1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/mmu_context.h>
59 #include <linux/percpu.h>
60 #include <linux/slab.h>
61 #include <linux/kthread.h>
62 #include <linux/blkdev.h>
63 #include <linux/bvec.h>
64 #include <linux/net.h>
66 #include <net/af_unix.h>
68 #include <linux/anon_inodes.h>
69 #include <linux/sched/mm.h>
70 #include <linux/uaccess.h>
71 #include <linux/nospec.h>
72 #include <linux/sizes.h>
73 #include <linux/hugetlb.h>
74 #include <linux/highmem.h>
75 #include <linux/namei.h>
76 #include <linux/fsnotify.h>
77 #include <linux/fadvise.h>
78 #include <linux/eventpoll.h>
79 #include <linux/fs_struct.h>
80 #include <linux/splice.h>
81 #include <linux/task_work.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)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp
;
104 u32 tail ____cacheline_aligned_in_smp
;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq
, cq
;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask
, cq_ring_mask
;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries
, cq_ring_entries
;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
155 * Number of completion events lost because the queue was full;
156 * this should be avoided by the application by making sure
157 * there are not more requests pending than there is space in
158 * the completion queue.
160 * Written by the kernel, shouldn't be modified by the
161 * application (i.e. get number of "new events" by comparing to
164 * As completion events come in out of order this counter is not
165 * ordered with any other data.
169 * Ring buffer of completion events.
171 * The kernel writes completion events fresh every time they are
172 * produced, so the application is allowed to modify pending
175 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
178 struct io_mapped_ubuf
{
181 struct bio_vec
*bvec
;
182 unsigned int nr_bvecs
;
185 struct fixed_file_table
{
189 struct fixed_file_ref_node
{
190 struct percpu_ref refs
;
191 struct list_head node
;
192 struct list_head file_list
;
193 struct fixed_file_data
*file_data
;
194 struct work_struct work
;
197 struct fixed_file_data
{
198 struct fixed_file_table
*table
;
199 struct io_ring_ctx
*ctx
;
201 struct percpu_ref
*cur_refs
;
202 struct percpu_ref refs
;
203 struct completion done
;
204 struct list_head ref_list
;
209 struct list_head list
;
217 struct percpu_ref refs
;
218 } ____cacheline_aligned_in_smp
;
222 unsigned int compat
: 1;
223 unsigned int account_mem
: 1;
224 unsigned int cq_overflow_flushed
: 1;
225 unsigned int drain_next
: 1;
226 unsigned int eventfd_async
: 1;
229 * Ring buffer of indices into array of io_uring_sqe, which is
230 * mmapped by the application using the IORING_OFF_SQES offset.
232 * This indirection could e.g. be used to assign fixed
233 * io_uring_sqe entries to operations and only submit them to
234 * the queue when needed.
236 * The kernel modifies neither the indices array nor the entries
240 unsigned cached_sq_head
;
243 unsigned sq_thread_idle
;
244 unsigned cached_sq_dropped
;
245 atomic_t cached_cq_overflow
;
246 unsigned long sq_check_overflow
;
248 struct list_head defer_list
;
249 struct list_head timeout_list
;
250 struct list_head cq_overflow_list
;
252 wait_queue_head_t inflight_wait
;
253 struct io_uring_sqe
*sq_sqes
;
254 } ____cacheline_aligned_in_smp
;
256 struct io_rings
*rings
;
260 struct task_struct
*sqo_thread
; /* if using sq thread polling */
261 struct mm_struct
*sqo_mm
;
262 wait_queue_head_t sqo_wait
;
265 * If used, fixed file set. Writers must ensure that ->refs is dead,
266 * readers must ensure that ->refs is alive as long as the file* is
267 * used. Only updated through io_uring_register(2).
269 struct fixed_file_data
*file_data
;
270 unsigned nr_user_files
;
272 struct file
*ring_file
;
274 /* if used, fixed mapped user buffers */
275 unsigned nr_user_bufs
;
276 struct io_mapped_ubuf
*user_bufs
;
278 struct user_struct
*user
;
280 const struct cred
*creds
;
282 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
283 struct completion
*completions
;
285 /* if all else fails... */
286 struct io_kiocb
*fallback_req
;
288 #if defined(CONFIG_UNIX)
289 struct socket
*ring_sock
;
292 struct idr io_buffer_idr
;
294 struct idr personality_idr
;
297 unsigned cached_cq_tail
;
300 atomic_t cq_timeouts
;
301 unsigned long cq_check_overflow
;
302 struct wait_queue_head cq_wait
;
303 struct fasync_struct
*cq_fasync
;
304 struct eventfd_ctx
*cq_ev_fd
;
305 } ____cacheline_aligned_in_smp
;
308 struct mutex uring_lock
;
309 wait_queue_head_t wait
;
310 } ____cacheline_aligned_in_smp
;
313 spinlock_t completion_lock
;
316 * ->poll_list is protected by the ctx->uring_lock for
317 * io_uring instances that don't use IORING_SETUP_SQPOLL.
318 * For SQPOLL, only the single threaded io_sq_thread() will
319 * manipulate the list, hence no extra locking is needed there.
321 struct list_head poll_list
;
322 struct hlist_head
*cancel_hash
;
323 unsigned cancel_hash_bits
;
324 bool poll_multi_file
;
326 spinlock_t inflight_lock
;
327 struct list_head inflight_list
;
328 } ____cacheline_aligned_in_smp
;
330 struct work_struct exit_work
;
334 * First field must be the file pointer in all the
335 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
337 struct io_poll_iocb
{
340 struct wait_queue_head
*head
;
346 struct wait_queue_entry wait
;
351 struct file
*put_file
;
355 struct io_timeout_data
{
356 struct io_kiocb
*req
;
357 struct hrtimer timer
;
358 struct timespec64 ts
;
359 enum hrtimer_mode mode
;
364 struct sockaddr __user
*addr
;
365 int __user
*addr_len
;
367 unsigned long nofile
;
391 /* NOTE: kiocb has the file as the first member, so don't do it here */
399 struct sockaddr __user
*addr
;
406 struct user_msghdr __user
*msg
;
412 struct io_buffer
*kbuf
;
421 struct filename
*filename
;
422 struct statx __user
*buffer
;
424 unsigned long nofile
;
427 struct io_files_update
{
453 struct epoll_event event
;
457 struct file
*file_out
;
458 struct file
*file_in
;
465 struct io_provide_buf
{
474 struct io_async_connect
{
475 struct sockaddr_storage address
;
478 struct io_async_msghdr
{
479 struct iovec fast_iov
[UIO_FASTIOV
];
481 struct sockaddr __user
*uaddr
;
483 struct sockaddr_storage addr
;
487 struct iovec fast_iov
[UIO_FASTIOV
];
493 struct io_async_ctx
{
495 struct io_async_rw rw
;
496 struct io_async_msghdr msg
;
497 struct io_async_connect connect
;
498 struct io_timeout_data timeout
;
503 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
504 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
505 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
506 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
507 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
508 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
516 REQ_F_IOPOLL_COMPLETED_BIT
,
517 REQ_F_LINK_TIMEOUT_BIT
,
521 REQ_F_TIMEOUT_NOSEQ_BIT
,
522 REQ_F_COMP_LOCKED_BIT
,
523 REQ_F_NEED_CLEANUP_BIT
,
526 REQ_F_BUFFER_SELECTED_BIT
,
527 REQ_F_NO_FILE_TABLE_BIT
,
529 /* not a real bit, just to check we're not overflowing the space */
535 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
536 /* drain existing IO first */
537 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
539 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
540 /* doesn't sever on completion < 0 */
541 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
543 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
544 /* IOSQE_BUFFER_SELECT */
545 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
548 REQ_F_LINK_HEAD
= BIT(REQ_F_LINK_HEAD_BIT
),
549 /* already grabbed next link */
550 REQ_F_LINK_NEXT
= BIT(REQ_F_LINK_NEXT_BIT
),
551 /* fail rest of links */
552 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
553 /* on inflight list */
554 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
555 /* read/write uses file position */
556 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
557 /* must not punt to workers */
558 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
559 /* polled IO has completed */
560 REQ_F_IOPOLL_COMPLETED
= BIT(REQ_F_IOPOLL_COMPLETED_BIT
),
561 /* has linked timeout */
562 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
563 /* timeout request */
564 REQ_F_TIMEOUT
= BIT(REQ_F_TIMEOUT_BIT
),
566 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
567 /* must be punted even for NONBLOCK */
568 REQ_F_MUST_PUNT
= BIT(REQ_F_MUST_PUNT_BIT
),
569 /* no timeout sequence */
570 REQ_F_TIMEOUT_NOSEQ
= BIT(REQ_F_TIMEOUT_NOSEQ_BIT
),
571 /* completion under lock */
572 REQ_F_COMP_LOCKED
= BIT(REQ_F_COMP_LOCKED_BIT
),
574 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
575 /* in overflow list */
576 REQ_F_OVERFLOW
= BIT(REQ_F_OVERFLOW_BIT
),
577 /* already went through poll handler */
578 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
579 /* buffer already selected */
580 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
581 /* doesn't need file table for this request */
582 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
586 struct io_poll_iocb poll
;
587 struct io_wq_work work
;
591 * NOTE! Each of the iocb union members has the file pointer
592 * as the first entry in their struct definition. So you can
593 * access the file pointer through any of the sub-structs,
594 * or directly as just 'ki_filp' in this struct.
600 struct io_poll_iocb poll
;
601 struct io_accept accept
;
603 struct io_cancel cancel
;
604 struct io_timeout timeout
;
605 struct io_connect connect
;
606 struct io_sr_msg sr_msg
;
608 struct io_close close
;
609 struct io_files_update files_update
;
610 struct io_fadvise fadvise
;
611 struct io_madvise madvise
;
612 struct io_epoll epoll
;
613 struct io_splice splice
;
614 struct io_provide_buf pbuf
;
617 struct io_async_ctx
*io
;
619 bool needs_fixed_file
;
622 struct io_ring_ctx
*ctx
;
623 struct list_head list
;
626 struct task_struct
*task
;
632 struct list_head link_list
;
634 struct list_head inflight_entry
;
636 struct percpu_ref
*fixed_file_refs
;
640 * Only commands that never go async can use the below fields,
641 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
642 * async armed poll handlers for regular commands. The latter
643 * restore the work, if needed.
646 struct callback_head task_work
;
647 struct hlist_node hash_node
;
648 struct async_poll
*apoll
;
650 struct io_wq_work work
;
654 #define IO_PLUG_THRESHOLD 2
655 #define IO_IOPOLL_BATCH 8
657 struct io_submit_state
{
658 struct blk_plug plug
;
661 * io_kiocb alloc cache
663 void *reqs
[IO_IOPOLL_BATCH
];
664 unsigned int free_reqs
;
667 * File reference cache
671 unsigned int has_refs
;
672 unsigned int used_refs
;
673 unsigned int ios_left
;
677 /* needs req->io allocated for deferral/async */
678 unsigned async_ctx
: 1;
679 /* needs current->mm setup, does mm access */
680 unsigned needs_mm
: 1;
681 /* needs req->file assigned */
682 unsigned needs_file
: 1;
683 /* hash wq insertion if file is a regular file */
684 unsigned hash_reg_file
: 1;
685 /* unbound wq insertion if file is a non-regular file */
686 unsigned unbound_nonreg_file
: 1;
687 /* opcode is not supported by this kernel */
688 unsigned not_supported
: 1;
689 /* needs file table */
690 unsigned file_table
: 1;
692 unsigned needs_fs
: 1;
693 /* set if opcode supports polled "wait" */
695 unsigned pollout
: 1;
696 /* op supports buffer selection */
697 unsigned buffer_select
: 1;
700 static const struct io_op_def io_op_defs
[] = {
701 [IORING_OP_NOP
] = {},
702 [IORING_OP_READV
] = {
706 .unbound_nonreg_file
= 1,
710 [IORING_OP_WRITEV
] = {
715 .unbound_nonreg_file
= 1,
718 [IORING_OP_FSYNC
] = {
721 [IORING_OP_READ_FIXED
] = {
723 .unbound_nonreg_file
= 1,
726 [IORING_OP_WRITE_FIXED
] = {
729 .unbound_nonreg_file
= 1,
732 [IORING_OP_POLL_ADD
] = {
734 .unbound_nonreg_file
= 1,
736 [IORING_OP_POLL_REMOVE
] = {},
737 [IORING_OP_SYNC_FILE_RANGE
] = {
740 [IORING_OP_SENDMSG
] = {
744 .unbound_nonreg_file
= 1,
748 [IORING_OP_RECVMSG
] = {
752 .unbound_nonreg_file
= 1,
757 [IORING_OP_TIMEOUT
] = {
761 [IORING_OP_TIMEOUT_REMOVE
] = {},
762 [IORING_OP_ACCEPT
] = {
765 .unbound_nonreg_file
= 1,
769 [IORING_OP_ASYNC_CANCEL
] = {},
770 [IORING_OP_LINK_TIMEOUT
] = {
774 [IORING_OP_CONNECT
] = {
778 .unbound_nonreg_file
= 1,
781 [IORING_OP_FALLOCATE
] = {
784 [IORING_OP_OPENAT
] = {
788 [IORING_OP_CLOSE
] = {
792 [IORING_OP_FILES_UPDATE
] = {
796 [IORING_OP_STATX
] = {
804 .unbound_nonreg_file
= 1,
808 [IORING_OP_WRITE
] = {
811 .unbound_nonreg_file
= 1,
814 [IORING_OP_FADVISE
] = {
817 [IORING_OP_MADVISE
] = {
823 .unbound_nonreg_file
= 1,
829 .unbound_nonreg_file
= 1,
833 [IORING_OP_OPENAT2
] = {
837 [IORING_OP_EPOLL_CTL
] = {
838 .unbound_nonreg_file
= 1,
841 [IORING_OP_SPLICE
] = {
844 .unbound_nonreg_file
= 1,
846 [IORING_OP_PROVIDE_BUFFERS
] = {},
847 [IORING_OP_REMOVE_BUFFERS
] = {},
850 static void io_wq_submit_work(struct io_wq_work
**workptr
);
851 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
852 static void io_put_req(struct io_kiocb
*req
);
853 static void __io_double_put_req(struct io_kiocb
*req
);
854 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
855 static void io_queue_linked_timeout(struct io_kiocb
*req
);
856 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
857 struct io_uring_files_update
*ip
,
859 static int io_grab_files(struct io_kiocb
*req
);
860 static void io_cleanup_req(struct io_kiocb
*req
);
861 static int io_file_get(struct io_submit_state
*state
, struct io_kiocb
*req
,
862 int fd
, struct file
**out_file
, bool fixed
);
863 static void __io_queue_sqe(struct io_kiocb
*req
,
864 const struct io_uring_sqe
*sqe
);
866 static struct kmem_cache
*req_cachep
;
868 static const struct file_operations io_uring_fops
;
870 struct sock
*io_uring_get_socket(struct file
*file
)
872 #if defined(CONFIG_UNIX)
873 if (file
->f_op
== &io_uring_fops
) {
874 struct io_ring_ctx
*ctx
= file
->private_data
;
876 return ctx
->ring_sock
->sk
;
881 EXPORT_SYMBOL(io_uring_get_socket
);
883 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
885 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
887 complete(&ctx
->completions
[0]);
890 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
892 struct io_ring_ctx
*ctx
;
895 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
899 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
900 if (!ctx
->fallback_req
)
903 ctx
->completions
= kmalloc(2 * sizeof(struct completion
), GFP_KERNEL
);
904 if (!ctx
->completions
)
908 * Use 5 bits less than the max cq entries, that should give us around
909 * 32 entries per hash list if totally full and uniformly spread.
911 hash_bits
= ilog2(p
->cq_entries
);
915 ctx
->cancel_hash_bits
= hash_bits
;
916 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
918 if (!ctx
->cancel_hash
)
920 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
922 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
923 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
926 ctx
->flags
= p
->flags
;
927 init_waitqueue_head(&ctx
->cq_wait
);
928 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
929 init_completion(&ctx
->completions
[0]);
930 init_completion(&ctx
->completions
[1]);
931 idr_init(&ctx
->io_buffer_idr
);
932 idr_init(&ctx
->personality_idr
);
933 mutex_init(&ctx
->uring_lock
);
934 init_waitqueue_head(&ctx
->wait
);
935 spin_lock_init(&ctx
->completion_lock
);
936 INIT_LIST_HEAD(&ctx
->poll_list
);
937 INIT_LIST_HEAD(&ctx
->defer_list
);
938 INIT_LIST_HEAD(&ctx
->timeout_list
);
939 init_waitqueue_head(&ctx
->inflight_wait
);
940 spin_lock_init(&ctx
->inflight_lock
);
941 INIT_LIST_HEAD(&ctx
->inflight_list
);
944 if (ctx
->fallback_req
)
945 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
946 kfree(ctx
->completions
);
947 kfree(ctx
->cancel_hash
);
952 static inline bool __req_need_defer(struct io_kiocb
*req
)
954 struct io_ring_ctx
*ctx
= req
->ctx
;
956 return req
->sequence
!= ctx
->cached_cq_tail
957 + atomic_read(&ctx
->cached_cq_overflow
);
960 static inline bool req_need_defer(struct io_kiocb
*req
)
962 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
))
963 return __req_need_defer(req
);
968 static struct io_kiocb
*io_get_deferred_req(struct io_ring_ctx
*ctx
)
970 struct io_kiocb
*req
;
972 req
= list_first_entry_or_null(&ctx
->defer_list
, struct io_kiocb
, list
);
973 if (req
&& !req_need_defer(req
)) {
974 list_del_init(&req
->list
);
981 static struct io_kiocb
*io_get_timeout_req(struct io_ring_ctx
*ctx
)
983 struct io_kiocb
*req
;
985 req
= list_first_entry_or_null(&ctx
->timeout_list
, struct io_kiocb
, list
);
987 if (req
->flags
& REQ_F_TIMEOUT_NOSEQ
)
989 if (!__req_need_defer(req
)) {
990 list_del_init(&req
->list
);
998 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1000 struct io_rings
*rings
= ctx
->rings
;
1002 /* order cqe stores with ring update */
1003 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1005 if (wq_has_sleeper(&ctx
->cq_wait
)) {
1006 wake_up_interruptible(&ctx
->cq_wait
);
1007 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1011 static inline void io_req_work_grab_env(struct io_kiocb
*req
,
1012 const struct io_op_def
*def
)
1014 if (!req
->work
.mm
&& def
->needs_mm
) {
1015 mmgrab(current
->mm
);
1016 req
->work
.mm
= current
->mm
;
1018 if (!req
->work
.creds
)
1019 req
->work
.creds
= get_current_cred();
1020 if (!req
->work
.fs
&& def
->needs_fs
) {
1021 spin_lock(¤t
->fs
->lock
);
1022 if (!current
->fs
->in_exec
) {
1023 req
->work
.fs
= current
->fs
;
1024 req
->work
.fs
->users
++;
1026 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1028 spin_unlock(¤t
->fs
->lock
);
1030 if (!req
->work
.task_pid
)
1031 req
->work
.task_pid
= task_pid_vnr(current
);
1034 static inline void io_req_work_drop_env(struct io_kiocb
*req
)
1037 mmdrop(req
->work
.mm
);
1038 req
->work
.mm
= NULL
;
1040 if (req
->work
.creds
) {
1041 put_cred(req
->work
.creds
);
1042 req
->work
.creds
= NULL
;
1045 struct fs_struct
*fs
= req
->work
.fs
;
1047 spin_lock(&req
->work
.fs
->lock
);
1050 spin_unlock(&req
->work
.fs
->lock
);
1056 static inline void io_prep_async_work(struct io_kiocb
*req
,
1057 struct io_kiocb
**link
)
1059 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1061 if (req
->flags
& REQ_F_ISREG
) {
1062 if (def
->hash_reg_file
)
1063 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1065 if (def
->unbound_nonreg_file
)
1066 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1069 io_req_work_grab_env(req
, def
);
1071 *link
= io_prep_linked_timeout(req
);
1074 static inline void io_queue_async_work(struct io_kiocb
*req
)
1076 struct io_ring_ctx
*ctx
= req
->ctx
;
1077 struct io_kiocb
*link
;
1079 io_prep_async_work(req
, &link
);
1081 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1082 &req
->work
, req
->flags
);
1083 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1086 io_queue_linked_timeout(link
);
1089 static void io_kill_timeout(struct io_kiocb
*req
)
1093 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
1095 atomic_inc(&req
->ctx
->cq_timeouts
);
1096 list_del_init(&req
->list
);
1097 req
->flags
|= REQ_F_COMP_LOCKED
;
1098 io_cqring_fill_event(req
, 0);
1103 static void io_kill_timeouts(struct io_ring_ctx
*ctx
)
1105 struct io_kiocb
*req
, *tmp
;
1107 spin_lock_irq(&ctx
->completion_lock
);
1108 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, list
)
1109 io_kill_timeout(req
);
1110 spin_unlock_irq(&ctx
->completion_lock
);
1113 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1115 struct io_kiocb
*req
;
1117 while ((req
= io_get_timeout_req(ctx
)) != NULL
)
1118 io_kill_timeout(req
);
1120 __io_commit_cqring(ctx
);
1122 while ((req
= io_get_deferred_req(ctx
)) != NULL
)
1123 io_queue_async_work(req
);
1126 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1128 struct io_rings
*rings
= ctx
->rings
;
1131 tail
= ctx
->cached_cq_tail
;
1133 * writes to the cq entry need to come after reading head; the
1134 * control dependency is enough as we're using WRITE_ONCE to
1137 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1140 ctx
->cached_cq_tail
++;
1141 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1144 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1148 if (!ctx
->eventfd_async
)
1150 return io_wq_current_is_worker();
1153 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1155 if (waitqueue_active(&ctx
->wait
))
1156 wake_up(&ctx
->wait
);
1157 if (waitqueue_active(&ctx
->sqo_wait
))
1158 wake_up(&ctx
->sqo_wait
);
1159 if (io_should_trigger_evfd(ctx
))
1160 eventfd_signal(ctx
->cq_ev_fd
, 1);
1163 /* Returns true if there are no backlogged entries after the flush */
1164 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1166 struct io_rings
*rings
= ctx
->rings
;
1167 struct io_uring_cqe
*cqe
;
1168 struct io_kiocb
*req
;
1169 unsigned long flags
;
1173 if (list_empty_careful(&ctx
->cq_overflow_list
))
1175 if ((ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
) ==
1176 rings
->cq_ring_entries
))
1180 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1182 /* if force is set, the ring is going away. always drop after that */
1184 ctx
->cq_overflow_flushed
= 1;
1187 while (!list_empty(&ctx
->cq_overflow_list
)) {
1188 cqe
= io_get_cqring(ctx
);
1192 req
= list_first_entry(&ctx
->cq_overflow_list
, struct io_kiocb
,
1194 list_move(&req
->list
, &list
);
1195 req
->flags
&= ~REQ_F_OVERFLOW
;
1197 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1198 WRITE_ONCE(cqe
->res
, req
->result
);
1199 WRITE_ONCE(cqe
->flags
, req
->cflags
);
1201 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1202 atomic_inc_return(&ctx
->cached_cq_overflow
));
1206 io_commit_cqring(ctx
);
1208 clear_bit(0, &ctx
->sq_check_overflow
);
1209 clear_bit(0, &ctx
->cq_check_overflow
);
1211 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1212 io_cqring_ev_posted(ctx
);
1214 while (!list_empty(&list
)) {
1215 req
= list_first_entry(&list
, struct io_kiocb
, list
);
1216 list_del(&req
->list
);
1223 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1225 struct io_ring_ctx
*ctx
= req
->ctx
;
1226 struct io_uring_cqe
*cqe
;
1228 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1231 * If we can't get a cq entry, userspace overflowed the
1232 * submission (by quite a lot). Increment the overflow count in
1235 cqe
= io_get_cqring(ctx
);
1237 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1238 WRITE_ONCE(cqe
->res
, res
);
1239 WRITE_ONCE(cqe
->flags
, cflags
);
1240 } else if (ctx
->cq_overflow_flushed
) {
1241 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1242 atomic_inc_return(&ctx
->cached_cq_overflow
));
1244 if (list_empty(&ctx
->cq_overflow_list
)) {
1245 set_bit(0, &ctx
->sq_check_overflow
);
1246 set_bit(0, &ctx
->cq_check_overflow
);
1248 req
->flags
|= REQ_F_OVERFLOW
;
1249 refcount_inc(&req
->refs
);
1251 req
->cflags
= cflags
;
1252 list_add_tail(&req
->list
, &ctx
->cq_overflow_list
);
1256 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1258 __io_cqring_fill_event(req
, res
, 0);
1261 static void __io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1263 struct io_ring_ctx
*ctx
= req
->ctx
;
1264 unsigned long flags
;
1266 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1267 __io_cqring_fill_event(req
, res
, cflags
);
1268 io_commit_cqring(ctx
);
1269 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1271 io_cqring_ev_posted(ctx
);
1274 static void io_cqring_add_event(struct io_kiocb
*req
, long res
)
1276 __io_cqring_add_event(req
, res
, 0);
1279 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1281 return req
== (struct io_kiocb
*)
1282 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1285 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1287 struct io_kiocb
*req
;
1289 req
= ctx
->fallback_req
;
1290 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1296 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1297 struct io_submit_state
*state
)
1299 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1300 struct io_kiocb
*req
;
1303 req
= kmem_cache_alloc(req_cachep
, gfp
);
1306 } else if (!state
->free_reqs
) {
1310 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1311 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
1314 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1315 * retry single alloc to be on the safe side.
