io_submit_state_start(&ctx
->submit_state
, nr
);
7313 while (submitted
< nr
) {
7314 const struct io_uring_sqe
*sqe
;
7315 struct io_kiocb
*req
;
7317 req
= io_alloc_req(ctx
);
7318 if (unlikely(!req
)) {
7320 submitted
= -EAGAIN
;
7323 sqe
= io_get_sqe(ctx
);
7324 if (unlikely(!sqe
)) {
7325 list_add(&req
->inflight_entry
, &ctx
->submit_state
.free_list
);
7328 /* will complete beyond this point, count as submitted */
7330 if (io_submit_sqe(ctx
, req
, sqe
))
7334 if (unlikely(submitted
!= nr
)) {
7335 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
7336 int unused
= nr
- ref_used
;
7338 current
->io_uring
->cached_refs
+= unused
;
7339 percpu_ref_put_many(&ctx
->refs
, unused
);
7342 io_submit_state_end(&ctx
->submit_state
, ctx
);
7343 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7344 io_commit_sqring(ctx
);
7349 static inline bool io_sqd_events_pending(struct io_sq_data
*sqd
)
7351 return READ_ONCE(sqd
->state
);
7354 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
7356 /* Tell userspace we may need a wakeup call */
7357 spin_lock(&ctx
->completion_lock
);
7358 WRITE_ONCE(ctx
->rings
->sq_flags
,
7359 ctx
->rings
->sq_flags
| IORING_SQ_NEED_WAKEUP
);
7360 spin_unlock(&ctx
->completion_lock
);
7363 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
7365 spin_lock(&ctx
->completion_lock
);
7366 WRITE_ONCE(ctx
->rings
->sq_flags
,
7367 ctx
->rings
->sq_flags
& ~IORING_SQ_NEED_WAKEUP
);
7368 spin_unlock(&ctx
->completion_lock
);
7371 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
7373 unsigned int to_submit
;
7376 to_submit
= io_sqring_entries(ctx
);
7377 /* if we're handling multiple rings, cap submit size for fairness */
7378 if (cap_entries
&& to_submit
> IORING_SQPOLL_CAP_ENTRIES_VALUE
)
7379 to_submit
= IORING_SQPOLL_CAP_ENTRIES_VALUE
;
7381 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
7382 unsigned nr_events
= 0;
7383 const struct cred
*creds
= NULL
;
7385 if (ctx
->sq_creds
!= current_cred())
7386 creds
= override_creds(ctx
->sq_creds
);
7388 mutex_lock(&ctx
->uring_lock
);
7389 if (!list_empty(&ctx
->iopoll_list
))
7390 io_do_iopoll(ctx
, &nr_events
, 0);
7393 * Don't submit if refs are dying, good for io_uring_register(),
7394 * but also it is relied upon by io_ring_exit_work()
7396 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
7397 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
7398 ret
= io_submit_sqes(ctx
, to_submit
);
7399 mutex_unlock(&ctx
->uring_lock
);
7401 if (to_submit
&& wq_has_sleeper(&ctx
->sqo_sq_wait
))
7402 wake_up(&ctx
->sqo_sq_wait
);
7404 revert_creds(creds
);
7410 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7412 struct io_ring_ctx
*ctx
;
7413 unsigned sq_thread_idle
= 0;
7415 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7416 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
7417 sqd
->sq_thread_idle
= sq_thread_idle
;
7420 static bool io_sqd_handle_event(struct io_sq_data
*sqd
)
7422 bool did_sig
= false;
7423 struct ksignal ksig
;
7425 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
7426 signal_pending(current
)) {
7427 mutex_unlock(&sqd
->lock
);
7428 if (signal_pending(current
))
7429 did_sig
= get_signal(&ksig
);
7431 mutex_lock(&sqd
->lock
);
7433 return did_sig
|| test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7436 static int io_sq_thread(void *data
)
7438 struct io_sq_data
*sqd
= data
;
7439 struct io_ring_ctx
*ctx
;
7440 unsigned long timeout
= 0;
7441 char buf
[TASK_COMM_LEN
];
7444 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
7445 set_task_comm(current
, buf
);
7447 if (sqd
->sq_cpu
!= -1)
7448 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
7450 set_cpus_allowed_ptr(current
, cpu_online_mask
);
7451 current
->flags
|= PF_NO_SETAFFINITY
;
7453 mutex_lock(&sqd
->lock
);
7455 bool cap_entries
, sqt_spin
= false;
7457 if (io_sqd_events_pending(sqd
) || signal_pending(current
)) {
7458 if (io_sqd_handle_event(sqd
))
7460 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7463 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7464 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7465 int ret
= __io_sq_thread(ctx
, cap_entries
);
7467 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7470 if (io_run_task_work())
7473 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7476 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7480 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7481 if (!io_sqd_events_pending(sqd
) && !current
->task_works
) {
7482 bool needs_sched
= true;
7484 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7485 io_ring_set_wakeup_flag(ctx
);
7487 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7488 !list_empty_careful(&ctx
->iopoll_list
)) {
7489 needs_sched
= false;
7492 if (io_sqring_entries(ctx
)) {
7493 needs_sched
= false;
7499 mutex_unlock(&sqd
->lock
);
7501 mutex_lock(&sqd
->lock
);
7503 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7504 io_ring_clear_wakeup_flag(ctx
);
7507 finish_wait(&sqd
->wait
, &wait
);
7508 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7511 io_uring_cancel_generic(true, sqd
);
7513 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7514 io_ring_set_wakeup_flag(ctx
);
7516 mutex_unlock(&sqd
->lock
);
7518 complete(&sqd
->exited
);
7522 struct io_wait_queue
{
7523 struct wait_queue_entry wq
;
7524 struct io_ring_ctx
*ctx
;
7526 unsigned nr_timeouts
;
7529 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7531 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7532 int dist
= ctx
->cached_cq_tail
- (int) iowq
->cq_tail
;
7535 * Wake up if we have enough events, or if a timeout occurred since we
7536 * started waiting. For timeouts, we always want to return to userspace,
7537 * regardless of event count.
7539 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7542 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7543 int wake_flags
, void *key
)
7545 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7549 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7550 * the task, and the next invocation will do it.
7552 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->check_cq_overflow
))
7553 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7557 static int io_run_task_work_sig(void)
7559 if (io_run_task_work())
7561 if (!signal_pending(current
))
7563 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
7564 return -ERESTARTSYS
;
7568 /* when returns >0, the caller should retry */
7569 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7570 struct io_wait_queue
*iowq
,
7571 signed long *timeout
)
7575 /* make sure we run task_work before checking for signals */
7576 ret
= io_run_task_work_sig();
7577 if (ret
|| io_should_wake(iowq
))
7579 /* let the caller flush overflows, retry */
7580 if (test_bit(0, &ctx
->check_cq_overflow
))
7583 *timeout
= schedule_timeout(*timeout
);
7584 return !*timeout
? -ETIME
: 1;
7588 * Wait until events become available, if we don't already have some. The
7589 * application must reap them itself, as they reside on the shared cq ring.
