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/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.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.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
185 struct io_mapped_ubuf
{
188 struct bio_vec
*bvec
;
189 unsigned int nr_bvecs
;
192 struct fixed_file_table
{
196 struct fixed_file_ref_node
{
197 struct percpu_ref refs
;
198 struct list_head node
;
199 struct list_head file_list
;
200 struct fixed_file_data
*file_data
;
201 struct llist_node llist
;
204 struct fixed_file_data
{
205 struct fixed_file_table
*table
;
206 struct io_ring_ctx
*ctx
;
208 struct percpu_ref
*cur_refs
;
209 struct percpu_ref refs
;
210 struct completion done
;
211 struct list_head ref_list
;
216 struct list_head list
;
224 struct percpu_ref refs
;
225 } ____cacheline_aligned_in_smp
;
229 unsigned int compat
: 1;
230 unsigned int limit_mem
: 1;
231 unsigned int cq_overflow_flushed
: 1;
232 unsigned int drain_next
: 1;
233 unsigned int eventfd_async
: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head
;
250 unsigned sq_thread_idle
;
251 unsigned cached_sq_dropped
;
252 atomic_t cached_cq_overflow
;
253 unsigned long sq_check_overflow
;
255 struct list_head defer_list
;
256 struct list_head timeout_list
;
257 struct list_head cq_overflow_list
;
259 wait_queue_head_t inflight_wait
;
260 struct io_uring_sqe
*sq_sqes
;
261 } ____cacheline_aligned_in_smp
;
263 struct io_rings
*rings
;
267 struct task_struct
*sqo_thread
; /* if using sq thread polling */
268 struct mm_struct
*sqo_mm
;
269 wait_queue_head_t sqo_wait
;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data
*file_data
;
277 unsigned nr_user_files
;
279 struct file
*ring_file
;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs
;
283 struct io_mapped_ubuf
*user_bufs
;
285 struct user_struct
*user
;
287 const struct cred
*creds
;
289 struct completion ref_comp
;
290 struct completion sq_thread_comp
;
292 /* if all else fails... */
293 struct io_kiocb
*fallback_req
;
295 #if defined(CONFIG_UNIX)
296 struct socket
*ring_sock
;
299 struct idr io_buffer_idr
;
301 struct idr personality_idr
;
304 unsigned cached_cq_tail
;
307 atomic_t cq_timeouts
;
308 unsigned long cq_check_overflow
;
309 struct wait_queue_head cq_wait
;
310 struct fasync_struct
*cq_fasync
;
311 struct eventfd_ctx
*cq_ev_fd
;
312 } ____cacheline_aligned_in_smp
;
315 struct mutex uring_lock
;
316 wait_queue_head_t wait
;
317 } ____cacheline_aligned_in_smp
;
320 spinlock_t completion_lock
;
323 * ->poll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head poll_list
;
329 struct hlist_head
*cancel_hash
;
330 unsigned cancel_hash_bits
;
331 bool poll_multi_file
;
333 spinlock_t inflight_lock
;
334 struct list_head inflight_list
;
335 } ____cacheline_aligned_in_smp
;
337 struct delayed_work file_put_work
;
338 struct llist_head file_put_llist
;
340 struct work_struct exit_work
;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb
{
350 struct wait_queue_head
*head
;
356 struct wait_queue_entry wait
;
361 struct file
*put_file
;
365 struct io_timeout_data
{
366 struct io_kiocb
*req
;
367 struct hrtimer timer
;
368 struct timespec64 ts
;
369 enum hrtimer_mode mode
;
374 struct sockaddr __user
*addr
;
375 int __user
*addr_len
;
377 unsigned long nofile
;
402 /* NOTE: kiocb has the file as the first member, so don't do it here */
410 struct sockaddr __user
*addr
;
417 struct user_msghdr __user
*msg
;
423 struct io_buffer
*kbuf
;
429 struct filename
*filename
;
431 unsigned long nofile
;
434 struct io_files_update
{
460 struct epoll_event event
;
464 struct file
*file_out
;
465 struct file
*file_in
;
472 struct io_provide_buf
{
486 const char __user
*filename
;
487 struct statx __user
*buffer
;
490 struct io_async_connect
{
491 struct sockaddr_storage address
;
494 struct io_async_msghdr
{
495 struct iovec fast_iov
[UIO_FASTIOV
];
497 struct sockaddr __user
*uaddr
;
499 struct sockaddr_storage addr
;
503 struct iovec fast_iov
[UIO_FASTIOV
];
507 struct wait_page_queue wpq
;
508 struct callback_head task_work
;
511 struct io_async_ctx
{
513 struct io_async_rw rw
;
514 struct io_async_msghdr msg
;
515 struct io_async_connect connect
;
516 struct io_timeout_data timeout
;
521 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
522 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
523 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
524 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
525 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
526 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
534 REQ_F_LINK_TIMEOUT_BIT
,
538 REQ_F_TIMEOUT_NOSEQ_BIT
,
539 REQ_F_COMP_LOCKED_BIT
,
540 REQ_F_NEED_CLEANUP_BIT
,
543 REQ_F_BUFFER_SELECTED_BIT
,
544 REQ_F_NO_FILE_TABLE_BIT
,
545 REQ_F_QUEUE_TIMEOUT_BIT
,
546 REQ_F_WORK_INITIALIZED_BIT
,
547 REQ_F_TASK_PINNED_BIT
,
549 /* not a real bit, just to check we're not overflowing the space */
555 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
556 /* drain existing IO first */
557 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
559 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
560 /* doesn't sever on completion < 0 */
561 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
563 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
564 /* IOSQE_BUFFER_SELECT */
565 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
568 REQ_F_LINK_HEAD
= BIT(REQ_F_LINK_HEAD_BIT
),
569 /* already grabbed next link */
570 REQ_F_LINK_NEXT
= BIT(REQ_F_LINK_NEXT_BIT
),
571 /* fail rest of links */
572 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
573 /* on inflight list */
574 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
575 /* read/write uses file position */
576 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
577 /* must not punt to workers */
578 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
579 /* has linked timeout */
580 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
581 /* timeout request */
582 REQ_F_TIMEOUT
= BIT(REQ_F_TIMEOUT_BIT
),
584 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
585 /* must be punted even for NONBLOCK */
586 REQ_F_MUST_PUNT
= BIT(REQ_F_MUST_PUNT_BIT
),
587 /* no timeout sequence */
588 REQ_F_TIMEOUT_NOSEQ
= BIT(REQ_F_TIMEOUT_NOSEQ_BIT
),
589 /* completion under lock */
590 REQ_F_COMP_LOCKED
= BIT(REQ_F_COMP_LOCKED_BIT
),
592 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
593 /* in overflow list */
594 REQ_F_OVERFLOW
= BIT(REQ_F_OVERFLOW_BIT
),
595 /* already went through poll handler */
596 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
597 /* buffer already selected */
598 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
599 /* doesn't need file table for this request */
600 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
601 /* needs to queue linked timeout */
602 REQ_F_QUEUE_TIMEOUT
= BIT(REQ_F_QUEUE_TIMEOUT_BIT
),
603 /* io_wq_work is initialized */
604 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
605 /* req->task is refcounted */
606 REQ_F_TASK_PINNED
= BIT(REQ_F_TASK_PINNED_BIT
),
610 struct io_poll_iocb poll
;
611 struct io_wq_work work
;
615 * NOTE! Each of the iocb union members has the file pointer
616 * as the first entry in their struct definition. So you can
617 * access the file pointer through any of the sub-structs,
618 * or directly as just 'ki_filp' in this struct.
624 struct io_poll_iocb poll
;
625 struct io_accept accept
;
627 struct io_cancel cancel
;
628 struct io_timeout timeout
;
629 struct io_connect connect
;
630 struct io_sr_msg sr_msg
;
632 struct io_close close
;
633 struct io_files_update files_update
;
634 struct io_fadvise fadvise
;
635 struct io_madvise madvise
;
636 struct io_epoll epoll
;
637 struct io_splice splice
;
638 struct io_provide_buf pbuf
;
639 struct io_statx statx
;
642 struct io_async_ctx
*io
;
645 /* polled IO has completed */
650 struct io_ring_ctx
*ctx
;
651 struct list_head list
;
654 struct task_struct
*task
;
660 struct list_head link_list
;
662 struct list_head inflight_entry
;
664 struct percpu_ref
*fixed_file_refs
;
668 * Only commands that never go async can use the below fields,
669 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
670 * async armed poll handlers for regular commands. The latter
671 * restore the work, if needed.
674 struct callback_head task_work
;
675 struct hlist_node hash_node
;
676 struct async_poll
*apoll
;
678 struct io_wq_work work
;
682 #define IO_IOPOLL_BATCH 8
684 struct io_submit_state
{
685 struct blk_plug plug
;
688 * io_kiocb alloc cache
690 void *reqs
[IO_IOPOLL_BATCH
];
691 unsigned int free_reqs
;
694 * File reference cache
698 unsigned int has_refs
;
699 unsigned int used_refs
;
700 unsigned int ios_left
;
704 /* needs req->io allocated for deferral/async */
705 unsigned async_ctx
: 1;
706 /* needs current->mm setup, does mm access */
707 unsigned needs_mm
: 1;
708 /* needs req->file assigned */
709 unsigned needs_file
: 1;
710 /* don't fail if file grab fails */
711 unsigned needs_file_no_error
: 1;
712 /* hash wq insertion if file is a regular file */
713 unsigned hash_reg_file
: 1;
714 /* unbound wq insertion if file is a non-regular file */
715 unsigned unbound_nonreg_file
: 1;
716 /* opcode is not supported by this kernel */
717 unsigned not_supported
: 1;
718 /* needs file table */
719 unsigned file_table
: 1;
721 unsigned needs_fs
: 1;
722 /* set if opcode supports polled "wait" */
724 unsigned pollout
: 1;
725 /* op supports buffer selection */
726 unsigned buffer_select
: 1;
729 static const struct io_op_def io_op_defs
[] = {
730 [IORING_OP_NOP
] = {},
731 [IORING_OP_READV
] = {
735 .unbound_nonreg_file
= 1,
739 [IORING_OP_WRITEV
] = {
744 .unbound_nonreg_file
= 1,
747 [IORING_OP_FSYNC
] = {
750 [IORING_OP_READ_FIXED
] = {
752 .unbound_nonreg_file
= 1,
755 [IORING_OP_WRITE_FIXED
] = {
758 .unbound_nonreg_file
= 1,
761 [IORING_OP_POLL_ADD
] = {
763 .unbound_nonreg_file
= 1,
765 [IORING_OP_POLL_REMOVE
] = {},
766 [IORING_OP_SYNC_FILE_RANGE
] = {
769 [IORING_OP_SENDMSG
] = {
773 .unbound_nonreg_file
= 1,
777 [IORING_OP_RECVMSG
] = {
781 .unbound_nonreg_file
= 1,
786 [IORING_OP_TIMEOUT
] = {
790 [IORING_OP_TIMEOUT_REMOVE
] = {},
791 [IORING_OP_ACCEPT
] = {
794 .unbound_nonreg_file
= 1,
798 [IORING_OP_ASYNC_CANCEL
] = {},
799 [IORING_OP_LINK_TIMEOUT
] = {
803 [IORING_OP_CONNECT
] = {
807 .unbound_nonreg_file
= 1,
810 [IORING_OP_FALLOCATE
] = {
813 [IORING_OP_OPENAT
] = {
817 [IORING_OP_CLOSE
] = {
819 .needs_file_no_error
= 1,
822 [IORING_OP_FILES_UPDATE
] = {
826 [IORING_OP_STATX
] = {
834 .unbound_nonreg_file
= 1,
838 [IORING_OP_WRITE
] = {
841 .unbound_nonreg_file
= 1,
844 [IORING_OP_FADVISE
] = {
847 [IORING_OP_MADVISE
] = {
853 .unbound_nonreg_file
= 1,
859 .unbound_nonreg_file
= 1,
863 [IORING_OP_OPENAT2
] = {
867 [IORING_OP_EPOLL_CTL
] = {
868 .unbound_nonreg_file
= 1,
871 [IORING_OP_SPLICE
] = {
874 .unbound_nonreg_file
= 1,
876 [IORING_OP_PROVIDE_BUFFERS
] = {},
877 [IORING_OP_REMOVE_BUFFERS
] = {},
881 .unbound_nonreg_file
= 1,
885 enum io_mem_account
{
890 static void io_wq_submit_work(struct io_wq_work
**workptr
);
891 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
892 static void io_put_req(struct io_kiocb
*req
);
893 static void __io_double_put_req(struct io_kiocb
*req
);
894 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
895 static void io_queue_linked_timeout(struct io_kiocb
*req
);
896 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
897 struct io_uring_files_update
*ip
,
899 static int io_grab_files(struct io_kiocb
*req
);
900 static void io_cleanup_req(struct io_kiocb
*req
);
901 static int io_file_get(struct io_submit_state
*state
, struct io_kiocb
*req
,
902 int fd
, struct file
**out_file
, bool fixed
);
903 static void __io_queue_sqe(struct io_kiocb
*req
,
904 const struct io_uring_sqe
*sqe
);
906 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
907 struct iovec
**iovec
, struct iov_iter
*iter
,
909 static int io_setup_async_rw(struct io_kiocb
*req
, ssize_t io_size
,
910 struct iovec
*iovec
, struct iovec
*fast_iov
,
911 struct iov_iter
*iter
);
913 static struct kmem_cache
*req_cachep
;
915 static const struct file_operations io_uring_fops
;
917 struct sock
*io_uring_get_socket(struct file
*file
)
919 #if defined(CONFIG_UNIX)
920 if (file
->f_op
== &io_uring_fops
) {
921 struct io_ring_ctx
*ctx
= file
->private_data
;
923 return ctx
->ring_sock
->sk
;
928 EXPORT_SYMBOL(io_uring_get_socket
);
930 static void io_get_req_task(struct io_kiocb
*req
)
932 if (req
->flags
& REQ_F_TASK_PINNED
)
934 get_task_struct(req
->task
);
935 req
->flags
|= REQ_F_TASK_PINNED
;
938 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
939 static void __io_put_req_task(struct io_kiocb
*req
)
941 if (req
->flags
& REQ_F_TASK_PINNED
)
942 put_task_struct(req
->task
);
945 static void io_file_put_work(struct work_struct
*work
);
948 * Note: must call io_req_init_async() for the first time you
949 * touch any members of io_wq_work.
951 static inline void io_req_init_async(struct io_kiocb
*req
)
953 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
956 memset(&req
->work
, 0, sizeof(req
->work
));
957 req
->flags
|= REQ_F_WORK_INITIALIZED
;
960 static inline bool io_async_submit(struct io_ring_ctx
*ctx
)
962 return ctx
->flags
& IORING_SETUP_SQPOLL
;
965 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
967 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
969 complete(&ctx
->ref_comp
);
972 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
974 struct io_ring_ctx
*ctx
;
977 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
981 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
982 if (!ctx
->fallback_req
)
986 * Use 5 bits less than the max cq entries, that should give us around
987 * 32 entries per hash list if totally full and uniformly spread.