1317 if (unlikely(ret
<= 0)) {
1318 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1319 if (!state
->reqs
[0])
1323 state
->free_reqs
= ret
- 1;
1324 req
= state
->reqs
[ret
- 1];
1327 req
= state
->reqs
[state
->free_reqs
];
1332 return io_get_fallback_req(ctx
);
1335 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
1339 percpu_ref_put(req
->fixed_file_refs
);
1344 static void __io_req_aux_free(struct io_kiocb
*req
)
1346 if (req
->flags
& REQ_F_NEED_CLEANUP
)
1347 io_cleanup_req(req
);
1351 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
1353 put_task_struct(req
->task
);
1355 io_req_work_drop_env(req
);
1358 static void __io_free_req(struct io_kiocb
*req
)
1360 __io_req_aux_free(req
);
1362 if (req
->flags
& REQ_F_INFLIGHT
) {
1363 struct io_ring_ctx
*ctx
= req
->ctx
;
1364 unsigned long flags
;
1366 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1367 list_del(&req
->inflight_entry
);
1368 if (waitqueue_active(&ctx
->inflight_wait
))
1369 wake_up(&ctx
->inflight_wait
);
1370 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1373 percpu_ref_put(&req
->ctx
->refs
);
1374 if (likely(!io_is_fallback_req(req
)))
1375 kmem_cache_free(req_cachep
, req
);
1377 clear_bit_unlock(0, (unsigned long *) &req
->ctx
->fallback_req
);
1381 void *reqs
[IO_IOPOLL_BATCH
];
1386 static void io_free_req_many(struct io_ring_ctx
*ctx
, struct req_batch
*rb
)
1390 if (rb
->need_iter
) {
1391 int i
, inflight
= 0;
1392 unsigned long flags
;
1394 for (i
= 0; i
< rb
->to_free
; i
++) {
1395 struct io_kiocb
*req
= rb
->reqs
[i
];
1397 if (req
->flags
& REQ_F_INFLIGHT
)
1399 __io_req_aux_free(req
);
1404 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1405 for (i
= 0; i
< rb
->to_free
; i
++) {
1406 struct io_kiocb
*req
= rb
->reqs
[i
];
1408 if (req
->flags
& REQ_F_INFLIGHT
) {
1409 list_del(&req
->inflight_entry
);
1414 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1416 if (waitqueue_active(&ctx
->inflight_wait
))
1417 wake_up(&ctx
->inflight_wait
);
1420 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
1421 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
1422 rb
->to_free
= rb
->need_iter
= 0;
1425 static bool io_link_cancel_timeout(struct io_kiocb
*req
)
1427 struct io_ring_ctx
*ctx
= req
->ctx
;
1430 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
1432 io_cqring_fill_event(req
, -ECANCELED
);
1433 io_commit_cqring(ctx
);
1434 req
->flags
&= ~REQ_F_LINK_HEAD
;
1442 static void io_req_link_next(struct io_kiocb
*req
, struct io_kiocb
**nxtptr
)
1444 struct io_ring_ctx
*ctx
= req
->ctx
;
1445 bool wake_ev
= false;
1447 /* Already got next link */
1448 if (req
->flags
& REQ_F_LINK_NEXT
)
1452 * The list should never be empty when we are called here. But could
1453 * potentially happen if the chain is messed up, check to be on the
1456 while (!list_empty(&req
->link_list
)) {
1457 struct io_kiocb
*nxt
= list_first_entry(&req
->link_list
,
1458 struct io_kiocb
, link_list
);
1460 if (unlikely((req
->flags
& REQ_F_LINK_TIMEOUT
) &&
1461 (nxt
->flags
& REQ_F_TIMEOUT
))) {
1462 list_del_init(&nxt
->link_list
);
1463 wake_ev
|= io_link_cancel_timeout(nxt
);
1464 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1468 list_del_init(&req
->link_list
);
1469 if (!list_empty(&nxt
->link_list
))
1470 nxt
->flags
|= REQ_F_LINK_HEAD
;
1475 req
->flags
|= REQ_F_LINK_NEXT
;
1477 io_cqring_ev_posted(ctx
);
1481 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1483 static void io_fail_links(struct io_kiocb
*req
)
1485 struct io_ring_ctx
*ctx
= req
->ctx
;
1486 unsigned long flags
;
1488 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1490 while (!list_empty(&req
->link_list
)) {
1491 struct io_kiocb
*link
= list_first_entry(&req
->link_list
,
1492 struct io_kiocb
, link_list
);
1494 list_del_init(&link
->link_list
);
1495 trace_io_uring_fail_link(req
, link
);
1497 if ((req
->flags
& REQ_F_LINK_TIMEOUT
) &&
1498 link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
1499 io_link_cancel_timeout(link
);
1501 io_cqring_fill_event(link
, -ECANCELED
);
1502 __io_double_put_req(link
);
1504 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1507 io_commit_cqring(ctx
);
1508 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1509 io_cqring_ev_posted(ctx
);
1512 static void io_req_find_next(struct io_kiocb
*req
, struct io_kiocb
**nxt
)
1514 if (likely(!(req
->flags
& REQ_F_LINK_HEAD
)))
1518 * If LINK is set, we have dependent requests in this chain. If we
1519 * didn't fail this request, queue the first one up, moving any other
1520 * dependencies to the next request. In case of failure, fail the rest
1523 if (req
->flags
& REQ_F_FAIL_LINK
) {
1525 } else if ((req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_COMP_LOCKED
)) ==
1526 REQ_F_LINK_TIMEOUT
) {
1527 struct io_ring_ctx
*ctx
= req
->ctx
;
1528 unsigned long flags
;
1531 * If this is a timeout link, we could be racing with the
1532 * timeout timer. Grab the completion lock for this case to
1533 * protect against that.
1535 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1536 io_req_link_next(req
, nxt
);
1537 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1539 io_req_link_next(req
, nxt
);
1543 static void io_free_req(struct io_kiocb
*req
)
1545 struct io_kiocb
*nxt
= NULL
;
1547 io_req_find_next(req
, &nxt
);
1551 io_queue_async_work(nxt
);
1554 static void io_link_work_cb(struct io_wq_work
**workptr
)
1556 struct io_kiocb
*req
= container_of(*workptr
, struct io_kiocb
, work
);
1557 struct io_kiocb
*link
;
1559 link
= list_first_entry(&req
->link_list
, struct io_kiocb
, link_list
);
1560 io_queue_linked_timeout(link
);
1561 io_wq_submit_work(workptr
);
1564 static void io_wq_assign_next(struct io_wq_work
**workptr
, struct io_kiocb
*nxt
)
1566 struct io_kiocb
*link
;
1567 const struct io_op_def
*def
= &io_op_defs
[nxt
->opcode
];
1569 if ((nxt
->flags
& REQ_F_ISREG
) && def
->hash_reg_file
)
1570 io_wq_hash_work(&nxt
->work
, file_inode(nxt
->file
));
1572 *workptr
= &nxt
->work
;
1573 link
= io_prep_linked_timeout(nxt
);
1575 nxt
->work
.func
= io_link_work_cb
;
1579 * Drop reference to request, return next in chain (if there is one) if this
1580 * was the last reference to this request.
1582 __attribute__((nonnull
))
1583 static void io_put_req_find_next(struct io_kiocb
*req
, struct io_kiocb
**nxtptr
)
1585 if (refcount_dec_and_test(&req
->refs
)) {
1586 io_req_find_next(req
, nxtptr
);
1591 static void io_put_req(struct io_kiocb
*req
)
1593 if (refcount_dec_and_test(&req
->refs
))
1597 static void io_steal_work(struct io_kiocb
*req
,
1598 struct io_wq_work
**workptr
)
1601 * It's in an io-wq worker, so there always should be at least
1602 * one reference, which will be dropped in io_put_work() just
1603 * after the current handler returns.
1605 * It also means, that if the counter dropped to 1, then there is
1606 * no asynchronous users left, so it's safe to steal the next work.
1608 if (refcount_read(&req
->refs
) == 1) {
1609 struct io_kiocb
*nxt
= NULL
;
1611 io_req_find_next(req
, &nxt
);
1613 io_wq_assign_next(workptr
, nxt
);
1618 * Must only be used if we don't need to care about links, usually from
1619 * within the completion handling itself.
1621 static void __io_double_put_req(struct io_kiocb
*req
)
1623 /* drop both submit and complete references */
1624 if (refcount_sub_and_test(2, &req
->refs
))
1628 static void io_double_put_req(struct io_kiocb
*req
)
1630 /* drop both submit and complete references */
1631 if (refcount_sub_and_test(2, &req
->refs
))
1635 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
, bool noflush
)
1637 struct io_rings
*rings
= ctx
->rings
;
1639 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1641 * noflush == true is from the waitqueue handler, just ensure
1642 * we wake up the task, and the next invocation will flush the
1643 * entries. We cannot safely to it from here.
1645 if (noflush
&& !list_empty(&ctx
->cq_overflow_list
))
1648 io_cqring_overflow_flush(ctx
, false);
1651 /* See comment at the top of this file */
1653 return ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
);
1656 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
1658 struct io_rings
*rings
= ctx
->rings
;
1660 /* make sure SQ entry isn't read before tail */
1661 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
1664 static inline bool io_req_multi_free(struct req_batch
*rb
, struct io_kiocb
*req
)
1666 if ((req
->flags
& REQ_F_LINK_HEAD
) || io_is_fallback_req(req
))
1669 if (req
->file
|| req
->io
)
1672 rb
->reqs
[rb
->to_free
++] = req
;
1673 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
1674 io_free_req_many(req
->ctx
, rb
);
1678 static int io_put_kbuf(struct io_kiocb
*req
)
1680 struct io_buffer
*kbuf
;
1683 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
1684 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
1685 cflags
|= IORING_CQE_F_BUFFER
;
1692 * Find and free completed poll iocbs
1694 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1695 struct list_head
*done
)
1697 struct req_batch rb
;
1698 struct io_kiocb
*req
;
1700 rb
.to_free
= rb
.need_iter
= 0;
1701 while (!list_empty(done
)) {
1704 req
= list_first_entry(done
, struct io_kiocb
, list
);
1705 list_del(&req
->list
);
1707 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
1708 cflags
= io_put_kbuf(req
);
1710 __io_cqring_fill_event(req
, req
->result
, cflags
);
1713 if (refcount_dec_and_test(&req
->refs
) &&
1714 !io_req_multi_free(&rb
, req
))
1718 io_commit_cqring(ctx
);
1719 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
1720 io_cqring_ev_posted(ctx
);
1721 io_free_req_many(ctx
, &rb
);
1724 static void io_iopoll_queue(struct list_head
*again
)
1726 struct io_kiocb
*req
;
1729 req
= list_first_entry(again
, struct io_kiocb
, list
);
1730 list_del(&req
->list
);
1731 refcount_inc(&req
->refs
);
1732 io_queue_async_work(req
);
1733 } while (!list_empty(again
));
1736 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1739 struct io_kiocb
*req
, *tmp
;
1746 * Only spin for completions if we don't have multiple devices hanging
1747 * off our complete list, and we're under the requested amount.
1749 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
1752 list_for_each_entry_safe(req
, tmp
, &ctx
->poll_list
, list
) {
1753 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
1756 * Move completed and retryable entries to our local lists.
1757 * If we find a request that requires polling, break out
1758 * and complete those lists first, if we have entries there.
1760 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
) {
1761 list_move_tail(&req
->list
, &done
);
1764 if (!list_empty(&done
))
1767 if (req
->result
== -EAGAIN
) {
1768 list_move_tail(&req
->list
, &again
);
1771 if (!list_empty(&again
))
1774 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
1783 if (!list_empty(&done
))
1784 io_iopoll_complete(ctx
, nr_events
, &done
);
1786 if (!list_empty(&again
))
1787 io_iopoll_queue(&again
);
1793 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1794 * non-spinning poll check - we'll still enter the driver poll loop, but only
1795 * as a non-spinning completion check.
1797 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1800 while (!list_empty(&ctx
->poll_list
) && !need_resched()) {
1803 ret
= io_do_iopoll(ctx
, nr_events
, min
);
1806 if (!min
|| *nr_events
>= min
)
1814 * We can't just wait for polled events to come to us, we have to actively
1815 * find and complete them.
1817 static void io_iopoll_reap_events(struct io_ring_ctx
*ctx
)
1819 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1822 mutex_lock(&ctx
->uring_lock
);
1823 while (!list_empty(&ctx
->poll_list
)) {
1824 unsigned int nr_events
= 0;
1826 io_iopoll_getevents(ctx
, &nr_events
, 1);
1829 * Ensure we allow local-to-the-cpu processing to take place,
1830 * in this case we need to ensure that we reap all events.
1834 mutex_unlock(&ctx
->uring_lock
);
1837 static int io_iopoll_check(struct io_ring_ctx
*ctx
, unsigned *nr_events
,
1840 int iters
= 0, ret
= 0;
1843 * We disallow the app entering submit/complete with polling, but we
1844 * still need to lock the ring to prevent racing with polled issue
1845 * that got punted to a workqueue.
1847 mutex_lock(&ctx
->uring_lock
);
1852 * Don't enter poll loop if we already have events pending.
1853 * If we do, we can potentially be spinning for commands that
1854 * already triggered a CQE (eg in error).
1856 if (io_cqring_events(ctx
, false))
1860 * If a submit got punted to a workqueue, we can have the
1861 * application entering polling for a command before it gets
1862 * issued. That app will hold the uring_lock for the duration
1863 * of the poll right here, so we need to take a breather every
1864 * now and then to ensure that the issue has a chance to add
1865 * the poll to the issued list. Otherwise we can spin here
1866 * forever, while the workqueue is stuck trying to acquire the
1869 if (!(++iters
& 7)) {
1870 mutex_unlock(&ctx
->uring_lock
);
1871 mutex_lock(&ctx
->uring_lock
);
1874 if (*nr_events
< min
)
1875 tmin
= min
- *nr_events
;
1877 ret
= io_iopoll_getevents(ctx
, nr_events
, tmin
);
1881 } while (min
&& !*nr_events
&& !need_resched());
1883 mutex_unlock(&ctx
->uring_lock
);
1887 static void kiocb_end_write(struct io_kiocb
*req
)
1890 * Tell lockdep we inherited freeze protection from submission
1893 if (req
->flags
& REQ_F_ISREG
) {
1894 struct inode
*inode
= file_inode(req
->file
);
1896 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
1898 file_end_write(req
->file
);
1901 static inline void req_set_fail_links(struct io_kiocb
*req
)
1903 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1904 req
->flags
|= REQ_F_FAIL_LINK
;
1907 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
)
1909 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
1912 if (kiocb
->ki_flags
& IOCB_WRITE
)
1913 kiocb_end_write(req
);
1915 if (res
!= req
->result
)
1916 req_set_fail_links(req
);
1917 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
1918 cflags
= io_put_kbuf(req
);
1919 __io_cqring_add_event(req
, res
, cflags
);
1922 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
1924 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
1926 io_complete_rw_common(kiocb
, res
);
1930 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
1932 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
1934 if (kiocb
->ki_flags
& IOCB_WRITE
)
1935 kiocb_end_write(req
);
1937 if (res
!= req
->result
)
1938 req_set_fail_links(req
);
1941 req
->flags
|= REQ_F_IOPOLL_COMPLETED
;
1945 * After the iocb has been issued, it's safe to be found on the poll list.
1946 * Adding the kiocb to the list AFTER submission ensures that we don't
1947 * find it from a io_iopoll_getevents() thread before the issuer is done
1948 * accessing the kiocb cookie.
1950 static void io_iopoll_req_issued(struct io_kiocb
*req
)
1952 struct io_ring_ctx
*ctx
= req
->ctx
;
1955 * Track whether we have multiple files in our lists. This will impact
1956 * how we do polling eventually, not spinning if we're on potentially
1957 * different devices.
1959 if (list_empty(&ctx
->poll_list
)) {
1960 ctx
->poll_multi_file
= false;
1961 } else if (!ctx
->poll_multi_file
) {
1962 struct io_kiocb
*list_req
;
1964 list_req
= list_first_entry(&ctx
->poll_list
, struct io_kiocb
,
1966 if (list_req
->file
!= req
->file
)
1967 ctx
->poll_multi_file
= true;
1971 * For fast devices, IO may have already completed. If it has, add
1972 * it to the front so we find it first.
1974 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
)
1975 list_add(&req
->list
, &ctx
->poll_list
);
1977 list_add_tail(&req
->list
, &ctx
->poll_list
);
1979 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
1980 wq_has_sleeper(&ctx
->sqo_wait
))
1981 wake_up(&ctx
->sqo_wait
);
1984 static void io_file_put(struct io_submit_state
*state
)
1987 int diff
= state
->has_refs
- state
->used_refs
;
1990 fput_many(state
->file
, diff
);
1996 * Get as many references to a file as we have IOs left in this submission,
1997 * assuming most submissions are for one file, or at least that each file
1998 * has more than one submission.
2000 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2006 if (state
->fd
== fd
) {
2013 state
->file
= fget_many(fd
, state
->ios_left
);
2018 state
->has_refs
= state
->ios_left
;
2019 state
->used_refs
= 1;
2025 * If we tracked the file through the SCM inflight mechanism, we could support
2026 * any file. For now, just ensure that anything potentially problematic is done
2029 static bool io_file_supports_async(struct file
*file
, int rw
)
2031 umode_t mode
= file_inode(file
)->i_mode
;
2033 if (S_ISBLK(mode
) || S_ISCHR(mode
) || S_ISSOCK(mode
))
2035 if (S_ISREG(mode
) && file
->f_op
!= &io_uring_fops
)
2038 if (!(file
->f_mode
& FMODE_NOWAIT
))
2042 return file
->f_op
->read_iter
!= NULL
;
2044 return file
->f_op
->write_iter
!= NULL
;
2047 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2048 bool force_nonblock
)
2050 struct io_ring_ctx
*ctx
= req
->ctx
;
2051 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2055 if (S_ISREG(file_inode(req
->file
)->i_mode
))
2056 req
->flags
|= REQ_F_ISREG
;
2058 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2059 if (kiocb
->ki_pos
== -1 && !(req
->file
->f_mode
& FMODE_STREAM
)) {
2060 req
->flags
|= REQ_F_CUR_POS
;
2061 kiocb
->ki_pos
= req
->file
->f_pos
;
2063 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2064 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2065 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2069 ioprio
= READ_ONCE(sqe
->ioprio
);
2071 ret
= ioprio_check_cap(ioprio
);
2075 kiocb
->ki_ioprio
= ioprio
;
2077 kiocb
->ki_ioprio
= get_current_ioprio();
2079 /* don't allow async punt if RWF_NOWAIT was requested */
2080 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) ||
2081 (req
->file
->f_flags
& O_NONBLOCK
))
2082 req
->flags
|= REQ_F_NOWAIT
;
2085 kiocb
->ki_flags
|= IOCB_NOWAIT
;
2087 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2088 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2089 !kiocb
->ki_filp
->f_op
->iopoll
)
2092 kiocb
->ki_flags
|= IOCB_HIPRI
;
2093 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2096 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2098 kiocb
->ki_complete
= io_complete_rw
;
2101 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2102 req
->rw
.len
= READ_ONCE(sqe
->len
);
2103 /* we own ->private, reuse it for the buffer index / buffer ID */
2104 req
->rw
.kiocb
.private = (void *) (unsigned long)
2105 READ_ONCE(sqe
->buf_index
);
2109 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2115 case -ERESTARTNOINTR
:
2116 case -ERESTARTNOHAND
:
2117 case -ERESTART_RESTARTBLOCK
:
2119 * We can't just restart the syscall, since previously
2120 * submitted sqes may already be in progress. Just fail this
2126 kiocb
->ki_complete(kiocb
, ret
, 0);
2130 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
)
2132 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2134 if (req
->flags
& REQ_F_CUR_POS
)
2135 req
->file
->f_pos
= kiocb
->ki_pos
;
2136 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2137 io_complete_rw(kiocb
, ret
, 0);
2139 io_rw_done(kiocb
, ret
);
2142 static ssize_t
io_import_fixed(struct io_kiocb
*req
, int rw
,
2143 struct iov_iter
*iter
)
2145 struct io_ring_ctx
*ctx
= req
->ctx
;
2146 size_t len
= req
->rw
.len
;
2147 struct io_mapped_ubuf
*imu
;
2148 unsigned index
, buf_index
;
2152 /* attempt to use fixed buffers without having provided iovecs */
2153 if (unlikely(!ctx
->user_bufs
))
2156 buf_index
= (unsigned long) req
->rw
.kiocb
.private;
2157 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
2160 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
2161 imu
= &ctx
->user_bufs
[index
];
2162 buf_addr
= req
->rw
.addr
;
2165 if (buf_addr
+ len
< buf_addr
)
2167 /* not inside the mapped region */
2168 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
2172 * May not be a start of buffer, set size appropriately
2173 * and advance us to the beginning.