7591 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7592 const sigset_t __user
*sig
, size_t sigsz
,
7593 struct __kernel_timespec __user
*uts
)
7595 struct io_wait_queue iowq
;
7596 struct io_rings
*rings
= ctx
->rings
;
7597 signed long timeout
= MAX_SCHEDULE_TIMEOUT
;
7601 io_cqring_overflow_flush(ctx
);
7602 if (io_cqring_events(ctx
) >= min_events
)
7604 if (!io_run_task_work())
7609 struct timespec64 ts
;
7611 if (get_timespec64(&ts
, uts
))
7613 timeout
= timespec64_to_jiffies(&ts
);
7617 #ifdef CONFIG_COMPAT
7618 if (in_compat_syscall())
7619 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7623 ret
= set_user_sigmask(sig
, sigsz
);
7629 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
7630 iowq
.wq
.private = current
;
7631 INIT_LIST_HEAD(&iowq
.wq
.entry
);
7633 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7634 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
7636 trace_io_uring_cqring_wait(ctx
, min_events
);
7638 /* if we can't even flush overflow, don't wait for more */
7639 if (!io_cqring_overflow_flush(ctx
)) {
7643 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
7644 TASK_INTERRUPTIBLE
);
7645 ret
= io_cqring_wait_schedule(ctx
, &iowq
, &timeout
);
7646 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
7650 restore_saved_sigmask_unless(ret
== -EINTR
);
7652 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7655 static void io_free_page_table(void **table
, size_t size
)
7657 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7659 for (i
= 0; i
< nr_tables
; i
++)
7664 static void **io_alloc_page_table(size_t size
)
7666 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7667 size_t init_size
= size
;
7670 table
= kcalloc(nr_tables
, sizeof(*table
), GFP_KERNEL_ACCOUNT
);
7674 for (i
= 0; i
< nr_tables
; i
++) {
7675 unsigned int this_size
= min_t(size_t, size
, PAGE_SIZE
);
7677 table
[i
] = kzalloc(this_size
, GFP_KERNEL_ACCOUNT
);
7679 io_free_page_table(table
, init_size
);
7687 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
7689 percpu_ref_exit(&ref_node
->refs
);
7693 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7695 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
7696 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
7697 unsigned long flags
;
7698 bool first_add
= false;
7700 spin_lock_irqsave(&ctx
->rsrc_ref_lock
, flags
);
7703 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7704 node
= list_first_entry(&ctx
->rsrc_ref_list
,
7705 struct io_rsrc_node
, node
);
7706 /* recycle ref nodes in order */
7709 list_del(&node
->node
);
7710 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
7712 spin_unlock_irqrestore(&ctx
->rsrc_ref_lock
, flags
);
7715 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, HZ
);
7718 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
)
7720 struct io_rsrc_node
*ref_node
;
7722 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7726 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7731 INIT_LIST_HEAD(&ref_node
->node
);
7732 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7733 ref_node
->done
= false;
7737 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
7738 struct io_rsrc_data
*data_to_kill
)
7740 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
7741 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
7744 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
7746 rsrc_node
->rsrc_data
= data_to_kill
;
7747 spin_lock_irq(&ctx
->rsrc_ref_lock
);
7748 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
7749 spin_unlock_irq(&ctx
->rsrc_ref_lock
);
7751 atomic_inc(&data_to_kill
->refs
);
7752 percpu_ref_kill(&rsrc_node
->refs
);
7753 ctx
->rsrc_node
= NULL
;
7756 if (!ctx
->rsrc_node
) {
7757 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
7758 ctx
->rsrc_backup_node
= NULL
;
7762 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
7764 if (ctx
->rsrc_backup_node
)
7766 ctx
->rsrc_backup_node
= io_rsrc_node_alloc(ctx
);
7767 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
7770 static int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
, struct io_ring_ctx
*ctx
)
7774 /* As we may drop ->uring_lock, other task may have started quiesce */
7778 data
->quiesce
= true;
7780 ret
= io_rsrc_node_switch_start(ctx
);
7783 io_rsrc_node_switch(ctx
, data
);
7785 /* kill initial ref, already quiesced if zero */
7786 if (atomic_dec_and_test(&data
->refs
))
7788 mutex_unlock(&ctx
->uring_lock
);
7789 flush_delayed_work(&ctx
->rsrc_put_work
);
7790 ret
= wait_for_completion_interruptible(&data
->done
);
7792 mutex_lock(&ctx
->uring_lock
);
7796 atomic_inc(&data
->refs
);
7797 /* wait for all works potentially completing data->done */
7798 flush_delayed_work(&ctx
->rsrc_put_work
);
7799 reinit_completion(&data
->done
);
7801 ret
= io_run_task_work_sig();
7802 mutex_lock(&ctx
->uring_lock
);
7804 data
->quiesce
= false;
7809 static u64
*io_get_tag_slot(struct io_rsrc_data
*data
, unsigned int idx
)
7811 unsigned int off
= idx
& IO_RSRC_TAG_TABLE_MASK
;
7812 unsigned int table_idx
= idx
>> IO_RSRC_TAG_TABLE_SHIFT
;
7814 return &data
->tags
[table_idx
][off
];
7817 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
7819 size_t size
= data
->nr
* sizeof(data
->tags
[0][0]);
7822 io_free_page_table((void **)data
->tags
, size
);
7826 static int io_rsrc_data_alloc(struct io_ring_ctx
*ctx
, rsrc_put_fn
*do_put
,
7827 u64 __user
*utags
, unsigned nr
,
7828 struct io_rsrc_data
**pdata
)
7830 struct io_rsrc_data
*data
;
7834 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7837 data
->tags
= (u64
**)io_alloc_page_table(nr
* sizeof(data
->tags
[0][0]));
7845 data
->do_put
= do_put
;
7848 for (i
= 0; i
< nr
; i
++) {
7849 u64
*tag_slot
= io_get_tag_slot(data
, i
);
7851 if (copy_from_user(tag_slot
, &utags
[i
],
7857 atomic_set(&data
->refs
, 1);
7858 init_completion(&data
->done
);
7862 io_rsrc_data_free(data
);
7866 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7868 table
->files
= kvcalloc(nr_files
, sizeof(table
->files
[0]),
7869 GFP_KERNEL_ACCOUNT
);
7870 return !!table
->files
;
7873 static void io_free_file_tables(struct io_file_table
*table
)
7875 kvfree(table
->files
);
7876 table
->files
= NULL
;
7879 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7881 #if defined(CONFIG_UNIX)
7882 if (ctx
->ring_sock
) {
7883 struct sock
*sock
= ctx
->ring_sock
->sk
;
7884 struct sk_buff
*skb
;
7886 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7892 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7895 file
= io_file_from_index(ctx
, i
);
7900 io_free_file_tables(&ctx
->file_table
);
7901 io_rsrc_data_free(ctx
->file_data
);
7902 ctx
->file_data
= NULL
;
7903 ctx
->nr_user_files
= 0;
7906 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7910 if (!ctx
->file_data
)
7912 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
7914 __io_sqe_files_unregister(ctx
);
7918 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7919 __releases(&sqd
->lock
)
7921 WARN_ON_ONCE(sqd
->thread
== current
);
7924 * Do the dance but not conditional clear_bit() because it'd race with
7925 * other threads incrementing park_pending and setting the bit.
7927 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7928 if (atomic_dec_return(&sqd
->park_pending
))
7929 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7930 mutex_unlock(&sqd
->lock
);
7933 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7934 __acquires(&sqd
->lock
)
7936 WARN_ON_ONCE(sqd
->thread
== current
);
7938 atomic_inc(&sqd
->park_pending
);
7939 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7940 mutex_lock(&sqd
->lock
);
7942 wake_up_process(sqd
->thread
);
7945 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7947 WARN_ON_ONCE(sqd
->thread
== current
);
7948 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
7950 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7951 mutex_lock(&sqd
->lock
);
7953 wake_up_process(sqd
->thread
);
7954 mutex_unlock(&sqd
->lock
);
7955 wait_for_completion(&sqd
->exited
);
7958 static void io_put_sq_data(struct io_sq_data
*sqd
)
7960 if (refcount_dec_and_test(&sqd
->refs
)) {
7961 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
7963 io_sq_thread_stop(sqd
);
7968 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
7970 struct io_sq_data
*sqd
= ctx
->sq_data
;
7973 io_sq_thread_park(sqd
);
7974 list_del_init(&ctx
->sqd_list
);
7975 io_sqd_update_thread_idle(sqd
);
7976 io_sq_thread_unpark(sqd
);
7978 io_put_sq_data(sqd
);
7979 ctx
->sq_data
= NULL
;
7983 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7985 struct io_ring_ctx
*ctx_attach
;
7986 struct io_sq_data
*sqd
;
7989 f
= fdget(p
->wq_fd
);
7991 return ERR_PTR(-ENXIO
);
7992 if (f
.file
->f_op
!= &io_uring_fops
) {
7994 return ERR_PTR(-EINVAL
);
7997 ctx_attach
= f
.file
->private_data
;
7998 sqd
= ctx_attach
->sq_data
;
8001 return ERR_PTR(-EINVAL
);
8003 if (sqd
->task_tgid
!= current
->tgid
) {
8005 return ERR_PTR(-EPERM
);
8008 refcount_inc(&sqd
->refs
);
8013 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
8016 struct io_sq_data
*sqd
;
8019 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
8020 sqd
= io_attach_sq_data(p
);
8025 /* fall through for EPERM case, setup new sqd/task */
8026 if (PTR_ERR(sqd
) != -EPERM
)
8030 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
8032 return ERR_PTR(-ENOMEM
);
8034 atomic_set(&sqd
->park_pending
, 0);
8035 refcount_set(&sqd
->refs
, 1);
8036 INIT_LIST_HEAD(&sqd
->ctx_list
);
8037 mutex_init(&sqd
->lock
);
8038 init_waitqueue_head(&sqd
->wait
);
8039 init_completion(&sqd
->exited
);
8043 #if defined(CONFIG_UNIX)
8045 * Ensure the UNIX gc is aware of our file set, so we are certain that
8046 * the io_uring can be safely unregistered on process exit, even if we have
8047 * loops in the file referencing.
8049 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
8051 struct sock
*sk
= ctx
->ring_sock
->sk
;
8052 struct scm_fp_list
*fpl
;
8053 struct sk_buff
*skb
;
8056 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
8060 skb
= alloc_skb(0, GFP_KERNEL
);
8069 fpl
->user
= get_uid(current_user());
8070 for (i
= 0; i
< nr
; i
++) {
8071 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8075 fpl
->fp
[nr_files
] = get_file(file
);
8076 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
8081 fpl
->max
= SCM_MAX_FD
;
8082 fpl
->count
= nr_files
;
8083 UNIXCB(skb
).fp
= fpl
;
8084 skb
->destructor
= unix_destruct_scm
;
8085 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
8086 skb_queue_head(&sk
->sk_receive_queue
, skb
);
8088 for (i
= 0; i
< nr_files
; i
++)
8099 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8100 * causes regular reference counting to break down. We rely on the UNIX
8101 * garbage collection to take care of this problem for us.
8103 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8105 unsigned left
, total
;
8109 left
= ctx
->nr_user_files
;
8111 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
8113 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
8117 total
+= this_files
;
8123 while (total
< ctx
->nr_user_files
) {
8124 struct file
*file
= io_file_from_index(ctx
, total
);
8134 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8140 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8142 struct file
*file
= prsrc
->file
;
8143 #if defined(CONFIG_UNIX)
8144 struct sock
*sock
= ctx
->ring_sock
->sk
;
8145 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
8146 struct sk_buff
*skb
;
8149 __skb_queue_head_init(&list
);
8152 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8153 * remove this entry and rearrange the file array.