989 hash_bits
= ilog2(p
->cq_entries
);
993 ctx
->cancel_hash_bits
= hash_bits
;
994 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
996 if (!ctx
->cancel_hash
)
998 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1000 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1001 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1004 ctx
->flags
= p
->flags
;
1005 init_waitqueue_head(&ctx
->sqo_wait
);
1006 init_waitqueue_head(&ctx
->cq_wait
);
1007 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1008 init_completion(&ctx
->ref_comp
);
1009 init_completion(&ctx
->sq_thread_comp
);
1010 idr_init(&ctx
->io_buffer_idr
);
1011 idr_init(&ctx
->personality_idr
);
1012 mutex_init(&ctx
->uring_lock
);
1013 init_waitqueue_head(&ctx
->wait
);
1014 spin_lock_init(&ctx
->completion_lock
);
1015 INIT_LIST_HEAD(&ctx
->poll_list
);
1016 INIT_LIST_HEAD(&ctx
->defer_list
);
1017 INIT_LIST_HEAD(&ctx
->timeout_list
);
1018 init_waitqueue_head(&ctx
->inflight_wait
);
1019 spin_lock_init(&ctx
->inflight_lock
);
1020 INIT_LIST_HEAD(&ctx
->inflight_list
);
1021 INIT_DELAYED_WORK(&ctx
->file_put_work
, io_file_put_work
);
1022 init_llist_head(&ctx
->file_put_llist
);
1025 if (ctx
->fallback_req
)
1026 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1027 kfree(ctx
->cancel_hash
);
1032 static inline bool __req_need_defer(struct io_kiocb
*req
)
1034 struct io_ring_ctx
*ctx
= req
->ctx
;
1036 return req
->sequence
!= ctx
->cached_cq_tail
1037 + atomic_read(&ctx
->cached_cq_overflow
);
1040 static inline bool req_need_defer(struct io_kiocb
*req
)
1042 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
))
1043 return __req_need_defer(req
);
1048 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1050 struct io_rings
*rings
= ctx
->rings
;
1052 /* order cqe stores with ring update */
1053 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1055 if (wq_has_sleeper(&ctx
->cq_wait
)) {
1056 wake_up_interruptible(&ctx
->cq_wait
);
1057 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1061 static inline void io_req_work_grab_env(struct io_kiocb
*req
,
1062 const struct io_op_def
*def
)
1064 if (!req
->work
.mm
&& def
->needs_mm
) {
1065 mmgrab(current
->mm
);
1066 req
->work
.mm
= current
->mm
;
1068 if (!req
->work
.creds
)
1069 req
->work
.creds
= get_current_cred();
1070 if (!req
->work
.fs
&& def
->needs_fs
) {
1071 spin_lock(¤t
->fs
->lock
);
1072 if (!current
->fs
->in_exec
) {
1073 req
->work
.fs
= current
->fs
;
1074 req
->work
.fs
->users
++;
1076 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1078 spin_unlock(¤t
->fs
->lock
);
1082 static inline void io_req_work_drop_env(struct io_kiocb
*req
)
1084 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1088 mmdrop(req
->work
.mm
);
1089 req
->work
.mm
= NULL
;
1091 if (req
->work
.creds
) {
1092 put_cred(req
->work
.creds
);
1093 req
->work
.creds
= NULL
;
1096 struct fs_struct
*fs
= req
->work
.fs
;
1098 spin_lock(&req
->work
.fs
->lock
);
1101 spin_unlock(&req
->work
.fs
->lock
);
1107 static inline void io_prep_async_work(struct io_kiocb
*req
,
1108 struct io_kiocb
**link
)
1110 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1112 if (req
->flags
& REQ_F_ISREG
) {
1113 if (def
->hash_reg_file
)
1114 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1116 if (def
->unbound_nonreg_file
)
1117 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1120 io_req_init_async(req
);
1121 io_req_work_grab_env(req
, def
);
1123 *link
= io_prep_linked_timeout(req
);
1126 static inline void io_queue_async_work(struct io_kiocb
*req
)
1128 struct io_ring_ctx
*ctx
= req
->ctx
;
1129 struct io_kiocb
*link
;
1131 io_prep_async_work(req
, &link
);
1133 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1134 &req
->work
, req
->flags
);
1135 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1138 io_queue_linked_timeout(link
);
1141 static void io_kill_timeout(struct io_kiocb
*req
)
1145 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
1147 atomic_inc(&req
->ctx
->cq_timeouts
);
1148 list_del_init(&req
->list
);
1149 req
->flags
|= REQ_F_COMP_LOCKED
;
1150 io_cqring_fill_event(req
, 0);
1155 static void io_kill_timeouts(struct io_ring_ctx
*ctx
)
1157 struct io_kiocb
*req
, *tmp
;
1159 spin_lock_irq(&ctx
->completion_lock
);
1160 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, list
)
1161 io_kill_timeout(req
);
1162 spin_unlock_irq(&ctx
->completion_lock
);
1165 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1168 struct io_kiocb
*req
= list_first_entry(&ctx
->defer_list
,
1169 struct io_kiocb
, list
);
1171 if (req_need_defer(req
))
1173 list_del_init(&req
->list
);
1174 io_queue_async_work(req
);
1175 } while (!list_empty(&ctx
->defer_list
));
1178 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1180 while (!list_empty(&ctx
->timeout_list
)) {
1181 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1182 struct io_kiocb
, list
);
1184 if (req
->flags
& REQ_F_TIMEOUT_NOSEQ
)
1186 if (req
->timeout
.target_seq
!= ctx
->cached_cq_tail
1187 - atomic_read(&ctx
->cq_timeouts
))
1190 list_del_init(&req
->list
);
1191 io_kill_timeout(req
);
1195 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1197 io_flush_timeouts(ctx
);
1198 __io_commit_cqring(ctx
);
1200 if (unlikely(!list_empty(&ctx
->defer_list
)))
1201 __io_queue_deferred(ctx
);
1204 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1206 struct io_rings
*rings
= ctx
->rings
;
1209 tail
= ctx
->cached_cq_tail
;
1211 * writes to the cq entry need to come after reading head; the
1212 * control dependency is enough as we're using WRITE_ONCE to
1215 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1218 ctx
->cached_cq_tail
++;
1219 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1222 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1226 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1228 if (!ctx
->eventfd_async
)
1230 return io_wq_current_is_worker();
1233 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1235 if (waitqueue_active(&ctx
->wait
))
1236 wake_up(&ctx
->wait
);
1237 if (waitqueue_active(&ctx
->sqo_wait
))
1238 wake_up(&ctx
->sqo_wait
);
1239 if (io_should_trigger_evfd(ctx
))
1240 eventfd_signal(ctx
->cq_ev_fd
, 1);
1243 /* Returns true if there are no backlogged entries after the flush */
1244 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1246 struct io_rings
*rings
= ctx
->rings
;
1247 struct io_uring_cqe
*cqe
;
1248 struct io_kiocb
*req
;
1249 unsigned long flags
;
1253 if (list_empty_careful(&ctx
->cq_overflow_list
))
1255 if ((ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
) ==
1256 rings
->cq_ring_entries
))
1260 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1262 /* if force is set, the ring is going away. always drop after that */
1264 ctx
->cq_overflow_flushed
= 1;
1267 while (!list_empty(&ctx
->cq_overflow_list
)) {
1268 cqe
= io_get_cqring(ctx
);
1272 req
= list_first_entry(&ctx
->cq_overflow_list
, struct io_kiocb
,
1274 list_move(&req
->list
, &list
);
1275 req
->flags
&= ~REQ_F_OVERFLOW
;
1277 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1278 WRITE_ONCE(cqe
->res
, req
->result
);
1279 WRITE_ONCE(cqe
->flags
, req
->cflags
);
1281 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1282 atomic_inc_return(&ctx
->cached_cq_overflow
));
1286 io_commit_cqring(ctx
);
1288 clear_bit(0, &ctx
->sq_check_overflow
);
1289 clear_bit(0, &ctx
->cq_check_overflow
);
1291 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1292 io_cqring_ev_posted(ctx
);
1294 while (!list_empty(&list
)) {
1295 req
= list_first_entry(&list
, struct io_kiocb
, list
);
1296 list_del(&req
->list
);
1303 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1305 struct io_ring_ctx
*ctx
= req
->ctx
;
1306 struct io_uring_cqe
*cqe
;
1308 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1311 * If we can't get a cq entry, userspace overflowed the
1312 * submission (by quite a lot). Increment the overflow count in
1315 cqe
= io_get_cqring(ctx
);
1317 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1318 WRITE_ONCE(cqe
->res
, res
);
1319 WRITE_ONCE(cqe
->flags
, cflags
);
1320 } else if (ctx
->cq_overflow_flushed
) {
1321 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1322 atomic_inc_return(&ctx
->cached_cq_overflow
));
1324 if (list_empty(&ctx
->cq_overflow_list
)) {
1325 set_bit(0, &ctx
->sq_check_overflow
);
1326 set_bit(0, &ctx
->cq_check_overflow
);
1328 req
->flags
|= REQ_F_OVERFLOW
;
1329 refcount_inc(&req
->refs
);
1331 req
->cflags
= cflags
;
1332 list_add_tail(&req
->list
, &ctx
->cq_overflow_list
);
1336 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1338 __io_cqring_fill_event(req
, res
, 0);
1341 static void __io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1343 struct io_ring_ctx
*ctx
= req
->ctx
;
1344 unsigned long flags
;
1346 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1347 __io_cqring_fill_event(req
, res
, cflags
);
1348 io_commit_cqring(ctx
);
1349 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1351 io_cqring_ev_posted(ctx
);
1354 static void io_cqring_add_event(struct io_kiocb
*req
, long res
)
1356 __io_cqring_add_event(req
, res
, 0);
1359 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1361 return req
== (struct io_kiocb
*)
1362 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1365 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1367 struct io_kiocb
*req
;
1369 req
= ctx
->fallback_req
;
1370 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1376 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1377 struct io_submit_state
*state
)
1379 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1380 struct io_kiocb
*req
;
1383 req
= kmem_cache_alloc(req_cachep
, gfp
);
1386 } else if (!state
->free_reqs
) {
1390 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1391 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
1394 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1395 * retry single alloc to be on the safe side.
1397 if (unlikely(ret
<= 0)) {
1398 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1399 if (!state
->reqs
[0])
1403 state
->free_reqs
= ret
- 1;
1404 req
= state
->reqs
[ret
- 1];
1407 req
= state
->reqs
[state
->free_reqs
];
1412 return io_get_fallback_req(ctx
);
1415 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
1419 percpu_ref_put(req
->fixed_file_refs
);
1424 static void __io_req_aux_free(struct io_kiocb
*req
)
1426 if (req
->flags
& REQ_F_NEED_CLEANUP
)
1427 io_cleanup_req(req
);
1431 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
1432 __io_put_req_task(req
);
1433 io_req_work_drop_env(req
);
1436 static void __io_free_req(struct io_kiocb
*req
)
1438 __io_req_aux_free(req
);
1440 if (req
->flags
& REQ_F_INFLIGHT
) {
1441 struct io_ring_ctx
*ctx
= req
->ctx
;
1442 unsigned long flags
;
1444 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1445 list_del(&req
->inflight_entry
);
1446 if (waitqueue_active(&ctx
->inflight_wait
))
1447 wake_up(&ctx
->inflight_wait
);
1448 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1451 percpu_ref_put(&req
->ctx
->refs
);
1452 if (likely(!io_is_fallback_req(req
)))
1453 kmem_cache_free(req_cachep
, req
);
1455 clear_bit_unlock(0, (unsigned long *) &req
->ctx
->fallback_req
);
1459 void *reqs
[IO_IOPOLL_BATCH
];
1464 static void io_free_req_many(struct io_ring_ctx
*ctx
, struct req_batch
*rb
)
1468 if (rb
->need_iter
) {
1469 int i
, inflight
= 0;
1470 unsigned long flags
;
1472 for (i
= 0; i
< rb
->to_free
; i
++) {
1473 struct io_kiocb
*req
= rb
->reqs
[i
];
1475 if (req
->flags
& REQ_F_INFLIGHT
)
1477 __io_req_aux_free(req
);
1482 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1483 for (i
= 0; i
< rb
->to_free
; i
++) {
1484 struct io_kiocb
*req
= rb
->reqs
[i
];
1486 if (req
->flags
& REQ_F_INFLIGHT
) {
1487 list_del(&req
->inflight_entry
);
1492 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1494 if (waitqueue_active(&ctx
->inflight_wait
))
1495 wake_up(&ctx
->inflight_wait
);
1498 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
1499 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
1500 rb
->to_free
= rb
->need_iter
= 0;
1503 static bool io_link_cancel_timeout(struct io_kiocb
*req
)
1505 struct io_ring_ctx
*ctx
= req
->ctx
;
1508 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
1510 io_cqring_fill_event(req
, -ECANCELED
);
1511 io_commit_cqring(ctx
);
1512 req
->flags
&= ~REQ_F_LINK_HEAD
;
1520 static void io_req_link_next(struct io_kiocb
*req
, struct io_kiocb
**nxtptr
)
1522 struct io_ring_ctx
*ctx
= req
->ctx
;
1523 bool wake_ev
= false;
1525 /* Already got next link */
1526 if (req
->flags
& REQ_F_LINK_NEXT
)
1530 * The list should never be empty when we are called here. But could
1531 * potentially happen if the chain is messed up, check to be on the
1534 while (!list_empty(&req
->link_list
)) {
1535 struct io_kiocb
*nxt
= list_first_entry(&req
->link_list
,
1536 struct io_kiocb
, link_list
);
1538 if (unlikely((req
->flags
& REQ_F_LINK_TIMEOUT
) &&
1539 (nxt
->flags
& REQ_F_TIMEOUT
))) {
1540 list_del_init(&nxt
->link_list
);
1541 wake_ev
|= io_link_cancel_timeout(nxt
);
1542 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1546 list_del_init(&req
->link_list
);
1547 if (!list_empty(&nxt
->link_list
))
1548 nxt
->flags
|= REQ_F_LINK_HEAD
;
1553 req
->flags
|= REQ_F_LINK_NEXT
;
1555 io_cqring_ev_posted(ctx
);
1559 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1561 static void io_fail_links(struct io_kiocb
*req
)
1563 struct io_ring_ctx
*ctx
= req
->ctx
;
1564 unsigned long flags
;
1566 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1568 while (!list_empty(&req
->link_list
)) {
1569 struct io_kiocb
*link
= list_first_entry(&req
->link_list
,
1570 struct io_kiocb
, link_list
);
1572 list_del_init(&link
->link_list
);
1573 trace_io_uring_fail_link(req
, link
);
1575 if ((req
->flags
& REQ_F_LINK_TIMEOUT
) &&
1576 link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
1577 io_link_cancel_timeout(link
);
1579 io_cqring_fill_event(link
, -ECANCELED
);
1580 __io_double_put_req(link
);
1582 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1585 io_commit_cqring(ctx
);
1586 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1587 io_cqring_ev_posted(ctx
);
1590 static void io_req_find_next(struct io_kiocb
*req
, struct io_kiocb
**nxt
)
1592 if (likely(!(req
->flags
& REQ_F_LINK_HEAD
)))
1596 * If LINK is set, we have dependent requests in this chain. If we
1597 * didn't fail this request, queue the first one up, moving any other
1598 * dependencies to the next request. In case of failure, fail the rest
1601 if (req
->flags
& REQ_F_FAIL_LINK
) {
1603 } else if ((req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_COMP_LOCKED
)) ==
1604 REQ_F_LINK_TIMEOUT
) {
1605 struct io_ring_ctx
*ctx
= req
->ctx
;
1606 unsigned long flags
;
1609 * If this is a timeout link, we could be racing with the
1610 * timeout timer. Grab the completion lock for this case to
1611 * protect against that.
1613 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1614 io_req_link_next(req
, nxt
);
1615 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1617 io_req_link_next(req
, nxt
);
1621 static void io_free_req(struct io_kiocb
*req
)
1623 struct io_kiocb
*nxt
= NULL
;
1625 io_req_find_next(req
, &nxt
);
1629 io_queue_async_work(nxt
);
1632 static void io_wq_assign_next(struct io_wq_work
**workptr
, struct io_kiocb
*nxt
)
1634 struct io_kiocb
*link
;
1635 const struct io_op_def
*def
= &io_op_defs
[nxt
->opcode
];
1637 if ((nxt
->flags
& REQ_F_ISREG
) && def
->hash_reg_file
)
1638 io_wq_hash_work(&nxt
->work
, file_inode(nxt
->file
));
1640 *workptr
= &nxt
->work
;
1641 link
= io_prep_linked_timeout(nxt
);
1643 nxt
->flags
|= REQ_F_QUEUE_TIMEOUT
;
1647 * Drop reference to request, return next in chain (if there is one) if this
1648 * was the last reference to this request.
1650 __attribute__((nonnull
))
1651 static void io_put_req_find_next(struct io_kiocb
*req
, struct io_kiocb
**nxtptr
)
1653 if (refcount_dec_and_test(&req
->refs
)) {
1654 io_req_find_next(req
, nxtptr
);
1659 static void io_put_req(struct io_kiocb
*req
)
1661 if (refcount_dec_and_test(&req
->refs
))
1665 static void io_steal_work(struct io_kiocb
*req
,
1666 struct io_wq_work
**workptr
)
1669 * It's in an io-wq worker, so there always should be at least
1670 * one reference, which will be dropped in io_put_work() just
1671 * after the current handler returns.
1673 * It also means, that if the counter dropped to 1, then there is
1674 * no asynchronous users left, so it's safe to steal the next work.
1676 if (refcount_read(&req
->refs
) == 1) {
1677 struct io_kiocb
*nxt
= NULL
;
1679 io_req_find_next(req
, &nxt
);
1681 io_wq_assign_next(workptr
, nxt
);
1686 * Must only be used if we don't need to care about links, usually from
1687 * within the completion handling itself.
1689 static void __io_double_put_req(struct io_kiocb
*req
)
1691 /* drop both submit and complete references */
1692 if (refcount_sub_and_test(2, &req
->refs
))
1696 static void io_double_put_req(struct io_kiocb
*req
)
1698 /* drop both submit and complete references */
1699 if (refcount_sub_and_test(2, &req
->refs
))
1703 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
, bool noflush
)
1705 struct io_rings
*rings
= ctx
->rings
;
1707 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1709 * noflush == true is from the waitqueue handler, just ensure
1710 * we wake up the task, and the next invocation will flush the
1711 * entries. We cannot safely to it from here.
1713 if (noflush
&& !list_empty(&ctx
->cq_overflow_list
))
1716 io_cqring_overflow_flush(ctx
, false);
1719 /* See comment at the top of this file */
1721 return ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
);
1724 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
1726 struct io_rings
*rings
= ctx
->rings
;
1728 /* make sure SQ entry isn't read before tail */
1729 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
1732 static inline bool io_req_multi_free(struct req_batch
*rb
, struct io_kiocb
*req
)
1734 if ((req
->flags
& REQ_F_LINK_HEAD
) || io_is_fallback_req(req
))
1737 if (req
->file
|| req
->io
)
1740 rb
->reqs
[rb
->to_free
++] = req
;
1741 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
1742 io_free_req_many(req
->ctx
, rb
);
1746 static int io_put_kbuf(struct io_kiocb
*req
)
1748 struct io_buffer
*kbuf
;
1751 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
1752 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
1753 cflags
|= IORING_CQE_F_BUFFER
;
1759 static void io_iopoll_queue(struct list_head
*again
)
1761 struct io_kiocb
*req
;
1764 req
= list_first_entry(again
, struct io_kiocb
, list
);
1765 list_del(&req
->list
);
1766 refcount_inc(&req
->refs
);
1767 io_queue_async_work(req
);
1768 } while (!list_empty(again
));
1772 * Find and free completed poll iocbs
1774 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1775 struct list_head
*done
)
1777 struct req_batch rb
;
1778 struct io_kiocb
*req
;
1781 /* order with ->result store in io_complete_rw_iopoll() */
1784 rb
.to_free
= rb
.need_iter
= 0;
1785 while (!list_empty(done
)) {
1788 req
= list_first_entry(done
, struct io_kiocb
, list
);
1789 if (READ_ONCE(req
->result
) == -EAGAIN
) {
1790 req
->iopoll_completed
= 0;
1791 list_move_tail(&req
->list
, &again
);
1794 list_del(&req
->list
);
1796 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
1797 cflags
= io_put_kbuf(req
);
1799 __io_cqring_fill_event(req
, req
->result
, cflags
);
1802 if (refcount_dec_and_test(&req
->refs
) &&
1803 !io_req_multi_free(&rb
, req
))
1807 io_commit_cqring(ctx
);
1808 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
1809 io_cqring_ev_posted(ctx
);
1810 io_free_req_many(ctx
, &rb
);
1812 if (!list_empty(&again
))
1813 io_iopoll_queue(&again
);
1816 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1819 struct io_kiocb
*req
, *tmp
;
1825 * Only spin for completions if we don't have multiple devices hanging
1826 * off our complete list, and we're under the requested amount.
1828 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
1831 list_for_each_entry_safe(req
, tmp
, &ctx
->poll_list
, list
) {
1832 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
1835 * Move completed and retryable entries to our local lists.
1836 * If we find a request that requires polling, break out
1837 * and complete those lists first, if we have entries there.
1839 if (READ_ONCE(req
->iopoll_completed
)) {
1840 list_move_tail(&req
->list
, &done
);
1843 if (!list_empty(&done
))
1846 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
1855 if (!list_empty(&done
))
1856 io_iopoll_complete(ctx
, nr_events
, &done
);
1862 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1863 * non-spinning poll check - we'll still enter the driver poll loop, but only
1864 * as a non-spinning completion check.
1866 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1869 while (!list_empty(&ctx
->poll_list
) && !need_resched()) {
1872 ret
= io_do_iopoll(ctx
, nr_events
, min
);
1875 if (!min
|| *nr_events
>= min
)
1883 * We can't just wait for polled events to come to us, we have to actively
1884 * find and complete them.
1886 static void io_iopoll_reap_events(struct io_ring_ctx
*ctx
)
1888 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1891 mutex_lock(&ctx
->uring_lock
);
1892 while (!list_empty(&ctx
->poll_list
)) {
1893 unsigned int nr_events
= 0;
1895 io_iopoll_getevents(ctx
, &nr_events
, 1);
1898 * Ensure we allow local-to-the-cpu processing to take place,
1899 * in this case we need to ensure that we reap all events.
1903 mutex_unlock(&ctx
->uring_lock
);
1906 static int io_iopoll_check(struct io_ring_ctx
*ctx
, unsigned *nr_events
,
1909 int iters
= 0, ret
= 0;
1912 * We disallow the app entering submit/complete with polling, but we
1913 * still need to lock the ring to prevent racing with polled issue
1914 * that got punted to a workqueue.
1916 mutex_lock(&ctx
->uring_lock
);
1921 * Don't enter poll loop if we already have events pending.
1922 * If we do, we can potentially be spinning for commands that
1923 * already triggered a CQE (eg in error).