2175 offset
= buf_addr
- imu
->ubuf
;
2176 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
2180 * Don't use iov_iter_advance() here, as it's really slow for
2181 * using the latter parts of a big fixed buffer - it iterates
2182 * over each segment manually. We can cheat a bit here, because
2185 * 1) it's a BVEC iter, we set it up
2186 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2187 * first and last bvec
2189 * So just find our index, and adjust the iterator afterwards.
2190 * If the offset is within the first bvec (or the whole first
2191 * bvec, just use iov_iter_advance(). This makes it easier
2192 * since we can just skip the first segment, which may not
2193 * be PAGE_SIZE aligned.
2195 const struct bio_vec
*bvec
= imu
->bvec
;
2197 if (offset
<= bvec
->bv_len
) {
2198 iov_iter_advance(iter
, offset
);
2200 unsigned long seg_skip
;
2202 /* skip first vec */
2203 offset
-= bvec
->bv_len
;
2204 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
2206 iter
->bvec
= bvec
+ seg_skip
;
2207 iter
->nr_segs
-= seg_skip
;
2208 iter
->count
-= bvec
->bv_len
+ offset
;
2209 iter
->iov_offset
= offset
& ~PAGE_MASK
;
2216 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2219 mutex_unlock(&ctx
->uring_lock
);
2222 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2225 * "Normal" inline submissions always hold the uring_lock, since we
2226 * grab it from the system call. Same is true for the SQPOLL offload.
2227 * The only exception is when we've detached the request and issue it
2228 * from an async worker thread, grab the lock for that case.
2231 mutex_lock(&ctx
->uring_lock
);
2234 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
2235 int bgid
, struct io_buffer
*kbuf
,
2238 struct io_buffer
*head
;
2240 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2243 io_ring_submit_lock(req
->ctx
, needs_lock
);
2245 lockdep_assert_held(&req
->ctx
->uring_lock
);
2247 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
2249 if (!list_empty(&head
->list
)) {
2250 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
2252 list_del(&kbuf
->list
);
2255 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
2257 if (*len
> kbuf
->len
)
2260 kbuf
= ERR_PTR(-ENOBUFS
);
2263 io_ring_submit_unlock(req
->ctx
, needs_lock
);
2268 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
2271 struct io_buffer
*kbuf
;
2274 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2275 bgid
= (int) (unsigned long) req
->rw
.kiocb
.private;
2276 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
2279 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
2280 req
->flags
|= REQ_F_BUFFER_SELECTED
;
2281 return u64_to_user_ptr(kbuf
->addr
);
2284 #ifdef CONFIG_COMPAT
2285 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
2288 struct compat_iovec __user
*uiov
;
2289 compat_ssize_t clen
;
2293 uiov
= u64_to_user_ptr(req
->rw
.addr
);
2294 if (!access_ok(uiov
, sizeof(*uiov
)))
2296 if (__get_user(clen
, &uiov
->iov_len
))
2302 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2304 return PTR_ERR(buf
);
2305 iov
[0].iov_base
= buf
;
2306 iov
[0].iov_len
= (compat_size_t
) len
;
2311 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2314 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
2318 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
2321 len
= iov
[0].iov_len
;
2324 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2326 return PTR_ERR(buf
);
2327 iov
[0].iov_base
= buf
;
2328 iov
[0].iov_len
= len
;
2332 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2335 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2339 else if (req
->rw
.len
> 1)
2342 #ifdef CONFIG_COMPAT
2343 if (req
->ctx
->compat
)
2344 return io_compat_import(req
, iov
, needs_lock
);
2347 return __io_iov_buffer_select(req
, iov
, needs_lock
);
2350 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
2351 struct iovec
**iovec
, struct iov_iter
*iter
,
2354 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
2355 size_t sqe_len
= req
->rw
.len
;
2359 opcode
= req
->opcode
;
2360 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
2362 return io_import_fixed(req
, rw
, iter
);
2365 /* buffer index only valid with fixed read/write, or buffer select */
2366 if (req
->rw
.kiocb
.private && !(req
->flags
& REQ_F_BUFFER_SELECT
))
2369 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
2370 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
2371 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
2374 return PTR_ERR(buf
);
2376 req
->rw
.len
= sqe_len
;
2379 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
2381 return ret
< 0 ? ret
: sqe_len
;
2385 struct io_async_rw
*iorw
= &req
->io
->rw
;
2388 iov_iter_init(iter
, rw
, *iovec
, iorw
->nr_segs
, iorw
->size
);
2389 if (iorw
->iov
== iorw
->fast_iov
)
2394 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
2395 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
2397 ret
= (*iovec
)->iov_len
;
2398 iov_iter_init(iter
, rw
, *iovec
, 1, ret
);
2404 #ifdef CONFIG_COMPAT
2405 if (req
->ctx
->compat
)
2406 return compat_import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
,
2410 return import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
);
2414 * For files that don't have ->read_iter() and ->write_iter(), handle them
2415 * by looping over ->read() or ->write() manually.
2417 static ssize_t
loop_rw_iter(int rw
, struct file
*file
, struct kiocb
*kiocb
,
2418 struct iov_iter
*iter
)
2423 * Don't support polled IO through this interface, and we can't
2424 * support non-blocking either. For the latter, this just causes
2425 * the kiocb to be handled from an async context.
2427 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2429 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2432 while (iov_iter_count(iter
)) {
2436 if (!iov_iter_is_bvec(iter
)) {
2437 iovec
= iov_iter_iovec(iter
);
2439 /* fixed buffers import bvec */
2440 iovec
.iov_base
= kmap(iter
->bvec
->bv_page
)
2442 iovec
.iov_len
= min(iter
->count
,
2443 iter
->bvec
->bv_len
- iter
->iov_offset
);
2447 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
2448 iovec
.iov_len
, &kiocb
->ki_pos
);
2450 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
2451 iovec
.iov_len
, &kiocb
->ki_pos
);
2454 if (iov_iter_is_bvec(iter
))
2455 kunmap(iter
->bvec
->bv_page
);
2463 if (nr
!= iovec
.iov_len
)
2465 iov_iter_advance(iter
, nr
);
2471 static void io_req_map_rw(struct io_kiocb
*req
, ssize_t io_size
,
2472 struct iovec
*iovec
, struct iovec
*fast_iov
,
2473 struct iov_iter
*iter
)
2475 req
->io
->rw
.nr_segs
= iter
->nr_segs
;
2476 req
->io
->rw
.size
= io_size
;
2477 req
->io
->rw
.iov
= iovec
;
2478 if (!req
->io
->rw
.iov
) {
2479 req
->io
->rw
.iov
= req
->io
->rw
.fast_iov
;
2480 if (req
->io
->rw
.iov
!= fast_iov
)
2481 memcpy(req
->io
->rw
.iov
, fast_iov
,
2482 sizeof(struct iovec
) * iter
->nr_segs
);
2484 req
->flags
|= REQ_F_NEED_CLEANUP
;
2488 static inline int __io_alloc_async_ctx(struct io_kiocb
*req
)
2490 req
->io
= kmalloc(sizeof(*req
->io
), GFP_KERNEL
);
2491 return req
->io
== NULL
;
2494 static int io_alloc_async_ctx(struct io_kiocb
*req
)
2496 if (!io_op_defs
[req
->opcode
].async_ctx
)
2499 return __io_alloc_async_ctx(req
);
2502 static int io_setup_async_rw(struct io_kiocb
*req
, ssize_t io_size
,
2503 struct iovec
*iovec
, struct iovec
*fast_iov
,
2504 struct iov_iter
*iter
)
2506 if (!io_op_defs
[req
->opcode
].async_ctx
)
2509 if (__io_alloc_async_ctx(req
))
2512 io_req_map_rw(req
, io_size
, iovec
, fast_iov
, iter
);
2517 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2518 bool force_nonblock
)
2520 struct io_async_ctx
*io
;
2521 struct iov_iter iter
;
2524 ret
= io_prep_rw(req
, sqe
, force_nonblock
);
2528 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
2531 /* either don't need iovec imported or already have it */
2532 if (!req
->io
|| req
->flags
& REQ_F_NEED_CLEANUP
)
2536 io
->rw
.iov
= io
->rw
.fast_iov
;
2538 ret
= io_import_iovec(READ
, req
, &io
->rw
.iov
, &iter
, !force_nonblock
);
2543 io_req_map_rw(req
, ret
, io
->rw
.iov
, io
->rw
.fast_iov
, &iter
);
2547 static int io_read(struct io_kiocb
*req
, bool force_nonblock
)
2549 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2550 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2551 struct iov_iter iter
;
2553 ssize_t io_size
, ret
;
2555 ret
= io_import_iovec(READ
, req
, &iovec
, &iter
, !force_nonblock
);
2559 /* Ensure we clear previously set non-block flag */
2560 if (!force_nonblock
)
2561 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
2565 if (req
->flags
& REQ_F_LINK_HEAD
)
2566 req
->result
= io_size
;
2569 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2570 * we know to async punt it even if it was opened O_NONBLOCK
2572 if (force_nonblock
&& !io_file_supports_async(req
->file
, READ
))
2575 iov_count
= iov_iter_count(&iter
);
2576 ret
= rw_verify_area(READ
, req
->file
, &kiocb
->ki_pos
, iov_count
);
2580 if (req
->file
->f_op
->read_iter
)
2581 ret2
= call_read_iter(req
->file
, kiocb
, &iter
);
2583 ret2
= loop_rw_iter(READ
, req
->file
, kiocb
, &iter
);
2585 /* Catch -EAGAIN return for forced non-blocking submission */
2586 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
2587 kiocb_done(kiocb
, ret2
);
2590 ret
= io_setup_async_rw(req
, io_size
, iovec
,
2591 inline_vecs
, &iter
);
2594 /* any defer here is final, must blocking retry */
2595 if (!(req
->flags
& REQ_F_NOWAIT
) &&
2596 !file_can_poll(req
->file
))
2597 req
->flags
|= REQ_F_MUST_PUNT
;
2603 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
2607 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2608 bool force_nonblock
)
2610 struct io_async_ctx
*io
;
2611 struct iov_iter iter
;
2614 ret
= io_prep_rw(req
, sqe
, force_nonblock
);
2618 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
2621 req
->fsize
= rlimit(RLIMIT_FSIZE
);
2623 /* either don't need iovec imported or already have it */
2624 if (!req
->io
|| req
->flags
& REQ_F_NEED_CLEANUP
)
2628 io
->rw
.iov
= io
->rw
.fast_iov
;
2630 ret
= io_import_iovec(WRITE
, req
, &io
->rw
.iov
, &iter
, !force_nonblock
);
2635 io_req_map_rw(req
, ret
, io
->rw
.iov
, io
->rw
.fast_iov
, &iter
);
2639 static int io_write(struct io_kiocb
*req
, bool force_nonblock
)
2641 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2642 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2643 struct iov_iter iter
;
2645 ssize_t ret
, io_size
;
2647 ret
= io_import_iovec(WRITE
, req
, &iovec
, &iter
, !force_nonblock
);
2651 /* Ensure we clear previously set non-block flag */
2652 if (!force_nonblock
)
2653 req
->rw
.kiocb
.ki_flags
&= ~IOCB_NOWAIT
;
2657 if (req
->flags
& REQ_F_LINK_HEAD
)
2658 req
->result
= io_size
;
2661 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2662 * we know to async punt it even if it was opened O_NONBLOCK
2664 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
2667 /* file path doesn't support NOWAIT for non-direct_IO */
2668 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
2669 (req
->flags
& REQ_F_ISREG
))
2672 iov_count
= iov_iter_count(&iter
);
2673 ret
= rw_verify_area(WRITE
, req
->file
, &kiocb
->ki_pos
, iov_count
);
2678 * Open-code file_start_write here to grab freeze protection,
2679 * which will be released by another thread in
2680 * io_complete_rw(). Fool lockdep by telling it the lock got
2681 * released so that it doesn't complain about the held lock when
2682 * we return to userspace.
2684 if (req
->flags
& REQ_F_ISREG
) {
2685 __sb_start_write(file_inode(req
->file
)->i_sb
,
2686 SB_FREEZE_WRITE
, true);
2687 __sb_writers_release(file_inode(req
->file
)->i_sb
,
2690 kiocb
->ki_flags
|= IOCB_WRITE
;
2692 if (!force_nonblock
)
2693 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= req
->fsize
;
2695 if (req
->file
->f_op
->write_iter
)
2696 ret2
= call_write_iter(req
->file
, kiocb
, &iter
);
2698 ret2
= loop_rw_iter(WRITE
, req
->file
, kiocb
, &iter
);
2700 if (!force_nonblock
)
2701 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= RLIM_INFINITY
;
2704 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2705 * retry them without IOCB_NOWAIT.
2707 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
2709 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
2710 kiocb_done(kiocb
, ret2
);
2713 ret
= io_setup_async_rw(req
, io_size
, iovec
,
2714 inline_vecs
, &iter
);
2717 /* any defer here is final, must blocking retry */
2718 if (!file_can_poll(req
->file
))
2719 req
->flags
|= REQ_F_MUST_PUNT
;
2724 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
2729 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2731 struct io_splice
* sp
= &req
->splice
;
2732 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
2735 if (req
->flags
& REQ_F_NEED_CLEANUP
)
2739 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
2740 sp
->off_out
= READ_ONCE(sqe
->off
);
2741 sp
->len
= READ_ONCE(sqe
->len
);
2742 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
2744 if (unlikely(sp
->flags
& ~valid_flags
))
2747 ret
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
), &sp
->file_in
,
2748 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
2751 req
->flags
|= REQ_F_NEED_CLEANUP
;
2753 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
))
2754 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
2759 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
2761 struct io_splice
*sp
= &req
->splice
;
2762 struct file
*in
= sp
->file_in
;
2763 struct file
*out
= sp
->file_out
;
2764 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
2765 loff_t
*poff_in
, *poff_out
;
2771 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
2772 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
2775 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
2776 if (force_nonblock
&& ret
== -EAGAIN
)
2780 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
2781 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
2783 io_cqring_add_event(req
, ret
);
2785 req_set_fail_links(req
);
2791 * IORING_OP_NOP just posts a completion event, nothing else.
2793 static int io_nop(struct io_kiocb
*req
)
2795 struct io_ring_ctx
*ctx
= req
->ctx
;
2797 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
2800 io_cqring_add_event(req
, 0);
2805 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2807 struct io_ring_ctx
*ctx
= req
->ctx
;
2812 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
2814 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
2817 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
2818 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
2821 req
->sync
.off
= READ_ONCE(sqe
->off
);
2822 req
->sync
.len
= READ_ONCE(sqe
->len
);
2826 static bool io_req_cancelled(struct io_kiocb
*req
)
2828 if (req
->work
.flags
& IO_WQ_WORK_CANCEL
) {
2829 req_set_fail_links(req
);
2830 io_cqring_add_event(req
, -ECANCELED
);
2838 static void __io_fsync(struct io_kiocb
*req
)
2840 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
2843 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
2844 end
> 0 ? end
: LLONG_MAX
,
2845 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
2847 req_set_fail_links(req
);
2848 io_cqring_add_event(req
, ret
);
2852 static void io_fsync_finish(struct io_wq_work
**workptr
)
2854 struct io_kiocb
*req
= container_of(*workptr
, struct io_kiocb
, work
);
2856 if (io_req_cancelled(req
))
2859 io_steal_work(req
, workptr
);
2862 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
2864 /* fsync always requires a blocking context */
2865 if (force_nonblock
) {
2866 req
->work
.func
= io_fsync_finish
;
2873 static void __io_fallocate(struct io_kiocb
*req
)
2877 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= req
->fsize
;
2878 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
2880 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= RLIM_INFINITY
;
2882 req_set_fail_links(req
);
2883 io_cqring_add_event(req
, ret
);
2887 static void io_fallocate_finish(struct io_wq_work
**workptr
)
2889 struct io_kiocb
*req
= container_of(*workptr
, struct io_kiocb
, work
);
2891 if (io_req_cancelled(req
))
2893 __io_fallocate(req
);
2894 io_steal_work(req
, workptr
);
2897 static int io_fallocate_prep(struct io_kiocb
*req
,
2898 const struct io_uring_sqe
*sqe
)
2900 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
2903 req
->sync
.off
= READ_ONCE(sqe
->off
);
2904 req
->sync
.len
= READ_ONCE(sqe
->addr
);
2905 req
->sync
.mode
= READ_ONCE(sqe
->len
);
2906 req
->fsize
= rlimit(RLIMIT_FSIZE
);
2910 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
2912 /* fallocate always requiring blocking context */
2913 if (force_nonblock
) {
2914 req
->work
.func
= io_fallocate_finish
;
2918 __io_fallocate(req
);
2922 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2924 const char __user
*fname
;
2927 if (sqe
->ioprio
|| sqe
->buf_index
)
2929 if (req
->flags
& REQ_F_FIXED_FILE
)
2931 if (req
->flags
& REQ_F_NEED_CLEANUP
)
2934 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
2935 req
->open
.how
.mode
= READ_ONCE(sqe
->len
);
2936 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
2937 req
->open
.how
.flags
= READ_ONCE(sqe
->open_flags
);
2938 if (force_o_largefile())
2939 req
->open
.how
.flags
|= O_LARGEFILE
;
2941 req
->open
.filename
= getname(fname
);
2942 if (IS_ERR(req
->open
.filename
)) {
2943 ret
= PTR_ERR(req
->open
.filename
);
2944 req
->open
.filename
= NULL
;
2948 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
2949 req
->flags
|= REQ_F_NEED_CLEANUP
;
2953 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
2955 struct open_how __user
*how
;
2956 const char __user
*fname
;
2960 if (sqe
->ioprio
|| sqe
->buf_index
)
2962 if (req
->flags
& REQ_F_FIXED_FILE
)
2964 if (req
->flags
& REQ_F_NEED_CLEANUP
)
2967 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
2968 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
2969 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
2970 len
= READ_ONCE(sqe
->len
);
2972 if (len
< OPEN_HOW_SIZE_VER0
)
2975 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
2980 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
2981 req
->open
.how
.flags
|= O_LARGEFILE
;
2983 req
->open
.filename
= getname(fname
);
2984 if (IS_ERR(req
->open
.filename
)) {
2985 ret
= PTR_ERR(req
->open
.filename
);
2986 req
->open
.filename
= NULL
;
2990 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
2991 req
->flags
|= REQ_F_NEED_CLEANUP
;
2995 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
2997 struct open_flags op
;
3004 ret
= build_open_flags(&req
->open
.how
, &op
);
3008 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
3012 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
3015 ret
= PTR_ERR(file
);
3017 fsnotify_open(file
);
3018 fd_install(ret
, file
);
3021 putname(req
->open
.filename
);
3022 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3024 req_set_fail_links(req
);
3025 io_cqring_add_event(req
, ret
);
3030 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
3032 req
->open
.how
= build_open_how(req
->open
.how
.flags
, req
->open
.how
.mode
);
3033 return io_openat2(req
, force_nonblock
);
3036 static int io_remove_buffers_prep(struct io_kiocb
*req
,
3037 const struct io_uring_sqe
*sqe
)
3039 struct io_provide_buf
*p
= &req
->pbuf
;
3042 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
3045 tmp
= READ_ONCE(sqe
->fd
);
3046 if (!tmp
|| tmp
> USHRT_MAX
)
3049 memset(p
, 0, sizeof(*p
));
3051 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3055 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
3056 int bgid
, unsigned nbufs
)
3060 /* shouldn't happen */
3064 /* the head kbuf is the list itself */
3065 while (!list_empty(&buf
->list
)) {
3066 struct io_buffer
*nxt
;
3068 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
3069 list_del(&nxt
->list
);
3076 idr_remove(&ctx
->io_buffer_idr
, bgid
);
3081 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
)
3083 struct io_provide_buf
*p
= &req
->pbuf
;
3084 struct io_ring_ctx
*ctx
= req
->ctx
;
3085 struct io_buffer
*head
;
3088 io_ring_submit_lock(ctx
, !force_nonblock
);
3090 lockdep_assert_held(&ctx
->uring_lock
);
3093 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3095 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
3097 io_ring_submit_lock(ctx
, !force_nonblock
);
3099 req_set_fail_links(req
);
3100 io_cqring_add_event(req
, ret
);
3105 static int io_provide_buffers_prep(struct io_kiocb
*req
,
3106 const struct io_uring_sqe
*sqe
)
3108 struct io_provide_buf
*p
= &req
->pbuf
;
3111 if (sqe
->ioprio
|| sqe
->rw_flags
)
3114 tmp
= READ_ONCE(sqe
->fd
);
3115 if (!tmp
|| tmp
> USHRT_MAX
)
3118 p
->addr
= READ_ONCE(sqe
->addr
);
3119 p
->len
= READ_ONCE(sqe
->len
);
3121 if (!access_ok(u64_to_user_ptr(p
->addr
), p
->len
))
3124 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3125 tmp
= READ_ONCE(sqe
->off
);
3126 if (tmp
> USHRT_MAX
)
3132 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
3134 struct io_buffer
*buf
;
3135 u64 addr
= pbuf
->addr
;
3136 int i
, bid
= pbuf
->bid
;
3138 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
3139 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
3144 buf
->len
= pbuf
->len
;
3149 INIT_LIST_HEAD(&buf
->list
);
3152 list_add_tail(&buf
->list
, &(*head
)->list
);
3156 return i
? i
: -ENOMEM
;
3159 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
)
3161 struct io_provide_buf
*p
= &req
->pbuf
;
3162 struct io_ring_ctx
*ctx
= req
->ctx
;
3163 struct io_buffer
*head
, *list
;
3166 io_ring_submit_lock(ctx
, !force_nonblock
);
3168 lockdep_assert_held(&ctx
->uring_lock
);
3170 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3172 ret
= io_add_buffers(p
, &head
);
3177 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
3180 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
3185 io_ring_submit_unlock(ctx
, !force_nonblock
);
3187 req_set_fail_links(req
);
3188 io_cqring_add_event(req
, ret
);
3193 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
3194 const struct io_uring_sqe
*sqe
)
3196 #if defined(CONFIG_EPOLL)
3197 if (sqe
->ioprio
|| sqe
->buf_index
)
3200 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
3201 req
->epoll
.op
= READ_ONCE(sqe
->len
);
3202 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
3204 if (ep_op_has_event(req
->epoll
.op
)) {
3205 struct epoll_event __user
*ev
;
3207 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3208 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
3218 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
)
3220 #if defined(CONFIG_EPOLL)
3221 struct io_epoll
*ie
= &req
->epoll
;
3224 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
3225 if (force_nonblock
&& ret
== -EAGAIN
)
3229 req_set_fail_links(req
);
3230 io_cqring_add_event(req
, ret
);
3238 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3240 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3241 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
3244 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
3245 req
->madvise
.len
= READ_ONCE(sqe
->len
);
3246 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
3253 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
3255 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3256 struct io_madvise
*ma
= &req
->madvise
;
3262 ret
= do_madvise(ma
->addr
, ma
->len
, ma
->advice
);
3264 req_set_fail_links(req
);
3265 io_cqring_add_event(req
, ret
);
3273 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3275 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
3278 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
3279 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
3280 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
3284 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
3286 struct io_fadvise
*fa
= &req
->fadvise
;
3289 if (force_nonblock
) {
3290 switch (fa
->advice
) {
3291 case POSIX_FADV_NORMAL
:
3292 case POSIX_FADV_RANDOM
:
3293 case POSIX_FADV_SEQUENTIAL
:
3300 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
3302 req_set_fail_links(req
);
3303 io_cqring_add_event(req
, ret
);
3308 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3310 const char __user
*fname
;
3311 unsigned lookup_flags
;
3314 if (sqe
->ioprio
|| sqe
->buf_index
)
3316 if (req
->flags
& REQ_F_FIXED_FILE
)
3318 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3321 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3322 req
->open
.mask
= READ_ONCE(sqe
->len
);
3323 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3324 req
->open
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3325 req
->open
.how
.flags
= READ_ONCE(sqe
->statx_flags
);
3327 if (vfs_stat_set_lookup_flags(&lookup_flags
, req
->open
.how
.flags
))
3330 req
->open
.filename
= getname_flags(fname
, lookup_flags
, NULL
);
3331 if (IS_ERR(req
->open
.filename
)) {
3332 ret
= PTR_ERR(req
->open
.filename
);
3333 req
->open
.filename
= NULL
;
3337 req
->flags
|= REQ_F_NEED_CLEANUP
;
3341 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
3343 struct io_open
*ctx
= &req
->open
;
3344 unsigned lookup_flags
;
3349 if (force_nonblock
) {
3350 /* only need file table for an actual valid fd */
3351 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
3352 req
->flags
|= REQ_F_NO_FILE_TABLE
;
3356 if (vfs_stat_set_lookup_flags(&lookup_flags
, ctx
->how
.flags
))
3360 /* filename_lookup() drops it, keep a reference */
3361 ctx
->filename
->refcnt
++;
3363 ret
= filename_lookup(ctx
->dfd
, ctx
->filename
, lookup_flags
, &path
,
3368 ret
= vfs_getattr(&path
, &stat
, ctx
->mask
, ctx
->how
.flags
);
3370 if (retry_estale(ret
, lookup_flags
)) {
3371 lookup_flags
|= LOOKUP_REVAL
;
3375 ret
= cp_statx(&stat
, ctx
->buffer
);
3377 putname(ctx
->filename
);
3378 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3380 req_set_fail_links(req
);
3381 io_cqring_add_event(req
, ret
);
3386 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3389 * If we queue this for async, it must not be cancellable. That would
3390 * leave the 'file' in an undeterminate state.