8155 skb
= skb_dequeue(head
);
8157 struct scm_fp_list
*fp
;
8159 fp
= UNIXCB(skb
).fp
;
8160 for (i
= 0; i
< fp
->count
; i
++) {
8163 if (fp
->fp
[i
] != file
)
8166 unix_notinflight(fp
->user
, fp
->fp
[i
]);
8167 left
= fp
->count
- 1 - i
;
8169 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
8170 left
* sizeof(struct file
*));
8177 __skb_queue_tail(&list
, skb
);
8187 __skb_queue_tail(&list
, skb
);
8189 skb
= skb_dequeue(head
);
8192 if (skb_peek(&list
)) {
8193 spin_lock_irq(&head
->lock
);
8194 while ((skb
= __skb_dequeue(&list
)) != NULL
)
8195 __skb_queue_tail(head
, skb
);
8196 spin_unlock_irq(&head
->lock
);
8203 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
8205 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
8206 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
8207 struct io_rsrc_put
*prsrc
, *tmp
;
8209 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
8210 list_del(&prsrc
->list
);
8213 bool lock_ring
= ctx
->flags
& IORING_SETUP_IOPOLL
;
8215 io_ring_submit_lock(ctx
, lock_ring
);
8216 spin_lock(&ctx
->completion_lock
);
8217 io_cqring_fill_event(ctx
, prsrc
->tag
, 0, 0);
8219 io_commit_cqring(ctx
);
8220 spin_unlock(&ctx
->completion_lock
);
8221 io_cqring_ev_posted(ctx
);
8222 io_ring_submit_unlock(ctx
, lock_ring
);
8225 rsrc_data
->do_put(ctx
, prsrc
);
8229 io_rsrc_node_destroy(ref_node
);
8230 if (atomic_dec_and_test(&rsrc_data
->refs
))
8231 complete(&rsrc_data
->done
);
8234 static void io_rsrc_put_work(struct work_struct
*work
)
8236 struct io_ring_ctx
*ctx
;
8237 struct llist_node
*node
;
8239 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
8240 node
= llist_del_all(&ctx
->rsrc_put_llist
);
8243 struct io_rsrc_node
*ref_node
;
8244 struct llist_node
*next
= node
->next
;
8246 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
8247 __io_rsrc_put_work(ref_node
);
8252 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8253 unsigned nr_args
, u64 __user
*tags
)
8255 __s32 __user
*fds
= (__s32 __user
*) arg
;
8264 if (nr_args
> IORING_MAX_FIXED_FILES
)
8266 if (nr_args
> rlimit(RLIMIT_NOFILE
))
8268 ret
= io_rsrc_node_switch_start(ctx
);
8271 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, tags
, nr_args
,
8277 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
))
8280 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
8281 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
8285 /* allow sparse sets */
8288 if (unlikely(*io_get_tag_slot(ctx
->file_data
, i
)))
8295 if (unlikely(!file
))
8299 * Don't allow io_uring instances to be registered. If UNIX
8300 * isn't enabled, then this causes a reference cycle and this
8301 * instance can never get freed. If UNIX is enabled we'll
8302 * handle it just fine, but there's still no point in allowing
8303 * a ring fd as it doesn't support regular read/write anyway.
8305 if (file
->f_op
== &io_uring_fops
) {
8309 io_fixed_file_set(io_fixed_file_slot(&ctx
->file_table
, i
), file
);
8312 ret
= io_sqe_files_scm(ctx
);
8314 __io_sqe_files_unregister(ctx
);
8318 io_rsrc_node_switch(ctx
, NULL
);
8321 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
8322 file
= io_file_from_index(ctx
, i
);
8326 io_free_file_tables(&ctx
->file_table
);
8327 ctx
->nr_user_files
= 0;
8329 io_rsrc_data_free(ctx
->file_data
);
8330 ctx
->file_data
= NULL
;
8334 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
8337 #if defined(CONFIG_UNIX)
8338 struct sock
*sock
= ctx
->ring_sock
->sk
;
8339 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
8340 struct sk_buff
*skb
;
8343 * See if we can merge this file into an existing skb SCM_RIGHTS
8344 * file set. If there's no room, fall back to allocating a new skb
8345 * and filling it in.
8347 spin_lock_irq(&head
->lock
);
8348 skb
= skb_peek(head
);
8350 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
8352 if (fpl
->count
< SCM_MAX_FD
) {
8353 __skb_unlink(skb
, head
);
8354 spin_unlock_irq(&head
->lock
);
8355 fpl
->fp
[fpl
->count
] = get_file(file
);
8356 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
8358 spin_lock_irq(&head
->lock
);
8359 __skb_queue_head(head
, skb
);
8364 spin_unlock_irq(&head
->lock
);
8371 return __io_sqe_files_scm(ctx
, 1, index
);
8377 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
8378 struct io_rsrc_node
*node
, void *rsrc
)
8380 struct io_rsrc_put
*prsrc
;
8382 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
8386 prsrc
->tag
= *io_get_tag_slot(data
, idx
);
8388 list_add(&prsrc
->list
, &node
->rsrc_list
);
8392 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
8393 unsigned int issue_flags
, u32 slot_index
)
8395 struct io_ring_ctx
*ctx
= req
->ctx
;
8396 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
8397 bool needs_switch
= false;
8398 struct io_fixed_file
*file_slot
;
8401 io_ring_submit_lock(ctx
, !force_nonblock
);
8402 if (file
->f_op
== &io_uring_fops
)
8405 if (!ctx
->file_data
)
8408 if (slot_index
>= ctx
->nr_user_files
)
8411 slot_index
= array_index_nospec(slot_index
, ctx
->nr_user_files
);
8412 file_slot
= io_fixed_file_slot(&ctx
->file_table
, slot_index
);
8414 if (file_slot
->file_ptr
) {
8415 struct file
*old_file
;
8417 ret
= io_rsrc_node_switch_start(ctx
);
8421 old_file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8422 ret
= io_queue_rsrc_removal(ctx
->file_data
, slot_index
,
8423 ctx
->rsrc_node
, old_file
);
8426 file_slot
->file_ptr
= 0;
8427 needs_switch
= true;
8430 *io_get_tag_slot(ctx
->file_data
, slot_index
) = 0;
8431 io_fixed_file_set(file_slot
, file
);
8432 ret
= io_sqe_file_register(ctx
, file
, slot_index
);
8434 file_slot
->file_ptr
= 0;
8441 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8442 io_ring_submit_unlock(ctx
, !force_nonblock
);
8448 static int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
)
8450 unsigned int offset
= req
->close
.file_slot
- 1;
8451 struct io_ring_ctx
*ctx
= req
->ctx
;
8452 struct io_fixed_file
*file_slot
;
8456 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8458 if (unlikely(!ctx
->file_data
))
8461 if (offset
>= ctx
->nr_user_files
)
8463 ret
= io_rsrc_node_switch_start(ctx
);
8467 i
= array_index_nospec(offset
, ctx
->nr_user_files
);
8468 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
8470 if (!file_slot
->file_ptr
)
8473 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8474 ret
= io_queue_rsrc_removal(ctx
->file_data
, offset
, ctx
->rsrc_node
, file
);
8478 file_slot
->file_ptr
= 0;
8479 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8482 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8486 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
8487 struct io_uring_rsrc_update2
*up
,
8490 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
8491 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
8492 struct io_rsrc_data
*data
= ctx
->file_data
;
8493 struct io_fixed_file
*file_slot
;
8497 bool needs_switch
= false;
8499 if (!ctx
->file_data
)
8501 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
8504 for (done
= 0; done
< nr_args
; done
++) {
8507 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
8508 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8512 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
8516 if (fd
== IORING_REGISTER_FILES_SKIP
)
8519 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
8520 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
8522 if (file_slot
->file_ptr
) {
8523 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8524 err
= io_queue_rsrc_removal(data
, up
->offset
+ done
,
8525 ctx
->rsrc_node
, file
);
8528 file_slot
->file_ptr
= 0;
8529 needs_switch
= true;
8538 * Don't allow io_uring instances to be registered. If
8539 * UNIX isn't enabled, then this causes a reference
8540 * cycle and this instance can never get freed. If UNIX
8541 * is enabled we'll handle it just fine, but there's
8542 * still no point in allowing a ring fd as it doesn't
8543 * support regular read/write anyway.