1925 if (io_cqring_events(ctx
, false))
1929 * If a submit got punted to a workqueue, we can have the
1930 * application entering polling for a command before it gets
1931 * issued. That app will hold the uring_lock for the duration
1932 * of the poll right here, so we need to take a breather every
1933 * now and then to ensure that the issue has a chance to add
1934 * the poll to the issued list. Otherwise we can spin here
1935 * forever, while the workqueue is stuck trying to acquire the
1938 if (!(++iters
& 7)) {
1939 mutex_unlock(&ctx
->uring_lock
);
1940 mutex_lock(&ctx
->uring_lock
);
1943 if (*nr_events
< min
)
1944 tmin
= min
- *nr_events
;
1946 ret
= io_iopoll_getevents(ctx
, nr_events
, tmin
);
1950 } while (min
&& !*nr_events
&& !need_resched());
1952 mutex_unlock(&ctx
->uring_lock
);
1956 static void kiocb_end_write(struct io_kiocb
*req
)
1959 * Tell lockdep we inherited freeze protection from submission
1962 if (req
->flags
& REQ_F_ISREG
) {
1963 struct inode
*inode
= file_inode(req
->file
);
1965 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
1967 file_end_write(req
->file
);
1970 static inline void req_set_fail_links(struct io_kiocb
*req
)
1972 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1973 req
->flags
|= REQ_F_FAIL_LINK
;
1976 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
)
1978 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
1981 if (kiocb
->ki_flags
& IOCB_WRITE
)
1982 kiocb_end_write(req
);
1984 if (res
!= req
->result
)
1985 req_set_fail_links(req
);
1986 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
1987 cflags
= io_put_kbuf(req
);
1988 __io_cqring_add_event(req
, res
, cflags
);
1991 static void io_sq_thread_drop_mm(struct io_ring_ctx
*ctx
)
1993 struct mm_struct
*mm
= current
->mm
;
1996 kthread_unuse_mm(mm
);
2001 static int io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
,
2002 struct io_kiocb
*req
)
2004 if (io_op_defs
[req
->opcode
].needs_mm
&& !current
->mm
) {
2005 if (unlikely(!mmget_not_zero(ctx
->sqo_mm
)))
2007 kthread_use_mm(ctx
->sqo_mm
);
2014 static bool io_resubmit_prep(struct io_kiocb
*req
, int error
)
2016 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2017 ssize_t ret
= -ECANCELED
;
2018 struct iov_iter iter
;
2026 switch (req
->opcode
) {
2027 case IORING_OP_READV
:
2028 case IORING_OP_READ_FIXED
:
2029 case IORING_OP_READ
:
2032 case IORING_OP_WRITEV
:
2033 case IORING_OP_WRITE_FIXED
:
2034 case IORING_OP_WRITE
:
2038 printk_once(KERN_WARNING
"io_uring: bad opcode in resubmit %d\n",
2043 ret
= io_import_iovec(rw
, req
, &iovec
, &iter
, false);
2046 ret
= io_setup_async_rw(req
, ret
, iovec
, inline_vecs
, &iter
);
2051 io_cqring_add_event(req
, ret
);
2052 req_set_fail_links(req
);
2057 static void io_rw_resubmit(struct callback_head
*cb
)
2059 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
2060 struct io_ring_ctx
*ctx
= req
->ctx
;
2063 __set_current_state(TASK_RUNNING
);
2065 err
= io_sq_thread_acquire_mm(ctx
, req
);
2067 if (io_resubmit_prep(req
, err
)) {
2068 refcount_inc(&req
->refs
);
2069 io_queue_async_work(req
);
2074 static bool io_rw_reissue(struct io_kiocb
*req
, long res
)
2077 struct task_struct
*tsk
;
2080 if ((res
!= -EAGAIN
&& res
!= -EOPNOTSUPP
) || io_wq_current_is_worker())
2084 init_task_work(&req
->task_work
, io_rw_resubmit
);
2085 ret
= task_work_add(tsk
, &req
->task_work
, true);
2092 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2094 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2096 if (!io_rw_reissue(req
, res
)) {
2097 io_complete_rw_common(kiocb
, res
);
2102 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2104 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2106 if (kiocb
->ki_flags
& IOCB_WRITE
)
2107 kiocb_end_write(req
);
2109 if (res
!= -EAGAIN
&& res
!= req
->result
)
2110 req_set_fail_links(req
);
2112 WRITE_ONCE(req
->result
, res
);
2113 /* order with io_poll_complete() checking ->result */
2114 if (res
!= -EAGAIN
) {
2116 WRITE_ONCE(req
->iopoll_completed
, 1);
2121 * After the iocb has been issued, it's safe to be found on the poll list.
2122 * Adding the kiocb to the list AFTER submission ensures that we don't
2123 * find it from a io_iopoll_getevents() thread before the issuer is done
2124 * accessing the kiocb cookie.
2126 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2128 struct io_ring_ctx
*ctx
= req
->ctx
;
2131 * Track whether we have multiple files in our lists. This will impact
2132 * how we do polling eventually, not spinning if we're on potentially
2133 * different devices.
2135 if (list_empty(&ctx
->poll_list
)) {
2136 ctx
->poll_multi_file
= false;
2137 } else if (!ctx
->poll_multi_file
) {
2138 struct io_kiocb
*list_req
;
2140 list_req
= list_first_entry(&ctx
->poll_list
, struct io_kiocb
,
2142 if (list_req
->file
!= req
->file
)
2143 ctx
->poll_multi_file
= true;
2147 * For fast devices, IO may have already completed. If it has, add
2148 * it to the front so we find it first.
2150 if (READ_ONCE(req
->iopoll_completed
))
2151 list_add(&req
->list
, &ctx
->poll_list
);
2153 list_add_tail(&req
->list
, &ctx
->poll_list
);
2155 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2156 wq_has_sleeper(&ctx
->sqo_wait
))
2157 wake_up(&ctx
->sqo_wait
);
2160 static void __io_state_file_put(struct io_submit_state
*state
)
2162 int diff
= state
->has_refs
- state
->used_refs
;
2165 fput_many(state
->file
, diff
);
2169 static inline void io_state_file_put(struct io_submit_state
*state
)
2172 __io_state_file_put(state
);
2176 * Get as many references to a file as we have IOs left in this submission,
2177 * assuming most submissions are for one file, or at least that each file
2178 * has more than one submission.
2180 static struct file
*__io_file_get(struct io_submit_state
*state
, int fd
)
2186 if (state
->fd
== fd
) {
2191 __io_state_file_put(state
);
2193 state
->file
= fget_many(fd
, state
->ios_left
);
2198 state
->has_refs
= state
->ios_left
;
2199 state
->used_refs
= 1;
2204 static bool io_bdev_nowait(struct block_device
*bdev
)
2207 return !bdev
|| queue_is_mq(bdev_get_queue(bdev
));
2214 * If we tracked the file through the SCM inflight mechanism, we could support
2215 * any file. For now, just ensure that anything potentially problematic is done
2218 static bool io_file_supports_async(struct file
*file
, int rw
)
2220 umode_t mode
= file_inode(file
)->i_mode
;
2222 if (S_ISBLK(mode
)) {
2223 if (io_bdev_nowait(file
->f_inode
->i_bdev
))
2227 if (S_ISCHR(mode
) || S_ISSOCK(mode
))
2229 if (S_ISREG(mode
)) {
2230 if (io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2231 file
->f_op
!= &io_uring_fops
)
2236 /* any ->read/write should understand O_NONBLOCK */
2237 if (file
->f_flags
& O_NONBLOCK
)
2240 if (!(file
->f_mode
& FMODE_NOWAIT
))
2244 return file
->f_op
->read_iter
!= NULL
;
2246 return file
->f_op
->write_iter
!= NULL
;
2249 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2250 bool force_nonblock
)
2252 struct io_ring_ctx
*ctx
= req
->ctx
;
2253 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2257 if (S_ISREG(file_inode(req
->file
)->i_mode
))
2258 req
->flags
|= REQ_F_ISREG
;
2260 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2261 if (kiocb
->ki_pos
== -1 && !(req
->file
->f_mode
& FMODE_STREAM
)) {
2262 req
->flags
|= REQ_F_CUR_POS
;
2263 kiocb
->ki_pos
= req
->file
->f_pos
;
2265 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2266 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2267 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2271 ioprio
= READ_ONCE(sqe
->ioprio
);
2273 ret
= ioprio_check_cap(ioprio
);
2277 kiocb
->ki_ioprio
= ioprio
;
2279 kiocb
->ki_ioprio
= get_current_ioprio();
2281 /* don't allow async punt if RWF_NOWAIT was requested */
2282 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2283 req
->flags
|= REQ_F_NOWAIT
;
2285 if (kiocb
->ki_flags
& IOCB_DIRECT
)
2286 io_get_req_task(req
);
2289 kiocb
->ki_flags
|= IOCB_NOWAIT
;
2291 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2292 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2293 !kiocb
->ki_filp
->f_op
->iopoll
)
2296 kiocb
->ki_flags
|= IOCB_HIPRI
;
2297 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2299 req
->iopoll_completed
= 0;
2301 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2303 kiocb
->ki_complete
= io_complete_rw
;
2306 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2307 req
->rw
.len
= READ_ONCE(sqe
->len
);
2308 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2312 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2318 case -ERESTARTNOINTR
:
2319 case -ERESTARTNOHAND
:
2320 case -ERESTART_RESTARTBLOCK
:
2322 * We can't just restart the syscall, since previously
2323 * submitted sqes may already be in progress. Just fail this
2329 kiocb
->ki_complete(kiocb
, ret
, 0);
2333 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
)
2335 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2337 if (req
->flags
& REQ_F_CUR_POS
)
2338 req
->file
->f_pos
= kiocb
->ki_pos
;
2339 if (ret
>= 0 && kiocb
->ki_complete
== io_complete_rw
)
2340 io_complete_rw(kiocb
, ret
, 0);
2342 io_rw_done(kiocb
, ret
);
2345 static ssize_t
io_import_fixed(struct io_kiocb
*req
, int rw
,
2346 struct iov_iter
*iter
)
2348 struct io_ring_ctx
*ctx
= req
->ctx
;
2349 size_t len
= req
->rw
.len
;
2350 struct io_mapped_ubuf
*imu
;
2351 u16 index
, buf_index
;
2355 /* attempt to use fixed buffers without having provided iovecs */
2356 if (unlikely(!ctx
->user_bufs
))
2359 buf_index
= req
->buf_index
;
2360 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
2363 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
2364 imu
= &ctx
->user_bufs
[index
];
2365 buf_addr
= req
->rw
.addr
;
2368 if (buf_addr
+ len
< buf_addr
)
2370 /* not inside the mapped region */
2371 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
2375 * May not be a start of buffer, set size appropriately
2376 * and advance us to the beginning.
2378 offset
= buf_addr
- imu
->ubuf
;
2379 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
2383 * Don't use iov_iter_advance() here, as it's really slow for
2384 * using the latter parts of a big fixed buffer - it iterates
2385 * over each segment manually. We can cheat a bit here, because
2388 * 1) it's a BVEC iter, we set it up
2389 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2390 * first and last bvec
2392 * So just find our index, and adjust the iterator afterwards.
2393 * If the offset is within the first bvec (or the whole first
2394 * bvec, just use iov_iter_advance(). This makes it easier
2395 * since we can just skip the first segment, which may not
2396 * be PAGE_SIZE aligned.
2398 const struct bio_vec
*bvec
= imu
->bvec
;
2400 if (offset
<= bvec
->bv_len
) {
2401 iov_iter_advance(iter
, offset
);
2403 unsigned long seg_skip
;
2405 /* skip first vec */
2406 offset
-= bvec
->bv_len
;
2407 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
2409 iter
->bvec
= bvec
+ seg_skip
;
2410 iter
->nr_segs
-= seg_skip
;
2411 iter
->count
-= bvec
->bv_len
+ offset
;
2412 iter
->iov_offset
= offset
& ~PAGE_MASK
;
2419 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2422 mutex_unlock(&ctx
->uring_lock
);
2425 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
2428 * "Normal" inline submissions always hold the uring_lock, since we
2429 * grab it from the system call. Same is true for the SQPOLL offload.
2430 * The only exception is when we've detached the request and issue it
2431 * from an async worker thread, grab the lock for that case.
2434 mutex_lock(&ctx
->uring_lock
);
2437 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
2438 int bgid
, struct io_buffer
*kbuf
,
2441 struct io_buffer
*head
;
2443 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
2446 io_ring_submit_lock(req
->ctx
, needs_lock
);
2448 lockdep_assert_held(&req
->ctx
->uring_lock
);
2450 head
= idr_find(&req
->ctx
->io_buffer_idr
, bgid
);
2452 if (!list_empty(&head
->list
)) {
2453 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
2455 list_del(&kbuf
->list
);
2458 idr_remove(&req
->ctx
->io_buffer_idr
, bgid
);
2460 if (*len
> kbuf
->len
)
2463 kbuf
= ERR_PTR(-ENOBUFS
);
2466 io_ring_submit_unlock(req
->ctx
, needs_lock
);
2471 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
2474 struct io_buffer
*kbuf
;
2477 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2478 bgid
= req
->buf_index
;
2479 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
2482 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
2483 req
->flags
|= REQ_F_BUFFER_SELECTED
;
2484 return u64_to_user_ptr(kbuf
->addr
);
2487 #ifdef CONFIG_COMPAT
2488 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
2491 struct compat_iovec __user
*uiov
;
2492 compat_ssize_t clen
;
2496 uiov
= u64_to_user_ptr(req
->rw
.addr
);
2497 if (!access_ok(uiov
, sizeof(*uiov
)))
2499 if (__get_user(clen
, &uiov
->iov_len
))
2505 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2507 return PTR_ERR(buf
);
2508 iov
[0].iov_base
= buf
;
2509 iov
[0].iov_len
= (compat_size_t
) len
;
2514 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2517 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
2521 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
2524 len
= iov
[0].iov_len
;
2527 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
2529 return PTR_ERR(buf
);
2530 iov
[0].iov_base
= buf
;
2531 iov
[0].iov_len
= len
;
2535 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
2538 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
2539 struct io_buffer
*kbuf
;
2541 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2542 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
2543 iov
[0].iov_len
= kbuf
->len
;
2548 else if (req
->rw
.len
> 1)
2551 #ifdef CONFIG_COMPAT
2552 if (req
->ctx
->compat
)
2553 return io_compat_import(req
, iov
, needs_lock
);
2556 return __io_iov_buffer_select(req
, iov
, needs_lock
);
2559 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
2560 struct iovec
**iovec
, struct iov_iter
*iter
,
2563 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
2564 size_t sqe_len
= req
->rw
.len
;
2568 opcode
= req
->opcode
;
2569 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
2571 return io_import_fixed(req
, rw
, iter
);
2574 /* buffer index only valid with fixed read/write, or buffer select */
2575 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
2578 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
2579 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
2580 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
2583 return PTR_ERR(buf
);
2585 req
->rw
.len
= sqe_len
;
2588 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
2590 return ret
< 0 ? ret
: sqe_len
;
2594 struct io_async_rw
*iorw
= &req
->io
->rw
;
2597 iov_iter_init(iter
, rw
, *iovec
, iorw
->nr_segs
, iorw
->size
);
2598 if (iorw
->iov
== iorw
->fast_iov
)
2603 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
2604 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
2606 ret
= (*iovec
)->iov_len
;
2607 iov_iter_init(iter
, rw
, *iovec
, 1, ret
);
2613 #ifdef CONFIG_COMPAT
2614 if (req
->ctx
->compat
)
2615 return compat_import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
,
2619 return import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
);
2623 * For files that don't have ->read_iter() and ->write_iter(), handle them
2624 * by looping over ->read() or ->write() manually.
2626 static ssize_t
loop_rw_iter(int rw
, struct file
*file
, struct kiocb
*kiocb
,
2627 struct iov_iter
*iter
)
2632 * Don't support polled IO through this interface, and we can't
2633 * support non-blocking either. For the latter, this just causes
2634 * the kiocb to be handled from an async context.