3392 req
->work
.flags
|= IO_WQ_WORK_NO_CANCEL
;
3394 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
3395 sqe
->rw_flags
|| sqe
->buf_index
)
3397 if (req
->flags
& REQ_F_FIXED_FILE
)
3400 req
->close
.fd
= READ_ONCE(sqe
->fd
);
3401 if (req
->file
->f_op
== &io_uring_fops
||
3402 req
->close
.fd
== req
->ctx
->ring_fd
)
3408 /* only called when __close_fd_get_file() is done */
3409 static void __io_close_finish(struct io_kiocb
*req
)
3413 ret
= filp_close(req
->close
.put_file
, req
->work
.files
);
3415 req_set_fail_links(req
);
3416 io_cqring_add_event(req
, ret
);
3417 fput(req
->close
.put_file
);
3421 static void io_close_finish(struct io_wq_work
**workptr
)
3423 struct io_kiocb
*req
= container_of(*workptr
, struct io_kiocb
, work
);
3425 /* not cancellable, don't do io_req_cancelled() */
3426 __io_close_finish(req
);
3427 io_steal_work(req
, workptr
);
3430 static int io_close(struct io_kiocb
*req
, bool force_nonblock
)
3434 req
->close
.put_file
= NULL
;
3435 ret
= __close_fd_get_file(req
->close
.fd
, &req
->close
.put_file
);
3439 /* if the file has a flush method, be safe and punt to async */
3440 if (req
->close
.put_file
->f_op
->flush
&& force_nonblock
) {
3441 /* submission ref will be dropped, take it for async */
3442 refcount_inc(&req
->refs
);
3444 req
->work
.func
= io_close_finish
;
3446 * Do manual async queue here to avoid grabbing files - we don't
3447 * need the files, and it'll cause io_close_finish() to close
3448 * the file again and cause a double CQE entry for this request
3450 io_queue_async_work(req
);
3455 * No ->flush(), safely close from here and just punt the
3456 * fput() to async context.
3458 __io_close_finish(req
);
3462 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3464 struct io_ring_ctx
*ctx
= req
->ctx
;
3469 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3471 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3474 req
->sync
.off
= READ_ONCE(sqe
->off
);
3475 req
->sync
.len
= READ_ONCE(sqe
->len
);
3476 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
3480 static void __io_sync_file_range(struct io_kiocb
*req
)
3484 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
3487 req_set_fail_links(req
);
3488 io_cqring_add_event(req
, ret
);
3493 static void io_sync_file_range_finish(struct io_wq_work
**workptr
)
3495 struct io_kiocb
*req
= container_of(*workptr
, struct io_kiocb
, work
);
3497 if (io_req_cancelled(req
))
3499 __io_sync_file_range(req
);
3500 io_steal_work(req
, workptr
);
3503 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
3505 /* sync_file_range always requires a blocking context */
3506 if (force_nonblock
) {
3507 req
->work
.func
= io_sync_file_range_finish
;
3511 __io_sync_file_range(req
);
3515 #if defined(CONFIG_NET)
3516 static int io_setup_async_msg(struct io_kiocb
*req
,
3517 struct io_async_msghdr
*kmsg
)
3521 if (io_alloc_async_ctx(req
)) {
3522 if (kmsg
->iov
!= kmsg
->fast_iov
)
3526 req
->flags
|= REQ_F_NEED_CLEANUP
;
3527 memcpy(&req
->io
->msg
, kmsg
, sizeof(*kmsg
));
3531 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3533 struct io_sr_msg
*sr
= &req
->sr_msg
;
3534 struct io_async_ctx
*io
= req
->io
;
3537 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
3538 sr
->msg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3539 sr
->len
= READ_ONCE(sqe
->len
);
3541 #ifdef CONFIG_COMPAT
3542 if (req
->ctx
->compat
)
3543 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
3546 if (!io
|| req
->opcode
== IORING_OP_SEND
)
3548 /* iovec is already imported */
3549 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3552 io
->msg
.iov
= io
->msg
.fast_iov
;
3553 ret
= sendmsg_copy_msghdr(&io
->msg
.msg
, sr
->msg
, sr
->msg_flags
,
3556 req
->flags
|= REQ_F_NEED_CLEANUP
;
3560 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
)
3562 struct io_async_msghdr
*kmsg
= NULL
;
3563 struct socket
*sock
;
3566 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3569 sock
= sock_from_file(req
->file
, &ret
);
3571 struct io_async_ctx io
;
3575 kmsg
= &req
->io
->msg
;
3576 kmsg
->msg
.msg_name
= &req
->io
->msg
.addr
;
3577 /* if iov is set, it's allocated already */
3579 kmsg
->iov
= kmsg
->fast_iov
;
3580 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
3582 struct io_sr_msg
*sr
= &req
->sr_msg
;
3585 kmsg
->msg
.msg_name
= &io
.msg
.addr
;
3587 io
.msg
.iov
= io
.msg
.fast_iov
;
3588 ret
= sendmsg_copy_msghdr(&io
.msg
.msg
, sr
->msg
,
3589 sr
->msg_flags
, &io
.msg
.iov
);
3594 flags
= req
->sr_msg
.msg_flags
;
3595 if (flags
& MSG_DONTWAIT
)
3596 req
->flags
|= REQ_F_NOWAIT
;
3597 else if (force_nonblock
)
3598 flags
|= MSG_DONTWAIT
;
3600 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
3601 if (force_nonblock
&& ret
== -EAGAIN
)
3602 return io_setup_async_msg(req
, kmsg
);
3603 if (ret
== -ERESTARTSYS
)
3607 if (kmsg
&& kmsg
->iov
!= kmsg
->fast_iov
)
3609 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3610 io_cqring_add_event(req
, ret
);
3612 req_set_fail_links(req
);
3617 static int io_send(struct io_kiocb
*req
, bool force_nonblock
)
3619 struct socket
*sock
;
3622 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3625 sock
= sock_from_file(req
->file
, &ret
);
3627 struct io_sr_msg
*sr
= &req
->sr_msg
;
3632 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
,
3637 msg
.msg_name
= NULL
;
3638 msg
.msg_control
= NULL
;
3639 msg
.msg_controllen
= 0;
3640 msg
.msg_namelen
= 0;
3642 flags
= req
->sr_msg
.msg_flags
;
3643 if (flags
& MSG_DONTWAIT
)
3644 req
->flags
|= REQ_F_NOWAIT
;
3645 else if (force_nonblock
)
3646 flags
|= MSG_DONTWAIT
;
3648 msg
.msg_flags
= flags
;
3649 ret
= sock_sendmsg(sock
, &msg
);
3650 if (force_nonblock
&& ret
== -EAGAIN
)
3652 if (ret
== -ERESTARTSYS
)
3656 io_cqring_add_event(req
, ret
);
3658 req_set_fail_links(req
);
3663 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
, struct io_async_ctx
*io
)
3665 struct io_sr_msg
*sr
= &req
->sr_msg
;
3666 struct iovec __user
*uiov
;
3670 ret
= __copy_msghdr_from_user(&io
->msg
.msg
, sr
->msg
, &io
->msg
.uaddr
,
3675 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3678 if (copy_from_user(io
->msg
.iov
, uiov
, sizeof(*uiov
)))
3680 sr
->len
= io
->msg
.iov
[0].iov_len
;
3681 iov_iter_init(&io
->msg
.msg
.msg_iter
, READ
, io
->msg
.iov
, 1,
3685 ret
= import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
3686 &io
->msg
.iov
, &io
->msg
.msg
.msg_iter
);
3694 #ifdef CONFIG_COMPAT
3695 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
3696 struct io_async_ctx
*io
)
3698 struct compat_msghdr __user
*msg_compat
;
3699 struct io_sr_msg
*sr
= &req
->sr_msg
;
3700 struct compat_iovec __user
*uiov
;
3705 msg_compat
= (struct compat_msghdr __user
*) sr
->msg
;
3706 ret
= __get_compat_msghdr(&io
->msg
.msg
, msg_compat
, &io
->msg
.uaddr
,
3711 uiov
= compat_ptr(ptr
);
3712 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3713 compat_ssize_t clen
;
3717 if (!access_ok(uiov
, sizeof(*uiov
)))
3719 if (__get_user(clen
, &uiov
->iov_len
))
3723 sr
->len
= io
->msg
.iov
[0].iov_len
;
3726 ret
= compat_import_iovec(READ
, uiov
, len
, UIO_FASTIOV
,
3728 &io
->msg
.msg
.msg_iter
);
3737 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
, struct io_async_ctx
*io
)
3739 io
->msg
.iov
= io
->msg
.fast_iov
;
3741 #ifdef CONFIG_COMPAT
3742 if (req
->ctx
->compat
)
3743 return __io_compat_recvmsg_copy_hdr(req
, io
);
3746 return __io_recvmsg_copy_hdr(req
, io
);
3749 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
3750 int *cflags
, bool needs_lock
)
3752 struct io_sr_msg
*sr
= &req
->sr_msg
;
3753 struct io_buffer
*kbuf
;
3755 if (!(req
->flags
& REQ_F_BUFFER_SELECT
))
3758 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
3763 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3765 *cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
3766 *cflags
|= IORING_CQE_F_BUFFER
;
3770 static int io_recvmsg_prep(struct io_kiocb
*req
,
3771 const struct io_uring_sqe
*sqe
)
3773 struct io_sr_msg
*sr
= &req
->sr_msg
;
3774 struct io_async_ctx
*io
= req
->io
;
3777 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
3778 sr
->msg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3779 sr
->len
= READ_ONCE(sqe
->len
);
3780 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
3782 #ifdef CONFIG_COMPAT
3783 if (req
->ctx
->compat
)
3784 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
3787 if (!io
|| req
->opcode
== IORING_OP_RECV
)
3789 /* iovec is already imported */
3790 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3793 ret
= io_recvmsg_copy_hdr(req
, io
);
3795 req
->flags
|= REQ_F_NEED_CLEANUP
;
3799 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
)
3801 struct io_async_msghdr
*kmsg
= NULL
;
3802 struct socket
*sock
;
3803 int ret
, cflags
= 0;
3805 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3808 sock
= sock_from_file(req
->file
, &ret
);
3810 struct io_buffer
*kbuf
;
3811 struct io_async_ctx io
;
3815 kmsg
= &req
->io
->msg
;
3816 kmsg
->msg
.msg_name
= &req
->io
->msg
.addr
;
3817 /* if iov is set, it's allocated already */
3819 kmsg
->iov
= kmsg
->fast_iov
;
3820 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
3823 kmsg
->msg
.msg_name
= &io
.msg
.addr
;
3825 ret
= io_recvmsg_copy_hdr(req
, &io
);
3830 kbuf
= io_recv_buffer_select(req
, &cflags
, !force_nonblock
);
3832 return PTR_ERR(kbuf
);
3834 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3835 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
3836 1, req
->sr_msg
.len
);
3839 flags
= req
->sr_msg
.msg_flags
;
3840 if (flags
& MSG_DONTWAIT
)
3841 req
->flags
|= REQ_F_NOWAIT
;
3842 else if (force_nonblock
)
3843 flags
|= MSG_DONTWAIT
;
3845 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.msg
,
3846 kmsg
->uaddr
, flags
);
3847 if (force_nonblock
&& ret
== -EAGAIN
)
3848 return io_setup_async_msg(req
, kmsg
);
3849 if (ret
== -ERESTARTSYS
)
3853 if (kmsg
&& kmsg
->iov
!= kmsg
->fast_iov
)
3855 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3856 __io_cqring_add_event(req
, ret
, cflags
);
3858 req_set_fail_links(req
);
3863 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
)
3865 struct io_buffer
*kbuf
= NULL
;
3866 struct socket
*sock
;
3867 int ret
, cflags
= 0;
3869 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3872 sock
= sock_from_file(req
->file
, &ret
);
3874 struct io_sr_msg
*sr
= &req
->sr_msg
;
3875 void __user
*buf
= sr
->buf
;
3880 kbuf
= io_recv_buffer_select(req
, &cflags
, !force_nonblock
);
3882 return PTR_ERR(kbuf
);
3884 buf
= u64_to_user_ptr(kbuf
->addr
);
3886 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
,
3893 req
->flags
|= REQ_F_NEED_CLEANUP
;
3894 msg
.msg_name
= NULL
;
3895 msg
.msg_control
= NULL
;
3896 msg
.msg_controllen
= 0;
3897 msg
.msg_namelen
= 0;
3898 msg
.msg_iocb
= NULL
;
3901 flags
= req
->sr_msg
.msg_flags
;
3902 if (flags
& MSG_DONTWAIT
)
3903 req
->flags
|= REQ_F_NOWAIT
;
3904 else if (force_nonblock
)
3905 flags
|= MSG_DONTWAIT
;
3907 ret
= sock_recvmsg(sock
, &msg
, flags
);
3908 if (force_nonblock
&& ret
== -EAGAIN
)
3910 if (ret
== -ERESTARTSYS
)
3915 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3916 __io_cqring_add_event(req
, ret
, cflags
);
3918 req_set_fail_links(req
);
3923 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3925 struct io_accept
*accept
= &req
->accept
;
3927 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3929 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
3932 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3933 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3934 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
3935 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
3939 static int __io_accept(struct io_kiocb
*req
, bool force_nonblock
)
3941 struct io_accept
*accept
= &req
->accept
;
3942 unsigned file_flags
;
3945 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
3946 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
3947 accept
->addr_len
, accept
->flags
,
3949 if (ret
== -EAGAIN
&& force_nonblock
)
3951 if (ret
== -ERESTARTSYS
)
3954 req_set_fail_links(req
);
3955 io_cqring_add_event(req
, ret
);
3960 static void io_accept_finish(struct io_wq_work
**workptr
)
3962 struct io_kiocb
*req
= container_of(*workptr
, struct io_kiocb
, work
);
3964 if (io_req_cancelled(req
))
3966 __io_accept(req
, false);
3967 io_steal_work(req
, workptr
);
3970 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
)
3974 ret
= __io_accept(req
, force_nonblock
);
3975 if (ret
== -EAGAIN
&& force_nonblock
) {
3976 req
->work
.func
= io_accept_finish
;
3982 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3984 struct io_connect
*conn
= &req
->connect
;
3985 struct io_async_ctx
*io
= req
->io
;
3987 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3989 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
3992 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3993 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
3998 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
3999 &io
->connect
.address
);
4002 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
)
4004 struct io_async_ctx __io
, *io
;
4005 unsigned file_flags
;
4011 ret
= move_addr_to_kernel(req
->connect
.addr
,
4012 req
->connect
.addr_len
,
4013 &__io
.connect
.address
);
4019 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4021 ret
= __sys_connect_file(req
->file
, &io
->connect
.address
,
4022 req
->connect
.addr_len
, file_flags
);
4023 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4026 if (io_alloc_async_ctx(req
)) {
4030 memcpy(&req
->io
->connect
, &__io
.connect
, sizeof(__io
.connect
));
4033 if (ret
== -ERESTARTSYS
)
4037 req_set_fail_links(req
);
4038 io_cqring_add_event(req
, ret
);
4042 #else /* !CONFIG_NET */
4043 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4048 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
)
4053 static int io_send(struct io_kiocb
*req
, bool force_nonblock
)
4058 static int io_recvmsg_prep(struct io_kiocb
*req
,
4059 const struct io_uring_sqe
*sqe
)
4064 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
)
4069 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
)
4074 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4079 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
)
4084 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4089 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
)
4093 #endif /* CONFIG_NET */
4095 struct io_poll_table
{
4096 struct poll_table_struct pt
;
4097 struct io_kiocb
*req
;
4101 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
4102 struct wait_queue_head
*head
)
4104 if (unlikely(poll
->head
)) {
4105 pt
->error
= -EINVAL
;
4111 add_wait_queue(head
, &poll
->wait
);
4114 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
4115 struct poll_table_struct
*p
)
4117 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
4119 __io_queue_proc(&pt
->req
->apoll
->poll
, pt
, head
);
4122 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
4123 __poll_t mask
, task_work_func_t func
)
4125 struct task_struct
*tsk
;
4128 /* for instances that support it check for an event match first: */
4129 if (mask
&& !(mask
& poll
->events
))
4132 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
4134 list_del_init(&poll
->wait
.entry
);
4138 init_task_work(&req
->task_work
, func
);
4140 * If this fails, then the task is exiting. Punt to one of the io-wq
4141 * threads to ensure the work gets run, we can't always rely on exit
4142 * cancelation taking care of this.