8545 if (file
->f_op
== &io_uring_fops
) {
8550 *io_get_tag_slot(data
, up
->offset
+ done
) = tag
;
8551 io_fixed_file_set(file_slot
, file
);
8552 err
= io_sqe_file_register(ctx
, file
, i
);
8554 file_slot
->file_ptr
= 0;
8562 io_rsrc_node_switch(ctx
, data
);
8563 return done
? done
: err
;
8566 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
8567 struct task_struct
*task
)
8569 struct io_wq_hash
*hash
;
8570 struct io_wq_data data
;
8571 unsigned int concurrency
;
8573 mutex_lock(&ctx
->uring_lock
);
8574 hash
= ctx
->hash_map
;
8576 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
8578 mutex_unlock(&ctx
->uring_lock
);
8579 return ERR_PTR(-ENOMEM
);
8581 refcount_set(&hash
->refs
, 1);
8582 init_waitqueue_head(&hash
->wait
);
8583 ctx
->hash_map
= hash
;
8585 mutex_unlock(&ctx
->uring_lock
);
8589 data
.free_work
= io_wq_free_work
;
8590 data
.do_work
= io_wq_submit_work
;
8592 /* Do QD, or 4 * CPUS, whatever is smallest */
8593 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8595 return io_wq_create(concurrency
, &data
);
8598 static int io_uring_alloc_task_context(struct task_struct
*task
,
8599 struct io_ring_ctx
*ctx
)
8601 struct io_uring_task
*tctx
;
8604 tctx
= kzalloc(sizeof(*tctx
), GFP_KERNEL
);
8605 if (unlikely(!tctx
))
8608 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8609 if (unlikely(ret
)) {
8614 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
8615 if (IS_ERR(tctx
->io_wq
)) {
8616 ret
= PTR_ERR(tctx
->io_wq
);
8617 percpu_counter_destroy(&tctx
->inflight
);
8623 init_waitqueue_head(&tctx
->wait
);
8624 atomic_set(&tctx
->in_idle
, 0);
8625 atomic_set(&tctx
->inflight_tracked
, 0);
8626 task
->io_uring
= tctx
;
8627 spin_lock_init(&tctx
->task_lock
);
8628 INIT_WQ_LIST(&tctx
->task_list
);
8629 init_task_work(&tctx
->task_work
, tctx_task_work
);
8633 void __io_uring_free(struct task_struct
*tsk
)
8635 struct io_uring_task
*tctx
= tsk
->io_uring
;
8637 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8638 WARN_ON_ONCE(tctx
->io_wq
);
8639 WARN_ON_ONCE(tctx
->cached_refs
);
8641 percpu_counter_destroy(&tctx
->inflight
);
8643 tsk
->io_uring
= NULL
;
8646 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8647 struct io_uring_params
*p
)
8651 /* Retain compatibility with failing for an invalid attach attempt */
8652 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
8653 IORING_SETUP_ATTACH_WQ
) {
8656 f
= fdget(p
->wq_fd
);
8659 if (f
.file
->f_op
!= &io_uring_fops
) {
8665 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8666 struct task_struct
*tsk
;
8667 struct io_sq_data
*sqd
;
8670 sqd
= io_get_sq_data(p
, &attached
);
8676 ctx
->sq_creds
= get_current_cred();
8678 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8679 if (!ctx
->sq_thread_idle
)
8680 ctx
->sq_thread_idle
= HZ
;
8682 io_sq_thread_park(sqd
);
8683 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
8684 io_sqd_update_thread_idle(sqd
);
8685 /* don't attach to a dying SQPOLL thread, would be racy */
8686 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
8687 io_sq_thread_unpark(sqd
);
8694 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8695 int cpu
= p
->sq_thread_cpu
;
8698 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
8705 sqd
->task_pid
= current
->pid
;
8706 sqd
->task_tgid
= current
->tgid
;
8707 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
8714 ret
= io_uring_alloc_task_context(tsk
, ctx
);
8715 wake_up_new_task(tsk
);
8718 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8719 /* Can't have SQ_AFF without SQPOLL */
8726 complete(&ctx
->sq_data
->exited
);
8728 io_sq_thread_finish(ctx
);
8732 static inline void __io_unaccount_mem(struct user_struct
*user
,
8733 unsigned long nr_pages
)
8735 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8738 static inline int __io_account_mem(struct user_struct
*user
,
8739 unsigned long nr_pages
)
8741 unsigned long page_limit
, cur_pages
, new_pages
;
8743 /* Don't allow more pages than we can safely lock */
8744 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8747 cur_pages
= atomic_long_read(&user
->locked_vm
);
8748 new_pages
= cur_pages
+ nr_pages
;
8749 if (new_pages
> page_limit
)
8751 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8752 new_pages
) != cur_pages
);
8757 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8760 __io_unaccount_mem(ctx
->user
, nr_pages
);
8762 if (ctx
->mm_account
)
8763 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8766 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8771 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8776 if (ctx
->mm_account
)
8777 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8782 static void io_mem_free(void *ptr
)
8789 page
= virt_to_head_page(ptr
);
8790 if (put_page_testzero(page
))
8791 free_compound_page(page
);
8794 static void *io_mem_alloc(size_t size
)
8796 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
8797 __GFP_NORETRY
| __GFP_ACCOUNT
;
8799 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
8802 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8805 struct io_rings
*rings
;
8806 size_t off
, sq_array_size
;
8808 off
= struct_size(rings
, cqes
, cq_entries
);
8809 if (off
== SIZE_MAX
)
8813 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8821 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8822 if (sq_array_size
== SIZE_MAX
)
8825 if (check_add_overflow(off
, sq_array_size
, &off
))
8831 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
8833 struct io_mapped_ubuf
*imu
= *slot
;
8836 if (imu
!= ctx
->dummy_ubuf
) {
8837 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
8838 unpin_user_page(imu
->bvec
[i
].bv_page
);
8839 if (imu
->acct_pages
)
8840 io_unaccount_mem(ctx
, imu
->acct_pages
);
8846 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8848 io_buffer_unmap(ctx
, &prsrc
->buf
);
8852 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8856 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
8857 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
8858 kfree(ctx
->user_bufs
);
8859 io_rsrc_data_free(ctx
->buf_data
);
8860 ctx
->user_bufs
= NULL
;
8861 ctx
->buf_data
= NULL
;
8862 ctx
->nr_user_bufs
= 0;
8865 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8872 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
8874 __io_sqe_buffers_unregister(ctx
);
8878 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8879 void __user
*arg
, unsigned index
)
8881 struct iovec __user
*src
;
8883 #ifdef CONFIG_COMPAT
8885 struct compat_iovec __user
*ciovs
;
8886 struct compat_iovec ciov
;
8888 ciovs
= (struct compat_iovec __user
*) arg
;
8889 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8892 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8893 dst
->iov_len
= ciov
.iov_len
;
8897 src
= (struct iovec __user
*) arg
;
8898 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8904 * Not super efficient, but this is just a registration time. And we do cache
8905 * the last compound head, so generally we'll only do a full search if we don't
8908 * We check if the given compound head page has already been accounted, to
8909 * avoid double accounting it. This allows us to account the full size of the
8910 * page, not just the constituent pages of a huge page.
8912 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8913 int nr_pages
, struct page
*hpage
)
8917 /* check current page array */
8918 for (i
= 0; i
< nr_pages
; i
++) {
8919 if (!PageCompound(pages
[i
]))
8921 if (compound_head(pages
[i
]) == hpage
)
8925 /* check previously registered pages */
8926 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8927 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
8929 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8930 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8932 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8940 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8941 int nr_pages
, struct io_mapped_ubuf
*imu
,
8942 struct page
**last_hpage
)
8946 imu
->acct_pages
= 0;
8947 for (i
= 0; i
< nr_pages
; i
++) {
8948 if (!PageCompound(pages
[i
])) {
8953 hpage
= compound_head(pages
[i
]);
8954 if (hpage
== *last_hpage
)
8956 *last_hpage
= hpage
;
8957 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8959 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8963 if (!imu
->acct_pages
)
8966 ret
= io_account_mem(ctx
, imu
->acct_pages
);
8968 imu
->acct_pages
= 0;
8972 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
8973 struct io_mapped_ubuf
**pimu
,
8974 struct page
**last_hpage
)
8976 struct io_mapped_ubuf
*imu
= NULL
;
8977 struct vm_area_struct
**vmas
= NULL
;
8978 struct page
**pages
= NULL
;
8979 unsigned long off
, start
, end
, ubuf
;
8981 int ret
, pret
, nr_pages
, i
;
8983 if (!iov
->iov_base
) {
8984 *pimu
= ctx
->dummy_ubuf
;
8988 ubuf
= (unsigned long) iov
->iov_base
;
8989 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8990 start
= ubuf
>> PAGE_SHIFT
;
8991 nr_pages
= end
- start
;
8996 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
9000 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
9005 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
9010 mmap_read_lock(current
->mm
);
9011 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
9013 if (pret
== nr_pages
) {
9014 /* don't support file backed memory */
9015 for (i
= 0; i
< nr_pages
; i
++) {
9016 struct vm_area_struct
*vma
= vmas
[i
];
9018 if (vma_is_shmem(vma
))
9021 !is_file_hugepages(vma
->vm_file
)) {
9027 ret
= pret
< 0 ? pret
: -EFAULT
;
9029 mmap_read_unlock(current
->mm
);
9032 * if we did partial map, or found file backed vmas,
9033 * release any pages we did get
9036 unpin_user_pages(pages
, pret
);
9040 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
9042 unpin_user_pages(pages
, pret
);
9046 off
= ubuf
& ~PAGE_MASK
;
9047 size
= iov
->iov_len
;
9048 for (i
= 0; i
< nr_pages
; i
++) {
9051 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
9052 imu
->bvec
[i
].bv_page
= pages
[i
];
9053 imu
->bvec
[i
].bv_len
= vec_len
;
9054 imu
->bvec
[i
].bv_offset
= off
;
9058 /* store original address for later verification */
9060 imu
->ubuf_end
= ubuf
+ iov
->iov_len
;
9061 imu
->nr_bvecs
= nr_pages
;
9072 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
9074 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
9075 return ctx
->user_bufs
? 0 : -ENOMEM
;
9078 static int io_buffer_validate(struct iovec
*iov
)
9080 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
9083 * Don't impose further limits on the size and buffer
9084 * constraints here, we'll -EINVAL later when IO is
9085 * submitted if they are wrong.