2636 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2638 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
2641 while (iov_iter_count(iter
)) {
2645 if (!iov_iter_is_bvec(iter
)) {
2646 iovec
= iov_iter_iovec(iter
);
2648 /* fixed buffers import bvec */
2649 iovec
.iov_base
= kmap(iter
->bvec
->bv_page
)
2651 iovec
.iov_len
= min(iter
->count
,
2652 iter
->bvec
->bv_len
- iter
->iov_offset
);
2656 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
2657 iovec
.iov_len
, &kiocb
->ki_pos
);
2659 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
2660 iovec
.iov_len
, &kiocb
->ki_pos
);
2663 if (iov_iter_is_bvec(iter
))
2664 kunmap(iter
->bvec
->bv_page
);
2672 if (nr
!= iovec
.iov_len
)
2674 iov_iter_advance(iter
, nr
);
2680 static void io_req_map_rw(struct io_kiocb
*req
, ssize_t io_size
,
2681 struct iovec
*iovec
, struct iovec
*fast_iov
,
2682 struct iov_iter
*iter
)
2684 req
->io
->rw
.nr_segs
= iter
->nr_segs
;
2685 req
->io
->rw
.size
= io_size
;
2686 req
->io
->rw
.iov
= iovec
;
2687 if (!req
->io
->rw
.iov
) {
2688 req
->io
->rw
.iov
= req
->io
->rw
.fast_iov
;
2689 if (req
->io
->rw
.iov
!= fast_iov
)
2690 memcpy(req
->io
->rw
.iov
, fast_iov
,
2691 sizeof(struct iovec
) * iter
->nr_segs
);
2693 req
->flags
|= REQ_F_NEED_CLEANUP
;
2697 static inline int __io_alloc_async_ctx(struct io_kiocb
*req
)
2699 req
->io
= kmalloc(sizeof(*req
->io
), GFP_KERNEL
);
2700 return req
->io
== NULL
;
2703 static int io_alloc_async_ctx(struct io_kiocb
*req
)
2705 if (!io_op_defs
[req
->opcode
].async_ctx
)
2708 return __io_alloc_async_ctx(req
);
2711 static int io_setup_async_rw(struct io_kiocb
*req
, ssize_t io_size
,
2712 struct iovec
*iovec
, struct iovec
*fast_iov
,
2713 struct iov_iter
*iter
)
2715 if (!io_op_defs
[req
->opcode
].async_ctx
)
2718 if (__io_alloc_async_ctx(req
))
2721 io_req_map_rw(req
, io_size
, iovec
, fast_iov
, iter
);
2726 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2727 bool force_nonblock
)
2729 struct io_async_ctx
*io
;
2730 struct iov_iter iter
;
2733 ret
= io_prep_rw(req
, sqe
, force_nonblock
);
2737 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
2740 /* either don't need iovec imported or already have it */
2741 if (!req
->io
|| req
->flags
& REQ_F_NEED_CLEANUP
)
2745 io
->rw
.iov
= io
->rw
.fast_iov
;
2747 ret
= io_import_iovec(READ
, req
, &io
->rw
.iov
, &iter
, !force_nonblock
);
2752 io_req_map_rw(req
, ret
, io
->rw
.iov
, io
->rw
.fast_iov
, &iter
);
2756 static void __io_async_buf_error(struct io_kiocb
*req
, int error
)
2758 struct io_ring_ctx
*ctx
= req
->ctx
;
2760 spin_lock_irq(&ctx
->completion_lock
);
2761 io_cqring_fill_event(req
, error
);
2762 io_commit_cqring(ctx
);
2763 spin_unlock_irq(&ctx
->completion_lock
);
2765 io_cqring_ev_posted(ctx
);
2766 req_set_fail_links(req
);
2767 io_double_put_req(req
);
2770 static void io_async_buf_cancel(struct callback_head
*cb
)
2772 struct io_async_rw
*rw
;
2773 struct io_kiocb
*req
;
2775 rw
= container_of(cb
, struct io_async_rw
, task_work
);
2776 req
= rw
->wpq
.wait
.private;
2777 __io_async_buf_error(req
, -ECANCELED
);
2780 static void io_async_buf_retry(struct callback_head
*cb
)
2782 struct io_async_rw
*rw
;
2783 struct io_ring_ctx
*ctx
;
2784 struct io_kiocb
*req
;
2786 rw
= container_of(cb
, struct io_async_rw
, task_work
);
2787 req
= rw
->wpq
.wait
.private;
2790 __set_current_state(TASK_RUNNING
);
2791 if (!io_sq_thread_acquire_mm(ctx
, req
)) {
2792 mutex_lock(&ctx
->uring_lock
);
2793 __io_queue_sqe(req
, NULL
);
2794 mutex_unlock(&ctx
->uring_lock
);
2796 __io_async_buf_error(req
, -EFAULT
);
2800 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
2801 int sync
, void *arg
)
2803 struct wait_page_queue
*wpq
;
2804 struct io_kiocb
*req
= wait
->private;
2805 struct io_async_rw
*rw
= &req
->io
->rw
;
2806 struct wait_page_key
*key
= arg
;
2807 struct task_struct
*tsk
;
2810 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
2812 ret
= wake_page_match(wpq
, key
);
2816 list_del_init(&wait
->entry
);
2818 init_task_work(&rw
->task_work
, io_async_buf_retry
);
2819 /* submit ref gets dropped, acquire a new one */
2820 refcount_inc(&req
->refs
);
2822 ret
= task_work_add(tsk
, &rw
->task_work
, true);
2823 if (unlikely(ret
)) {
2824 /* queue just for cancelation */
2825 init_task_work(&rw
->task_work
, io_async_buf_cancel
);
2826 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
2827 task_work_add(tsk
, &rw
->task_work
, true);
2829 wake_up_process(tsk
);
2833 static bool io_rw_should_retry(struct io_kiocb
*req
)
2835 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2838 /* never retry for NOWAIT, we just complete with -EAGAIN */
2839 if (req
->flags
& REQ_F_NOWAIT
)
2842 /* already tried, or we're doing O_DIRECT */
2843 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_WAITQ
))
2846 * just use poll if we can, and don't attempt if the fs doesn't
2847 * support callback based unlocks
2849 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
2853 * If request type doesn't require req->io to defer in general,
2854 * we need to allocate it here
2856 if (!req
->io
&& __io_alloc_async_ctx(req
))
2859 ret
= kiocb_wait_page_queue_init(kiocb
, &req
->io
->rw
.wpq
,
2860 io_async_buf_func
, req
);
2862 io_get_req_task(req
);
2869 static int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
2871 if (req
->file
->f_op
->read_iter
)
2872 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
2873 return loop_rw_iter(READ
, req
->file
, &req
->rw
.kiocb
, iter
);
2876 static int io_read(struct io_kiocb
*req
, bool force_nonblock
)
2878 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2879 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2880 struct iov_iter iter
;
2882 ssize_t io_size
, ret
;
2884 ret
= io_import_iovec(READ
, req
, &iovec
, &iter
, !force_nonblock
);
2888 /* Ensure we clear previously set non-block flag */
2889 if (!force_nonblock
)
2890 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
2894 if (req
->flags
& REQ_F_LINK_HEAD
)
2895 req
->result
= io_size
;
2898 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2899 * we know to async punt it even if it was opened O_NONBLOCK
2901 if (force_nonblock
&& !io_file_supports_async(req
->file
, READ
))
2904 iov_count
= iov_iter_count(&iter
);
2905 ret
= rw_verify_area(READ
, req
->file
, &kiocb
->ki_pos
, iov_count
);
2907 unsigned long nr_segs
= iter
.nr_segs
;
2910 ret2
= io_iter_do_read(req
, &iter
);
2912 /* Catch -EAGAIN return for forced non-blocking submission */
2913 if (!force_nonblock
|| (ret2
!= -EAGAIN
&& ret2
!= -EIO
)) {
2914 kiocb_done(kiocb
, ret2
);
2916 iter
.count
= iov_count
;
2917 iter
.nr_segs
= nr_segs
;
2919 ret
= io_setup_async_rw(req
, io_size
, iovec
,
2920 inline_vecs
, &iter
);
2923 /* if we can retry, do so with the callbacks armed */
2924 if (io_rw_should_retry(req
)) {
2925 ret2
= io_iter_do_read(req
, &iter
);
2926 if (ret2
== -EIOCBQUEUED
) {
2928 } else if (ret2
!= -EAGAIN
) {
2929 kiocb_done(kiocb
, ret2
);
2933 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
2938 if (!(req
->flags
& REQ_F_NEED_CLEANUP
))
2943 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2944 bool force_nonblock
)
2946 struct io_async_ctx
*io
;
2947 struct iov_iter iter
;
2950 ret
= io_prep_rw(req
, sqe
, force_nonblock
);
2954 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
2957 req
->fsize
= rlimit(RLIMIT_FSIZE
);
2959 /* either don't need iovec imported or already have it */
2960 if (!req
->io
|| req
->flags
& REQ_F_NEED_CLEANUP
)
2964 io
->rw
.iov
= io
->rw
.fast_iov
;
2966 ret
= io_import_iovec(WRITE
, req
, &io
->rw
.iov
, &iter
, !force_nonblock
);
2971 io_req_map_rw(req
, ret
, io
->rw
.iov
, io
->rw
.fast_iov
, &iter
);
2975 static int io_write(struct io_kiocb
*req
, bool force_nonblock
)
2977 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2978 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2979 struct iov_iter iter
;
2981 ssize_t ret
, io_size
;
2983 ret
= io_import_iovec(WRITE
, req
, &iovec
, &iter
, !force_nonblock
);
2987 /* Ensure we clear previously set non-block flag */
2988 if (!force_nonblock
)
2989 req
->rw
.kiocb
.ki_flags
&= ~IOCB_NOWAIT
;
2993 if (req
->flags
& REQ_F_LINK_HEAD
)
2994 req
->result
= io_size
;
2997 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2998 * we know to async punt it even if it was opened O_NONBLOCK
3000 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
3003 /* file path doesn't support NOWAIT for non-direct_IO */
3004 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3005 (req
->flags
& REQ_F_ISREG
))
3008 iov_count
= iov_iter_count(&iter
);
3009 ret
= rw_verify_area(WRITE
, req
->file
, &kiocb
->ki_pos
, iov_count
);
3011 unsigned long nr_segs
= iter
.nr_segs
;
3015 * Open-code file_start_write here to grab freeze protection,
3016 * which will be released by another thread in
3017 * io_complete_rw(). Fool lockdep by telling it the lock got
3018 * released so that it doesn't complain about the held lock when
3019 * we return to userspace.
3021 if (req
->flags
& REQ_F_ISREG
) {
3022 __sb_start_write(file_inode(req
->file
)->i_sb
,
3023 SB_FREEZE_WRITE
, true);
3024 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3027 kiocb
->ki_flags
|= IOCB_WRITE
;
3029 if (!force_nonblock
)
3030 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= req
->fsize
;
3032 if (req
->file
->f_op
->write_iter
)
3033 ret2
= call_write_iter(req
->file
, kiocb
, &iter
);
3035 ret2
= loop_rw_iter(WRITE
, req
->file
, kiocb
, &iter
);
3037 if (!force_nonblock
)
3038 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= RLIM_INFINITY
;
3041 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3042 * retry them without IOCB_NOWAIT.
3044 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3046 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3047 kiocb_done(kiocb
, ret2
);
3049 iter
.count
= iov_count
;
3050 iter
.nr_segs
= nr_segs
;
3052 ret
= io_setup_async_rw(req
, io_size
, iovec
,
3053 inline_vecs
, &iter
);
3060 if (!(req
->flags
& REQ_F_NEED_CLEANUP
))
3065 static int __io_splice_prep(struct io_kiocb
*req
,
3066 const struct io_uring_sqe
*sqe
)
3068 struct io_splice
* sp
= &req
->splice
;
3069 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3072 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3074 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3078 sp
->len
= READ_ONCE(sqe
->len
);
3079 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3081 if (unlikely(sp
->flags
& ~valid_flags
))
3084 ret
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
), &sp
->file_in
,
3085 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3088 req
->flags
|= REQ_F_NEED_CLEANUP
;
3090 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3092 * Splice operation will be punted aync, and here need to
3093 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3095 io_req_init_async(req
);
3096 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3102 static int io_tee_prep(struct io_kiocb
*req
,
3103 const struct io_uring_sqe
*sqe
)
3105 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3107 return __io_splice_prep(req
, sqe
);
3110 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3112 struct io_splice
*sp
= &req
->splice
;
3113 struct file
*in
= sp
->file_in
;
3114 struct file
*out
= sp
->file_out
;
3115 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3121 ret
= do_tee(in
, out
, sp
->len
, flags
);
3123 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3124 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3126 io_cqring_add_event(req
, ret
);
3128 req_set_fail_links(req
);
3133 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3135 struct io_splice
* sp
= &req
->splice
;
3137 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3138 sp
->off_out
= READ_ONCE(sqe
->off
);
3139 return __io_splice_prep(req
, sqe
);
3142 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3144 struct io_splice
*sp
= &req
->splice
;
3145 struct file
*in
= sp
->file_in
;
3146 struct file
*out
= sp
->file_out
;
3147 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3148 loff_t
*poff_in
, *poff_out
;
3154 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3155 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3158 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3160 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3161 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3163 io_cqring_add_event(req
, ret
);
3165 req_set_fail_links(req
);
3171 * IORING_OP_NOP just posts a completion event, nothing else.
3173 static int io_nop(struct io_kiocb
*req
)
3175 struct io_ring_ctx
*ctx
= req
->ctx
;
3177 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3180 io_cqring_add_event(req
, 0);
3185 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3187 struct io_ring_ctx
*ctx
= req
->ctx
;
3192 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3194 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3197 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3198 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3201 req
->sync
.off
= READ_ONCE(sqe
->off
);
3202 req
->sync
.len
= READ_ONCE(sqe
->len
);
3206 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3208 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3211 /* fsync always requires a blocking context */
3215 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3216 end
> 0 ? end
: LLONG_MAX
,
3217 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3219 req_set_fail_links(req
);
3220 io_cqring_add_event(req
, ret
);
3225 static int io_fallocate_prep(struct io_kiocb
*req
,
3226 const struct io_uring_sqe
*sqe
)
3228 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3230 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3233 req
->sync
.off
= READ_ONCE(sqe
->off
);
3234 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3235 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3236 req
->fsize
= rlimit(RLIMIT_FSIZE
);
3240 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
3244 /* fallocate always requiring blocking context */
3248 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= req
->fsize
;
3249 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3251 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= RLIM_INFINITY
;
3253 req_set_fail_links(req
);
3254 io_cqring_add_event(req
, ret
);
3259 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3261 const char __user
*fname
;
3264 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3266 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3268 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3271 /* open.how should be already initialised */
3272 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
3273 req
->open
.how
.flags
|= O_LARGEFILE
;
3275 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3276 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3277 req
->open
.filename
= getname(fname
);
3278 if (IS_ERR(req
->open
.filename
)) {
3279 ret
= PTR_ERR(req
->open
.filename
);
3280 req
->open
.filename
= NULL
;
3283 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
3284 req
->flags
|= REQ_F_NEED_CLEANUP
;
3288 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3292 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3294 mode
= READ_ONCE(sqe
->len
);
3295 flags
= READ_ONCE(sqe
->open_flags
);
3296 req
->open
.how
= build_open_how(flags
, mode
);
3297 return __io_openat_prep(req
, sqe
);
3300 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3302 struct open_how __user
*how
;
3306 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3308 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3309 len
= READ_ONCE(sqe
->len
);
3310 if (len
< OPEN_HOW_SIZE_VER0
)
3313 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
3318 return __io_openat_prep(req
, sqe
);
3321 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
3323 struct open_flags op
;
3330 ret
= build_open_flags(&req
->open
.how
, &op
);
3334 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
3338 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
3341 ret
= PTR_ERR(file
);
3343 fsnotify_open(file
);
3344 fd_install(ret
, file
);
3347 putname(req
->open
.filename
);
3348 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3350 req_set_fail_links(req
);
3351 io_cqring_add_event(req
, ret
);
3356 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
3358 return io_openat2(req
, force_nonblock
);
3361 static int io_remove_buffers_prep(struct io_kiocb
*req
,
3362 const struct io_uring_sqe
*sqe
)
3364 struct io_provide_buf
*p
= &req
->pbuf
;
3367 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
3370 tmp
= READ_ONCE(sqe
->fd
);
3371 if (!tmp
|| tmp
> USHRT_MAX
)
3374 memset(p
, 0, sizeof(*p
));
3376 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3380 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
3381 int bgid
, unsigned nbufs
)
3385 /* shouldn't happen */
3389 /* the head kbuf is the list itself */
3390 while (!list_empty(&buf
->list
)) {
3391 struct io_buffer
*nxt
;
3393 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
3394 list_del(&nxt
->list
);
3401 idr_remove(&ctx
->io_buffer_idr
, bgid
);
3406 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
)
3408 struct io_provide_buf
*p
= &req
->pbuf
;
3409 struct io_ring_ctx
*ctx
= req
->ctx
;
3410 struct io_buffer
*head
;
3413 io_ring_submit_lock(ctx
, !force_nonblock
);
3415 lockdep_assert_held(&ctx
->uring_lock
);
3418 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3420 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
3422 io_ring_submit_lock(ctx
, !force_nonblock
);
3424 req_set_fail_links(req
);
3425 io_cqring_add_event(req
, ret
);
3430 static int io_provide_buffers_prep(struct io_kiocb
*req
,
3431 const struct io_uring_sqe
*sqe
)
3433 struct io_provide_buf
*p
= &req
->pbuf
;
3436 if (sqe
->ioprio
|| sqe
->rw_flags
)
3439 tmp
= READ_ONCE(sqe
->fd
);
3440 if (!tmp
|| tmp
> USHRT_MAX
)
3443 p
->addr
= READ_ONCE(sqe
->addr
);
3444 p
->len
= READ_ONCE(sqe
->len
);
3446 if (!access_ok(u64_to_user_ptr(p
->addr
), (p
->len
* p
->nbufs
)))
3449 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3450 tmp
= READ_ONCE(sqe
->off
);
3451 if (tmp
> USHRT_MAX
)
3457 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
3459 struct io_buffer
*buf
;
3460 u64 addr
= pbuf
->addr
;
3461 int i
, bid
= pbuf
->bid
;
3463 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
3464 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
3469 buf
->len
= pbuf
->len
;
3474 INIT_LIST_HEAD(&buf
->list
);
3477 list_add_tail(&buf
->list
, &(*head
)->list
);
3481 return i
? i
: -ENOMEM
;
3484 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
)
3486 struct io_provide_buf
*p
= &req
->pbuf
;
3487 struct io_ring_ctx
*ctx
= req
->ctx
;
3488 struct io_buffer
*head
, *list
;
3491 io_ring_submit_lock(ctx
, !force_nonblock
);
3493 lockdep_assert_held(&ctx
->uring_lock
);
3495 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3497 ret
= io_add_buffers(p
, &head
);
3502 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
3505 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
3510 io_ring_submit_unlock(ctx
, !force_nonblock
);
3512 req_set_fail_links(req
);
3513 io_cqring_add_event(req
, ret
);
3518 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
3519 const struct io_uring_sqe
*sqe
)
3521 #if defined(CONFIG_EPOLL)
3522 if (sqe
->ioprio
|| sqe
->buf_index
)
3524 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3527 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
3528 req
->epoll
.op
= READ_ONCE(sqe
->len
);
3529 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
3531 if (ep_op_has_event(req
->epoll
.op
)) {
3532 struct epoll_event __user
*ev
;
3534 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3535 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
3545 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
)
3547 #if defined(CONFIG_EPOLL)
3548 struct io_epoll
*ie
= &req
->epoll
;
3551 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
3552 if (force_nonblock
&& ret
== -EAGAIN
)
3556 req_set_fail_links(req
);
3557 io_cqring_add_event(req
, ret
);
3565 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3567 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3568 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
3570 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3573 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
3574 req
->madvise
.len
= READ_ONCE(sqe
->len
);
3575 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
3582 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
3584 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3585 struct io_madvise
*ma
= &req
->madvise
;
3591 ret
= do_madvise(ma
->addr
, ma
->len
, ma
->advice
);
3593 req_set_fail_links(req
);
3594 io_cqring_add_event(req
, ret
);
3602 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3604 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
3606 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3609 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
3610 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
3611 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
3615 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
3617 struct io_fadvise
*fa
= &req
->fadvise
;
3620 if (force_nonblock
) {
3621 switch (fa
->advice
) {
3622 case POSIX_FADV_NORMAL
:
3623 case POSIX_FADV_RANDOM
:
3624 case POSIX_FADV_SEQUENTIAL
:
3631 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
3633 req_set_fail_links(req
);
3634 io_cqring_add_event(req
, ret
);
3639 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3641 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3643 if (sqe
->ioprio
|| sqe
->buf_index
)
3645 if (req
->flags
& REQ_F_FIXED_FILE
)
3648 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
3649 req
->statx
.mask
= READ_ONCE(sqe
->len
);
3650 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3651 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3652 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
3657 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
3659 struct io_statx
*ctx
= &req
->statx
;
3662 if (force_nonblock
) {
3663 /* only need file table for an actual valid fd */
3664 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
3665 req
->flags
|= REQ_F_NO_FILE_TABLE
;
3669 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
3673 req_set_fail_links(req
);
3674 io_cqring_add_event(req
, ret
);
3679 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3682 * If we queue this for async, it must not be cancellable. That would
3683 * leave the 'file' in an undeterminate state, and here need to modify
3684 * io_wq_work.flags, so initialize io_wq_work firstly.