4144 ret
= task_work_add(tsk
, &req
->task_work
, true);
4145 if (unlikely(ret
)) {
4146 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
4147 task_work_add(tsk
, &req
->task_work
, true);
4149 wake_up_process(tsk
);
4153 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
4154 __acquires(&req
->ctx
->completion_lock
)
4156 struct io_ring_ctx
*ctx
= req
->ctx
;
4158 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4159 struct poll_table_struct pt
= { ._key
= poll
->events
};
4161 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
4164 spin_lock_irq(&ctx
->completion_lock
);
4165 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4166 add_wait_queue(poll
->head
, &poll
->wait
);
4173 static void io_async_task_func(struct callback_head
*cb
)
4175 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4176 struct async_poll
*apoll
= req
->apoll
;
4177 struct io_ring_ctx
*ctx
= req
->ctx
;
4180 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
4182 if (io_poll_rewait(req
, &apoll
->poll
)) {
4183 spin_unlock_irq(&ctx
->completion_lock
);
4187 if (hash_hashed(&req
->hash_node
))
4188 hash_del(&req
->hash_node
);
4190 canceled
= READ_ONCE(apoll
->poll
.canceled
);
4192 io_cqring_fill_event(req
, -ECANCELED
);
4193 io_commit_cqring(ctx
);
4196 spin_unlock_irq(&ctx
->completion_lock
);
4198 /* restore ->work in case we need to retry again */
4199 memcpy(&req
->work
, &apoll
->work
, sizeof(req
->work
));
4203 io_cqring_ev_posted(ctx
);
4204 req_set_fail_links(req
);
4205 io_double_put_req(req
);
4209 __set_current_state(TASK_RUNNING
);
4210 mutex_lock(&ctx
->uring_lock
);
4211 __io_queue_sqe(req
, NULL
);
4212 mutex_unlock(&ctx
->uring_lock
);
4217 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
4220 struct io_kiocb
*req
= wait
->private;
4221 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
4223 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
4226 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
4229 static void io_poll_req_insert(struct io_kiocb
*req
)
4231 struct io_ring_ctx
*ctx
= req
->ctx
;
4232 struct hlist_head
*list
;
4234 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
4235 hlist_add_head(&req
->hash_node
, list
);
4238 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
4239 struct io_poll_iocb
*poll
,
4240 struct io_poll_table
*ipt
, __poll_t mask
,
4241 wait_queue_func_t wake_func
)
4242 __acquires(&ctx
->completion_lock
)
4244 struct io_ring_ctx
*ctx
= req
->ctx
;
4245 bool cancel
= false;
4247 poll
->file
= req
->file
;
4249 poll
->done
= poll
->canceled
= false;
4250 poll
->events
= mask
;
4252 ipt
->pt
._key
= mask
;
4254 ipt
->error
= -EINVAL
;
4256 INIT_LIST_HEAD(&poll
->wait
.entry
);
4257 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
4258 poll
->wait
.private = req
;
4260 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
4262 spin_lock_irq(&ctx
->completion_lock
);
4263 if (likely(poll
->head
)) {
4264 spin_lock(&poll
->head
->lock
);
4265 if (unlikely(list_empty(&poll
->wait
.entry
))) {
4271 if (mask
|| ipt
->error
)
4272 list_del_init(&poll
->wait
.entry
);
4274 WRITE_ONCE(poll
->canceled
, true);
4275 else if (!poll
->done
) /* actually waiting for an event */
4276 io_poll_req_insert(req
);
4277 spin_unlock(&poll
->head
->lock
);
4283 static bool io_arm_poll_handler(struct io_kiocb
*req
)
4285 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
4286 struct io_ring_ctx
*ctx
= req
->ctx
;
4287 struct async_poll
*apoll
;
4288 struct io_poll_table ipt
;
4291 if (!req
->file
|| !file_can_poll(req
->file
))
4293 if (req
->flags
& (REQ_F_MUST_PUNT
| REQ_F_POLLED
))
4295 if (!def
->pollin
&& !def
->pollout
)
4298 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
4299 if (unlikely(!apoll
))
4302 req
->flags
|= REQ_F_POLLED
;
4303 memcpy(&apoll
->work
, &req
->work
, sizeof(req
->work
));
4305 get_task_struct(current
);
4306 req
->task
= current
;
4308 INIT_HLIST_NODE(&req
->hash_node
);
4312 mask
|= POLLIN
| POLLRDNORM
;
4314 mask
|= POLLOUT
| POLLWRNORM
;
4315 mask
|= POLLERR
| POLLPRI
;
4317 ipt
.pt
._qproc
= io_async_queue_proc
;
4319 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
4323 apoll
->poll
.done
= true;
4324 spin_unlock_irq(&ctx
->completion_lock
);
4325 memcpy(&req
->work
, &apoll
->work
, sizeof(req
->work
));
4329 spin_unlock_irq(&ctx
->completion_lock
);
4330 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
4331 apoll
->poll
.events
);
4335 static bool __io_poll_remove_one(struct io_kiocb
*req
,
4336 struct io_poll_iocb
*poll
)
4338 bool do_complete
= false;
4340 spin_lock(&poll
->head
->lock
);
4341 WRITE_ONCE(poll
->canceled
, true);
4342 if (!list_empty(&poll
->wait
.entry
)) {
4343 list_del_init(&poll
->wait
.entry
);
4346 spin_unlock(&poll
->head
->lock
);
4350 static bool io_poll_remove_one(struct io_kiocb
*req
)
4352 struct async_poll
*apoll
= NULL
;
4355 if (req
->opcode
== IORING_OP_POLL_ADD
) {
4356 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
4359 /* non-poll requests have submit ref still */
4360 do_complete
= __io_poll_remove_one(req
, &req
->apoll
->poll
);
4365 hash_del(&req
->hash_node
);
4367 if (do_complete
&& apoll
) {
4369 * restore ->work because we need to call io_req_work_drop_env.
4371 memcpy(&req
->work
, &apoll
->work
, sizeof(req
->work
));
4376 io_cqring_fill_event(req
, -ECANCELED
);
4377 io_commit_cqring(req
->ctx
);
4378 req
->flags
|= REQ_F_COMP_LOCKED
;
4385 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
4387 struct hlist_node
*tmp
;
4388 struct io_kiocb
*req
;
4391 spin_lock_irq(&ctx
->completion_lock
);
4392 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
4393 struct hlist_head
*list
;
4395 list
= &ctx
->cancel_hash
[i
];
4396 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
)
4397 posted
+= io_poll_remove_one(req
);
4399 spin_unlock_irq(&ctx
->completion_lock
);
4402 io_cqring_ev_posted(ctx
);
4405 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
4407 struct hlist_head
*list
;
4408 struct io_kiocb
*req
;
4410 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
4411 hlist_for_each_entry(req
, list
, hash_node
) {
4412 if (sqe_addr
!= req
->user_data
)
4414 if (io_poll_remove_one(req
))
4422 static int io_poll_remove_prep(struct io_kiocb
*req
,
4423 const struct io_uring_sqe
*sqe
)
4425 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4427 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
4431 req
->poll
.addr
= READ_ONCE(sqe
->addr
);
4436 * Find a running poll command that matches one specified in sqe->addr,
4437 * and remove it if found.
4439 static int io_poll_remove(struct io_kiocb
*req
)
4441 struct io_ring_ctx
*ctx
= req
->ctx
;
4445 addr
= req
->poll
.addr
;
4446 spin_lock_irq(&ctx
->completion_lock
);
4447 ret
= io_poll_cancel(ctx
, addr
);
4448 spin_unlock_irq(&ctx
->completion_lock
);
4450 io_cqring_add_event(req
, ret
);
4452 req_set_fail_links(req
);
4457 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
4459 struct io_ring_ctx
*ctx
= req
->ctx
;
4461 req
->poll
.done
= true;
4462 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
4463 io_commit_cqring(ctx
);
4466 static void io_poll_task_handler(struct io_kiocb
*req
, struct io_kiocb
**nxt
)
4468 struct io_ring_ctx
*ctx
= req
->ctx
;
4469 struct io_poll_iocb
*poll
= &req
->poll
;
4471 if (io_poll_rewait(req
, poll
)) {
4472 spin_unlock_irq(&ctx
->completion_lock
);
4476 hash_del(&req
->hash_node
);
4477 io_poll_complete(req
, req
->result
, 0);
4478 req
->flags
|= REQ_F_COMP_LOCKED
;
4479 io_put_req_find_next(req
, nxt
);
4480 spin_unlock_irq(&ctx
->completion_lock
);
4482 io_cqring_ev_posted(ctx
);
4485 static void io_poll_task_func(struct callback_head
*cb
)
4487 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4488 struct io_kiocb
*nxt
= NULL
;
4490 io_poll_task_handler(req
, &nxt
);
4492 struct io_ring_ctx
*ctx
= nxt
->ctx
;
4494 mutex_lock(&ctx
->uring_lock
);
4495 __io_queue_sqe(nxt
, NULL
);
4496 mutex_unlock(&ctx
->uring_lock
);
4500 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
4503 struct io_kiocb
*req
= wait
->private;
4504 struct io_poll_iocb
*poll
= &req
->poll
;
4506 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
4509 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
4510 struct poll_table_struct
*p
)
4512 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
4514 __io_queue_proc(&pt
->req
->poll
, pt
, head
);
4517 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4519 struct io_poll_iocb
*poll
= &req
->poll
;
4522 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4524 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
4529 events
= READ_ONCE(sqe
->poll_events
);
4530 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
;
4532 get_task_struct(current
);
4533 req
->task
= current
;
4537 static int io_poll_add(struct io_kiocb
*req
)
4539 struct io_poll_iocb
*poll
= &req
->poll
;
4540 struct io_ring_ctx
*ctx
= req
->ctx
;
4541 struct io_poll_table ipt
;
4544 INIT_HLIST_NODE(&req
->hash_node
);
4545 INIT_LIST_HEAD(&req
->list
);
4546 ipt
.pt
._qproc
= io_poll_queue_proc
;
4548 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
4551 if (mask
) { /* no async, we'd stolen it */
4553 io_poll_complete(req
, mask
, 0);
4555 spin_unlock_irq(&ctx
->completion_lock
);
4558 io_cqring_ev_posted(ctx
);
4564 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
4566 struct io_timeout_data
*data
= container_of(timer
,
4567 struct io_timeout_data
, timer
);
4568 struct io_kiocb
*req
= data
->req
;
4569 struct io_ring_ctx
*ctx
= req
->ctx
;
4570 unsigned long flags
;
4572 atomic_inc(&ctx
->cq_timeouts
);
4574 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
4576 * We could be racing with timeout deletion. If the list is empty,
4577 * then timeout lookup already found it and will be handling it.
4579 if (!list_empty(&req
->list
)) {
4580 struct io_kiocb
*prev
;
4583 * Adjust the reqs sequence before the current one because it
4584 * will consume a slot in the cq_ring and the cq_tail
4585 * pointer will be increased, otherwise other timeout reqs may
4586 * return in advance without waiting for enough wait_nr.
4589 list_for_each_entry_continue_reverse(prev
, &ctx
->timeout_list
, list
)
4591 list_del_init(&req
->list
);
4594 io_cqring_fill_event(req
, -ETIME
);
4595 io_commit_cqring(ctx
);
4596 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
4598 io_cqring_ev_posted(ctx
);
4599 req_set_fail_links(req
);
4601 return HRTIMER_NORESTART
;
4604 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
4606 struct io_kiocb
*req
;
4609 list_for_each_entry(req
, &ctx
->timeout_list
, list
) {
4610 if (user_data
== req
->user_data
) {
4611 list_del_init(&req
->list
);
4620 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
4624 req_set_fail_links(req
);
4625 io_cqring_fill_event(req
, -ECANCELED
);
4630 static int io_timeout_remove_prep(struct io_kiocb
*req
,
4631 const struct io_uring_sqe
*sqe
)
4633 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4635 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
4638 req
->timeout
.addr
= READ_ONCE(sqe
->addr
);
4639 req
->timeout
.flags
= READ_ONCE(sqe
->timeout_flags
);
4640 if (req
->timeout
.flags
)
4647 * Remove or update an existing timeout command
4649 static int io_timeout_remove(struct io_kiocb
*req
)
4651 struct io_ring_ctx
*ctx
= req
->ctx
;
4654 spin_lock_irq(&ctx
->completion_lock
);
4655 ret
= io_timeout_cancel(ctx
, req
->timeout
.addr
);
4657 io_cqring_fill_event(req
, ret
);
4658 io_commit_cqring(ctx
);
4659 spin_unlock_irq(&ctx
->completion_lock
);
4660 io_cqring_ev_posted(ctx
);
4662 req_set_fail_links(req
);
4667 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
4668 bool is_timeout_link
)
4670 struct io_timeout_data
*data
;
4673 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4675 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
4677 if (sqe
->off
&& is_timeout_link
)
4679 flags
= READ_ONCE(sqe
->timeout_flags
);
4680 if (flags
& ~IORING_TIMEOUT_ABS
)
4683 req
->timeout
.count
= READ_ONCE(sqe
->off
);
4685 if (!req
->io
&& io_alloc_async_ctx(req
))
4688 data
= &req
->io
->timeout
;
4690 req
->flags
|= REQ_F_TIMEOUT
;
4692 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
4695 if (flags
& IORING_TIMEOUT_ABS
)
4696 data
->mode
= HRTIMER_MODE_ABS
;
4698 data
->mode
= HRTIMER_MODE_REL
;
4700 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
4704 static int io_timeout(struct io_kiocb
*req
)
4706 struct io_ring_ctx
*ctx
= req
->ctx
;
4707 struct io_timeout_data
*data
;
4708 struct list_head
*entry
;
4710 u32 count
= req
->timeout
.count
;
4711 u32 seq
= req
->sequence
;
4713 data
= &req
->io
->timeout
;
4716 * sqe->off holds how many events that need to occur for this
4717 * timeout event to be satisfied. If it isn't set, then this is
4718 * a pure timeout request, sequence isn't used.
4721 req
->flags
|= REQ_F_TIMEOUT_NOSEQ
;
4722 spin_lock_irq(&ctx
->completion_lock
);
4723 entry
= ctx
->timeout_list
.prev
;
4727 req
->sequence
= seq
+ count
;
4730 * Insertion sort, ensuring the first entry in the list is always
4731 * the one we need first.
4733 spin_lock_irq(&ctx
->completion_lock
);
4734 list_for_each_prev(entry
, &ctx
->timeout_list
) {
4735 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
, list
);
4737 long long tmp
, tmp_nxt
;
4738 u32 nxt_offset
= nxt
->timeout
.count
;
4740 if (nxt
->flags
& REQ_F_TIMEOUT_NOSEQ
)
4744 * Since seq + count can overflow, use type long
4747 tmp
= (long long)seq
+ count
;
4748 nxt_seq
= nxt
->sequence
- nxt_offset
;
4749 tmp_nxt
= (long long)nxt_seq
+ nxt_offset
;
4752 * cached_sq_head may overflow, and it will never overflow twice
4753 * once there is some timeout req still be valid.
4762 * Sequence of reqs after the insert one and itself should
4763 * be adjusted because each timeout req consumes a slot.
4768 req
->sequence
-= span
;
4770 list_add(&req
->list
, entry
);
4771 data
->timer
.function
= io_timeout_fn
;
4772 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
4773 spin_unlock_irq(&ctx
->completion_lock
);
4777 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
4779 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
4781 return req
->user_data
== (unsigned long) data
;
4784 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
4786 enum io_wq_cancel cancel_ret
;
4789 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
);
4790 switch (cancel_ret
) {
4791 case IO_WQ_CANCEL_OK
:
4794 case IO_WQ_CANCEL_RUNNING
:
4797 case IO_WQ_CANCEL_NOTFOUND
:
4805 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
4806 struct io_kiocb
*req
, __u64 sqe_addr
,
4809 unsigned long flags
;
4812 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
4813 if (ret
!= -ENOENT
) {
4814 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
4818 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
4819 ret
= io_timeout_cancel(ctx
, sqe_addr
);
4822 ret
= io_poll_cancel(ctx
, sqe_addr
);
4826 io_cqring_fill_event(req
, ret
);
4827 io_commit_cqring(ctx
);
4828 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
4829 io_cqring_ev_posted(ctx
);
4832 req_set_fail_links(req
);
4836 static int io_async_cancel_prep(struct io_kiocb
*req
,
4837 const struct io_uring_sqe
*sqe
)
4839 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4841 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
||
4845 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
4849 static int io_async_cancel(struct io_kiocb
*req
)
4851 struct io_ring_ctx
*ctx
= req
->ctx
;
4853 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
4857 static int io_files_update_prep(struct io_kiocb
*req
,
4858 const struct io_uring_sqe
*sqe
)
4860 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->rw_flags
)
4863 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
4864 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
4865 if (!req
->files_update
.nr_args
)
4867 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
4871 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
)
4873 struct io_ring_ctx
*ctx
= req
->ctx
;
4874 struct io_uring_files_update up
;
4880 up
.offset
= req
->files_update
.offset
;
4881 up
.fds
= req
->files_update
.arg
;
4883 mutex_lock(&ctx
->uring_lock
);
4884 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
4885 mutex_unlock(&ctx
->uring_lock
);
4888 req_set_fail_links(req
);
4889 io_cqring_add_event(req
, ret
);
4894 static int io_req_defer_prep(struct io_kiocb
*req
,
4895 const struct io_uring_sqe
*sqe
)
4902 if (io_op_defs
[req
->opcode
].file_table
) {
4903 ret
= io_grab_files(req
);
4908 io_req_work_grab_env(req
, &io_op_defs
[req
->opcode
]);
4910 switch (req
->opcode
) {
4913 case IORING_OP_READV
:
4914 case IORING_OP_READ_FIXED
:
4915 case IORING_OP_READ
:
4916 ret
= io_read_prep(req
, sqe
, true);
4918 case IORING_OP_WRITEV
:
4919 case IORING_OP_WRITE_FIXED
:
4920 case IORING_OP_WRITE
:
4921 ret
= io_write_prep(req
, sqe
, true);
4923 case IORING_OP_POLL_ADD
:
4924 ret
= io_poll_add_prep(req
, sqe
);
4926 case IORING_OP_POLL_REMOVE
:
4927 ret
= io_poll_remove_prep(req
, sqe
);
4929 case IORING_OP_FSYNC
:
4930 ret
= io_prep_fsync(req
, sqe
);
4932 case IORING_OP_SYNC_FILE_RANGE
:
4933 ret
= io_prep_sfr(req
, sqe
);
4935 case IORING_OP_SENDMSG
:
4936 case IORING_OP_SEND
:
4937 ret
= io_sendmsg_prep(req
, sqe
);
4939 case IORING_OP_RECVMSG
:
4940 case IORING_OP_RECV
:
4941 ret
= io_recvmsg_prep(req
, sqe
);
4943 case IORING_OP_CONNECT
:
4944 ret
= io_connect_prep(req
, sqe
);
4946 case IORING_OP_TIMEOUT
:
4947 ret
= io_timeout_prep(req
, sqe
, false);
4949 case IORING_OP_TIMEOUT_REMOVE
:
4950 ret
= io_timeout_remove_prep(req
, sqe
);
4952 case IORING_OP_ASYNC_CANCEL
:
4953 ret
= io_async_cancel_prep(req
, sqe
);
4955 case IORING_OP_LINK_TIMEOUT
:
4956 ret
= io_timeout_prep(req
, sqe
, true);
4958 case IORING_OP_ACCEPT
:
4959 ret
= io_accept_prep(req
, sqe
);
4961 case IORING_OP_FALLOCATE
:
4962 ret
= io_fallocate_prep(req
, sqe
);
4964 case IORING_OP_OPENAT
:
4965 ret
= io_openat_prep(req
, sqe
);
4967 case IORING_OP_CLOSE
:
4968 ret
= io_close_prep(req
, sqe
);
4970 case IORING_OP_FILES_UPDATE
:
4971 ret
= io_files_update_prep(req
, sqe
);
4973 case IORING_OP_STATX
:
4974 ret
= io_statx_prep(req
, sqe
);
4976 case IORING_OP_FADVISE
:
4977 ret
= io_fadvise_prep(req
, sqe
);
4979 case IORING_OP_MADVISE
:
4980 ret
= io_madvise_prep(req
, sqe
);
4982 case IORING_OP_OPENAT2
:
4983 ret
= io_openat2_prep(req
, sqe
);
4985 case IORING_OP_EPOLL_CTL
:
4986 ret
= io_epoll_ctl_prep(req
, sqe
);
4988 case IORING_OP_SPLICE
:
4989 ret
= io_splice_prep(req
, sqe
);
4991 case IORING_OP_PROVIDE_BUFFERS
:
4992 ret
= io_provide_buffers_prep(req
, sqe
);
4994 case IORING_OP_REMOVE_BUFFERS
:
4995 ret
= io_remove_buffers_prep(req
, sqe
);
4998 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5007 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5009 struct io_ring_ctx
*ctx
= req
->ctx
;
5012 /* Still need defer if there is pending req in defer list. */
5013 if (!req_need_defer(req
) && list_empty_careful(&ctx
->defer_list
))
5016 if (!req
->io
&& io_alloc_async_ctx(req
))
5019 ret
= io_req_defer_prep(req
, sqe
);
5023 spin_lock_irq(&ctx
->completion_lock
);
5024 if (!req_need_defer(req
) && list_empty(&ctx
->defer_list
)) {
5025 spin_unlock_irq(&ctx
->completion_lock
);
5029 trace_io_uring_defer(ctx
, req
, req
->user_data
);
5030 list_add_tail(&req
->list
, &ctx
->defer_list
);
5031 spin_unlock_irq(&ctx
->completion_lock
);
5032 return -EIOCBQUEUED
;
5035 static void io_cleanup_req(struct io_kiocb
*req
)
5037 struct io_async_ctx
*io
= req
->io
;
5039 switch (req
->opcode
) {
5040 case IORING_OP_READV
:
5041 case IORING_OP_READ_FIXED
:
5042 case IORING_OP_READ
:
5043 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5044 kfree((void *)(unsigned long)req
->rw
.addr
);
5046 case IORING_OP_WRITEV
:
5047 case IORING_OP_WRITE_FIXED
:
5048 case IORING_OP_WRITE
:
5049 if (io
->rw
.iov
!= io
->rw
.fast_iov
)
5052 case IORING_OP_RECVMSG
:
5053 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5054 kfree(req
->sr_msg
.kbuf
);
5056 case IORING_OP_SENDMSG
:
5057 if (io
->msg
.iov
!= io
->msg
.fast_iov
)
5060 case IORING_OP_RECV
:
5061 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5062 kfree(req
->sr_msg
.kbuf
);
5064 case IORING_OP_OPENAT
:
5065 case IORING_OP_OPENAT2
:
5066 case IORING_OP_STATX
:
5067 putname(req
->open
.filename
);
5069 case IORING_OP_SPLICE
:
5070 io_put_file(req
, req
->splice
.file_in
,
5071 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
5075 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5078 static int io_issue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5079 bool force_nonblock
)
5081 struct io_ring_ctx
*ctx
= req
->ctx
;
5084 switch (req
->opcode
) {
5088 case IORING_OP_READV
:
5089 case IORING_OP_READ_FIXED
:
5090 case IORING_OP_READ
:
5092 ret
= io_read_prep(req
, sqe
, force_nonblock
);
5096 ret
= io_read(req
, force_nonblock
);
5098 case IORING_OP_WRITEV
:
5099 case IORING_OP_WRITE_FIXED
:
5100 case IORING_OP_WRITE
:
5102 ret
= io_write_prep(req
, sqe
, force_nonblock
);
5106 ret
= io_write(req
, force_nonblock
);
5108 case IORING_OP_FSYNC
:
5110 ret
= io_prep_fsync(req
, sqe
);
5114 ret
= io_fsync(req
, force_nonblock
);
5116 case IORING_OP_POLL_ADD
:
5118 ret
= io_poll_add_prep(req
, sqe
);
5122 ret
= io_poll_add(req
);
5124 case IORING_OP_POLL_REMOVE
:
5126 ret
= io_poll_remove_prep(req
, sqe
);
5130 ret
= io_poll_remove(req
);
5132 case IORING_OP_SYNC_FILE_RANGE
:
5134 ret
= io_prep_sfr(req
, sqe
);
5138 ret
= io_sync_file_range(req
, force_nonblock
);
5140 case IORING_OP_SENDMSG
:
5141 case IORING_OP_SEND
:
5143 ret
= io_sendmsg_prep(req
, sqe
);
5147 if (req
->opcode
== IORING_OP_SENDMSG
)
5148 ret
= io_sendmsg(req
, force_nonblock
);
5150 ret
= io_send(req
, force_nonblock
);
5152 case IORING_OP_RECVMSG
:
5153 case IORING_OP_RECV
:
5155 ret
= io_recvmsg_prep(req
, sqe
);
5159 if (req
->opcode
== IORING_OP_RECVMSG
)
5160 ret
= io_recvmsg(req
, force_nonblock
);
5162 ret
= io_recv(req
, force_nonblock
);
5164 case IORING_OP_TIMEOUT
:
5166 ret
= io_timeout_prep(req
, sqe
, false);
5170 ret
= io_timeout(req
);
5172 case IORING_OP_TIMEOUT_REMOVE
:
5174 ret
= io_timeout_remove_prep(req
, sqe
);
5178 ret
= io_timeout_remove(req
);
5180 case IORING_OP_ACCEPT
:
5182 ret
= io_accept_prep(req
, sqe
);
5186 ret
= io_accept(req
, force_nonblock
);
5188 case IORING_OP_CONNECT
:
5190 ret
= io_connect_prep(req
, sqe
);
5194 ret
= io_connect(req
, force_nonblock
);
5196 case IORING_OP_ASYNC_CANCEL
:
5198 ret
= io_async_cancel_prep(req
, sqe
);
5202 ret
= io_async_cancel(req
);
5204 case IORING_OP_FALLOCATE
:
5206 ret
= io_fallocate_prep(req
, sqe
);
5210 ret
= io_fallocate(req
, force_nonblock
);
5212 case IORING_OP_OPENAT
:
5214 ret
= io_openat_prep(req
, sqe
);
5218 ret
= io_openat(req
, force_nonblock
);
5220 case IORING_OP_CLOSE
:
5222 ret
= io_close_prep(req
, sqe
);
5226 ret
= io_close(req
, force_nonblock
);
5228 case IORING_OP_FILES_UPDATE
:
5230 ret
= io_files_update_prep(req
, sqe
);
5234 ret
= io_files_update(req
, force_nonblock
);
5236 case IORING_OP_STATX
:
5238 ret
= io_statx_prep(req
, sqe
);
5242 ret
= io_statx(req
, force_nonblock
);
5244 case IORING_OP_FADVISE
:
5246 ret
= io_fadvise_prep(req
, sqe
);
5250 ret
= io_fadvise(req
, force_nonblock
);
5252 case IORING_OP_MADVISE
:
5254 ret
= io_madvise_prep(req
, sqe
);
5258 ret
= io_madvise(req
, force_nonblock
);
5260 case IORING_OP_OPENAT2
:
5262 ret
= io_openat2_prep(req
, sqe
);
5266 ret
= io_openat2(req
, force_nonblock
);
5268 case IORING_OP_EPOLL_CTL
:
5270 ret
= io_epoll_ctl_prep(req
, sqe
);
5274 ret
= io_epoll_ctl(req
, force_nonblock
);
5276 case IORING_OP_SPLICE
:
5278 ret
= io_splice_prep(req
, sqe
);
5282 ret
= io_splice(req
, force_nonblock
);
5284 case IORING_OP_PROVIDE_BUFFERS
:
5286 ret
= io_provide_buffers_prep(req
, sqe
);
5290 ret
= io_provide_buffers(req
, force_nonblock
);
5292 case IORING_OP_REMOVE_BUFFERS
:
5294 ret
= io_remove_buffers_prep(req
, sqe
);
5298 ret
= io_remove_buffers(req
, force_nonblock
);
5308 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
5309 const bool in_async
= io_wq_current_is_worker();
5311 if (req
->result
== -EAGAIN
)
5314 /* workqueue context doesn't hold uring_lock, grab it now */
5316 mutex_lock(&ctx
->uring_lock
);
5318 io_iopoll_req_issued(req
);
5321 mutex_unlock(&ctx
->uring_lock
);
5327 static void io_wq_submit_work(struct io_wq_work
**workptr
)
5329 struct io_wq_work
*work
= *workptr
;
5330 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5333 /* if NO_CANCEL is set, we must still run the work */
5334 if ((work
->flags
& (IO_WQ_WORK_CANCEL
|IO_WQ_WORK_NO_CANCEL
)) ==
5335 IO_WQ_WORK_CANCEL
) {
5341 ret
= io_issue_sqe(req
, NULL
, false);
5343 * We can get EAGAIN for polled IO even though we're
5344 * forcing a sync submission from here, since we can't
5345 * wait for request slots on the block side.