9088 return iov
->iov_len
? -EFAULT
: 0;
9092 /* arbitrary limit, but we need something */
9093 if (iov
->iov_len
> SZ_1G
)
9096 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
9102 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
9103 unsigned int nr_args
, u64 __user
*tags
)
9105 struct page
*last_hpage
= NULL
;
9106 struct io_rsrc_data
*data
;
9112 if (!nr_args
|| nr_args
> IORING_MAX_REG_BUFFERS
)
9114 ret
= io_rsrc_node_switch_start(ctx
);
9117 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, tags
, nr_args
, &data
);
9120 ret
= io_buffers_map_alloc(ctx
, nr_args
);
9122 io_rsrc_data_free(data
);
9126 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
9127 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
9130 ret
= io_buffer_validate(&iov
);
9133 if (!iov
.iov_base
&& *io_get_tag_slot(data
, i
)) {
9138 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
9144 WARN_ON_ONCE(ctx
->buf_data
);
9146 ctx
->buf_data
= data
;
9148 __io_sqe_buffers_unregister(ctx
);
9150 io_rsrc_node_switch(ctx
, NULL
);
9154 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
9155 struct io_uring_rsrc_update2
*up
,
9156 unsigned int nr_args
)
9158 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
9159 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
9160 struct page
*last_hpage
= NULL
;
9161 bool needs_switch
= false;
9167 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
9170 for (done
= 0; done
< nr_args
; done
++) {
9171 struct io_mapped_ubuf
*imu
;
9172 int offset
= up
->offset
+ done
;
9175 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
9178 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
9182 err
= io_buffer_validate(&iov
);
9185 if (!iov
.iov_base
&& tag
) {
9189 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
9193 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
9194 if (ctx
->user_bufs
[i
] != ctx
->dummy_ubuf
) {
9195 err
= io_queue_rsrc_removal(ctx
->buf_data
, offset
,
9196 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
9197 if (unlikely(err
)) {
9198 io_buffer_unmap(ctx
, &imu
);
9201 ctx
->user_bufs
[i
] = NULL
;
9202 needs_switch
= true;
9205 ctx
->user_bufs
[i
] = imu
;
9206 *io_get_tag_slot(ctx
->buf_data
, offset
) = tag
;
9210 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
9211 return done
? done
: err
;
9214 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
9216 __s32 __user
*fds
= arg
;
9222 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
9225 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
9226 if (IS_ERR(ctx
->cq_ev_fd
)) {
9227 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
9229 ctx
->cq_ev_fd
= NULL
;
9236 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
9238 if (ctx
->cq_ev_fd
) {
9239 eventfd_ctx_put(ctx
->cq_ev_fd
);
9240 ctx
->cq_ev_fd
= NULL
;
9247 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
9249 struct io_buffer
*buf
;
9250 unsigned long index
;
9252 xa_for_each(&ctx
->io_buffers
, index
, buf
) {
9253 __io_remove_buffers(ctx
, buf
, index
, -1U);
9258 static void io_req_cache_free(struct list_head
*list
)
9260 struct io_kiocb
*req
, *nxt
;
9262 list_for_each_entry_safe(req
, nxt
, list
, inflight_entry
) {
9263 list_del(&req
->inflight_entry
);
9264 kmem_cache_free(req_cachep
, req
);
9268 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
9270 struct io_submit_state
*state
= &ctx
->submit_state
;
9272 mutex_lock(&ctx
->uring_lock
);
9274 if (state
->free_reqs
) {
9275 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
9276 state
->free_reqs
= 0;
9279 io_flush_cached_locked_reqs(ctx
, state
);
9280 io_req_cache_free(&state
->free_list
);
9281 mutex_unlock(&ctx
->uring_lock
);
9284 static void io_wait_rsrc_data(struct io_rsrc_data
*data
)
9286 if (data
&& !atomic_dec_and_test(&data
->refs
))
9287 wait_for_completion(&data
->done
);
9290 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
9292 io_sq_thread_finish(ctx
);
9294 if (ctx
->mm_account
) {
9295 mmdrop(ctx
->mm_account
);
9296 ctx
->mm_account
= NULL
;
9299 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9300 io_wait_rsrc_data(ctx
->buf_data
);
9301 io_wait_rsrc_data(ctx
->file_data
);
9303 mutex_lock(&ctx
->uring_lock
);
9305 __io_sqe_buffers_unregister(ctx
);
9307 __io_sqe_files_unregister(ctx
);
9309 __io_cqring_overflow_flush(ctx
, true);
9310 mutex_unlock(&ctx
->uring_lock
);
9311 io_eventfd_unregister(ctx
);
9312 io_destroy_buffers(ctx
);
9314 put_cred(ctx
->sq_creds
);
9316 /* there are no registered resources left, nobody uses it */
9318 io_rsrc_node_destroy(ctx
->rsrc_node
);
9319 if (ctx
->rsrc_backup_node
)
9320 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
9321 flush_delayed_work(&ctx
->rsrc_put_work
);
9323 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
9324 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
9326 #if defined(CONFIG_UNIX)
9327 if (ctx
->ring_sock
) {
9328 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
9329 sock_release(ctx
->ring_sock
);
9332 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
9334 io_mem_free(ctx
->rings
);
9335 io_mem_free(ctx
->sq_sqes
);
9337 percpu_ref_exit(&ctx
->refs
);
9338 free_uid(ctx
->user
);
9339 io_req_caches_free(ctx
);
9341 io_wq_put_hash(ctx
->hash_map
);
9342 kfree(ctx
->cancel_hash
);
9343 kfree(ctx
->dummy_ubuf
);
9347 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
9349 struct io_ring_ctx
*ctx
= file
->private_data
;
9352 poll_wait(file
, &ctx
->poll_wait
, wait
);
9354 * synchronizes with barrier from wq_has_sleeper call in
9358 if (!io_sqring_full(ctx
))
9359 mask
|= EPOLLOUT
| EPOLLWRNORM
;
9362 * Don't flush cqring overflow list here, just do a simple check.
9363 * Otherwise there could possible be ABBA deadlock:
9366 * lock(&ctx->uring_lock);
9368 * lock(&ctx->uring_lock);
9371 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9372 * pushs them to do the flush.
9374 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->check_cq_overflow
))
9375 mask
|= EPOLLIN
| EPOLLRDNORM
;
9380 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9382 const struct cred
*creds
;
9384 creds
= xa_erase(&ctx
->personalities
, id
);
9393 struct io_tctx_exit
{
9394 struct callback_head task_work
;
9395 struct completion completion
;
9396 struct io_ring_ctx
*ctx
;
9399 static void io_tctx_exit_cb(struct callback_head
*cb
)
9401 struct io_uring_task
*tctx
= current
->io_uring
;
9402 struct io_tctx_exit
*work
;
9404 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
9406 * When @in_idle, we're in cancellation and it's racy to remove the
9407 * node. It'll be removed by the end of cancellation, just ignore it.
9409 if (!atomic_read(&tctx
->in_idle
))
9410 io_uring_del_tctx_node((unsigned long)work
->ctx
);
9411 complete(&work
->completion
);
9414 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
9416 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9418 return req
->ctx
== data
;
9421 static void io_ring_exit_work(struct work_struct
*work
)
9423 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
9424 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
9425 unsigned long interval
= HZ
/ 20;
9426 struct io_tctx_exit exit
;
9427 struct io_tctx_node
*node
;
9431 * If we're doing polled IO and end up having requests being
9432 * submitted async (out-of-line), then completions can come in while
9433 * we're waiting for refs to drop. We need to reap these manually,
9434 * as nobody else will be looking for them.
9437 io_uring_try_cancel_requests(ctx
, NULL
, true);
9439 struct io_sq_data
*sqd
= ctx
->sq_data
;
9440 struct task_struct
*tsk
;
9442 io_sq_thread_park(sqd
);
9444 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
9445 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
9446 io_cancel_ctx_cb
, ctx
, true);
9447 io_sq_thread_unpark(sqd
);
9450 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
9451 /* there is little hope left, don't run it too often */
9454 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
9456 init_completion(&exit
.completion
);
9457 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
9460 * Some may use context even when all refs and requests have been put,
9461 * and they are free to do so while still holding uring_lock or
9462 * completion_lock, see io_req_task_submit(). Apart from other work,
9463 * this lock/unlock section also waits them to finish.
9465 mutex_lock(&ctx
->uring_lock
);
9466 while (!list_empty(&ctx
->tctx_list
)) {
9467 WARN_ON_ONCE(time_after(jiffies
, timeout
));
9469 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
9471 /* don't spin on a single task if cancellation failed */
9472 list_rotate_left(&ctx
->tctx_list
);
9473 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
9474 if (WARN_ON_ONCE(ret
))
9476 wake_up_process(node
->task
);
9478 mutex_unlock(&ctx
->uring_lock
);
9479 wait_for_completion(&exit
.completion
);
9480 mutex_lock(&ctx
->uring_lock
);
9482 mutex_unlock(&ctx
->uring_lock
);
9483 spin_lock(&ctx
->completion_lock
);
9484 spin_unlock(&ctx
->completion_lock
);
9486 io_ring_ctx_free(ctx
);
9489 /* Returns true if we found and killed one or more timeouts */
9490 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
9493 struct io_kiocb
*req
, *tmp
;
9496 spin_lock(&ctx
->completion_lock
);
9497 spin_lock_irq(&ctx
->timeout_lock
);
9498 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
9499 if (io_match_task(req
, tsk
, cancel_all
)) {
9500 io_kill_timeout(req
, -ECANCELED
);
9504 spin_unlock_irq(&ctx
->timeout_lock
);
9506 io_commit_cqring(ctx
);
9507 spin_unlock(&ctx
->completion_lock
);
9509 io_cqring_ev_posted(ctx
);
9510 return canceled
!= 0;
9513 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
9515 unsigned long index
;
9516 struct creds
*creds
;
9518 mutex_lock(&ctx
->uring_lock
);
9519 percpu_ref_kill(&ctx
->refs
);
9521 __io_cqring_overflow_flush(ctx
, true);
9522 xa_for_each(&ctx
->personalities
, index
, creds
)
9523 io_unregister_personality(ctx
, index
);
9524 mutex_unlock(&ctx
->uring_lock
);
9526 io_kill_timeouts(ctx
, NULL
, true);
9527 io_poll_remove_all(ctx
, NULL
, true);
9529 /* if we failed setting up the ctx, we might not have any rings */
9530 io_iopoll_try_reap_events(ctx
);
9532 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
9534 * Use system_unbound_wq to avoid spawning tons of event kworkers
9535 * if we're exiting a ton of rings at the same time. It just adds
9536 * noise and overhead, there's no discernable change in runtime
9537 * over using system_wq.