3686 io_req_init_async(req
);
3687 req
->work
.flags
|= IO_WQ_WORK_NO_CANCEL
;
3689 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3691 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
3692 sqe
->rw_flags
|| sqe
->buf_index
)
3694 if (req
->flags
& REQ_F_FIXED_FILE
)
3697 req
->close
.fd
= READ_ONCE(sqe
->fd
);
3698 if ((req
->file
&& req
->file
->f_op
== &io_uring_fops
) ||
3699 req
->close
.fd
== req
->ctx
->ring_fd
)
3702 req
->close
.put_file
= NULL
;
3706 static int io_close(struct io_kiocb
*req
, bool force_nonblock
)
3708 struct io_close
*close
= &req
->close
;
3711 /* might be already done during nonblock submission */
3712 if (!close
->put_file
) {
3713 ret
= __close_fd_get_file(close
->fd
, &close
->put_file
);
3715 return (ret
== -ENOENT
) ? -EBADF
: ret
;
3718 /* if the file has a flush method, be safe and punt to async */
3719 if (close
->put_file
->f_op
->flush
&& force_nonblock
) {
3720 /* avoid grabbing files - we don't need the files */
3721 req
->flags
|= REQ_F_NO_FILE_TABLE
| REQ_F_MUST_PUNT
;
3725 /* No ->flush() or already async, safely close from here */
3726 ret
= filp_close(close
->put_file
, req
->work
.files
);
3728 req_set_fail_links(req
);
3729 io_cqring_add_event(req
, ret
);
3730 fput(close
->put_file
);
3731 close
->put_file
= NULL
;
3736 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3738 struct io_ring_ctx
*ctx
= req
->ctx
;
3743 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3745 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3748 req
->sync
.off
= READ_ONCE(sqe
->off
);
3749 req
->sync
.len
= READ_ONCE(sqe
->len
);
3750 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
3754 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
3758 /* sync_file_range always requires a blocking context */
3762 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
3765 req_set_fail_links(req
);
3766 io_cqring_add_event(req
, ret
);
3771 #if defined(CONFIG_NET)
3772 static int io_setup_async_msg(struct io_kiocb
*req
,
3773 struct io_async_msghdr
*kmsg
)
3777 if (io_alloc_async_ctx(req
)) {
3778 if (kmsg
->iov
!= kmsg
->fast_iov
)
3782 req
->flags
|= REQ_F_NEED_CLEANUP
;
3783 memcpy(&req
->io
->msg
, kmsg
, sizeof(*kmsg
));
3787 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3789 struct io_sr_msg
*sr
= &req
->sr_msg
;
3790 struct io_async_ctx
*io
= req
->io
;
3793 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3796 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
3797 sr
->msg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3798 sr
->len
= READ_ONCE(sqe
->len
);
3800 #ifdef CONFIG_COMPAT
3801 if (req
->ctx
->compat
)
3802 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
3805 if (!io
|| req
->opcode
== IORING_OP_SEND
)
3807 /* iovec is already imported */
3808 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3811 io
->msg
.iov
= io
->msg
.fast_iov
;
3812 ret
= sendmsg_copy_msghdr(&io
->msg
.msg
, sr
->msg
, sr
->msg_flags
,
3815 req
->flags
|= REQ_F_NEED_CLEANUP
;
3819 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
)
3821 struct io_async_msghdr
*kmsg
= NULL
;
3822 struct socket
*sock
;
3825 sock
= sock_from_file(req
->file
, &ret
);
3827 struct io_async_ctx io
;
3831 kmsg
= &req
->io
->msg
;
3832 kmsg
->msg
.msg_name
= &req
->io
->msg
.addr
;
3833 /* if iov is set, it's allocated already */
3835 kmsg
->iov
= kmsg
->fast_iov
;
3836 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
3838 struct io_sr_msg
*sr
= &req
->sr_msg
;
3841 kmsg
->msg
.msg_name
= &io
.msg
.addr
;
3843 io
.msg
.iov
= io
.msg
.fast_iov
;
3844 ret
= sendmsg_copy_msghdr(&io
.msg
.msg
, sr
->msg
,
3845 sr
->msg_flags
, &io
.msg
.iov
);
3850 flags
= req
->sr_msg
.msg_flags
;
3851 if (flags
& MSG_DONTWAIT
)
3852 req
->flags
|= REQ_F_NOWAIT
;
3853 else if (force_nonblock
)
3854 flags
|= MSG_DONTWAIT
;
3856 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
3857 if (force_nonblock
&& ret
== -EAGAIN
)
3858 return io_setup_async_msg(req
, kmsg
);
3859 if (ret
== -ERESTARTSYS
)
3863 if (kmsg
&& kmsg
->iov
!= kmsg
->fast_iov
)
3865 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3866 io_cqring_add_event(req
, ret
);
3868 req_set_fail_links(req
);
3873 static int io_send(struct io_kiocb
*req
, bool force_nonblock
)
3875 struct socket
*sock
;
3878 sock
= sock_from_file(req
->file
, &ret
);
3880 struct io_sr_msg
*sr
= &req
->sr_msg
;
3885 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
,
3890 msg
.msg_name
= NULL
;
3891 msg
.msg_control
= NULL
;
3892 msg
.msg_controllen
= 0;
3893 msg
.msg_namelen
= 0;
3895 flags
= req
->sr_msg
.msg_flags
;
3896 if (flags
& MSG_DONTWAIT
)
3897 req
->flags
|= REQ_F_NOWAIT
;
3898 else if (force_nonblock
)
3899 flags
|= MSG_DONTWAIT
;
3901 msg
.msg_flags
= flags
;
3902 ret
= sock_sendmsg(sock
, &msg
);
3903 if (force_nonblock
&& ret
== -EAGAIN
)
3905 if (ret
== -ERESTARTSYS
)
3909 io_cqring_add_event(req
, ret
);
3911 req_set_fail_links(req
);
3916 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
, struct io_async_ctx
*io
)
3918 struct io_sr_msg
*sr
= &req
->sr_msg
;
3919 struct iovec __user
*uiov
;
3923 ret
= __copy_msghdr_from_user(&io
->msg
.msg
, sr
->msg
, &io
->msg
.uaddr
,
3928 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3931 if (copy_from_user(io
->msg
.iov
, uiov
, sizeof(*uiov
)))
3933 sr
->len
= io
->msg
.iov
[0].iov_len
;
3934 iov_iter_init(&io
->msg
.msg
.msg_iter
, READ
, io
->msg
.iov
, 1,
3938 ret
= import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
3939 &io
->msg
.iov
, &io
->msg
.msg
.msg_iter
);
3947 #ifdef CONFIG_COMPAT
3948 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
3949 struct io_async_ctx
*io
)
3951 struct compat_msghdr __user
*msg_compat
;
3952 struct io_sr_msg
*sr
= &req
->sr_msg
;
3953 struct compat_iovec __user
*uiov
;
3958 msg_compat
= (struct compat_msghdr __user
*) sr
->msg
;
3959 ret
= __get_compat_msghdr(&io
->msg
.msg
, msg_compat
, &io
->msg
.uaddr
,
3964 uiov
= compat_ptr(ptr
);
3965 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3966 compat_ssize_t clen
;
3970 if (!access_ok(uiov
, sizeof(*uiov
)))
3972 if (__get_user(clen
, &uiov
->iov_len
))
3976 sr
->len
= io
->msg
.iov
[0].iov_len
;
3979 ret
= compat_import_iovec(READ
, uiov
, len
, UIO_FASTIOV
,
3981 &io
->msg
.msg
.msg_iter
);
3990 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
, struct io_async_ctx
*io
)
3992 io
->msg
.iov
= io
->msg
.fast_iov
;
3994 #ifdef CONFIG_COMPAT
3995 if (req
->ctx
->compat
)
3996 return __io_compat_recvmsg_copy_hdr(req
, io
);
3999 return __io_recvmsg_copy_hdr(req
, io
);
4002 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4003 int *cflags
, bool needs_lock
)
4005 struct io_sr_msg
*sr
= &req
->sr_msg
;
4006 struct io_buffer
*kbuf
;
4008 if (!(req
->flags
& REQ_F_BUFFER_SELECT
))
4011 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4016 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4018 *cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
4019 *cflags
|= IORING_CQE_F_BUFFER
;
4023 static int io_recvmsg_prep(struct io_kiocb
*req
,
4024 const struct io_uring_sqe
*sqe
)
4026 struct io_sr_msg
*sr
= &req
->sr_msg
;
4027 struct io_async_ctx
*io
= req
->io
;
4030 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4033 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4034 sr
->msg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4035 sr
->len
= READ_ONCE(sqe
->len
);
4036 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4038 #ifdef CONFIG_COMPAT
4039 if (req
->ctx
->compat
)
4040 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4043 if (!io
|| req
->opcode
== IORING_OP_RECV
)
4045 /* iovec is already imported */
4046 if (req
->flags
& REQ_F_NEED_CLEANUP
)
4049 ret
= io_recvmsg_copy_hdr(req
, io
);
4051 req
->flags
|= REQ_F_NEED_CLEANUP
;
4055 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
)
4057 struct io_async_msghdr
*kmsg
= NULL
;
4058 struct socket
*sock
;
4059 int ret
, cflags
= 0;
4061 sock
= sock_from_file(req
->file
, &ret
);
4063 struct io_buffer
*kbuf
;
4064 struct io_async_ctx io
;
4068 kmsg
= &req
->io
->msg
;
4069 kmsg
->msg
.msg_name
= &req
->io
->msg
.addr
;
4070 /* if iov is set, it's allocated already */
4072 kmsg
->iov
= kmsg
->fast_iov
;
4073 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4076 kmsg
->msg
.msg_name
= &io
.msg
.addr
;
4078 ret
= io_recvmsg_copy_hdr(req
, &io
);
4083 kbuf
= io_recv_buffer_select(req
, &cflags
, !force_nonblock
);
4085 return PTR_ERR(kbuf
);
4087 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4088 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4089 1, req
->sr_msg
.len
);
4092 flags
= req
->sr_msg
.msg_flags
;
4093 if (flags
& MSG_DONTWAIT
)
4094 req
->flags
|= REQ_F_NOWAIT
;
4095 else if (force_nonblock
)
4096 flags
|= MSG_DONTWAIT
;
4098 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.msg
,
4099 kmsg
->uaddr
, flags
);
4100 if (force_nonblock
&& ret
== -EAGAIN
)
4101 return io_setup_async_msg(req
, kmsg
);
4102 if (ret
== -ERESTARTSYS
)
4106 if (kmsg
&& kmsg
->iov
!= kmsg
->fast_iov
)
4108 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4109 __io_cqring_add_event(req
, ret
, cflags
);
4111 req_set_fail_links(req
);
4116 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
)
4118 struct io_buffer
*kbuf
= NULL
;
4119 struct socket
*sock
;
4120 int ret
, cflags
= 0;
4122 sock
= sock_from_file(req
->file
, &ret
);
4124 struct io_sr_msg
*sr
= &req
->sr_msg
;
4125 void __user
*buf
= sr
->buf
;
4130 kbuf
= io_recv_buffer_select(req
, &cflags
, !force_nonblock
);
4132 return PTR_ERR(kbuf
);
4134 buf
= u64_to_user_ptr(kbuf
->addr
);
4136 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
,
4143 req
->flags
|= REQ_F_NEED_CLEANUP
;
4144 msg
.msg_name
= NULL
;
4145 msg
.msg_control
= NULL
;
4146 msg
.msg_controllen
= 0;
4147 msg
.msg_namelen
= 0;
4148 msg
.msg_iocb
= NULL
;
4151 flags
= req
->sr_msg
.msg_flags
;
4152 if (flags
& MSG_DONTWAIT
)
4153 req
->flags
|= REQ_F_NOWAIT
;
4154 else if (force_nonblock
)
4155 flags
|= MSG_DONTWAIT
;
4157 ret
= sock_recvmsg(sock
, &msg
, flags
);
4158 if (force_nonblock
&& ret
== -EAGAIN
)
4160 if (ret
== -ERESTARTSYS
)
4165 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4166 __io_cqring_add_event(req
, ret
, cflags
);
4168 req_set_fail_links(req
);
4173 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4175 struct io_accept
*accept
= &req
->accept
;
4177 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
4179 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4182 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4183 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4184 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4185 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4189 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
)
4191 struct io_accept
*accept
= &req
->accept
;
4192 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4195 if (req
->file
->f_flags
& O_NONBLOCK
)
4196 req
->flags
|= REQ_F_NOWAIT
;
4198 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4199 accept
->addr_len
, accept
->flags
,
4201 if (ret
== -EAGAIN
&& force_nonblock
)
4204 if (ret
== -ERESTARTSYS
)
4206 req_set_fail_links(req
);
4208 io_cqring_add_event(req
, ret
);
4213 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4215 struct io_connect
*conn
= &req
->connect
;
4216 struct io_async_ctx
*io
= req
->io
;
4218 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
4220 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4223 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4224 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4229 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
4230 &io
->connect
.address
);
4233 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
)
4235 struct io_async_ctx __io
, *io
;
4236 unsigned file_flags
;
4242 ret
= move_addr_to_kernel(req
->connect
.addr
,
4243 req
->connect
.addr_len
,
4244 &__io
.connect
.address
);
4250 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4252 ret
= __sys_connect_file(req
->file
, &io
->connect
.address
,
4253 req
->connect
.addr_len
, file_flags
);
4254 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4257 if (io_alloc_async_ctx(req
)) {
4261 memcpy(&req
->io
->connect
, &__io
.connect
, sizeof(__io
.connect
));
4264 if (ret
== -ERESTARTSYS
)
4268 req_set_fail_links(req
);
4269 io_cqring_add_event(req
, ret
);
4273 #else /* !CONFIG_NET */
4274 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4279 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
)
4284 static int io_send(struct io_kiocb
*req
, bool force_nonblock
)
4289 static int io_recvmsg_prep(struct io_kiocb
*req
,
4290 const struct io_uring_sqe
*sqe
)
4295 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
)
4300 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
)
4305 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4310 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
)
4315 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4320 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
)
4324 #endif /* CONFIG_NET */
4326 struct io_poll_table
{
4327 struct poll_table_struct pt
;
4328 struct io_kiocb
*req
;
4332 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
4333 __poll_t mask
, task_work_func_t func
)
4335 struct task_struct
*tsk
;
4338 /* for instances that support it check for an event match first: */
4339 if (mask
&& !(mask
& poll
->events
))
4342 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
4344 list_del_init(&poll
->wait
.entry
);
4348 init_task_work(&req
->task_work
, func
);
4350 * If this fails, then the task is exiting. When a task exits, the
4351 * work gets canceled, so just cancel this request as well instead
4352 * of executing it. We can't safely execute it anyway, as we may not
4353 * have the needed state needed for it anyway.
4355 ret
= task_work_add(tsk
, &req
->task_work
, true);
4356 if (unlikely(ret
)) {
4357 WRITE_ONCE(poll
->canceled
, true);
4358 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
4359 task_work_add(tsk
, &req
->task_work
, true);
4361 wake_up_process(tsk
);
4365 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
4366 __acquires(&req
->ctx
->completion_lock
)
4368 struct io_ring_ctx
*ctx
= req
->ctx
;
4370 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4371 struct poll_table_struct pt
= { ._key
= poll
->events
};
4373 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
4376 spin_lock_irq(&ctx
->completion_lock
);
4377 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4378 add_wait_queue(poll
->head
, &poll
->wait
);
4385 static void io_poll_remove_double(struct io_kiocb
*req
)
4387 struct io_poll_iocb
*poll
= (struct io_poll_iocb
*) req
->io
;
4389 lockdep_assert_held(&req
->ctx
->completion_lock
);
4391 if (poll
&& poll
->head
) {
4392 struct wait_queue_head
*head
= poll
->head
;
4394 spin_lock(&head
->lock
);
4395 list_del_init(&poll
->wait
.entry
);
4396 if (poll
->wait
.private)
4397 refcount_dec(&req
->refs
);
4399 spin_unlock(&head
->lock
);
4403 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
4405 struct io_ring_ctx
*ctx
= req
->ctx
;
4407 io_poll_remove_double(req
);
4408 req
->poll
.done
= true;
4409 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
4410 io_commit_cqring(ctx
);
4413 static void io_poll_task_handler(struct io_kiocb
*req
, struct io_kiocb
**nxt
)
4415 struct io_ring_ctx
*ctx
= req
->ctx
;
4417 if (io_poll_rewait(req
, &req
->poll
)) {
4418 spin_unlock_irq(&ctx
->completion_lock
);
4422 hash_del(&req
->hash_node
);
4423 io_poll_complete(req
, req
->result
, 0);
4424 req
->flags
|= REQ_F_COMP_LOCKED
;
4425 io_put_req_find_next(req
, nxt
);
4426 spin_unlock_irq(&ctx
->completion_lock
);
4428 io_cqring_ev_posted(ctx
);
4431 static void io_poll_task_func(struct callback_head
*cb
)
4433 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4434 struct io_kiocb
*nxt
= NULL
;
4436 io_poll_task_handler(req
, &nxt
);
4438 struct io_ring_ctx
*ctx
= nxt
->ctx
;
4440 mutex_lock(&ctx
->uring_lock
);
4441 __io_queue_sqe(nxt
, NULL
);
4442 mutex_unlock(&ctx
->uring_lock
);
4446 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
4447 int sync
, void *key
)
4449 struct io_kiocb
*req
= wait
->private;
4450 struct io_poll_iocb
*poll
= (struct io_poll_iocb
*) req
->io
;
4451 __poll_t mask
= key_to_poll(key
);
4453 /* for instances that support it check for an event match first: */
4454 if (mask
&& !(mask
& poll
->events
))
4457 if (req
->poll
.head
) {
4460 spin_lock(&req
->poll
.head
->lock
);
4461 done
= list_empty(&req
->poll
.wait
.entry
);
4463 list_del_init(&req
->poll
.wait
.entry
);
4464 spin_unlock(&req
->poll
.head
->lock
);
4466 __io_async_wake(req
, poll
, mask
, io_poll_task_func
);
4468 refcount_dec(&req
->refs
);
4472 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
4473 wait_queue_func_t wake_func
)
4477 poll
->canceled
= false;
4478 poll
->events
= events
;
4479 INIT_LIST_HEAD(&poll
->wait
.entry
);
4480 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
4483 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
4484 struct wait_queue_head
*head
)
4486 struct io_kiocb
*req
= pt
->req
;
4489 * If poll->head is already set, it's because the file being polled
4490 * uses multiple waitqueues for poll handling (eg one for read, one
4491 * for write). Setup a separate io_poll_iocb if this happens.