5354 req_set_fail_links(req
);
5355 io_cqring_add_event(req
, ret
);
5359 io_steal_work(req
, workptr
);
5362 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
5365 struct fixed_file_table
*table
;
5367 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
5368 return table
->files
[index
& IORING_FILE_TABLE_MASK
];;
5371 static int io_file_get(struct io_submit_state
*state
, struct io_kiocb
*req
,
5372 int fd
, struct file
**out_file
, bool fixed
)
5374 struct io_ring_ctx
*ctx
= req
->ctx
;
5378 if (unlikely(!ctx
->file_data
||
5379 (unsigned) fd
>= ctx
->nr_user_files
))
5381 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
5382 file
= io_file_from_index(ctx
, fd
);
5385 req
->fixed_file_refs
= ctx
->file_data
->cur_refs
;
5386 percpu_ref_get(req
->fixed_file_refs
);
5388 trace_io_uring_file_get(ctx
, fd
);
5389 file
= __io_file_get(state
, fd
);
5390 if (unlikely(!file
))
5398 static int io_req_set_file(struct io_submit_state
*state
, struct io_kiocb
*req
,
5403 fixed
= (req
->flags
& REQ_F_FIXED_FILE
) != 0;
5404 if (unlikely(!fixed
&& req
->needs_fixed_file
))
5407 return io_file_get(state
, req
, fd
, &req
->file
, fixed
);
5410 static int io_grab_files(struct io_kiocb
*req
)
5413 struct io_ring_ctx
*ctx
= req
->ctx
;
5415 if (req
->work
.files
|| (req
->flags
& REQ_F_NO_FILE_TABLE
))
5417 if (!ctx
->ring_file
)
5421 spin_lock_irq(&ctx
->inflight_lock
);
5423 * We use the f_ops->flush() handler to ensure that we can flush
5424 * out work accessing these files if the fd is closed. Check if
5425 * the fd has changed since we started down this path, and disallow
5426 * this operation if it has.
5428 if (fcheck(ctx
->ring_fd
) == ctx
->ring_file
) {
5429 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
5430 req
->flags
|= REQ_F_INFLIGHT
;
5431 req
->work
.files
= current
->files
;
5434 spin_unlock_irq(&ctx
->inflight_lock
);
5440 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
5442 struct io_timeout_data
*data
= container_of(timer
,
5443 struct io_timeout_data
, timer
);
5444 struct io_kiocb
*req
= data
->req
;
5445 struct io_ring_ctx
*ctx
= req
->ctx
;
5446 struct io_kiocb
*prev
= NULL
;
5447 unsigned long flags
;
5449 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5452 * We don't expect the list to be empty, that will only happen if we
5453 * race with the completion of the linked work.
5455 if (!list_empty(&req
->link_list
)) {
5456 prev
= list_entry(req
->link_list
.prev
, struct io_kiocb
,
5458 if (refcount_inc_not_zero(&prev
->refs
)) {
5459 list_del_init(&req
->link_list
);
5460 prev
->flags
&= ~REQ_F_LINK_TIMEOUT
;
5465 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5468 req_set_fail_links(prev
);
5469 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
5472 io_cqring_add_event(req
, -ETIME
);
5475 return HRTIMER_NORESTART
;
5478 static void io_queue_linked_timeout(struct io_kiocb
*req
)
5480 struct io_ring_ctx
*ctx
= req
->ctx
;
5483 * If the list is now empty, then our linked request finished before
5484 * we got a chance to setup the timer
5486 spin_lock_irq(&ctx
->completion_lock
);
5487 if (!list_empty(&req
->link_list
)) {
5488 struct io_timeout_data
*data
= &req
->io
->timeout
;
5490 data
->timer
.function
= io_link_timeout_fn
;
5491 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
5494 spin_unlock_irq(&ctx
->completion_lock
);
5496 /* drop submission reference */
5500 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
5502 struct io_kiocb
*nxt
;
5504 if (!(req
->flags
& REQ_F_LINK_HEAD
))
5506 /* for polled retry, if flag is set, we already went through here */
5507 if (req
->flags
& REQ_F_POLLED
)
5510 nxt
= list_first_entry_or_null(&req
->link_list
, struct io_kiocb
,
5512 if (!nxt
|| nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
5515 req
->flags
|= REQ_F_LINK_TIMEOUT
;
5519 static void __io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5521 struct io_kiocb
*linked_timeout
;
5522 struct io_kiocb
*nxt
;
5523 const struct cred
*old_creds
= NULL
;
5527 linked_timeout
= io_prep_linked_timeout(req
);
5529 if (req
->work
.creds
&& req
->work
.creds
!= current_cred()) {
5531 revert_creds(old_creds
);
5532 if (old_creds
== req
->work
.creds
)
5533 old_creds
= NULL
; /* restored original creds */
5535 old_creds
= override_creds(req
->work
.creds
);
5538 ret
= io_issue_sqe(req
, sqe
, true);
5541 * We async punt it if the file wasn't marked NOWAIT, or if the file
5542 * doesn't support non-blocking read/write attempts
5544 if (ret
== -EAGAIN
&& (!(req
->flags
& REQ_F_NOWAIT
) ||
5545 (req
->flags
& REQ_F_MUST_PUNT
))) {
5546 if (io_arm_poll_handler(req
)) {
5548 io_queue_linked_timeout(linked_timeout
);
5552 if (io_op_defs
[req
->opcode
].file_table
) {
5553 ret
= io_grab_files(req
);
5559 * Queued up for async execution, worker will release
5560 * submit reference when the iocb is actually submitted.
5562 io_queue_async_work(req
);
5568 /* drop submission reference */
5569 io_put_req_find_next(req
, &nxt
);
5571 if (linked_timeout
) {
5573 io_queue_linked_timeout(linked_timeout
);
5575 io_put_req(linked_timeout
);
5578 /* and drop final reference, if we failed */
5580 io_cqring_add_event(req
, ret
);
5581 req_set_fail_links(req
);
5587 if (req
->flags
& REQ_F_FORCE_ASYNC
)
5593 revert_creds(old_creds
);
5596 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5600 ret
= io_req_defer(req
, sqe
);
5602 if (ret
!= -EIOCBQUEUED
) {
5604 io_cqring_add_event(req
, ret
);
5605 req_set_fail_links(req
);
5606 io_double_put_req(req
);
5608 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
5609 ret
= io_req_defer_prep(req
, sqe
);
5610 if (unlikely(ret
< 0))
5613 * Never try inline submit of IOSQE_ASYNC is set, go straight
5614 * to async execution.
5616 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
5617 io_queue_async_work(req
);
5619 __io_queue_sqe(req
, sqe
);
5623 static inline void io_queue_link_head(struct io_kiocb
*req
)
5625 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
5626 io_cqring_add_event(req
, -ECANCELED
);
5627 io_double_put_req(req
);
5629 io_queue_sqe(req
, NULL
);
5632 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5633 struct io_submit_state
*state
, struct io_kiocb
**link
)
5635 struct io_ring_ctx
*ctx
= req
->ctx
;
5639 * If we already have a head request, queue this one for async
5640 * submittal once the head completes. If we don't have a head but
5641 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5642 * submitted sync once the chain is complete. If none of those
5643 * conditions are true (normal request), then just queue it.
5646 struct io_kiocb
*head
= *link
;
5649 * Taking sequential execution of a link, draining both sides
5650 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5651 * requests in the link. So, it drains the head and the
5652 * next after the link request. The last one is done via
5653 * drain_next flag to persist the effect across calls.
5655 if (req
->flags
& REQ_F_IO_DRAIN
) {
5656 head
->flags
|= REQ_F_IO_DRAIN
;
5657 ctx
->drain_next
= 1;
5659 if (io_alloc_async_ctx(req
))
5662 ret
= io_req_defer_prep(req
, sqe
);
5664 /* fail even hard links since we don't submit */
5665 head
->flags
|= REQ_F_FAIL_LINK
;
5668 trace_io_uring_link(ctx
, req
, head
);
5669 list_add_tail(&req
->link_list
, &head
->link_list
);
5671 /* last request of a link, enqueue the link */
5672 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
5673 io_queue_link_head(head
);
5677 if (unlikely(ctx
->drain_next
)) {
5678 req
->flags
|= REQ_F_IO_DRAIN
;
5679 ctx
->drain_next
= 0;
5681 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
5682 req
->flags
|= REQ_F_LINK_HEAD
;
5683 INIT_LIST_HEAD(&req
->link_list
);
5685 if (io_alloc_async_ctx(req
))
5688 ret
= io_req_defer_prep(req
, sqe
);
5690 req
->flags
|= REQ_F_FAIL_LINK
;
5693 io_queue_sqe(req
, sqe
);
5701 * Batched submission is done, ensure local IO is flushed out.
5703 static void io_submit_state_end(struct io_submit_state
*state
)
5705 blk_finish_plug(&state
->plug
);
5707 if (state
->free_reqs
)
5708 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
5712 * Start submission side cache.
5714 static void io_submit_state_start(struct io_submit_state
*state
,
5715 unsigned int max_ios
)
5717 blk_start_plug(&state
->plug
);
5718 state
->free_reqs
= 0;
5720 state
->ios_left
= max_ios
;
5723 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
5725 struct io_rings
*rings
= ctx
->rings
;
5728 * Ensure any loads from the SQEs are done at this point,
5729 * since once we write the new head, the application could
5730 * write new data to them.
5732 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
5736 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5737 * that is mapped by userspace. This means that care needs to be taken to
5738 * ensure that reads are stable, as we cannot rely on userspace always
5739 * being a good citizen. If members of the sqe are validated and then later
5740 * used, it's important that those reads are done through READ_ONCE() to
5741 * prevent a re-load down the line.
5743 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
5745 u32
*sq_array
= ctx
->sq_array
;
5749 * The cached sq head (or cq tail) serves two purposes:
5751 * 1) allows us to batch the cost of updating the user visible
5753 * 2) allows the kernel side to track the head on its own, even
5754 * though the application is the one updating it.
5756 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
5757 if (likely(head
< ctx
->sq_entries
))
5758 return &ctx
->sq_sqes
[head
];
5760 /* drop invalid entries */
5761 ctx
->cached_sq_dropped
++;
5762 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
5766 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
5768 ctx
->cached_sq_head
++;
5771 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5772 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5773 IOSQE_BUFFER_SELECT)
5775 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
5776 const struct io_uring_sqe
*sqe
,
5777 struct io_submit_state
*state
, bool async
)
5779 unsigned int sqe_flags
;
5783 * All io need record the previous position, if LINK vs DARIN,
5784 * it can be used to mark the position of the first IO in the
5787 req
->sequence
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
5788 req
->opcode
= READ_ONCE(sqe
->opcode
);
5789 req
->user_data
= READ_ONCE(sqe
->user_data
);
5794 /* one is dropped after submission, the other at completion */
5795 refcount_set(&req
->refs
, 2);
5798 req
->needs_fixed_file
= async
;
5799 INIT_IO_WORK(&req
->work
, io_wq_submit_work
);
5801 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
5804 if (io_op_defs
[req
->opcode
].needs_mm
&& !current
->mm
) {
5805 if (unlikely(!mmget_not_zero(ctx
->sqo_mm
)))
5807 use_mm(ctx
->sqo_mm
);
5810 sqe_flags
= READ_ONCE(sqe
->flags
);
5811 /* enforce forwards compatibility on users */
5812 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
5815 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
5816 !io_op_defs
[req
->opcode
].buffer_select
)
5819 id
= READ_ONCE(sqe
->personality
);
5821 req
->work
.creds
= idr_find(&ctx
->personality_idr
, id
);
5822 if (unlikely(!req
->work
.creds
))
5824 get_cred(req
->work
.creds
);
5827 /* same numerical values with corresponding REQ_F_*, safe to copy */
5828 req
->flags
|= sqe_flags
& (IOSQE_IO_DRAIN
| IOSQE_IO_HARDLINK
|
5829 IOSQE_ASYNC
| IOSQE_FIXED_FILE
|
5830 IOSQE_BUFFER_SELECT
| IOSQE_IO_LINK
);
5832 if (!io_op_defs
[req
->opcode
].needs_file
)
5835 return io_req_set_file(state
, req
, READ_ONCE(sqe
->fd
));
5838 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
,
5839 struct file
*ring_file
, int ring_fd
, bool async
)
5841 struct io_submit_state state
, *statep
= NULL
;
5842 struct io_kiocb
*link
= NULL
;
5843 int i
, submitted
= 0;
5845 /* if we have a backlog and couldn't flush it all, return BUSY */
5846 if (test_bit(0, &ctx
->sq_check_overflow
)) {
5847 if (!list_empty(&ctx
->cq_overflow_list
) &&
5848 !io_cqring_overflow_flush(ctx
, false))
5852 /* make sure SQ entry isn't read before tail */
5853 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
5855 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
5858 if (nr
> IO_PLUG_THRESHOLD
) {
5859 io_submit_state_start(&state
, nr
);
5863 ctx
->ring_fd
= ring_fd
;
5864 ctx
->ring_file
= ring_file
;
5866 for (i
= 0; i
< nr
; i
++) {
5867 const struct io_uring_sqe
*sqe
;
5868 struct io_kiocb
*req
;
5871 sqe
= io_get_sqe(ctx
);
5872 if (unlikely(!sqe
)) {
5873 io_consume_sqe(ctx
);
5876 req
= io_alloc_req(ctx
, statep
);
5877 if (unlikely(!req
)) {
5879 submitted
= -EAGAIN
;
5883 err
= io_init_req(ctx
, req
, sqe
, statep
, async
);
5884 io_consume_sqe(ctx
);
5885 /* will complete beyond this point, count as submitted */
5888 if (unlikely(err
)) {
5890 io_cqring_add_event(req
, err
);
5891 io_double_put_req(req
);
5895 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
5897 err
= io_submit_sqe(req
, sqe
, statep
, &link
);
5902 if (unlikely(submitted
!= nr
)) {
5903 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
5905 percpu_ref_put_many(&ctx
->refs
, nr
- ref_used
);
5908 io_queue_link_head(link
);
5910 io_submit_state_end(&state
);
5912 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5913 io_commit_sqring(ctx
);
5918 static inline void io_sq_thread_drop_mm(struct io_ring_ctx
*ctx
)
5920 struct mm_struct
*mm
= current
->mm
;
5928 static int io_sq_thread(void *data
)
5930 struct io_ring_ctx
*ctx
= data
;
5931 const struct cred
*old_cred
;
5932 mm_segment_t old_fs
;
5934 unsigned long timeout
;
5937 complete(&ctx
->completions
[1]);
5941 old_cred
= override_creds(ctx
->creds
);
5943 timeout
= jiffies
+ ctx
->sq_thread_idle
;
5944 while (!kthread_should_park()) {
5945 unsigned int to_submit
;
5947 if (!list_empty(&ctx
->poll_list
)) {
5948 unsigned nr_events
= 0;
5950 mutex_lock(&ctx
->uring_lock
);
5951 if (!list_empty(&ctx
->poll_list
))
5952 io_iopoll_getevents(ctx
, &nr_events
, 0);
5954 timeout
= jiffies
+ ctx
->sq_thread_idle
;
5955 mutex_unlock(&ctx
->uring_lock
);
5958 to_submit
= io_sqring_entries(ctx
);
5961 * If submit got -EBUSY, flag us as needing the application
5962 * to enter the kernel to reap and flush events.
5964 if (!to_submit
|| ret
== -EBUSY
) {
5966 * Drop cur_mm before scheduling, we can't hold it for
5967 * long periods (or over schedule()). Do this before
5968 * adding ourselves to the waitqueue, as the unuse/drop
5971 io_sq_thread_drop_mm(ctx
);
5974 * We're polling. If we're within the defined idle
5975 * period, then let us spin without work before going
5976 * to sleep. The exception is if we got EBUSY doing
5977 * more IO, we should wait for the application to
5978 * reap events and wake us up.
5980 if (!list_empty(&ctx
->poll_list
) ||
5981 (!time_after(jiffies
, timeout
) && ret
!= -EBUSY
&&
5982 !percpu_ref_is_dying(&ctx
->refs
))) {
5983 if (current
->task_works
)
5989 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
5990 TASK_INTERRUPTIBLE
);
5993 * While doing polled IO, before going to sleep, we need
5994 * to check if there are new reqs added to poll_list, it
5995 * is because reqs may have been punted to io worker and
5996 * will be added to poll_list later, hence check the
5999 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6000 !list_empty_careful(&ctx
->poll_list
)) {
6001 finish_wait(&ctx
->sqo_wait
, &wait
);
6005 /* Tell userspace we may need a wakeup call */
6006 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6007 /* make sure to read SQ tail after writing flags */
6010 to_submit
= io_sqring_entries(ctx
);
6011 if (!to_submit
|| ret
== -EBUSY
) {
6012 if (kthread_should_park()) {
6013 finish_wait(&ctx
->sqo_wait
, &wait
);
6016 if (current
->task_works
) {
6018 finish_wait(&ctx
->sqo_wait
, &wait
);
6021 if (signal_pending(current
))
6022 flush_signals(current
);
6024 finish_wait(&ctx
->sqo_wait
, &wait
);
6026 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6029 finish_wait(&ctx
->sqo_wait
, &wait
);
6031 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6034 mutex_lock(&ctx
->uring_lock
);
6035 ret
= io_submit_sqes(ctx
, to_submit
, NULL
, -1, true);
6036 mutex_unlock(&ctx
->uring_lock
);
6037 timeout
= jiffies
+ ctx
->sq_thread_idle
;
6040 if (current
->task_works
)
6044 io_sq_thread_drop_mm(ctx
);
6045 revert_creds(old_cred
);
6052 struct io_wait_queue
{
6053 struct wait_queue_entry wq
;
6054 struct io_ring_ctx
*ctx
;
6056 unsigned nr_timeouts
;
6059 static inline bool io_should_wake(struct io_wait_queue
*iowq
, bool noflush
)
6061 struct io_ring_ctx
*ctx
= iowq
->ctx
;
6064 * Wake up if we have enough events, or if a timeout occurred since we
6065 * started waiting. For timeouts, we always want to return to userspace,
6066 * regardless of event count.