9539 queue_work(system_unbound_wq
, &ctx
->exit_work
);
9542 static int io_uring_release(struct inode
*inode
, struct file
*file
)
9544 struct io_ring_ctx
*ctx
= file
->private_data
;
9546 file
->private_data
= NULL
;
9547 io_ring_ctx_wait_and_kill(ctx
);
9551 struct io_task_cancel
{
9552 struct task_struct
*task
;
9556 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
9558 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9559 struct io_task_cancel
*cancel
= data
;
9561 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
9564 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
9565 struct task_struct
*task
, bool cancel_all
)
9567 struct io_defer_entry
*de
;
9570 spin_lock(&ctx
->completion_lock
);
9571 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
9572 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
9573 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
9577 spin_unlock(&ctx
->completion_lock
);
9578 if (list_empty(&list
))
9581 while (!list_empty(&list
)) {
9582 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
9583 list_del_init(&de
->list
);
9584 io_req_complete_failed(de
->req
, -ECANCELED
);
9590 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
9592 struct io_tctx_node
*node
;
9593 enum io_wq_cancel cret
;
9596 mutex_lock(&ctx
->uring_lock
);
9597 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
9598 struct io_uring_task
*tctx
= node
->task
->io_uring
;
9601 * io_wq will stay alive while we hold uring_lock, because it's
9602 * killed after ctx nodes, which requires to take the lock.
9604 if (!tctx
|| !tctx
->io_wq
)
9606 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
9607 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9609 mutex_unlock(&ctx
->uring_lock
);
9614 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
9615 struct task_struct
*task
,
9618 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
9619 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
9622 enum io_wq_cancel cret
;
9626 ret
|= io_uring_try_cancel_iowq(ctx
);
9627 } else if (tctx
&& tctx
->io_wq
) {
9629 * Cancels requests of all rings, not only @ctx, but
9630 * it's fine as the task is in exit/exec.
9632 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
9634 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9637 /* SQPOLL thread does its own polling */
9638 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
9639 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
9640 while (!list_empty_careful(&ctx
->iopoll_list
)) {
9641 io_iopoll_try_reap_events(ctx
);
9646 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
9647 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
9648 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
9650 ret
|= io_run_task_work();
9657 static int __io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9659 struct io_uring_task
*tctx
= current
->io_uring
;
9660 struct io_tctx_node
*node
;
9663 if (unlikely(!tctx
)) {
9664 ret
= io_uring_alloc_task_context(current
, ctx
);
9668 tctx
= current
->io_uring
;
9669 if (ctx
->iowq_limits_set
) {
9670 unsigned int limits
[2] = { ctx
->iowq_limits
[0],
9671 ctx
->iowq_limits
[1], };
9673 ret
= io_wq_max_workers(tctx
->io_wq
, limits
);
9678 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
9679 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
9683 node
->task
= current
;
9685 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
9692 mutex_lock(&ctx
->uring_lock
);
9693 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
9694 mutex_unlock(&ctx
->uring_lock
);
9701 * Note that this task has used io_uring. We use it for cancelation purposes.
9703 static inline int io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9705 struct io_uring_task
*tctx
= current
->io_uring
;
9707 if (likely(tctx
&& tctx
->last
== ctx
))
9709 return __io_uring_add_tctx_node(ctx
);
9713 * Remove this io_uring_file -> task mapping.
9715 static void io_uring_del_tctx_node(unsigned long index
)
9717 struct io_uring_task
*tctx
= current
->io_uring
;
9718 struct io_tctx_node
*node
;
9722 node
= xa_erase(&tctx
->xa
, index
);
9726 WARN_ON_ONCE(current
!= node
->task
);
9727 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
9729 mutex_lock(&node
->ctx
->uring_lock
);
9730 list_del(&node
->ctx_node
);
9731 mutex_unlock(&node
->ctx
->uring_lock
);
9733 if (tctx
->last
== node
->ctx
)
9738 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
9740 struct io_wq
*wq
= tctx
->io_wq
;
9741 struct io_tctx_node
*node
;
9742 unsigned long index
;
9744 xa_for_each(&tctx
->xa
, index
, node
) {
9745 io_uring_del_tctx_node(index
);
9750 * Must be after io_uring_del_task_file() (removes nodes under
9751 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9753 io_wq_put_and_exit(wq
);
9758 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
9761 return atomic_read(&tctx
->inflight_tracked
);
9762 return percpu_counter_sum(&tctx
->inflight
);
9765 static void io_uring_drop_tctx_refs(struct task_struct
*task
)
9767 struct io_uring_task
*tctx
= task
->io_uring
;
9768 unsigned int refs
= tctx
->cached_refs
;
9771 tctx
->cached_refs
= 0;
9772 percpu_counter_sub(&tctx
->inflight
, refs
);
9773 put_task_struct_many(task
, refs
);
9778 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9779 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9781 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
9783 struct io_uring_task
*tctx
= current
->io_uring
;
9784 struct io_ring_ctx
*ctx
;
9788 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
9790 if (!current
->io_uring
)
9793 io_wq_exit_start(tctx
->io_wq
);
9795 atomic_inc(&tctx
->in_idle
);
9797 io_uring_drop_tctx_refs(current
);
9798 /* read completions before cancelations */
9799 inflight
= tctx_inflight(tctx
, !cancel_all
);
9804 struct io_tctx_node
*node
;
9805 unsigned long index
;
9807 xa_for_each(&tctx
->xa
, index
, node
) {
9808 /* sqpoll task will cancel all its requests */
9809 if (node
->ctx
->sq_data
)
9811 io_uring_try_cancel_requests(node
->ctx
, current
,
9815 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9816 io_uring_try_cancel_requests(ctx
, current
,
9820 prepare_to_wait(&tctx
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
9821 io_uring_drop_tctx_refs(current
);
9823 * If we've seen completions, retry without waiting. This
9824 * avoids a race where a completion comes in before we did
9825 * prepare_to_wait().
9827 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
9829 finish_wait(&tctx
->wait
, &wait
);
9831 atomic_dec(&tctx
->in_idle
);
9833 io_uring_clean_tctx(tctx
);
9835 /* for exec all current's requests should be gone, kill tctx */
9836 __io_uring_free(current
);
9840 void __io_uring_cancel(bool cancel_all
)
9842 io_uring_cancel_generic(cancel_all
, NULL
);
9845 static void *io_uring_validate_mmap_request(struct file
*file
,
9846 loff_t pgoff
, size_t sz
)
9848 struct io_ring_ctx
*ctx
= file
->private_data
;
9849 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9854 case IORING_OFF_SQ_RING
:
9855 case IORING_OFF_CQ_RING
:
9858 case IORING_OFF_SQES
:
9862 return ERR_PTR(-EINVAL
);
9865 page
= virt_to_head_page(ptr
);
9866 if (sz
> page_size(page
))
9867 return ERR_PTR(-EINVAL
);
9874 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9876 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9880 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9882 return PTR_ERR(ptr
);
9884 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9885 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9888 #else /* !CONFIG_MMU */
9890 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9892 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9895 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9897 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9900 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9901 unsigned long addr
, unsigned long len
,
9902 unsigned long pgoff
, unsigned long flags
)
9906 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9908 return PTR_ERR(ptr
);
9910 return (unsigned long) ptr
;
9913 #endif /* !CONFIG_MMU */
9915 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9920 if (!io_sqring_full(ctx
))
9922 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9924 if (!io_sqring_full(ctx
))
9927 } while (!signal_pending(current
));
9929 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9933 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9934 struct __kernel_timespec __user
**ts
,
9935 const sigset_t __user
**sig
)
9937 struct io_uring_getevents_arg arg
;
9940 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9941 * is just a pointer to the sigset_t.
9943 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9944 *sig
= (const sigset_t __user
*) argp
;
9950 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9951 * timespec and sigset_t pointers if good.
9953 if (*argsz
!= sizeof(arg
))
9955 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9957 *sig
= u64_to_user_ptr(arg
.sigmask
);
9958 *argsz
= arg
.sigmask_sz
;
9959 *ts
= u64_to_user_ptr(arg
.ts
);
9963 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9964 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9967 struct io_ring_ctx
*ctx
;
9974 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9975 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
)))
9979 if (unlikely(!f
.file
))
9983 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
9987 ctx
= f
.file
->private_data
;
9988 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
9992 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9996 * For SQ polling, the thread will do all submissions and completions.
9997 * Just return the requested submit count, and wake the thread if
10001 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10002 io_cqring_overflow_flush(ctx
);
10004 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
10008 if (flags
& IORING_ENTER_SQ_WAKEUP
)
10009 wake_up(&ctx
->sq_data
->wait
);
10010 if (flags
& IORING_ENTER_SQ_WAIT
) {
10011 ret
= io_sqpoll_wait_sq(ctx
);
10015 submitted
= to_submit
;
10016 } else if (to_submit
) {
10017 ret
= io_uring_add_tctx_node(ctx
);
10020 mutex_lock(&ctx
->uring_lock
);
10021 submitted
= io_submit_sqes(ctx
, to_submit
);
10022 mutex_unlock(&ctx
->uring_lock
);
10024 if (submitted
!= to_submit
)
10027 if (flags
& IORING_ENTER_GETEVENTS
) {
10028 const sigset_t __user
*sig
;
10029 struct __kernel_timespec __user
*ts
;
10031 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
10035 min_complete
= min(min_complete
, ctx
->cq_entries
);
10038 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10039 * space applications don't need to do io completion events
10040 * polling again, they can rely on io_sq_thread to do polling
10041 * work, which can reduce cpu usage and uring_lock contention.
10043 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
10044 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10045 ret
= io_iopoll_check(ctx
, min_complete
);
10047 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
10052 percpu_ref_put(&ctx
->refs
);
10055 return submitted
? submitted
: ret
;
10058 #ifdef CONFIG_PROC_FS
10059 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
10060 const struct cred
*cred
)
10062 struct user_namespace
*uns
= seq_user_ns(m
);
10063 struct group_info
*gi
;
10068 seq_printf(m
, "%5d\n", id
);
10069 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
10070 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
10071 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
10072 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
10073 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
10074 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
10075 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
10076 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
10077 seq_puts(m
, "\n\tGroups:\t");
10078 gi
= cred
->group_info
;
10079 for (g
= 0; g
< gi
->ngroups
; g
++) {
10080 seq_put_decimal_ull(m
, g
? " " : "",
10081 from_kgid_munged(uns
, gi
->gid
[g
]));
10083 seq_puts(m
, "\n\tCapEff:\t");
10084 cap
= cred
->cap_effective
;
10085 CAP_FOR_EACH_U32(__capi
)
10086 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
10091 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
10093 struct io_sq_data
*sq
= NULL
;
10098 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10099 * since fdinfo case grabs it in the opposite direction of normal use
10100 * cases. If we fail to get the lock, we just don't iterate any
10101 * structures that could be going away outside the io_uring mutex.