4493 if (unlikely(poll
->head
)) {
4494 /* already have a 2nd entry, fail a third attempt */
4496 pt
->error
= -EINVAL
;
4499 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
4501 pt
->error
= -ENOMEM
;
4504 io_init_poll_iocb(poll
, req
->poll
.events
, io_poll_double_wake
);
4505 refcount_inc(&req
->refs
);
4506 poll
->wait
.private = req
;
4507 req
->io
= (void *) poll
;
4513 if (poll
->events
& EPOLLEXCLUSIVE
)
4514 add_wait_queue_exclusive(head
, &poll
->wait
);
4516 add_wait_queue(head
, &poll
->wait
);
4519 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
4520 struct poll_table_struct
*p
)
4522 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
4524 __io_queue_proc(&pt
->req
->apoll
->poll
, pt
, head
);
4527 static void io_async_task_func(struct callback_head
*cb
)
4529 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4530 struct async_poll
*apoll
= req
->apoll
;
4531 struct io_ring_ctx
*ctx
= req
->ctx
;
4532 bool canceled
= false;
4534 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
4536 if (io_poll_rewait(req
, &apoll
->poll
)) {
4537 spin_unlock_irq(&ctx
->completion_lock
);
4541 /* If req is still hashed, it cannot have been canceled. Don't check. */
4542 if (hash_hashed(&req
->hash_node
)) {
4543 hash_del(&req
->hash_node
);
4545 canceled
= READ_ONCE(apoll
->poll
.canceled
);
4547 io_cqring_fill_event(req
, -ECANCELED
);
4548 io_commit_cqring(ctx
);
4552 spin_unlock_irq(&ctx
->completion_lock
);
4554 /* restore ->work in case we need to retry again */
4555 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4556 memcpy(&req
->work
, &apoll
->work
, sizeof(req
->work
));
4560 __set_current_state(TASK_RUNNING
);
4561 if (io_sq_thread_acquire_mm(ctx
, req
)) {
4562 io_cqring_add_event(req
, -EFAULT
);
4565 mutex_lock(&ctx
->uring_lock
);
4566 __io_queue_sqe(req
, NULL
);
4567 mutex_unlock(&ctx
->uring_lock
);
4569 io_cqring_ev_posted(ctx
);
4571 req_set_fail_links(req
);
4572 io_double_put_req(req
);
4576 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
4579 struct io_kiocb
*req
= wait
->private;
4580 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
4582 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
4585 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
4588 static void io_poll_req_insert(struct io_kiocb
*req
)
4590 struct io_ring_ctx
*ctx
= req
->ctx
;
4591 struct hlist_head
*list
;
4593 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
4594 hlist_add_head(&req
->hash_node
, list
);
4597 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
4598 struct io_poll_iocb
*poll
,
4599 struct io_poll_table
*ipt
, __poll_t mask
,
4600 wait_queue_func_t wake_func
)
4601 __acquires(&ctx
->completion_lock
)
4603 struct io_ring_ctx
*ctx
= req
->ctx
;
4604 bool cancel
= false;
4606 poll
->file
= req
->file
;
4607 io_init_poll_iocb(poll
, mask
, wake_func
);
4608 poll
->wait
.private = req
;
4610 ipt
->pt
._key
= mask
;
4612 ipt
->error
= -EINVAL
;
4614 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
4616 spin_lock_irq(&ctx
->completion_lock
);
4617 if (likely(poll
->head
)) {
4618 spin_lock(&poll
->head
->lock
);
4619 if (unlikely(list_empty(&poll
->wait
.entry
))) {
4625 if (mask
|| ipt
->error
)
4626 list_del_init(&poll
->wait
.entry
);
4628 WRITE_ONCE(poll
->canceled
, true);
4629 else if (!poll
->done
) /* actually waiting for an event */
4630 io_poll_req_insert(req
);
4631 spin_unlock(&poll
->head
->lock
);
4637 static bool io_arm_poll_handler(struct io_kiocb
*req
)
4639 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
4640 struct io_ring_ctx
*ctx
= req
->ctx
;
4641 struct async_poll
*apoll
;
4642 struct io_poll_table ipt
;
4646 if (!req
->file
|| !file_can_poll(req
->file
))
4648 if (req
->flags
& (REQ_F_MUST_PUNT
| REQ_F_POLLED
))
4650 if (!def
->pollin
&& !def
->pollout
)
4653 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
4654 if (unlikely(!apoll
))
4657 req
->flags
|= REQ_F_POLLED
;
4658 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4659 memcpy(&apoll
->work
, &req
->work
, sizeof(req
->work
));
4660 had_io
= req
->io
!= NULL
;
4662 io_get_req_task(req
);
4664 INIT_HLIST_NODE(&req
->hash_node
);
4668 mask
|= POLLIN
| POLLRDNORM
;
4670 mask
|= POLLOUT
| POLLWRNORM
;
4671 mask
|= POLLERR
| POLLPRI
;
4673 ipt
.pt
._qproc
= io_async_queue_proc
;
4675 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
4679 /* only remove double add if we did it here */
4681 io_poll_remove_double(req
);
4682 spin_unlock_irq(&ctx
->completion_lock
);
4683 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4684 memcpy(&req
->work
, &apoll
->work
, sizeof(req
->work
));
4688 spin_unlock_irq(&ctx
->completion_lock
);
4689 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
4690 apoll
->poll
.events
);
4694 static bool __io_poll_remove_one(struct io_kiocb
*req
,
4695 struct io_poll_iocb
*poll
)
4697 bool do_complete
= false;
4699 spin_lock(&poll
->head
->lock
);
4700 WRITE_ONCE(poll
->canceled
, true);
4701 if (!list_empty(&poll
->wait
.entry
)) {
4702 list_del_init(&poll
->wait
.entry
);
4705 spin_unlock(&poll
->head
->lock
);
4706 hash_del(&req
->hash_node
);
4710 static bool io_poll_remove_one(struct io_kiocb
*req
)
4714 if (req
->opcode
== IORING_OP_POLL_ADD
) {
4715 io_poll_remove_double(req
);
4716 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
4718 struct async_poll
*apoll
= req
->apoll
;
4720 /* non-poll requests have submit ref still */
4721 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
4725 * restore ->work because we will call
4726 * io_req_work_drop_env below when dropping the
4729 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4730 memcpy(&req
->work
, &apoll
->work
,
4737 io_cqring_fill_event(req
, -ECANCELED
);
4738 io_commit_cqring(req
->ctx
);
4739 req
->flags
|= REQ_F_COMP_LOCKED
;
4746 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
4748 struct hlist_node
*tmp
;
4749 struct io_kiocb
*req
;
4752 spin_lock_irq(&ctx
->completion_lock
);
4753 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
4754 struct hlist_head
*list
;
4756 list
= &ctx
->cancel_hash
[i
];
4757 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
)
4758 posted
+= io_poll_remove_one(req
);
4760 spin_unlock_irq(&ctx
->completion_lock
);
4763 io_cqring_ev_posted(ctx
);
4766 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
4768 struct hlist_head
*list
;
4769 struct io_kiocb
*req
;
4771 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
4772 hlist_for_each_entry(req
, list
, hash_node
) {
4773 if (sqe_addr
!= req
->user_data
)
4775 if (io_poll_remove_one(req
))
4783 static int io_poll_remove_prep(struct io_kiocb
*req
,
4784 const struct io_uring_sqe
*sqe
)
4786 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4788 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
4792 req
->poll
.addr
= READ_ONCE(sqe
->addr
);
4797 * Find a running poll command that matches one specified in sqe->addr,
4798 * and remove it if found.
4800 static int io_poll_remove(struct io_kiocb
*req
)
4802 struct io_ring_ctx
*ctx
= req
->ctx
;
4806 addr
= req
->poll
.addr
;
4807 spin_lock_irq(&ctx
->completion_lock
);
4808 ret
= io_poll_cancel(ctx
, addr
);
4809 spin_unlock_irq(&ctx
->completion_lock
);
4811 io_cqring_add_event(req
, ret
);
4813 req_set_fail_links(req
);
4818 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
4821 struct io_kiocb
*req
= wait
->private;
4822 struct io_poll_iocb
*poll
= &req
->poll
;
4824 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
4827 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
4828 struct poll_table_struct
*p
)
4830 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
4832 __io_queue_proc(&pt
->req
->poll
, pt
, head
);
4835 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4837 struct io_poll_iocb
*poll
= &req
->poll
;
4840 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4842 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
4847 events
= READ_ONCE(sqe
->poll32_events
);
4849 events
= swahw32(events
);
4851 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
4852 (events
& EPOLLEXCLUSIVE
);
4854 io_get_req_task(req
);
4858 static int io_poll_add(struct io_kiocb
*req
)
4860 struct io_poll_iocb
*poll
= &req
->poll
;
4861 struct io_ring_ctx
*ctx
= req
->ctx
;
4862 struct io_poll_table ipt
;
4865 INIT_HLIST_NODE(&req
->hash_node
);
4866 INIT_LIST_HEAD(&req
->list
);
4867 ipt
.pt
._qproc
= io_poll_queue_proc
;
4869 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
4872 if (mask
) { /* no async, we'd stolen it */
4874 io_poll_complete(req
, mask
, 0);
4876 spin_unlock_irq(&ctx
->completion_lock
);
4879 io_cqring_ev_posted(ctx
);
4885 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
4887 struct io_timeout_data
*data
= container_of(timer
,
4888 struct io_timeout_data
, timer
);
4889 struct io_kiocb
*req
= data
->req
;
4890 struct io_ring_ctx
*ctx
= req
->ctx
;
4891 unsigned long flags
;
4893 atomic_inc(&ctx
->cq_timeouts
);
4895 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
4897 * We could be racing with timeout deletion. If the list is empty,
4898 * then timeout lookup already found it and will be handling it.
4900 if (!list_empty(&req
->list
))
4901 list_del_init(&req
->list
);
4903 io_cqring_fill_event(req
, -ETIME
);
4904 io_commit_cqring(ctx
);
4905 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
4907 io_cqring_ev_posted(ctx
);
4908 req_set_fail_links(req
);
4910 return HRTIMER_NORESTART
;
4913 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
4915 struct io_kiocb
*req
;
4918 list_for_each_entry(req
, &ctx
->timeout_list
, list
) {
4919 if (user_data
== req
->user_data
) {
4920 list_del_init(&req
->list
);
4929 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
4933 req_set_fail_links(req
);
4934 io_cqring_fill_event(req
, -ECANCELED
);
4939 static int io_timeout_remove_prep(struct io_kiocb
*req
,
4940 const struct io_uring_sqe
*sqe
)
4942 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4944 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
4947 req
->timeout
.addr
= READ_ONCE(sqe
->addr
);
4948 req
->timeout
.flags
= READ_ONCE(sqe
->timeout_flags
);
4949 if (req
->timeout
.flags
)
4956 * Remove or update an existing timeout command
4958 static int io_timeout_remove(struct io_kiocb
*req
)
4960 struct io_ring_ctx
*ctx
= req
->ctx
;
4963 spin_lock_irq(&ctx
->completion_lock
);
4964 ret
= io_timeout_cancel(ctx
, req
->timeout
.addr
);
4966 io_cqring_fill_event(req
, ret
);
4967 io_commit_cqring(ctx
);
4968 spin_unlock_irq(&ctx
->completion_lock
);
4969 io_cqring_ev_posted(ctx
);
4971 req_set_fail_links(req
);
4976 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
4977 bool is_timeout_link
)
4979 struct io_timeout_data
*data
;
4981 u32 off
= READ_ONCE(sqe
->off
);
4983 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4985 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
4987 if (off
&& is_timeout_link
)
4989 flags
= READ_ONCE(sqe
->timeout_flags
);
4990 if (flags
& ~IORING_TIMEOUT_ABS
)
4993 req
->timeout
.off
= off
;
4995 if (!req
->io
&& io_alloc_async_ctx(req
))
4998 data
= &req
->io
->timeout
;
5000 req
->flags
|= REQ_F_TIMEOUT
;
5002 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5005 if (flags
& IORING_TIMEOUT_ABS
)
5006 data
->mode
= HRTIMER_MODE_ABS
;
5008 data
->mode
= HRTIMER_MODE_REL
;
5010 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5014 static int io_timeout(struct io_kiocb
*req
)
5016 struct io_ring_ctx
*ctx
= req
->ctx
;
5017 struct io_timeout_data
*data
= &req
->io
->timeout
;
5018 struct list_head
*entry
;
5019 u32 tail
, off
= req
->timeout
.off
;
5021 spin_lock_irq(&ctx
->completion_lock
);
5024 * sqe->off holds how many events that need to occur for this
5025 * timeout event to be satisfied. If it isn't set, then this is
5026 * a pure timeout request, sequence isn't used.
5029 req
->flags
|= REQ_F_TIMEOUT_NOSEQ
;
5030 entry
= ctx
->timeout_list
.prev
;
5034 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5035 req
->timeout
.target_seq
= tail
+ off
;
5038 * Insertion sort, ensuring the first entry in the list is always
5039 * the one we need first.
5041 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5042 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
, list
);
5044 if (nxt
->flags
& REQ_F_TIMEOUT_NOSEQ
)
5046 /* nxt.seq is behind @tail, otherwise would've been completed */
5047 if (off
>= nxt
->timeout
.target_seq
- tail
)
5051 list_add(&req
->list
, entry
);
5052 data
->timer
.function
= io_timeout_fn
;
5053 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5054 spin_unlock_irq(&ctx
->completion_lock
);
5058 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5060 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5062 return req
->user_data
== (unsigned long) data
;
5065 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5067 enum io_wq_cancel cancel_ret
;
5070 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5071 switch (cancel_ret
) {
5072 case IO_WQ_CANCEL_OK
:
5075 case IO_WQ_CANCEL_RUNNING
:
5078 case IO_WQ_CANCEL_NOTFOUND
:
5086 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5087 struct io_kiocb
*req
, __u64 sqe_addr
,
5090 unsigned long flags
;
5093 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5094 if (ret
!= -ENOENT
) {
5095 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5099 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5100 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5103 ret
= io_poll_cancel(ctx
, sqe_addr
);
5107 io_cqring_fill_event(req
, ret
);
5108 io_commit_cqring(ctx
);
5109 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5110 io_cqring_ev_posted(ctx
);
5113 req_set_fail_links(req
);
5117 static int io_async_cancel_prep(struct io_kiocb
*req
,
5118 const struct io_uring_sqe
*sqe
)
5120 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5122 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
||
5126 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5130 static int io_async_cancel(struct io_kiocb
*req
)
5132 struct io_ring_ctx
*ctx
= req
->ctx
;
5134 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
5138 static int io_files_update_prep(struct io_kiocb
*req
,
5139 const struct io_uring_sqe
*sqe
)
5141 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->rw_flags
)
5144 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
5145 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
5146 if (!req
->files_update
.nr_args
)
5148 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
5152 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
)
5154 struct io_ring_ctx
*ctx
= req
->ctx
;
5155 struct io_uring_files_update up
;
5161 up
.offset
= req
->files_update
.offset
;
5162 up
.fds
= req
->files_update
.arg
;
5164 mutex_lock(&ctx
->uring_lock
);
5165 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
5166 mutex_unlock(&ctx
->uring_lock
);
5169 req_set_fail_links(req
);
5170 io_cqring_add_event(req
, ret
);
5175 static int io_req_defer_prep(struct io_kiocb
*req
,
5176 const struct io_uring_sqe
*sqe
)
5183 io_req_init_async(req
);
5185 if (io_op_defs
[req
->opcode
].file_table
) {
5186 ret
= io_grab_files(req
);
5191 io_req_work_grab_env(req
, &io_op_defs
[req
->opcode
]);
5193 switch (req
->opcode
) {
5196 case IORING_OP_READV
:
5197 case IORING_OP_READ_FIXED
:
5198 case IORING_OP_READ
:
5199 ret
= io_read_prep(req
, sqe
, true);
5201 case IORING_OP_WRITEV
:
5202 case IORING_OP_WRITE_FIXED
:
5203 case IORING_OP_WRITE
:
5204 ret
= io_write_prep(req
, sqe
, true);
5206 case IORING_OP_POLL_ADD
:
5207 ret
= io_poll_add_prep(req
, sqe
);
5209 case IORING_OP_POLL_REMOVE
:
5210 ret
= io_poll_remove_prep(req
, sqe
);
5212 case IORING_OP_FSYNC
:
5213 ret
= io_prep_fsync(req
, sqe
);
5215 case IORING_OP_SYNC_FILE_RANGE
:
5216 ret
= io_prep_sfr(req
, sqe
);
5218 case IORING_OP_SENDMSG
:
5219 case IORING_OP_SEND
:
5220 ret
= io_sendmsg_prep(req
, sqe
);
5222 case IORING_OP_RECVMSG
:
5223 case IORING_OP_RECV
:
5224 ret
= io_recvmsg_prep(req
, sqe
);
5226 case IORING_OP_CONNECT
:
5227 ret
= io_connect_prep(req
, sqe
);
5229 case IORING_OP_TIMEOUT
:
5230 ret
= io_timeout_prep(req
, sqe
, false);
5232 case IORING_OP_TIMEOUT_REMOVE
:
5233 ret
= io_timeout_remove_prep(req
, sqe
);
5235 case IORING_OP_ASYNC_CANCEL
:
5236 ret
= io_async_cancel_prep(req
, sqe
);
5238 case IORING_OP_LINK_TIMEOUT
:
5239 ret
= io_timeout_prep(req
, sqe
, true);
5241 case IORING_OP_ACCEPT
:
5242 ret
= io_accept_prep(req
, sqe
);
5244 case IORING_OP_FALLOCATE
:
5245 ret
= io_fallocate_prep(req
, sqe
);
5247 case IORING_OP_OPENAT
:
5248 ret
= io_openat_prep(req
, sqe
);
5250 case IORING_OP_CLOSE
:
5251 ret
= io_close_prep(req
, sqe
);
5253 case IORING_OP_FILES_UPDATE
:
5254 ret
= io_files_update_prep(req
, sqe
);
5256 case IORING_OP_STATX
:
5257 ret
= io_statx_prep(req
, sqe
);
5259 case IORING_OP_FADVISE
:
5260 ret
= io_fadvise_prep(req
, sqe
);
5262 case IORING_OP_MADVISE
:
5263 ret
= io_madvise_prep(req
, sqe
);
5265 case IORING_OP_OPENAT2
:
5266 ret
= io_openat2_prep(req
, sqe
);
5268 case IORING_OP_EPOLL_CTL
:
5269 ret
= io_epoll_ctl_prep(req
, sqe
);
5271 case IORING_OP_SPLICE
:
5272 ret
= io_splice_prep(req
, sqe
);
5274 case IORING_OP_PROVIDE_BUFFERS
:
5275 ret
= io_provide_buffers_prep(req
, sqe
);
5277 case IORING_OP_REMOVE_BUFFERS
:
5278 ret
= io_remove_buffers_prep(req
, sqe
);
5281 ret
= io_tee_prep(req
, sqe
);
5284 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5293 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5295 struct io_ring_ctx
*ctx
= req
->ctx
;
5298 /* Still need defer if there is pending req in defer list. */
5299 if (!req_need_defer(req
) && list_empty_careful(&ctx
->defer_list
))
5303 if (io_alloc_async_ctx(req
))
5305 ret
= io_req_defer_prep(req
, sqe
);
5310 spin_lock_irq(&ctx
->completion_lock
);
5311 if (!req_need_defer(req
) && list_empty(&ctx
->defer_list
)) {
5312 spin_unlock_irq(&ctx
->completion_lock
);
5316 trace_io_uring_defer(ctx
, req
, req
->user_data
);
5317 list_add_tail(&req
->list
, &ctx
->defer_list
);
5318 spin_unlock_irq(&ctx
->completion_lock
);
5319 return -EIOCBQUEUED
;
5322 static void io_cleanup_req(struct io_kiocb
*req
)
5324 struct io_async_ctx
*io
= req
->io
;
5326 switch (req
->opcode
) {
5327 case IORING_OP_READV
:
5328 case IORING_OP_READ_FIXED
:
5329 case IORING_OP_READ
:
5330 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5331 kfree((void *)(unsigned long)req
->rw
.addr
);
5333 case IORING_OP_WRITEV
:
5334 case IORING_OP_WRITE_FIXED
:
5335 case IORING_OP_WRITE
:
5336 if (io
->rw
.iov
!= io
->rw
.fast_iov
)
5339 case IORING_OP_RECVMSG
:
5340 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5341 kfree(req
->sr_msg
.kbuf
);
5343 case IORING_OP_SENDMSG
:
5344 if (io
->msg
.iov
!= io
->msg
.fast_iov
)
5347 case IORING_OP_RECV
:
5348 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5349 kfree(req
->sr_msg
.kbuf
);
5351 case IORING_OP_OPENAT
:
5352 case IORING_OP_OPENAT2
:
5354 case IORING_OP_SPLICE
:
5356 io_put_file(req
, req
->splice
.file_in
,
5357 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
5361 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5364 static int io_issue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5365 bool force_nonblock
)
5367 struct io_ring_ctx
*ctx
= req
->ctx
;
5370 switch (req
->opcode
) {
5374 case IORING_OP_READV
:
5375 case IORING_OP_READ_FIXED
:
5376 case IORING_OP_READ
:
5378 ret
= io_read_prep(req
, sqe
, force_nonblock
);
5382 ret
= io_read(req
, force_nonblock
);
5384 case IORING_OP_WRITEV
:
5385 case IORING_OP_WRITE_FIXED
:
5386 case IORING_OP_WRITE
:
5388 ret
= io_write_prep(req
, sqe
, force_nonblock
);
5392 ret
= io_write(req
, force_nonblock
);
5394 case IORING_OP_FSYNC
:
5396 ret
= io_prep_fsync(req
, sqe
);
5400 ret
= io_fsync(req
, force_nonblock
);
5402 case IORING_OP_POLL_ADD
:
5404 ret
= io_poll_add_prep(req
, sqe
);
5408 ret
= io_poll_add(req
);
5410 case IORING_OP_POLL_REMOVE
:
5412 ret
= io_poll_remove_prep(req
, sqe
);
5416 ret
= io_poll_remove(req
);
5418 case IORING_OP_SYNC_FILE_RANGE
:
5420 ret
= io_prep_sfr(req
, sqe
);
5424 ret
= io_sync_file_range(req
, force_nonblock
);
5426 case IORING_OP_SENDMSG
:
5427 case IORING_OP_SEND
:
5429 ret
= io_sendmsg_prep(req
, sqe
);
5433 if (req
->opcode
== IORING_OP_SENDMSG
)
5434 ret
= io_sendmsg(req
, force_nonblock
);
5436 ret
= io_send(req
, force_nonblock
);
5438 case IORING_OP_RECVMSG
:
5439 case IORING_OP_RECV
:
5441 ret
= io_recvmsg_prep(req
, sqe
);
5445 if (req
->opcode
== IORING_OP_RECVMSG
)
5446 ret
= io_recvmsg(req
, force_nonblock
);
5448 ret
= io_recv(req
, force_nonblock
);
5450 case IORING_OP_TIMEOUT
:
5452 ret
= io_timeout_prep(req
, sqe
, false);
5456 ret
= io_timeout(req
);
5458 case IORING_OP_TIMEOUT_REMOVE
:
5460 ret
= io_timeout_remove_prep(req
, sqe
);
5464 ret
= io_timeout_remove(req
);
5466 case IORING_OP_ACCEPT
:
5468 ret
= io_accept_prep(req
, sqe
);
5472 ret
= io_accept(req
, force_nonblock
);
5474 case IORING_OP_CONNECT
:
5476 ret
= io_connect_prep(req
, sqe
);
5480 ret
= io_connect(req
, force_nonblock
);
5482 case IORING_OP_ASYNC_CANCEL
:
5484 ret
= io_async_cancel_prep(req
, sqe
);
5488 ret
= io_async_cancel(req
);
5490 case IORING_OP_FALLOCATE
:
5492 ret
= io_fallocate_prep(req
, sqe
);
5496 ret
= io_fallocate(req
, force_nonblock
);
5498 case IORING_OP_OPENAT
:
5500 ret
= io_openat_prep(req
, sqe
);
5504 ret
= io_openat(req
, force_nonblock
);
5506 case IORING_OP_CLOSE
:
5508 ret
= io_close_prep(req
, sqe
);
5512 ret
= io_close(req
, force_nonblock
);
5514 case IORING_OP_FILES_UPDATE
:
5516 ret
= io_files_update_prep(req
, sqe
);
5520 ret
= io_files_update(req
, force_nonblock
);
5522 case IORING_OP_STATX
:
5524 ret
= io_statx_prep(req
, sqe
);
5528 ret
= io_statx(req
, force_nonblock
);
5530 case IORING_OP_FADVISE
:
5532 ret
= io_fadvise_prep(req
, sqe
);
5536 ret
= io_fadvise(req
, force_nonblock
);
5538 case IORING_OP_MADVISE
:
5540 ret
= io_madvise_prep(req
, sqe
);
5544 ret
= io_madvise(req
, force_nonblock
);
5546 case IORING_OP_OPENAT2
:
5548 ret
= io_openat2_prep(req
, sqe
);
5552 ret
= io_openat2(req
, force_nonblock
);
5554 case IORING_OP_EPOLL_CTL
:
5556 ret
= io_epoll_ctl_prep(req
, sqe
);
5560 ret
= io_epoll_ctl(req
, force_nonblock
);
5562 case IORING_OP_SPLICE
:
5564 ret
= io_splice_prep(req
, sqe
);
5568 ret
= io_splice(req
, force_nonblock
);
5570 case IORING_OP_PROVIDE_BUFFERS
:
5572 ret
= io_provide_buffers_prep(req
, sqe
);
5576 ret
= io_provide_buffers(req
, force_nonblock
);
5578 case IORING_OP_REMOVE_BUFFERS
:
5580 ret
= io_remove_buffers_prep(req
, sqe
);
5584 ret
= io_remove_buffers(req
, force_nonblock
);
5588 ret
= io_tee_prep(req
, sqe
);
5592 ret
= io_tee(req
, force_nonblock
);
5602 /* If the op doesn't have a file, we're not polling for it */
5603 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
5604 const bool in_async
= io_wq_current_is_worker();
5606 if (req
->result
== -EAGAIN
)
5609 /* workqueue context doesn't hold uring_lock, grab it now */
5611 mutex_lock(&ctx
->uring_lock
);
5613 io_iopoll_req_issued(req
);
5616 mutex_unlock(&ctx
->uring_lock
);
5622 static void io_arm_async_linked_timeout(struct io_kiocb
*req
)
5624 struct io_kiocb
*link
;
5626 /* link head's timeout is queued in io_queue_async_work() */
5627 if (!(req
->flags
& REQ_F_QUEUE_TIMEOUT
))
5630 link
= list_first_entry(&req
->link_list
, struct io_kiocb
, link_list
);
5631 io_queue_linked_timeout(link
);
5634 static void io_wq_submit_work(struct io_wq_work
**workptr
)
5636 struct io_wq_work
*work
= *workptr
;
5637 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5640 io_arm_async_linked_timeout(req
);
5642 /* if NO_CANCEL is set, we must still run the work */
5643 if ((work
->flags
& (IO_WQ_WORK_CANCEL
|IO_WQ_WORK_NO_CANCEL
)) ==
5644 IO_WQ_WORK_CANCEL
) {
5650 ret
= io_issue_sqe(req
, NULL
, false);
5652 * We can get EAGAIN for polled IO even though we're
5653 * forcing a sync submission from here, since we can't
5654 * wait for request slots on the block side.