6068 return io_cqring_events(ctx
, noflush
) >= iowq
->to_wait
||
6069 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
6072 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
6073 int wake_flags
, void *key
)
6075 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
6078 /* use noflush == true, as we can't safely rely on locking context */
6079 if (!io_should_wake(iowq
, true))
6082 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
6086 * Wait until events become available, if we don't already have some. The
6087 * application must reap them itself, as they reside on the shared cq ring.
6089 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
6090 const sigset_t __user
*sig
, size_t sigsz
)
6092 struct io_wait_queue iowq
= {
6095 .func
= io_wake_function
,
6096 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
6099 .to_wait
= min_events
,
6101 struct io_rings
*rings
= ctx
->rings
;
6105 if (io_cqring_events(ctx
, false) >= min_events
)
6107 if (!current
->task_works
)
6113 #ifdef CONFIG_COMPAT
6114 if (in_compat_syscall())
6115 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
6119 ret
= set_user_sigmask(sig
, sigsz
);
6125 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
6126 trace_io_uring_cqring_wait(ctx
, min_events
);
6128 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
6129 TASK_INTERRUPTIBLE
);
6130 if (current
->task_works
)
6132 if (io_should_wake(&iowq
, false))
6135 if (signal_pending(current
)) {
6140 finish_wait(&ctx
->wait
, &iowq
.wq
);
6142 restore_saved_sigmask_unless(ret
== -EINTR
);
6144 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
6147 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6149 #if defined(CONFIG_UNIX)
6150 if (ctx
->ring_sock
) {
6151 struct sock
*sock
= ctx
->ring_sock
->sk
;
6152 struct sk_buff
*skb
;
6154 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
6160 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6163 file
= io_file_from_index(ctx
, i
);
6170 static void io_file_ref_kill(struct percpu_ref
*ref
)
6172 struct fixed_file_data
*data
;
6174 data
= container_of(ref
, struct fixed_file_data
, refs
);
6175 complete(&data
->done
);
6178 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6180 struct fixed_file_data
*data
= ctx
->file_data
;
6181 struct fixed_file_ref_node
*ref_node
= NULL
;
6182 unsigned nr_tables
, i
;
6183 unsigned long flags
;
6188 spin_lock_irqsave(&data
->lock
, flags
);
6189 if (!list_empty(&data
->ref_list
))
6190 ref_node
= list_first_entry(&data
->ref_list
,
6191 struct fixed_file_ref_node
, node
);
6192 spin_unlock_irqrestore(&data
->lock
, flags
);
6194 percpu_ref_kill(&ref_node
->refs
);
6196 percpu_ref_kill(&data
->refs
);
6198 /* wait for all refs nodes to complete */
6199 wait_for_completion(&data
->done
);
6201 __io_sqe_files_unregister(ctx
);
6202 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
6203 for (i
= 0; i
< nr_tables
; i
++)
6204 kfree(data
->table
[i
].files
);
6206 percpu_ref_exit(&data
->refs
);
6208 ctx
->file_data
= NULL
;
6209 ctx
->nr_user_files
= 0;
6213 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
6215 if (ctx
->sqo_thread
) {
6216 wait_for_completion(&ctx
->completions
[1]);
6218 * The park is a bit of a work-around, without it we get
6219 * warning spews on shutdown with SQPOLL set and affinity
6220 * set to a single CPU.
6222 kthread_park(ctx
->sqo_thread
);
6223 kthread_stop(ctx
->sqo_thread
);
6224 ctx
->sqo_thread
= NULL
;
6228 static void io_finish_async(struct io_ring_ctx
*ctx
)
6230 io_sq_thread_stop(ctx
);
6233 io_wq_destroy(ctx
->io_wq
);
6238 #if defined(CONFIG_UNIX)
6240 * Ensure the UNIX gc is aware of our file set, so we are certain that
6241 * the io_uring can be safely unregistered on process exit, even if we have
6242 * loops in the file referencing.
6244 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
6246 struct sock
*sk
= ctx
->ring_sock
->sk
;
6247 struct scm_fp_list
*fpl
;
6248 struct sk_buff
*skb
;
6251 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
6255 skb
= alloc_skb(0, GFP_KERNEL
);
6264 fpl
->user
= get_uid(ctx
->user
);
6265 for (i
= 0; i
< nr
; i
++) {
6266 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
6270 fpl
->fp
[nr_files
] = get_file(file
);
6271 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
6276 fpl
->max
= SCM_MAX_FD
;
6277 fpl
->count
= nr_files
;
6278 UNIXCB(skb
).fp
= fpl
;
6279 skb
->destructor
= unix_destruct_scm
;
6280 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
6281 skb_queue_head(&sk
->sk_receive_queue
, skb
);
6283 for (i
= 0; i
< nr_files
; i
++)
6294 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6295 * causes regular reference counting to break down. We rely on the UNIX
6296 * garbage collection to take care of this problem for us.
6298 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
6300 unsigned left
, total
;
6304 left
= ctx
->nr_user_files
;
6306 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
6308 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
6312 total
+= this_files
;
6318 while (total
< ctx
->nr_user_files
) {
6319 struct file
*file
= io_file_from_index(ctx
, total
);
6329 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
6335 static int io_sqe_alloc_file_tables(struct io_ring_ctx
*ctx
, unsigned nr_tables
,
6340 for (i
= 0; i
< nr_tables
; i
++) {
6341 struct fixed_file_table
*table
= &ctx
->file_data
->table
[i
];
6342 unsigned this_files
;
6344 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
6345 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
6349 nr_files
-= this_files
;
6355 for (i
= 0; i
< nr_tables
; i
++) {
6356 struct fixed_file_table
*table
= &ctx
->file_data
->table
[i
];
6357 kfree(table
->files
);
6362 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
6364 #if defined(CONFIG_UNIX)
6365 struct sock
*sock
= ctx
->ring_sock
->sk
;
6366 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
6367 struct sk_buff
*skb
;
6370 __skb_queue_head_init(&list
);
6373 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6374 * remove this entry and rearrange the file array.
6376 skb
= skb_dequeue(head
);
6378 struct scm_fp_list
*fp
;
6380 fp
= UNIXCB(skb
).fp
;
6381 for (i
= 0; i
< fp
->count
; i
++) {
6384 if (fp
->fp
[i
] != file
)
6387 unix_notinflight(fp
->user
, fp
->fp
[i
]);
6388 left
= fp
->count
- 1 - i
;
6390 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
6391 left
* sizeof(struct file
*));
6398 __skb_queue_tail(&list
, skb
);
6408 __skb_queue_tail(&list
, skb
);
6410 skb
= skb_dequeue(head
);
6413 if (skb_peek(&list
)) {
6414 spin_lock_irq(&head
->lock
);
6415 while ((skb
= __skb_dequeue(&list
)) != NULL
)
6416 __skb_queue_tail(head
, skb
);
6417 spin_unlock_irq(&head
->lock
);
6424 struct io_file_put
{
6425 struct list_head list
;
6429 static void io_file_put_work(struct work_struct
*work
)
6431 struct fixed_file_ref_node
*ref_node
;
6432 struct fixed_file_data
*file_data
;
6433 struct io_ring_ctx
*ctx
;
6434 struct io_file_put
*pfile
, *tmp
;
6435 unsigned long flags
;
6437 ref_node
= container_of(work
, struct fixed_file_ref_node
, work
);
6438 file_data
= ref_node
->file_data
;
6439 ctx
= file_data
->ctx
;
6441 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
6442 list_del_init(&pfile
->list
);
6443 io_ring_file_put(ctx
, pfile
->file
);
6447 spin_lock_irqsave(&file_data
->lock
, flags
);
6448 list_del_init(&ref_node
->node
);
6449 spin_unlock_irqrestore(&file_data
->lock
, flags
);
6451 percpu_ref_exit(&ref_node
->refs
);
6453 percpu_ref_put(&file_data
->refs
);
6456 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
6458 struct fixed_file_ref_node
*ref_node
;
6460 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
6462 queue_work(system_wq
, &ref_node
->work
);
6465 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
6466 struct io_ring_ctx
*ctx
)
6468 struct fixed_file_ref_node
*ref_node
;
6470 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
6472 return ERR_PTR(-ENOMEM
);
6474 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
6477 return ERR_PTR(-ENOMEM
);
6479 INIT_LIST_HEAD(&ref_node
->node
);
6480 INIT_LIST_HEAD(&ref_node
->file_list
);
6481 INIT_WORK(&ref_node
->work
, io_file_put_work
);
6482 ref_node
->file_data
= ctx
->file_data
;
6487 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
6489 percpu_ref_exit(&ref_node
->refs
);
6493 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
6496 __s32 __user
*fds
= (__s32 __user
*) arg
;
6501 struct fixed_file_ref_node
*ref_node
;
6502 unsigned long flags
;
6508 if (nr_args
> IORING_MAX_FIXED_FILES
)
6511 ctx
->file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL
);
6512 if (!ctx
->file_data
)
6514 ctx
->file_data
->ctx
= ctx
;
6515 init_completion(&ctx
->file_data
->done
);
6516 INIT_LIST_HEAD(&ctx
->file_data
->ref_list
);
6517 spin_lock_init(&ctx
->file_data
->lock
);
6519 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
6520 ctx
->file_data
->table
= kcalloc(nr_tables
,
6521 sizeof(struct fixed_file_table
),
6523 if (!ctx
->file_data
->table
) {
6524 kfree(ctx
->file_data
);
6525 ctx
->file_data
= NULL
;
6529 if (percpu_ref_init(&ctx
->file_data
->refs
, io_file_ref_kill
,
6530 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
)) {
6531 kfree(ctx
->file_data
->table
);
6532 kfree(ctx
->file_data
);
6533 ctx
->file_data
= NULL
;
6537 if (io_sqe_alloc_file_tables(ctx
, nr_tables
, nr_args
)) {
6538 percpu_ref_exit(&ctx
->file_data
->refs
);
6539 kfree(ctx
->file_data
->table
);
6540 kfree(ctx
->file_data
);
6541 ctx
->file_data
= NULL
;
6545 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
6546 struct fixed_file_table
*table
;
6550 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
6552 /* allow sparse sets */
6558 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
6559 index
= i
& IORING_FILE_TABLE_MASK
;
6567 * Don't allow io_uring instances to be registered. If UNIX
6568 * isn't enabled, then this causes a reference cycle and this
6569 * instance can never get freed. If UNIX is enabled we'll
6570 * handle it just fine, but there's still no point in allowing
6571 * a ring fd as it doesn't support regular read/write anyway.
6573 if (file
->f_op
== &io_uring_fops
) {
6578 table
->files
[index
] = file
;
6582 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6583 file
= io_file_from_index(ctx
, i
);
6587 for (i
= 0; i
< nr_tables
; i
++)
6588 kfree(ctx
->file_data
->table
[i
].files
);
6590 kfree(ctx
->file_data
->table
);
6591 kfree(ctx
->file_data
);
6592 ctx
->file_data
= NULL
;
6593 ctx
->nr_user_files
= 0;
6597 ret
= io_sqe_files_scm(ctx
);
6599 io_sqe_files_unregister(ctx
);
6603 ref_node
= alloc_fixed_file_ref_node(ctx
);
6604 if (IS_ERR(ref_node
)) {
6605 io_sqe_files_unregister(ctx
);
6606 return PTR_ERR(ref_node
);
6609 ctx
->file_data
->cur_refs
= &ref_node
->refs
;
6610 spin_lock_irqsave(&ctx
->file_data
->lock
, flags
);
6611 list_add(&ref_node
->node
, &ctx
->file_data
->ref_list
);
6612 spin_unlock_irqrestore(&ctx
->file_data
->lock
, flags
);
6613 percpu_ref_get(&ctx
->file_data
->refs
);
6617 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
6620 #if defined(CONFIG_UNIX)
6621 struct sock
*sock
= ctx
->ring_sock
->sk
;
6622 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
6623 struct sk_buff
*skb
;
6626 * See if we can merge this file into an existing skb SCM_RIGHTS
6627 * file set. If there's no room, fall back to allocating a new skb
6628 * and filling it in.
6630 spin_lock_irq(&head
->lock
);
6631 skb
= skb_peek(head
);
6633 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
6635 if (fpl
->count
< SCM_MAX_FD
) {
6636 __skb_unlink(skb
, head
);
6637 spin_unlock_irq(&head
->lock
);
6638 fpl
->fp
[fpl
->count
] = get_file(file
);
6639 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
6641 spin_lock_irq(&head
->lock
);
6642 __skb_queue_head(head
, skb
);
6647 spin_unlock_irq(&head
->lock
);
6654 return __io_sqe_files_scm(ctx
, 1, index
);
6660 static int io_queue_file_removal(struct fixed_file_data
*data
,
6663 struct io_file_put
*pfile
;
6664 struct percpu_ref
*refs
= data
->cur_refs
;
6665 struct fixed_file_ref_node
*ref_node
;
6667 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
6671 ref_node
= container_of(refs
, struct fixed_file_ref_node
, refs
);
6673 list_add(&pfile
->list
, &ref_node
->file_list
);
6678 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
6679 struct io_uring_files_update
*up
,
6682 struct fixed_file_data
*data
= ctx
->file_data
;
6683 struct fixed_file_ref_node
*ref_node
;
6688 unsigned long flags
;
6689 bool needs_switch
= false;
6691 if (check_add_overflow(up
->offset
, nr_args
, &done
))
6693 if (done
> ctx
->nr_user_files
)
6696 ref_node
= alloc_fixed_file_ref_node(ctx
);
6697 if (IS_ERR(ref_node
))
6698 return PTR_ERR(ref_node
);
6701 fds
= u64_to_user_ptr(up
->fds
);
6703 struct fixed_file_table
*table
;
6707 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
6711 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
6712 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
6713 index
= i
& IORING_FILE_TABLE_MASK
;
6714 if (table
->files
[index
]) {
6715 file
= io_file_from_index(ctx
, index
);
6716 err
= io_queue_file_removal(data
, file
);
6719 table
->files
[index
] = NULL
;
6720 needs_switch
= true;
6729 * Don't allow io_uring instances to be registered. If
6730 * UNIX isn't enabled, then this causes a reference
6731 * cycle and this instance can never get freed. If UNIX
6732 * is enabled we'll handle it just fine, but there's
6733 * still no point in allowing a ring fd as it doesn't
6734 * support regular read/write anyway.
6736 if (file
->f_op
== &io_uring_fops
) {
6741 table
->files
[index
] = file
;
6742 err
= io_sqe_file_register(ctx
, file
, i
);
6752 percpu_ref_kill(data
->cur_refs
);
6753 spin_lock_irqsave(&data
->lock
, flags
);
6754 list_add(&ref_node
->node
, &data
->ref_list
);
6755 data
->cur_refs
= &ref_node
->refs
;
6756 spin_unlock_irqrestore(&data
->lock
, flags
);
6757 percpu_ref_get(&ctx
->file_data
->refs
);
6759 destroy_fixed_file_ref_node(ref_node
);
6761 return done
? done
: err
;
6764 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
6767 struct io_uring_files_update up
;
6769 if (!ctx
->file_data
)
6773 if (copy_from_user(&up
, arg
, sizeof(up
)))
6778 return __io_sqe_files_update(ctx
, &up
, nr_args
);
6781 static void io_free_work(struct io_wq_work
*work
)
6783 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6785 /* Consider that io_steal_work() relies on this ref */
6789 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
6790 struct io_uring_params
*p
)
6792 struct io_wq_data data
;
6794 struct io_ring_ctx
*ctx_attach
;
6795 unsigned int concurrency
;
6798 data
.user
= ctx
->user
;
6799 data
.free_work
= io_free_work
;
6801 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
6802 /* Do QD, or 4 * CPUS, whatever is smallest */
6803 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
6805 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
6806 if (IS_ERR(ctx
->io_wq
)) {
6807 ret
= PTR_ERR(ctx
->io_wq
);
6813 f
= fdget(p
->wq_fd
);
6817 if (f
.file
->f_op
!= &io_uring_fops
) {
6822 ctx_attach
= f
.file
->private_data
;
6823 /* @io_wq is protected by holding the fd */
6824 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
6829 ctx
->io_wq
= ctx_attach
->io_wq
;
6835 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
6836 struct io_uring_params
*p
)
6840 init_waitqueue_head(&ctx
->sqo_wait
);
6841 mmgrab(current
->mm
);
6842 ctx
->sqo_mm
= current
->mm
;
6844 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
6846 if (!capable(CAP_SYS_ADMIN
))
6849 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
6850 if (!ctx
->sq_thread_idle
)
6851 ctx
->sq_thread_idle
= HZ
;
6853 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
6854 int cpu
= p
->sq_thread_cpu
;
6857 if (cpu
>= nr_cpu_ids
)
6859 if (!cpu_online(cpu
))
6862 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
6866 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
6869 if (IS_ERR(ctx
->sqo_thread
)) {
6870 ret
= PTR_ERR(ctx
->sqo_thread
);
6871 ctx
->sqo_thread
= NULL
;
6874 wake_up_process(ctx
->sqo_thread
);
6875 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
6876 /* Can't have SQ_AFF without SQPOLL */
6881 ret
= io_init_wq_offload(ctx
, p
);
6887 io_finish_async(ctx
);
6888 mmdrop(ctx
->sqo_mm
);
6893 static void io_unaccount_mem(struct user_struct
*user
, unsigned long nr_pages
)
6895 atomic_long_sub(nr_pages
, &user
->locked_vm
);
6898 static int io_account_mem(struct user_struct
*user
, unsigned long nr_pages
)
6900 unsigned long page_limit
, cur_pages
, new_pages
;
6902 /* Don't allow more pages than we can safely lock */
6903 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
6906 cur_pages
= atomic_long_read(&user
->locked_vm
);
6907 new_pages
= cur_pages
+ nr_pages
;
6908 if (new_pages
> page_limit
)
6910 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
6911 new_pages
) != cur_pages
);
6916 static void io_mem_free(void *ptr
)
6923 page
= virt_to_head_page(ptr
);
6924 if (put_page_testzero(page
))
6925 free_compound_page(page
);
6928 static void *io_mem_alloc(size_t size
)
6930 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
6933 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
6936 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
6939 struct io_rings
*rings
;
6940 size_t off
, sq_array_size
;
6942 off
= struct_size(rings
, cqes
, cq_entries
);
6943 if (off
== SIZE_MAX
)
6947 off
= ALIGN(off
, SMP_CACHE_BYTES
);
6952 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
6953 if (sq_array_size
== SIZE_MAX
)
6956 if (check_add_overflow(off
, sq_array_size
, &off
))
6965 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
6969 pages
= (size_t)1 << get_order(
6970 rings_size(sq_entries
, cq_entries
, NULL
));
6971 pages
+= (size_t)1 << get_order(
6972 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
6977 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
6981 if (!ctx
->user_bufs
)
6984 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
6985 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
6987 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
6988 unpin_user_page(imu
->bvec
[j
].bv_page
);
6990 if (ctx
->account_mem
)
6991 io_unaccount_mem(ctx
->user
, imu
->nr_bvecs
);
6996 kfree(ctx
->user_bufs
);
6997 ctx
->user_bufs
= NULL
;
6998 ctx
->nr_user_bufs
= 0;
7002 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
7003 void __user
*arg
, unsigned index
)
7005 struct iovec __user
*src
;
7007 #ifdef CONFIG_COMPAT
7009 struct compat_iovec __user
*ciovs
;
7010 struct compat_iovec ciov
;
7012 ciovs
= (struct compat_iovec __user
*) arg
;
7013 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
7016 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
7017 dst
->iov_len
= ciov
.iov_len
;
7021 src
= (struct iovec __user
*) arg
;
7022 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
7027 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7030 struct vm_area_struct
**vmas
= NULL
;
7031 struct page
**pages
= NULL
;
7032 int i
, j
, got_pages
= 0;
7037 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
7040 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
7042 if (!ctx
->user_bufs
)
7045 for (i
= 0; i
< nr_args
; i
++) {
7046 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
7047 unsigned long off
, start
, end
, ubuf
;
7052 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
7057 * Don't impose further limits on the size and buffer
7058 * constraints here, we'll -EINVAL later when IO is
7059 * submitted if they are wrong.