10103 has_lock
= mutex_trylock(&ctx
->uring_lock
);
10105 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10111 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
10112 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
10113 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
10114 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
10115 struct file
*f
= io_file_from_index(ctx
, i
);
10118 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
10120 seq_printf(m
, "%5u: <none>\n", i
);
10122 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
10123 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
10124 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
10125 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
10127 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
10129 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
10130 unsigned long index
;
10131 const struct cred
*cred
;
10133 seq_printf(m
, "Personalities:\n");
10134 xa_for_each(&ctx
->personalities
, index
, cred
)
10135 io_uring_show_cred(m
, index
, cred
);
10137 seq_printf(m
, "PollList:\n");
10138 spin_lock(&ctx
->completion_lock
);
10139 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
10140 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
10141 struct io_kiocb
*req
;
10143 hlist_for_each_entry(req
, list
, hash_node
)
10144 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
10145 req
->task
->task_works
!= NULL
);
10147 spin_unlock(&ctx
->completion_lock
);
10149 mutex_unlock(&ctx
->uring_lock
);
10152 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
10154 struct io_ring_ctx
*ctx
= f
->private_data
;
10156 if (percpu_ref_tryget(&ctx
->refs
)) {
10157 __io_uring_show_fdinfo(ctx
, m
);
10158 percpu_ref_put(&ctx
->refs
);
10163 static const struct file_operations io_uring_fops
= {
10164 .release
= io_uring_release
,
10165 .mmap
= io_uring_mmap
,
10167 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
10168 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
10170 .poll
= io_uring_poll
,
10171 #ifdef CONFIG_PROC_FS
10172 .show_fdinfo
= io_uring_show_fdinfo
,
10176 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
10177 struct io_uring_params
*p
)
10179 struct io_rings
*rings
;
10180 size_t size
, sq_array_offset
;
10182 /* make sure these are sane, as we already accounted them */
10183 ctx
->sq_entries
= p
->sq_entries
;
10184 ctx
->cq_entries
= p
->cq_entries
;
10186 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
10187 if (size
== SIZE_MAX
)
10190 rings
= io_mem_alloc(size
);
10194 ctx
->rings
= rings
;
10195 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
10196 rings
->sq_ring_mask
= p
->sq_entries
- 1;
10197 rings
->cq_ring_mask
= p
->cq_entries
- 1;
10198 rings
->sq_ring_entries
= p
->sq_entries
;
10199 rings
->cq_ring_entries
= p
->cq_entries
;
10201 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
10202 if (size
== SIZE_MAX
) {
10203 io_mem_free(ctx
->rings
);
10208 ctx
->sq_sqes
= io_mem_alloc(size
);
10209 if (!ctx
->sq_sqes
) {
10210 io_mem_free(ctx
->rings
);
10218 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
10222 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
10226 ret
= io_uring_add_tctx_node(ctx
);
10231 fd_install(fd
, file
);
10236 * Allocate an anonymous fd, this is what constitutes the application
10237 * visible backing of an io_uring instance. The application mmaps this
10238 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10239 * we have to tie this fd to a socket for file garbage collection purposes.
10241 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
10244 #if defined(CONFIG_UNIX)
10247 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
10250 return ERR_PTR(ret
);
10253 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
10254 O_RDWR
| O_CLOEXEC
);
10255 #if defined(CONFIG_UNIX)
10256 if (IS_ERR(file
)) {
10257 sock_release(ctx
->ring_sock
);
10258 ctx
->ring_sock
= NULL
;
10260 ctx
->ring_sock
->file
= file
;
10266 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
10267 struct io_uring_params __user
*params
)
10269 struct io_ring_ctx
*ctx
;
10275 if (entries
> IORING_MAX_ENTRIES
) {
10276 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10278 entries
= IORING_MAX_ENTRIES
;
10282 * Use twice as many entries for the CQ ring. It's possible for the
10283 * application to drive a higher depth than the size of the SQ ring,
10284 * since the sqes are only used at submission time. This allows for
10285 * some flexibility in overcommitting a bit. If the application has
10286 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10287 * of CQ ring entries manually.
10289 p
->sq_entries
= roundup_pow_of_two(entries
);
10290 if (p
->flags
& IORING_SETUP_CQSIZE
) {
10292 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10293 * to a power-of-two, if it isn't already. We do NOT impose
10294 * any cq vs sq ring sizing.
10296 if (!p
->cq_entries
)
10298 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
10299 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10301 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
10303 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
10304 if (p
->cq_entries
< p
->sq_entries
)
10307 p
->cq_entries
= 2 * p
->sq_entries
;
10310 ctx
= io_ring_ctx_alloc(p
);
10313 ctx
->compat
= in_compat_syscall();
10314 if (!capable(CAP_IPC_LOCK
))
10315 ctx
->user
= get_uid(current_user());
10318 * This is just grabbed for accounting purposes. When a process exits,
10319 * the mm is exited and dropped before the files, hence we need to hang
10320 * on to this mm purely for the purposes of being able to unaccount
10321 * memory (locked/pinned vm). It's not used for anything else.
10323 mmgrab(current
->mm
);
10324 ctx
->mm_account
= current
->mm
;
10326 ret
= io_allocate_scq_urings(ctx
, p
);
10330 ret
= io_sq_offload_create(ctx
, p
);
10333 /* always set a rsrc node */
10334 ret
= io_rsrc_node_switch_start(ctx
);
10337 io_rsrc_node_switch(ctx
, NULL
);
10339 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
10340 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
10341 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
10342 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
10343 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
10344 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
10345 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
10346 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
10348 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
10349 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
10350 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
10351 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
10352 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
10353 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
10354 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
10355 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
10357 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
10358 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
10359 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
10360 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
10361 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
10362 IORING_FEAT_RSRC_TAGS
;
10364 if (copy_to_user(params
, p
, sizeof(*p
))) {
10369 file
= io_uring_get_file(ctx
);
10370 if (IS_ERR(file
)) {
10371 ret
= PTR_ERR(file
);
10376 * Install ring fd as the very last thing, so we don't risk someone
10377 * having closed it before we finish setup
10379 ret
= io_uring_install_fd(ctx
, file
);
10381 /* fput will clean it up */
10386 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
10389 io_ring_ctx_wait_and_kill(ctx
);
10394 * Sets up an aio uring context, and returns the fd. Applications asks for a
10395 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10396 * params structure passed in.
10398 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
10400 struct io_uring_params p
;
10403 if (copy_from_user(&p
, params
, sizeof(p
)))
10405 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
10410 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
10411 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
10412 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
10413 IORING_SETUP_R_DISABLED
))
10416 return io_uring_create(entries
, &p
, params
);
10419 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
10420 struct io_uring_params __user
*, params
)
10422 return io_uring_setup(entries
, params
);
10425 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
10427 struct io_uring_probe
*p
;
10431 size
= struct_size(p
, ops
, nr_args
);
10432 if (size
== SIZE_MAX
)
10434 p
= kzalloc(size
, GFP_KERNEL
);
10439 if (copy_from_user(p
, arg
, size
))
10442 if (memchr_inv(p
, 0, size
))
10445 p
->last_op
= IORING_OP_LAST
- 1;
10446 if (nr_args
> IORING_OP_LAST
)
10447 nr_args
= IORING_OP_LAST
;
10449 for (i
= 0; i
< nr_args
; i
++) {
10451 if (!io_op_defs
[i
].not_supported
)
10452 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
10457 if (copy_to_user(arg
, p
, size
))
10464 static int io_register_personality(struct io_ring_ctx
*ctx
)
10466 const struct cred
*creds
;
10470 creds
= get_current_cred();
10472 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
10473 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
10481 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
10482 unsigned int nr_args
)
10484 struct io_uring_restriction
*res
;
10488 /* Restrictions allowed only if rings started disabled */
10489 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10492 /* We allow only a single restrictions registration */
10493 if (ctx
->restrictions
.registered
)
10496 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
10499 size
= array_size(nr_args
, sizeof(*res
));
10500 if (size
== SIZE_MAX
)
10503 res
= memdup_user(arg
, size
);
10505 return PTR_ERR(res
);
10509 for (i
= 0; i
< nr_args
; i
++) {
10510 switch (res
[i
].opcode
) {
10511 case IORING_RESTRICTION_REGISTER_OP
:
10512 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
10517 __set_bit(res
[i
].register_op
,
10518 ctx
->restrictions
.register_op
);
10520 case IORING_RESTRICTION_SQE_OP
:
10521 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
10526 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
10528 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
10529 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
10531 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
10532 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
10541 /* Reset all restrictions if an error happened */
10543 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
10545 ctx
->restrictions
.registered
= true;
10551 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10553 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10556 if (ctx
->restrictions
.registered
)
10557 ctx
->restricted
= 1;
10559 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
10560 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
10561 wake_up(&ctx
->sq_data
->wait
);
10565 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
10566 struct io_uring_rsrc_update2
*up
,
10574 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
10576 err
= io_rsrc_node_switch_start(ctx
);
10581 case IORING_RSRC_FILE
:
10582 return __io_sqe_files_update(ctx
, up
, nr_args
);
10583 case IORING_RSRC_BUFFER
:
10584 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
10589 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10592 struct io_uring_rsrc_update2 up
;
10596 memset(&up
, 0, sizeof(up
));
10597 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
10599 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
10602 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10603 unsigned size
, unsigned type
)
10605 struct io_uring_rsrc_update2 up
;
10607 if (size
!= sizeof(up
))
10609 if (copy_from_user(&up
, arg
, sizeof(up
)))
10611 if (!up
.nr
|| up
.resv
)
10613 return __io_register_rsrc_update(ctx
, type
, &up
, up
.nr
);
10616 static int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
10617 unsigned int size
, unsigned int type
)
10619 struct io_uring_rsrc_register rr
;
10621 /* keep it extendible */
10622 if (size
!= sizeof(rr
))
10625 memset(&rr
, 0, sizeof(rr
));
10626 if (copy_from_user(&rr
, arg
, size
))
10628 if (!rr
.nr
|| rr
.resv
|| rr
.resv2
)
10632 case IORING_RSRC_FILE
:
10633 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
10634 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10635 case IORING_RSRC_BUFFER
:
10636 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
10637 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10642 static int io_register_iowq_aff(struct io_ring_ctx
*ctx
, void __user
*arg
,
10645 struct io_uring_task
*tctx
= current
->io_uring
;
10646 cpumask_var_t new_mask
;
10649 if (!tctx
|| !tctx
->io_wq
)
10652 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
10655 cpumask_clear(new_mask
);
10656 if (len
> cpumask_size())
10657 len
= cpumask_size();
10659 if (copy_from_user(new_mask
, arg
, len
)) {
10660 free_cpumask_var(new_mask
);
10664 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
10665 free_cpumask_var(new_mask
);
10669 static int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
10671 struct io_uring_task
*tctx
= current
->io_uring
;
10673 if (!tctx
|| !tctx
->io_wq
)
10676 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
10679 static int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
10681 __must_hold(&ctx
->uring_lock
)
10683 struct io_tctx_node
*node
;
10684 struct io_uring_task
*tctx
= NULL
;
10685 struct io_sq_data
*sqd
= NULL
;
10686 __u32 new_count
[2];
10689 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
10691 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10692 if (new_count
[i
] > INT_MAX
)
10695 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10696 sqd
= ctx
->sq_data
;
10699 * Observe the correct sqd->lock -> ctx->uring_lock
10700 * ordering. Fine to drop uring_lock here, we hold
10701 * a ref to the ctx.