5663 req_set_fail_links(req
);
5664 io_cqring_add_event(req
, ret
);
5668 io_steal_work(req
, workptr
);
5671 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
5674 struct fixed_file_table
*table
;
5676 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
5677 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
5680 static int io_file_get(struct io_submit_state
*state
, struct io_kiocb
*req
,
5681 int fd
, struct file
**out_file
, bool fixed
)
5683 struct io_ring_ctx
*ctx
= req
->ctx
;
5687 if (unlikely(!ctx
->file_data
||
5688 (unsigned) fd
>= ctx
->nr_user_files
))
5690 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
5691 file
= io_file_from_index(ctx
, fd
);
5693 req
->fixed_file_refs
= ctx
->file_data
->cur_refs
;
5694 percpu_ref_get(req
->fixed_file_refs
);
5697 trace_io_uring_file_get(ctx
, fd
);
5698 file
= __io_file_get(state
, fd
);
5701 if (file
|| io_op_defs
[req
->opcode
].needs_file_no_error
) {
5708 static int io_req_set_file(struct io_submit_state
*state
, struct io_kiocb
*req
,
5713 fixed
= (req
->flags
& REQ_F_FIXED_FILE
) != 0;
5714 if (unlikely(!fixed
&& io_async_submit(req
->ctx
)))
5717 return io_file_get(state
, req
, fd
, &req
->file
, fixed
);
5720 static int io_grab_files(struct io_kiocb
*req
)
5723 struct io_ring_ctx
*ctx
= req
->ctx
;
5725 if (req
->work
.files
|| (req
->flags
& REQ_F_NO_FILE_TABLE
))
5727 if (!ctx
->ring_file
)
5731 spin_lock_irq(&ctx
->inflight_lock
);
5733 * We use the f_ops->flush() handler to ensure that we can flush
5734 * out work accessing these files if the fd is closed. Check if
5735 * the fd has changed since we started down this path, and disallow
5736 * this operation if it has.
5738 if (fcheck(ctx
->ring_fd
) == ctx
->ring_file
) {
5739 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
5740 req
->flags
|= REQ_F_INFLIGHT
;
5741 req
->work
.files
= current
->files
;
5744 spin_unlock_irq(&ctx
->inflight_lock
);
5750 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
5752 struct io_timeout_data
*data
= container_of(timer
,
5753 struct io_timeout_data
, timer
);
5754 struct io_kiocb
*req
= data
->req
;
5755 struct io_ring_ctx
*ctx
= req
->ctx
;
5756 struct io_kiocb
*prev
= NULL
;
5757 unsigned long flags
;
5759 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5762 * We don't expect the list to be empty, that will only happen if we
5763 * race with the completion of the linked work.
5765 if (!list_empty(&req
->link_list
)) {
5766 prev
= list_entry(req
->link_list
.prev
, struct io_kiocb
,
5768 if (refcount_inc_not_zero(&prev
->refs
)) {
5769 list_del_init(&req
->link_list
);
5770 prev
->flags
&= ~REQ_F_LINK_TIMEOUT
;
5775 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5778 req_set_fail_links(prev
);
5779 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
5782 io_cqring_add_event(req
, -ETIME
);
5785 return HRTIMER_NORESTART
;
5788 static void io_queue_linked_timeout(struct io_kiocb
*req
)
5790 struct io_ring_ctx
*ctx
= req
->ctx
;
5793 * If the list is now empty, then our linked request finished before
5794 * we got a chance to setup the timer
5796 spin_lock_irq(&ctx
->completion_lock
);
5797 if (!list_empty(&req
->link_list
)) {
5798 struct io_timeout_data
*data
= &req
->io
->timeout
;
5800 data
->timer
.function
= io_link_timeout_fn
;
5801 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
5804 spin_unlock_irq(&ctx
->completion_lock
);
5806 /* drop submission reference */
5810 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
5812 struct io_kiocb
*nxt
;
5814 if (!(req
->flags
& REQ_F_LINK_HEAD
))
5816 /* for polled retry, if flag is set, we already went through here */
5817 if (req
->flags
& REQ_F_POLLED
)
5820 nxt
= list_first_entry_or_null(&req
->link_list
, struct io_kiocb
,
5822 if (!nxt
|| nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
5825 req
->flags
|= REQ_F_LINK_TIMEOUT
;
5829 static void __io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5831 struct io_kiocb
*linked_timeout
;
5832 struct io_kiocb
*nxt
;
5833 const struct cred
*old_creds
= NULL
;
5837 linked_timeout
= io_prep_linked_timeout(req
);
5839 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) && req
->work
.creds
&&
5840 req
->work
.creds
!= current_cred()) {
5842 revert_creds(old_creds
);
5843 if (old_creds
== req
->work
.creds
)
5844 old_creds
= NULL
; /* restored original creds */
5846 old_creds
= override_creds(req
->work
.creds
);
5849 ret
= io_issue_sqe(req
, sqe
, true);
5852 * We async punt it if the file wasn't marked NOWAIT, or if the file
5853 * doesn't support non-blocking read/write attempts
5855 if (ret
== -EAGAIN
&& (!(req
->flags
& REQ_F_NOWAIT
) ||
5856 (req
->flags
& REQ_F_MUST_PUNT
))) {
5857 if (io_arm_poll_handler(req
)) {
5859 io_queue_linked_timeout(linked_timeout
);
5863 io_req_init_async(req
);
5865 if (io_op_defs
[req
->opcode
].file_table
) {
5866 ret
= io_grab_files(req
);
5872 * Queued up for async execution, worker will release
5873 * submit reference when the iocb is actually submitted.
5875 io_queue_async_work(req
);
5881 /* drop submission reference */
5882 io_put_req_find_next(req
, &nxt
);
5884 if (linked_timeout
) {
5886 io_queue_linked_timeout(linked_timeout
);
5888 io_put_req(linked_timeout
);
5891 /* and drop final reference, if we failed */
5893 io_cqring_add_event(req
, ret
);
5894 req_set_fail_links(req
);
5900 if (req
->flags
& REQ_F_FORCE_ASYNC
)
5906 revert_creds(old_creds
);
5909 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5913 ret
= io_req_defer(req
, sqe
);
5915 if (ret
!= -EIOCBQUEUED
) {
5917 io_cqring_add_event(req
, ret
);
5918 req_set_fail_links(req
);
5919 io_double_put_req(req
);
5921 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
5924 if (io_alloc_async_ctx(req
))
5926 ret
= io_req_defer_prep(req
, sqe
);
5927 if (unlikely(ret
< 0))
5932 * Never try inline submit of IOSQE_ASYNC is set, go straight
5933 * to async execution.
5935 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
5936 io_queue_async_work(req
);
5938 __io_queue_sqe(req
, sqe
);
5942 static inline void io_queue_link_head(struct io_kiocb
*req
)
5944 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
5945 io_cqring_add_event(req
, -ECANCELED
);
5946 io_double_put_req(req
);
5948 io_queue_sqe(req
, NULL
);
5951 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5952 struct io_kiocb
**link
)
5954 struct io_ring_ctx
*ctx
= req
->ctx
;
5958 * If we already have a head request, queue this one for async
5959 * submittal once the head completes. If we don't have a head but
5960 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5961 * submitted sync once the chain is complete. If none of those
5962 * conditions are true (normal request), then just queue it.
5965 struct io_kiocb
*head
= *link
;
5968 * Taking sequential execution of a link, draining both sides
5969 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5970 * requests in the link. So, it drains the head and the
5971 * next after the link request. The last one is done via
5972 * drain_next flag to persist the effect across calls.
5974 if (req
->flags
& REQ_F_IO_DRAIN
) {
5975 head
->flags
|= REQ_F_IO_DRAIN
;
5976 ctx
->drain_next
= 1;
5978 if (io_alloc_async_ctx(req
))
5981 ret
= io_req_defer_prep(req
, sqe
);
5983 /* fail even hard links since we don't submit */
5984 head
->flags
|= REQ_F_FAIL_LINK
;
5987 trace_io_uring_link(ctx
, req
, head
);
5988 list_add_tail(&req
->link_list
, &head
->link_list
);
5990 /* last request of a link, enqueue the link */
5991 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
5992 io_queue_link_head(head
);
5996 if (unlikely(ctx
->drain_next
)) {
5997 req
->flags
|= REQ_F_IO_DRAIN
;
5998 ctx
->drain_next
= 0;
6000 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
6001 req
->flags
|= REQ_F_LINK_HEAD
;
6002 INIT_LIST_HEAD(&req
->link_list
);
6004 if (io_alloc_async_ctx(req
))
6007 ret
= io_req_defer_prep(req
, sqe
);
6009 req
->flags
|= REQ_F_FAIL_LINK
;
6012 io_queue_sqe(req
, sqe
);
6020 * Batched submission is done, ensure local IO is flushed out.
6022 static void io_submit_state_end(struct io_submit_state
*state
)
6024 blk_finish_plug(&state
->plug
);
6025 io_state_file_put(state
);
6026 if (state
->free_reqs
)
6027 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6031 * Start submission side cache.
6033 static void io_submit_state_start(struct io_submit_state
*state
,
6034 unsigned int max_ios
)
6036 blk_start_plug(&state
->plug
);
6038 state
->plug
.nowait
= true;
6040 state
->free_reqs
= 0;
6042 state
->ios_left
= max_ios
;
6045 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6047 struct io_rings
*rings
= ctx
->rings
;
6050 * Ensure any loads from the SQEs are done at this point,
6051 * since once we write the new head, the application could
6052 * write new data to them.
6054 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6058 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6059 * that is mapped by userspace. This means that care needs to be taken to
6060 * ensure that reads are stable, as we cannot rely on userspace always
6061 * being a good citizen. If members of the sqe are validated and then later
6062 * used, it's important that those reads are done through READ_ONCE() to
6063 * prevent a re-load down the line.
6065 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6067 u32
*sq_array
= ctx
->sq_array
;
6071 * The cached sq head (or cq tail) serves two purposes:
6073 * 1) allows us to batch the cost of updating the user visible
6075 * 2) allows the kernel side to track the head on its own, even
6076 * though the application is the one updating it.
6078 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6079 if (likely(head
< ctx
->sq_entries
))
6080 return &ctx
->sq_sqes
[head
];
6082 /* drop invalid entries */
6083 ctx
->cached_sq_dropped
++;
6084 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6088 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6090 ctx
->cached_sq_head
++;
6093 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6094 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6095 IOSQE_BUFFER_SELECT)
6097 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6098 const struct io_uring_sqe
*sqe
,
6099 struct io_submit_state
*state
)
6101 unsigned int sqe_flags
;
6105 * All io need record the previous position, if LINK vs DARIN,
6106 * it can be used to mark the position of the first IO in the
6109 req
->sequence
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
6110 req
->opcode
= READ_ONCE(sqe
->opcode
);
6111 req
->user_data
= READ_ONCE(sqe
->user_data
);
6116 /* one is dropped after submission, the other at completion */
6117 refcount_set(&req
->refs
, 2);
6118 req
->task
= current
;
6121 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6124 if (unlikely(io_sq_thread_acquire_mm(ctx
, req
)))
6127 sqe_flags
= READ_ONCE(sqe
->flags
);
6128 /* enforce forwards compatibility on users */
6129 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6132 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6133 !io_op_defs
[req
->opcode
].buffer_select
)
6136 id
= READ_ONCE(sqe
->personality
);
6138 io_req_init_async(req
);
6139 req
->work
.creds
= idr_find(&ctx
->personality_idr
, id
);
6140 if (unlikely(!req
->work
.creds
))
6142 get_cred(req
->work
.creds
);
6145 /* same numerical values with corresponding REQ_F_*, safe to copy */
6146 req
->flags
|= sqe_flags
;
6148 if (!io_op_defs
[req
->opcode
].needs_file
)
6151 return io_req_set_file(state
, req
, READ_ONCE(sqe
->fd
));
6154 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
,
6155 struct file
*ring_file
, int ring_fd
)
6157 struct io_submit_state state
;
6158 struct io_kiocb
*link
= NULL
;
6159 int i
, submitted
= 0;
6161 /* if we have a backlog and couldn't flush it all, return BUSY */
6162 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6163 if (!list_empty(&ctx
->cq_overflow_list
) &&
6164 !io_cqring_overflow_flush(ctx
, false))
6168 /* make sure SQ entry isn't read before tail */
6169 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6171 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6174 io_submit_state_start(&state
, nr
);
6176 ctx
->ring_fd
= ring_fd
;
6177 ctx
->ring_file
= ring_file
;
6179 for (i
= 0; i
< nr
; i
++) {
6180 const struct io_uring_sqe
*sqe
;
6181 struct io_kiocb
*req
;
6184 sqe
= io_get_sqe(ctx
);
6185 if (unlikely(!sqe
)) {
6186 io_consume_sqe(ctx
);
6189 req
= io_alloc_req(ctx
, &state
);
6190 if (unlikely(!req
)) {
6192 submitted
= -EAGAIN
;
6196 err
= io_init_req(ctx
, req
, sqe
, &state
);
6197 io_consume_sqe(ctx
);
6198 /* will complete beyond this point, count as submitted */
6201 if (unlikely(err
)) {
6203 io_cqring_add_event(req
, err
);
6204 io_double_put_req(req
);
6208 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6209 true, io_async_submit(ctx
));
6210 err
= io_submit_sqe(req
, sqe
, &link
);
6215 if (unlikely(submitted
!= nr
)) {
6216 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6218 percpu_ref_put_many(&ctx
->refs
, nr
- ref_used
);
6221 io_queue_link_head(link
);
6222 io_submit_state_end(&state
);
6224 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6225 io_commit_sqring(ctx
);
6230 static int io_sq_thread(void *data
)
6232 struct io_ring_ctx
*ctx
= data
;
6233 const struct cred
*old_cred
;
6235 unsigned long timeout
;
6238 complete(&ctx
->sq_thread_comp
);
6240 old_cred
= override_creds(ctx
->creds
);
6242 timeout
= jiffies
+ ctx
->sq_thread_idle
;
6243 while (!kthread_should_park()) {
6244 unsigned int to_submit
;
6246 if (!list_empty(&ctx
->poll_list
)) {
6247 unsigned nr_events
= 0;
6249 mutex_lock(&ctx
->uring_lock
);
6250 if (!list_empty(&ctx
->poll_list
))
6251 io_iopoll_getevents(ctx
, &nr_events
, 0);
6253 timeout
= jiffies
+ ctx
->sq_thread_idle
;
6254 mutex_unlock(&ctx
->uring_lock
);
6257 to_submit
= io_sqring_entries(ctx
);
6260 * If submit got -EBUSY, flag us as needing the application
6261 * to enter the kernel to reap and flush events.