7062 if (!iov
.iov_base
|| !iov
.iov_len
)
7065 /* arbitrary limit, but we need something */
7066 if (iov
.iov_len
> SZ_1G
)
7069 ubuf
= (unsigned long) iov
.iov_base
;
7070 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
7071 start
= ubuf
>> PAGE_SHIFT
;
7072 nr_pages
= end
- start
;
7074 if (ctx
->account_mem
) {
7075 ret
= io_account_mem(ctx
->user
, nr_pages
);
7081 if (!pages
|| nr_pages
> got_pages
) {
7084 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
7086 vmas
= kvmalloc_array(nr_pages
,
7087 sizeof(struct vm_area_struct
*),
7089 if (!pages
|| !vmas
) {
7091 if (ctx
->account_mem
)
7092 io_unaccount_mem(ctx
->user
, nr_pages
);
7095 got_pages
= nr_pages
;
7098 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
7102 if (ctx
->account_mem
)
7103 io_unaccount_mem(ctx
->user
, nr_pages
);
7108 down_read(¤t
->mm
->mmap_sem
);
7109 pret
= pin_user_pages(ubuf
, nr_pages
,
7110 FOLL_WRITE
| FOLL_LONGTERM
,
7112 if (pret
== nr_pages
) {
7113 /* don't support file backed memory */
7114 for (j
= 0; j
< nr_pages
; j
++) {
7115 struct vm_area_struct
*vma
= vmas
[j
];
7118 !is_file_hugepages(vma
->vm_file
)) {
7124 ret
= pret
< 0 ? pret
: -EFAULT
;
7126 up_read(¤t
->mm
->mmap_sem
);
7129 * if we did partial map, or found file backed vmas,
7130 * release any pages we did get
7133 unpin_user_pages(pages
, pret
);
7134 if (ctx
->account_mem
)
7135 io_unaccount_mem(ctx
->user
, nr_pages
);
7140 off
= ubuf
& ~PAGE_MASK
;
7142 for (j
= 0; j
< nr_pages
; j
++) {
7145 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
7146 imu
->bvec
[j
].bv_page
= pages
[j
];
7147 imu
->bvec
[j
].bv_len
= vec_len
;
7148 imu
->bvec
[j
].bv_offset
= off
;
7152 /* store original address for later verification */
7154 imu
->len
= iov
.iov_len
;
7155 imu
->nr_bvecs
= nr_pages
;
7157 ctx
->nr_user_bufs
++;
7165 io_sqe_buffer_unregister(ctx
);
7169 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
7171 __s32 __user
*fds
= arg
;
7177 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
7180 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
7181 if (IS_ERR(ctx
->cq_ev_fd
)) {
7182 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
7183 ctx
->cq_ev_fd
= NULL
;
7190 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
7192 if (ctx
->cq_ev_fd
) {
7193 eventfd_ctx_put(ctx
->cq_ev_fd
);
7194 ctx
->cq_ev_fd
= NULL
;
7201 static int __io_destroy_buffers(int id
, void *p
, void *data
)
7203 struct io_ring_ctx
*ctx
= data
;
7204 struct io_buffer
*buf
= p
;
7206 __io_remove_buffers(ctx
, buf
, id
, -1U);
7210 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
7212 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
7213 idr_destroy(&ctx
->io_buffer_idr
);
7216 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
7218 io_finish_async(ctx
);
7220 mmdrop(ctx
->sqo_mm
);
7222 io_iopoll_reap_events(ctx
);
7223 io_sqe_buffer_unregister(ctx
);
7224 io_sqe_files_unregister(ctx
);
7225 io_eventfd_unregister(ctx
);
7226 io_destroy_buffers(ctx
);
7227 idr_destroy(&ctx
->personality_idr
);
7229 #if defined(CONFIG_UNIX)
7230 if (ctx
->ring_sock
) {
7231 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
7232 sock_release(ctx
->ring_sock
);
7236 io_mem_free(ctx
->rings
);
7237 io_mem_free(ctx
->sq_sqes
);
7239 percpu_ref_exit(&ctx
->refs
);
7240 if (ctx
->account_mem
)
7241 io_unaccount_mem(ctx
->user
,
7242 ring_pages(ctx
->sq_entries
, ctx
->cq_entries
));
7243 free_uid(ctx
->user
);
7244 put_cred(ctx
->creds
);
7245 kfree(ctx
->completions
);
7246 kfree(ctx
->cancel_hash
);
7247 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
7251 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
7253 struct io_ring_ctx
*ctx
= file
->private_data
;
7256 poll_wait(file
, &ctx
->cq_wait
, wait
);
7258 * synchronizes with barrier from wq_has_sleeper call in
7262 if (READ_ONCE(ctx
->rings
->sq
.tail
) - ctx
->cached_sq_head
!=
7263 ctx
->rings
->sq_ring_entries
)
7264 mask
|= EPOLLOUT
| EPOLLWRNORM
;
7265 if (io_cqring_events(ctx
, false))
7266 mask
|= EPOLLIN
| EPOLLRDNORM
;
7271 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
7273 struct io_ring_ctx
*ctx
= file
->private_data
;
7275 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
7278 static int io_remove_personalities(int id
, void *p
, void *data
)
7280 struct io_ring_ctx
*ctx
= data
;
7281 const struct cred
*cred
;
7283 cred
= idr_remove(&ctx
->personality_idr
, id
);
7289 static void io_ring_exit_work(struct work_struct
*work
)
7291 struct io_ring_ctx
*ctx
;
7293 ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
7295 io_cqring_overflow_flush(ctx
, true);
7297 wait_for_completion(&ctx
->completions
[0]);
7298 io_ring_ctx_free(ctx
);
7301 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
7303 mutex_lock(&ctx
->uring_lock
);
7304 percpu_ref_kill(&ctx
->refs
);
7305 mutex_unlock(&ctx
->uring_lock
);
7308 * Wait for sq thread to idle, if we have one. It won't spin on new
7309 * work after we've killed the ctx ref above. This is important to do
7310 * before we cancel existing commands, as the thread could otherwise
7311 * be queueing new work post that. If that's work we need to cancel,
7312 * it could cause shutdown to hang.
7314 while (ctx
->sqo_thread
&& !wq_has_sleeper(&ctx
->sqo_wait
))
7317 io_kill_timeouts(ctx
);
7318 io_poll_remove_all(ctx
);
7321 io_wq_cancel_all(ctx
->io_wq
);
7323 io_iopoll_reap_events(ctx
);
7324 /* if we failed setting up the ctx, we might not have any rings */
7326 io_cqring_overflow_flush(ctx
, true);
7327 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
7328 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
7329 queue_work(system_wq
, &ctx
->exit_work
);
7332 static int io_uring_release(struct inode
*inode
, struct file
*file
)
7334 struct io_ring_ctx
*ctx
= file
->private_data
;
7336 file
->private_data
= NULL
;
7337 io_ring_ctx_wait_and_kill(ctx
);
7341 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
7342 struct files_struct
*files
)
7344 while (!list_empty_careful(&ctx
->inflight_list
)) {
7345 struct io_kiocb
*cancel_req
= NULL
, *req
;
7348 spin_lock_irq(&ctx
->inflight_lock
);
7349 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
) {
7350 if (req
->work
.files
!= files
)
7352 /* req is being completed, ignore */
7353 if (!refcount_inc_not_zero(&req
->refs
))
7359 prepare_to_wait(&ctx
->inflight_wait
, &wait
,
7360 TASK_UNINTERRUPTIBLE
);
7361 spin_unlock_irq(&ctx
->inflight_lock
);
7363 /* We need to keep going until we don't find a matching req */
7367 if (cancel_req
->flags
& REQ_F_OVERFLOW
) {
7368 spin_lock_irq(&ctx
->completion_lock
);
7369 list_del(&cancel_req
->list
);
7370 cancel_req
->flags
&= ~REQ_F_OVERFLOW
;
7371 if (list_empty(&ctx
->cq_overflow_list
)) {
7372 clear_bit(0, &ctx
->sq_check_overflow
);
7373 clear_bit(0, &ctx
->cq_check_overflow
);
7375 spin_unlock_irq(&ctx
->completion_lock
);
7377 WRITE_ONCE(ctx
->rings
->cq_overflow
,
7378 atomic_inc_return(&ctx
->cached_cq_overflow
));
7381 * Put inflight ref and overflow ref. If that's
7382 * all we had, then we're done with this request.
7384 if (refcount_sub_and_test(2, &cancel_req
->refs
)) {
7385 io_put_req(cancel_req
);
7386 finish_wait(&ctx
->inflight_wait
, &wait
);
7391 io_wq_cancel_work(ctx
->io_wq
, &cancel_req
->work
);
7392 io_put_req(cancel_req
);
7394 finish_wait(&ctx
->inflight_wait
, &wait
);
7398 static int io_uring_flush(struct file
*file
, void *data
)
7400 struct io_ring_ctx
*ctx
= file
->private_data
;
7402 io_uring_cancel_files(ctx
, data
);
7405 * If the task is going away, cancel work it may have pending
7407 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
))
7408 io_wq_cancel_pid(ctx
->io_wq
, task_pid_vnr(current
));
7413 static void *io_uring_validate_mmap_request(struct file
*file
,
7414 loff_t pgoff
, size_t sz
)
7416 struct io_ring_ctx
*ctx
= file
->private_data
;
7417 loff_t offset
= pgoff
<< PAGE_SHIFT
;
7422 case IORING_OFF_SQ_RING
:
7423 case IORING_OFF_CQ_RING
:
7426 case IORING_OFF_SQES
:
7430 return ERR_PTR(-EINVAL
);
7433 page
= virt_to_head_page(ptr
);
7434 if (sz
> page_size(page
))
7435 return ERR_PTR(-EINVAL
);
7442 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
7444 size_t sz
= vma
->vm_end
- vma
->vm_start
;
7448 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
7450 return PTR_ERR(ptr
);
7452 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
7453 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
7456 #else /* !CONFIG_MMU */
7458 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
7460 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
7463 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
7465 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
7468 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
7469 unsigned long addr
, unsigned long len
,
7470 unsigned long pgoff
, unsigned long flags
)
7474 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
7476 return PTR_ERR(ptr
);
7478 return (unsigned long) ptr
;
7481 #endif /* !CONFIG_MMU */
7483 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
7484 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
7487 struct io_ring_ctx
*ctx
;
7492 if (current
->task_works
)
7495 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
7503 if (f
.file
->f_op
!= &io_uring_fops
)
7507 ctx
= f
.file
->private_data
;
7508 if (!percpu_ref_tryget(&ctx
->refs
))
7512 * For SQ polling, the thread will do all submissions and completions.
7513 * Just return the requested submit count, and wake the thread if
7517 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7518 if (!list_empty_careful(&ctx
->cq_overflow_list
))
7519 io_cqring_overflow_flush(ctx
, false);
7520 if (flags
& IORING_ENTER_SQ_WAKEUP
)
7521 wake_up(&ctx
->sqo_wait
);
7522 submitted
= to_submit
;
7523 } else if (to_submit
) {
7524 mutex_lock(&ctx
->uring_lock
);
7525 submitted
= io_submit_sqes(ctx
, to_submit
, f
.file
, fd
, false);
7526 mutex_unlock(&ctx
->uring_lock
);
7528 if (submitted
!= to_submit
)
7531 if (flags
& IORING_ENTER_GETEVENTS
) {
7532 unsigned nr_events
= 0;
7534 min_complete
= min(min_complete
, ctx
->cq_entries
);
7537 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7538 * space applications don't need to do io completion events
7539 * polling again, they can rely on io_sq_thread to do polling
7540 * work, which can reduce cpu usage and uring_lock contention.
7542 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
7543 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
7544 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
7546 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
7551 percpu_ref_put(&ctx
->refs
);
7554 return submitted
? submitted
: ret
;
7557 #ifdef CONFIG_PROC_FS
7558 static int io_uring_show_cred(int id
, void *p
, void *data
)
7560 const struct cred
*cred
= p
;
7561 struct seq_file
*m
= data
;
7562 struct user_namespace
*uns
= seq_user_ns(m
);
7563 struct group_info
*gi
;
7568 seq_printf(m
, "%5d\n", id
);
7569 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
7570 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
7571 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
7572 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
7573 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
7574 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
7575 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
7576 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
7577 seq_puts(m
, "\n\tGroups:\t");
7578 gi
= cred
->group_info
;
7579 for (g
= 0; g
< gi
->ngroups
; g
++) {
7580 seq_put_decimal_ull(m
, g
? " " : "",
7581 from_kgid_munged(uns
, gi
->gid
[g
]));
7583 seq_puts(m
, "\n\tCapEff:\t");
7584 cap
= cred
->cap_effective
;
7585 CAP_FOR_EACH_U32(__capi
)
7586 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
7591 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
7595 mutex_lock(&ctx
->uring_lock
);
7596 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
7597 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7598 struct fixed_file_table
*table
;
7601 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7602 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
7604 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
7606 seq_printf(m
, "%5u: <none>\n", i
);
7608 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
7609 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
7610 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
7612 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
7613 (unsigned int) buf
->len
);
7615 if (!idr_is_empty(&ctx
->personality_idr
)) {
7616 seq_printf(m
, "Personalities:\n");
7617 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
7619 seq_printf(m
, "PollList:\n");
7620 spin_lock_irq(&ctx
->completion_lock
);
7621 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
7622 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
7623 struct io_kiocb
*req
;
7625 hlist_for_each_entry(req
, list
, hash_node
)
7626 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
7627 req
->task
->task_works
!= NULL
);
7629 spin_unlock_irq(&ctx
->completion_lock
);
7630 mutex_unlock(&ctx
->uring_lock
);
7633 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
7635 struct io_ring_ctx
*ctx
= f
->private_data
;
7637 if (percpu_ref_tryget(&ctx
->refs
)) {
7638 __io_uring_show_fdinfo(ctx
, m
);
7639 percpu_ref_put(&ctx
->refs
);
7644 static const struct file_operations io_uring_fops
= {
7645 .release
= io_uring_release
,
7646 .flush
= io_uring_flush
,
7647 .mmap
= io_uring_mmap
,
7649 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
7650 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
7652 .poll
= io_uring_poll
,
7653 .fasync
= io_uring_fasync
,
7654 #ifdef CONFIG_PROC_FS
7655 .show_fdinfo
= io_uring_show_fdinfo
,
7659 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
7660 struct io_uring_params
*p
)
7662 struct io_rings
*rings
;
7663 size_t size
, sq_array_offset
;
7665 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
7666 if (size
== SIZE_MAX
)
7669 rings
= io_mem_alloc(size
);
7674 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
7675 rings
->sq_ring_mask
= p
->sq_entries
- 1;
7676 rings
->cq_ring_mask
= p
->cq_entries
- 1;
7677 rings
->sq_ring_entries
= p
->sq_entries
;
7678 rings
->cq_ring_entries
= p
->cq_entries
;
7679 ctx
->sq_mask
= rings
->sq_ring_mask
;
7680 ctx
->cq_mask
= rings
->cq_ring_mask
;
7681 ctx
->sq_entries
= rings
->sq_ring_entries
;
7682 ctx
->cq_entries
= rings
->cq_ring_entries
;
7684 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
7685 if (size
== SIZE_MAX
) {
7686 io_mem_free(ctx
->rings
);
7691 ctx
->sq_sqes
= io_mem_alloc(size
);
7692 if (!ctx
->sq_sqes
) {
7693 io_mem_free(ctx
->rings
);
7702 * Allocate an anonymous fd, this is what constitutes the application
7703 * visible backing of an io_uring instance. The application mmaps this
7704 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7705 * we have to tie this fd to a socket for file garbage collection purposes.
7707 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
7712 #if defined(CONFIG_UNIX)
7713 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
7719 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
7723 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
7724 O_RDWR
| O_CLOEXEC
);
7727 ret
= PTR_ERR(file
);
7731 #if defined(CONFIG_UNIX)
7732 ctx
->ring_sock
->file
= file
;
7734 fd_install(ret
, file
);
7737 #if defined(CONFIG_UNIX)
7738 sock_release(ctx
->ring_sock
);
7739 ctx
->ring_sock
= NULL
;
7744 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
7745 struct io_uring_params __user
*params
)
7747 struct user_struct
*user
= NULL
;
7748 struct io_ring_ctx
*ctx
;
7754 if (entries
> IORING_MAX_ENTRIES
) {
7755 if (!(p
->flags
& IORING_SETUP_CLAMP
))
7757 entries
= IORING_MAX_ENTRIES
;
7761 * Use twice as many entries for the CQ ring. It's possible for the
7762 * application to drive a higher depth than the size of the SQ ring,
7763 * since the sqes are only used at submission time. This allows for
7764 * some flexibility in overcommitting a bit. If the application has
7765 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7766 * of CQ ring entries manually.
7768 p
->sq_entries
= roundup_pow_of_two(entries
);
7769 if (p
->flags
& IORING_SETUP_CQSIZE
) {
7771 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7772 * to a power-of-two, if it isn't already. We do NOT impose
7773 * any cq vs sq ring sizing.
7775 if (p
->cq_entries
< p
->sq_entries
)
7777 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
7778 if (!(p
->flags
& IORING_SETUP_CLAMP
))
7780 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
7782 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
7784 p
->cq_entries
= 2 * p
->sq_entries
;
7787 user
= get_uid(current_user());
7788 account_mem
= !capable(CAP_IPC_LOCK
);
7791 ret
= io_account_mem(user
,
7792 ring_pages(p
->sq_entries
, p
->cq_entries
));
7799 ctx
= io_ring_ctx_alloc(p
);
7802 io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
7807 ctx
->compat
= in_compat_syscall();
7808 ctx
->account_mem
= account_mem
;
7810 ctx
->creds
= get_current_cred();
7812 ret
= io_allocate_scq_urings(ctx
, p
);
7816 ret
= io_sq_offload_start(ctx
, p
);
7820 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
7821 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
7822 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
7823 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
7824 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
7825 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
7826 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
7827 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
7829 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
7830 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
7831 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
7832 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
7833 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
7834 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
7835 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
7837 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
7838 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
7839 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
;
7841 if (copy_to_user(params
, p
, sizeof(*p
))) {
7846 * Install ring fd as the very last thing, so we don't risk someone
7847 * having closed it before we finish setup
7849 ret
= io_uring_get_fd(ctx
);
7853 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
7856 io_ring_ctx_wait_and_kill(ctx
);
7861 * Sets up an aio uring context, and returns the fd. Applications asks for a
7862 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7863 * params structure passed in.
7865 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
7867 struct io_uring_params p
;
7870 if (copy_from_user(&p
, params
, sizeof(p
)))
7872 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
7877 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
7878 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
7879 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
))
7882 return io_uring_create(entries
, &p
, params
);
7885 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
7886 struct io_uring_params __user
*, params
)
7888 return io_uring_setup(entries
, params
);
7891 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
7893 struct io_uring_probe
*p
;
7897 size
= struct_size(p
, ops
, nr_args
);
7898 if (size
== SIZE_MAX
)
7900 p
= kzalloc(size
, GFP_KERNEL
);
7905 if (copy_from_user(p
, arg
, size
))
7908 if (memchr_inv(p
, 0, size
))
7911 p
->last_op
= IORING_OP_LAST
- 1;
7912 if (nr_args
> IORING_OP_LAST
)
7913 nr_args
= IORING_OP_LAST
;
7915 for (i
= 0; i
< nr_args
; i
++) {
7917 if (!io_op_defs
[i
].not_supported
)
7918 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
7923 if (copy_to_user(arg
, p
, size
))
7930 static int io_register_personality(struct io_ring_ctx
*ctx
)
7932 const struct cred
*creds
= get_current_cred();
7935 id
= idr_alloc_cyclic(&ctx
->personality_idr
, (void *) creds
, 1,
7936 USHRT_MAX
, GFP_KERNEL
);
7942 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
7944 const struct cred
*old_creds
;
7946 old_creds
= idr_remove(&ctx
->personality_idr
, id
);
7948 put_cred(old_creds
);
7955 static bool io_register_op_must_quiesce(int op
)
7958 case IORING_UNREGISTER_FILES
:
7959 case IORING_REGISTER_FILES_UPDATE
:
7960 case IORING_REGISTER_PROBE
:
7961 case IORING_REGISTER_PERSONALITY
:
7962 case IORING_UNREGISTER_PERSONALITY
:
7969 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
7970 void __user
*arg
, unsigned nr_args
)
7971 __releases(ctx
->uring_lock
)
7972 __acquires(ctx
->uring_lock
)
7977 * We're inside the ring mutex, if the ref is already dying, then
7978 * someone else killed the ctx or is already going through
7979 * io_uring_register().
7981 if (percpu_ref_is_dying(&ctx
->refs
))
7984 if (io_register_op_must_quiesce(opcode
)) {
7985 percpu_ref_kill(&ctx
->refs
);
7988 * Drop uring mutex before waiting for references to exit. If
7989 * another thread is currently inside io_uring_enter() it might
7990 * need to grab the uring_lock to make progress. If we hold it
7991 * here across the drain wait, then we can deadlock. It's safe
7992 * to drop the mutex here, since no new references will come in
7993 * after we've killed the percpu ref.
7995 mutex_unlock(&ctx
->uring_lock
);
7996 ret
= wait_for_completion_interruptible(&ctx
->completions
[0]);
7997 mutex_lock(&ctx
->uring_lock
);
7999 percpu_ref_resurrect(&ctx
->refs
);
8006 case IORING_REGISTER_BUFFERS
:
8007 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
8009 case IORING_UNREGISTER_BUFFERS
:
8013 ret
= io_sqe_buffer_unregister(ctx
);
8015 case IORING_REGISTER_FILES
:
8016 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
8018 case IORING_UNREGISTER_FILES
:
8022 ret
= io_sqe_files_unregister(ctx
);
8024 case IORING_REGISTER_FILES_UPDATE
:
8025 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
8027 case IORING_REGISTER_EVENTFD
:
8028 case IORING_REGISTER_EVENTFD_ASYNC
:
8032 ret
= io_eventfd_register(ctx
, arg
);
8035 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
8036 ctx
->eventfd_async
= 1;
8038 ctx
->eventfd_async
= 0;
8040 case IORING_UNREGISTER_EVENTFD
:
8044 ret
= io_eventfd_unregister(ctx
);
8046 case IORING_REGISTER_PROBE
:
8048 if (!arg
|| nr_args
> 256)
8050 ret
= io_probe(ctx
, arg
, nr_args
);
8052 case IORING_REGISTER_PERSONALITY
:
8056 ret
= io_register_personality(ctx
);
8058 case IORING_UNREGISTER_PERSONALITY
:
8062 ret
= io_unregister_personality(ctx
, nr_args
);
8069 if (io_register_op_must_quiesce(opcode
)) {
8070 /* bring the ctx back to life */
8071 percpu_ref_reinit(&ctx
->refs
);
8073 reinit_completion(&ctx
->completions
[0]);
8078 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
8079 void __user
*, arg
, unsigned int, nr_args
)
8081 struct io_ring_ctx
*ctx
;
8090 if (f
.file
->f_op
!= &io_uring_fops
)
8093 ctx
= f
.file
->private_data
;
8095 mutex_lock(&ctx
->uring_lock
);
8096 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
8097 mutex_unlock(&ctx
->uring_lock
);
8098 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
8099 ctx
->cq_ev_fd
!= NULL
, ret
);
8105 static int __init
io_uring_init(void)
8107 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8108 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8109 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8112 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8113 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8114 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
8115 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
8116 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
8117 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
8118 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
8119 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
8120 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
8121 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
8122 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
8123 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
8124 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
8125 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
8126 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
8127 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
8128 BUILD_BUG_SQE_ELEM(28, __u16
, poll_events
);
8129 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
8130 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
8131 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
8132 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
8133 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
8134 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
8135 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
8136 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
8137 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
8138 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
8139 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
8140 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
8141 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
8143 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
8144 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
8145 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
8148 __initcall(io_uring_init
);