10703 refcount_inc(&sqd
->refs
);
10704 mutex_unlock(&ctx
->uring_lock
);
10705 mutex_lock(&sqd
->lock
);
10706 mutex_lock(&ctx
->uring_lock
);
10708 tctx
= sqd
->thread
->io_uring
;
10711 tctx
= current
->io_uring
;
10714 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
10716 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10718 ctx
->iowq_limits
[i
] = new_count
[i
];
10719 ctx
->iowq_limits_set
= true;
10722 if (tctx
&& tctx
->io_wq
) {
10723 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
10727 memset(new_count
, 0, sizeof(new_count
));
10731 mutex_unlock(&sqd
->lock
);
10732 io_put_sq_data(sqd
);
10735 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
10738 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10742 /* now propagate the restriction to all registered users */
10743 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
10744 struct io_uring_task
*tctx
= node
->task
->io_uring
;
10746 if (WARN_ON_ONCE(!tctx
->io_wq
))
10749 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10750 new_count
[i
] = ctx
->iowq_limits
[i
];
10751 /* ignore errors, it always returns zero anyway */
10752 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
10757 mutex_unlock(&sqd
->lock
);
10758 io_put_sq_data(sqd
);
10763 static bool io_register_op_must_quiesce(int op
)
10766 case IORING_REGISTER_BUFFERS
:
10767 case IORING_UNREGISTER_BUFFERS
:
10768 case IORING_REGISTER_FILES
:
10769 case IORING_UNREGISTER_FILES
:
10770 case IORING_REGISTER_FILES_UPDATE
:
10771 case IORING_REGISTER_PROBE
:
10772 case IORING_REGISTER_PERSONALITY
:
10773 case IORING_UNREGISTER_PERSONALITY
:
10774 case IORING_REGISTER_FILES2
:
10775 case IORING_REGISTER_FILES_UPDATE2
:
10776 case IORING_REGISTER_BUFFERS2
:
10777 case IORING_REGISTER_BUFFERS_UPDATE
:
10778 case IORING_REGISTER_IOWQ_AFF
:
10779 case IORING_UNREGISTER_IOWQ_AFF
:
10780 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10787 static int io_ctx_quiesce(struct io_ring_ctx
*ctx
)
10791 percpu_ref_kill(&ctx
->refs
);
10794 * Drop uring mutex before waiting for references to exit. If another
10795 * thread is currently inside io_uring_enter() it might need to grab the
10796 * uring_lock to make progress. If we hold it here across the drain
10797 * wait, then we can deadlock. It's safe to drop the mutex here, since
10798 * no new references will come in after we've killed the percpu ref.
10800 mutex_unlock(&ctx
->uring_lock
);
10802 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10805 ret
= io_run_task_work_sig();
10806 } while (ret
>= 0);
10807 mutex_lock(&ctx
->uring_lock
);
10810 io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
);
10814 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10815 void __user
*arg
, unsigned nr_args
)
10816 __releases(ctx
->uring_lock
)
10817 __acquires(ctx
->uring_lock
)
10822 * We're inside the ring mutex, if the ref is already dying, then
10823 * someone else killed the ctx or is already going through
10824 * io_uring_register().
10826 if (percpu_ref_is_dying(&ctx
->refs
))
10829 if (ctx
->restricted
) {
10830 if (opcode
>= IORING_REGISTER_LAST
)
10832 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
10833 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
10837 if (io_register_op_must_quiesce(opcode
)) {
10838 ret
= io_ctx_quiesce(ctx
);
10844 case IORING_REGISTER_BUFFERS
:
10845 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
10847 case IORING_UNREGISTER_BUFFERS
:
10849 if (arg
|| nr_args
)
10851 ret
= io_sqe_buffers_unregister(ctx
);
10853 case IORING_REGISTER_FILES
:
10854 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
10856 case IORING_UNREGISTER_FILES
:
10858 if (arg
|| nr_args
)
10860 ret
= io_sqe_files_unregister(ctx
);
10862 case IORING_REGISTER_FILES_UPDATE
:
10863 ret
= io_register_files_update(ctx
, arg
, nr_args
);
10865 case IORING_REGISTER_EVENTFD
:
10866 case IORING_REGISTER_EVENTFD_ASYNC
:
10870 ret
= io_eventfd_register(ctx
, arg
);
10873 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10874 ctx
->eventfd_async
= 1;
10876 ctx
->eventfd_async
= 0;
10878 case IORING_UNREGISTER_EVENTFD
:
10880 if (arg
|| nr_args
)
10882 ret
= io_eventfd_unregister(ctx
);
10884 case IORING_REGISTER_PROBE
:
10886 if (!arg
|| nr_args
> 256)
10888 ret
= io_probe(ctx
, arg
, nr_args
);
10890 case IORING_REGISTER_PERSONALITY
:
10892 if (arg
|| nr_args
)
10894 ret
= io_register_personality(ctx
);
10896 case IORING_UNREGISTER_PERSONALITY
:
10900 ret
= io_unregister_personality(ctx
, nr_args
);
10902 case IORING_REGISTER_ENABLE_RINGS
:
10904 if (arg
|| nr_args
)
10906 ret
= io_register_enable_rings(ctx
);
10908 case IORING_REGISTER_RESTRICTIONS
:
10909 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10911 case IORING_REGISTER_FILES2
:
10912 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
10914 case IORING_REGISTER_FILES_UPDATE2
:
10915 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10918 case IORING_REGISTER_BUFFERS2
:
10919 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
10921 case IORING_REGISTER_BUFFERS_UPDATE
:
10922 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10923 IORING_RSRC_BUFFER
);
10925 case IORING_REGISTER_IOWQ_AFF
:
10927 if (!arg
|| !nr_args
)
10929 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
10931 case IORING_UNREGISTER_IOWQ_AFF
:
10933 if (arg
|| nr_args
)
10935 ret
= io_unregister_iowq_aff(ctx
);
10937 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10939 if (!arg
|| nr_args
!= 2)
10941 ret
= io_register_iowq_max_workers(ctx
, arg
);
10948 if (io_register_op_must_quiesce(opcode
)) {
10949 /* bring the ctx back to life */
10950 percpu_ref_reinit(&ctx
->refs
);
10951 reinit_completion(&ctx
->ref_comp
);
10956 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10957 void __user
*, arg
, unsigned int, nr_args
)
10959 struct io_ring_ctx
*ctx
;
10968 if (f
.file
->f_op
!= &io_uring_fops
)
10971 ctx
= f
.file
->private_data
;
10973 io_run_task_work();
10975 mutex_lock(&ctx
->uring_lock
);
10976 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10977 mutex_unlock(&ctx
->uring_lock
);
10978 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10979 ctx
->cq_ev_fd
!= NULL
, ret
);
10985 static int __init
io_uring_init(void)
10987 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10988 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10989 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10992 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10993 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10994 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10995 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10996 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10997 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10998 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10999 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
11000 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
11001 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
11002 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
11003 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
11004 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
11005 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
11006 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
11007 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
11008 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
11009 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
11010 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
11011 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
11012 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
11013 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
11014 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
11015 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
11016 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
11017 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
11018 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
11019 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
11020 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
11021 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
11022 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
11023 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
11024 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
11026 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
11027 sizeof(struct io_uring_rsrc_update
));
11028 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
11029 sizeof(struct io_uring_rsrc_update2
));
11031 /* ->buf_index is u16 */
11032 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS
>= (1u << 16));
11034 /* should fit into one byte */
11035 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
11037 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
11038 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
11040 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
11044 __initcall(io_uring_init
);