6263 if (!to_submit
|| ret
== -EBUSY
) {
6265 * Drop cur_mm before scheduling, we can't hold it for
6266 * long periods (or over schedule()). Do this before
6267 * adding ourselves to the waitqueue, as the unuse/drop
6270 io_sq_thread_drop_mm(ctx
);
6273 * We're polling. If we're within the defined idle
6274 * period, then let us spin without work before going
6275 * to sleep. The exception is if we got EBUSY doing
6276 * more IO, we should wait for the application to
6277 * reap events and wake us up.
6279 if (!list_empty(&ctx
->poll_list
) ||
6280 (!time_after(jiffies
, timeout
) && ret
!= -EBUSY
&&
6281 !percpu_ref_is_dying(&ctx
->refs
))) {
6282 if (current
->task_works
)
6288 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
6289 TASK_INTERRUPTIBLE
);
6292 * While doing polled IO, before going to sleep, we need
6293 * to check if there are new reqs added to poll_list, it
6294 * is because reqs may have been punted to io worker and
6295 * will be added to poll_list later, hence check the
6298 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6299 !list_empty_careful(&ctx
->poll_list
)) {
6300 finish_wait(&ctx
->sqo_wait
, &wait
);
6304 /* Tell userspace we may need a wakeup call */
6305 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6306 /* make sure to read SQ tail after writing flags */
6309 to_submit
= io_sqring_entries(ctx
);
6310 if (!to_submit
|| ret
== -EBUSY
) {
6311 if (kthread_should_park()) {
6312 finish_wait(&ctx
->sqo_wait
, &wait
);
6315 if (current
->task_works
) {
6317 finish_wait(&ctx
->sqo_wait
, &wait
);
6320 if (signal_pending(current
))
6321 flush_signals(current
);
6323 finish_wait(&ctx
->sqo_wait
, &wait
);
6325 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6329 finish_wait(&ctx
->sqo_wait
, &wait
);
6331 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6334 mutex_lock(&ctx
->uring_lock
);
6335 if (likely(!percpu_ref_is_dying(&ctx
->refs
)))
6336 ret
= io_submit_sqes(ctx
, to_submit
, NULL
, -1);
6337 mutex_unlock(&ctx
->uring_lock
);
6338 timeout
= jiffies
+ ctx
->sq_thread_idle
;
6341 if (current
->task_works
)
6344 io_sq_thread_drop_mm(ctx
);
6345 revert_creds(old_cred
);
6352 struct io_wait_queue
{
6353 struct wait_queue_entry wq
;
6354 struct io_ring_ctx
*ctx
;
6356 unsigned nr_timeouts
;
6359 static inline bool io_should_wake(struct io_wait_queue
*iowq
, bool noflush
)
6361 struct io_ring_ctx
*ctx
= iowq
->ctx
;
6364 * Wake up if we have enough events, or if a timeout occurred since we
6365 * started waiting. For timeouts, we always want to return to userspace,
6366 * regardless of event count.
6368 return io_cqring_events(ctx
, noflush
) >= iowq
->to_wait
||
6369 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
6372 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
6373 int wake_flags
, void *key
)
6375 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
6378 /* use noflush == true, as we can't safely rely on locking context */
6379 if (!io_should_wake(iowq
, true))
6382 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
6386 * Wait until events become available, if we don't already have some. The
6387 * application must reap them itself, as they reside on the shared cq ring.
6389 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
6390 const sigset_t __user
*sig
, size_t sigsz
)
6392 struct io_wait_queue iowq
= {
6395 .func
= io_wake_function
,
6396 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
6399 .to_wait
= min_events
,
6401 struct io_rings
*rings
= ctx
->rings
;
6405 if (io_cqring_events(ctx
, false) >= min_events
)
6407 if (!current
->task_works
)
6413 #ifdef CONFIG_COMPAT
6414 if (in_compat_syscall())
6415 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
6419 ret
= set_user_sigmask(sig
, sigsz
);
6425 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
6426 trace_io_uring_cqring_wait(ctx
, min_events
);
6428 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
6429 TASK_INTERRUPTIBLE
);
6430 if (current
->task_works
)
6432 if (io_should_wake(&iowq
, false))
6435 if (signal_pending(current
)) {
6440 finish_wait(&ctx
->wait
, &iowq
.wq
);
6442 restore_saved_sigmask_unless(ret
== -EINTR
);
6444 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
6447 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6449 #if defined(CONFIG_UNIX)
6450 if (ctx
->ring_sock
) {
6451 struct sock
*sock
= ctx
->ring_sock
->sk
;
6452 struct sk_buff
*skb
;
6454 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
6460 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6463 file
= io_file_from_index(ctx
, i
);
6470 static void io_file_ref_kill(struct percpu_ref
*ref
)
6472 struct fixed_file_data
*data
;
6474 data
= container_of(ref
, struct fixed_file_data
, refs
);
6475 complete(&data
->done
);
6478 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6480 struct fixed_file_data
*data
= ctx
->file_data
;
6481 struct fixed_file_ref_node
*ref_node
= NULL
;
6482 unsigned nr_tables
, i
;
6487 spin_lock(&data
->lock
);
6488 if (!list_empty(&data
->ref_list
))
6489 ref_node
= list_first_entry(&data
->ref_list
,
6490 struct fixed_file_ref_node
, node
);
6491 spin_unlock(&data
->lock
);
6493 percpu_ref_kill(&ref_node
->refs
);
6495 percpu_ref_kill(&data
->refs
);
6497 /* wait for all refs nodes to complete */
6498 flush_delayed_work(&ctx
->file_put_work
);
6499 wait_for_completion(&data
->done
);
6501 __io_sqe_files_unregister(ctx
);
6502 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
6503 for (i
= 0; i
< nr_tables
; i
++)
6504 kfree(data
->table
[i
].files
);
6506 percpu_ref_exit(&data
->refs
);
6508 ctx
->file_data
= NULL
;
6509 ctx
->nr_user_files
= 0;
6513 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
6515 if (ctx
->sqo_thread
) {
6516 wait_for_completion(&ctx
->sq_thread_comp
);
6518 * The park is a bit of a work-around, without it we get
6519 * warning spews on shutdown with SQPOLL set and affinity
6520 * set to a single CPU.
6522 kthread_park(ctx
->sqo_thread
);
6523 kthread_stop(ctx
->sqo_thread
);
6524 ctx
->sqo_thread
= NULL
;
6528 static void io_finish_async(struct io_ring_ctx
*ctx
)
6530 io_sq_thread_stop(ctx
);
6533 io_wq_destroy(ctx
->io_wq
);
6538 #if defined(CONFIG_UNIX)
6540 * Ensure the UNIX gc is aware of our file set, so we are certain that
6541 * the io_uring can be safely unregistered on process exit, even if we have
6542 * loops in the file referencing.
6544 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
6546 struct sock
*sk
= ctx
->ring_sock
->sk
;
6547 struct scm_fp_list
*fpl
;
6548 struct sk_buff
*skb
;
6551 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
6555 skb
= alloc_skb(0, GFP_KERNEL
);
6564 fpl
->user
= get_uid(ctx
->user
);
6565 for (i
= 0; i
< nr
; i
++) {
6566 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
6570 fpl
->fp
[nr_files
] = get_file(file
);
6571 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
6576 fpl
->max
= SCM_MAX_FD
;
6577 fpl
->count
= nr_files
;
6578 UNIXCB(skb
).fp
= fpl
;
6579 skb
->destructor
= unix_destruct_scm
;
6580 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
6581 skb_queue_head(&sk
->sk_receive_queue
, skb
);
6583 for (i
= 0; i
< nr_files
; i
++)
6594 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6595 * causes regular reference counting to break down. We rely on the UNIX
6596 * garbage collection to take care of this problem for us.
6598 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
6600 unsigned left
, total
;
6604 left
= ctx
->nr_user_files
;
6606 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
6608 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
6612 total
+= this_files
;
6618 while (total
< ctx
->nr_user_files
) {
6619 struct file
*file
= io_file_from_index(ctx
, total
);
6629 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
6635 static int io_sqe_alloc_file_tables(struct io_ring_ctx
*ctx
, unsigned nr_tables
,
6640 for (i
= 0; i
< nr_tables
; i
++) {
6641 struct fixed_file_table
*table
= &ctx
->file_data
->table
[i
];
6642 unsigned this_files
;
6644 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
6645 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
6649 nr_files
-= this_files
;
6655 for (i
= 0; i
< nr_tables
; i
++) {
6656 struct fixed_file_table
*table
= &ctx
->file_data
->table
[i
];
6657 kfree(table
->files
);
6662 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
6664 #if defined(CONFIG_UNIX)
6665 struct sock
*sock
= ctx
->ring_sock
->sk
;
6666 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
6667 struct sk_buff
*skb
;
6670 __skb_queue_head_init(&list
);
6673 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6674 * remove this entry and rearrange the file array.
6676 skb
= skb_dequeue(head
);
6678 struct scm_fp_list
*fp
;
6680 fp
= UNIXCB(skb
).fp
;
6681 for (i
= 0; i
< fp
->count
; i
++) {
6684 if (fp
->fp
[i
] != file
)
6687 unix_notinflight(fp
->user
, fp
->fp
[i
]);
6688 left
= fp
->count
- 1 - i
;
6690 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
6691 left
* sizeof(struct file
*));
6698 __skb_queue_tail(&list
, skb
);
6708 __skb_queue_tail(&list
, skb
);
6710 skb
= skb_dequeue(head
);
6713 if (skb_peek(&list
)) {
6714 spin_lock_irq(&head
->lock
);
6715 while ((skb
= __skb_dequeue(&list
)) != NULL
)
6716 __skb_queue_tail(head
, skb
);
6717 spin_unlock_irq(&head
->lock
);
6724 struct io_file_put
{
6725 struct list_head list
;
6729 static void __io_file_put_work(struct fixed_file_ref_node
*ref_node
)
6731 struct fixed_file_data
*file_data
= ref_node
->file_data
;
6732 struct io_ring_ctx
*ctx
= file_data
->ctx
;
6733 struct io_file_put
*pfile
, *tmp
;
6735 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
6736 list_del(&pfile
->list
);
6737 io_ring_file_put(ctx
, pfile
->file
);
6741 spin_lock(&file_data
->lock
);
6742 list_del(&ref_node
->node
);
6743 spin_unlock(&file_data
->lock
);
6745 percpu_ref_exit(&ref_node
->refs
);
6747 percpu_ref_put(&file_data
->refs
);
6750 static void io_file_put_work(struct work_struct
*work
)
6752 struct io_ring_ctx
*ctx
;
6753 struct llist_node
*node
;
6755 ctx
= container_of(work
, struct io_ring_ctx
, file_put_work
.work
);
6756 node
= llist_del_all(&ctx
->file_put_llist
);
6759 struct fixed_file_ref_node
*ref_node
;
6760 struct llist_node
*next
= node
->next
;
6762 ref_node
= llist_entry(node
, struct fixed_file_ref_node
, llist
);
6763 __io_file_put_work(ref_node
);
6768 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
6770 struct fixed_file_ref_node
*ref_node
;
6771 struct io_ring_ctx
*ctx
;
6775 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
6776 ctx
= ref_node
->file_data
->ctx
;
6778 if (percpu_ref_is_dying(&ctx
->file_data
->refs
))
6781 first_add
= llist_add(&ref_node
->llist
, &ctx
->file_put_llist
);
6783 mod_delayed_work(system_wq
, &ctx
->file_put_work
, 0);
6785 queue_delayed_work(system_wq
, &ctx
->file_put_work
, delay
);
6788 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
6789 struct io_ring_ctx
*ctx
)
6791 struct fixed_file_ref_node
*ref_node
;
6793 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
6795 return ERR_PTR(-ENOMEM
);
6797 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
6800 return ERR_PTR(-ENOMEM
);
6802 INIT_LIST_HEAD(&ref_node
->node
);
6803 INIT_LIST_HEAD(&ref_node
->file_list
);
6804 ref_node
->file_data
= ctx
->file_data
;
6808 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
6810 percpu_ref_exit(&ref_node
->refs
);
6814 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
6817 __s32 __user
*fds
= (__s32 __user
*) arg
;
6822 struct fixed_file_ref_node
*ref_node
;
6828 if (nr_args
> IORING_MAX_FIXED_FILES
)
6831 ctx
->file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL
);
6832 if (!ctx
->file_data
)
6834 ctx
->file_data
->ctx
= ctx
;
6835 init_completion(&ctx
->file_data
->done
);
6836 INIT_LIST_HEAD(&ctx
->file_data
->ref_list
);
6837 spin_lock_init(&ctx
->file_data
->lock
);
6839 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
6840 ctx
->file_data
->table
= kcalloc(nr_tables
,
6841 sizeof(struct fixed_file_table
),
6843 if (!ctx
->file_data
->table
) {
6844 kfree(ctx
->file_data
);
6845 ctx
->file_data
= NULL
;
6849 if (percpu_ref_init(&ctx
->file_data
->refs
, io_file_ref_kill
,
6850 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
)) {
6851 kfree(ctx
->file_data
->table
);
6852 kfree(ctx
->file_data
);
6853 ctx
->file_data
= NULL
;
6857 if (io_sqe_alloc_file_tables(ctx
, nr_tables
, nr_args
)) {
6858 percpu_ref_exit(&ctx
->file_data
->refs
);
6859 kfree(ctx
->file_data
->table
);
6860 kfree(ctx
->file_data
);
6861 ctx
->file_data
= NULL
;
6865 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
6866 struct fixed_file_table
*table
;
6870 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
6872 /* allow sparse sets */
6878 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
6879 index
= i
& IORING_FILE_TABLE_MASK
;
6887 * Don't allow io_uring instances to be registered. If UNIX
6888 * isn't enabled, then this causes a reference cycle and this
6889 * instance can never get freed. If UNIX is enabled we'll
6890 * handle it just fine, but there's still no point in allowing
6891 * a ring fd as it doesn't support regular read/write anyway.
6893 if (file
->f_op
== &io_uring_fops
) {
6898 table
->files
[index
] = file
;
6902 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6903 file
= io_file_from_index(ctx
, i
);
6907 for (i
= 0; i
< nr_tables
; i
++)
6908 kfree(ctx
->file_data
->table
[i
].files
);
6910 kfree(ctx
->file_data
->table
);
6911 kfree(ctx
->file_data
);
6912 ctx
->file_data
= NULL
;
6913 ctx
->nr_user_files
= 0;
6917 ret
= io_sqe_files_scm(ctx
);
6919 io_sqe_files_unregister(ctx
);
6923 ref_node
= alloc_fixed_file_ref_node(ctx
);
6924 if (IS_ERR(ref_node
)) {
6925 io_sqe_files_unregister(ctx
);
6926 return PTR_ERR(ref_node
);
6929 ctx
->file_data
->cur_refs
= &ref_node
->refs
;
6930 spin_lock(&ctx
->file_data
->lock
);
6931 list_add(&ref_node
->node
, &ctx
->file_data
->ref_list
);
6932 spin_unlock(&ctx
->file_data
->lock
);
6933 percpu_ref_get(&ctx
->file_data
->refs
);
6937 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
6940 #if defined(CONFIG_UNIX)
6941 struct sock
*sock
= ctx
->ring_sock
->sk
;
6942 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
6943 struct sk_buff
*skb
;
6946 * See if we can merge this file into an existing skb SCM_RIGHTS
6947 * file set. If there's no room, fall back to allocating a new skb
6948 * and filling it in.
6950 spin_lock_irq(&head
->lock
);
6951 skb
= skb_peek(head
);
6953 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
6955 if (fpl
->count
< SCM_MAX_FD
) {
6956 __skb_unlink(skb
, head
);
6957 spin_unlock_irq(&head
->lock
);
6958 fpl
->fp
[fpl
->count
] = get_file(file
);
6959 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
6961 spin_lock_irq(&head
->lock
);
6962 __skb_queue_head(head
, skb
);
6967 spin_unlock_irq(&head
->lock
);
6974 return __io_sqe_files_scm(ctx
, 1, index
);
6980 static int io_queue_file_removal(struct fixed_file_data
*data
,
6983 struct io_file_put
*pfile
;
6984 struct percpu_ref
*refs
= data
->cur_refs
;
6985 struct fixed_file_ref_node
*ref_node
;
6987 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
6991 ref_node
= container_of(refs
, struct fixed_file_ref_node
, refs
);
6993 list_add(&pfile
->list
, &ref_node
->file_list
);
6998 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
6999 struct io_uring_files_update
*up
,
7002 struct fixed_file_data
*data
= ctx
->file_data
;
7003 struct fixed_file_ref_node
*ref_node
;
7008 bool needs_switch
= false;
7010 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7012 if (done
> ctx
->nr_user_files
)
7015 ref_node
= alloc_fixed_file_ref_node(ctx
);
7016 if (IS_ERR(ref_node
))
7017 return PTR_ERR(ref_node
);
7020 fds
= u64_to_user_ptr(up
->fds
);
7022 struct fixed_file_table
*table
;
7026 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7030 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
7031 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7032 index
= i
& IORING_FILE_TABLE_MASK
;
7033 if (table
->files
[index
]) {
7034 file
= io_file_from_index(ctx
, index
);
7035 err
= io_queue_file_removal(data
, file
);
7038 table
->files
[index
] = NULL
;
7039 needs_switch
= true;
7048 * Don't allow io_uring instances to be registered. If
7049 * UNIX isn't enabled, then this causes a reference
7050 * cycle and this instance can never get freed. If UNIX
7051 * is enabled we'll handle it just fine, but there's
7052 * still no point in allowing a ring fd as it doesn't
7053 * support regular read/write anyway.
7055 if (file
->f_op
== &io_uring_fops
) {
7060 table
->files
[index
] = file
;
7061 err
= io_sqe_file_register(ctx
, file
, i
);
7071 percpu_ref_kill(data
->cur_refs
);
7072 spin_lock(&data
->lock
);
7073 list_add(&ref_node
->node
, &data
->ref_list
);
7074 data
->cur_refs
= &ref_node
->refs
;
7075 spin_unlock(&data
->lock
);
7076 percpu_ref_get(&ctx
->file_data
->refs
);
7078 destroy_fixed_file_ref_node(ref_node
);
7080 return done
? done
: err
;
7083 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
7086 struct io_uring_files_update up
;
7088 if (!ctx
->file_data
)
7092 if (copy_from_user(&up
, arg
, sizeof(up
)))
7097 return __io_sqe_files_update(ctx
, &up
, nr_args
);
7100 static void io_free_work(struct io_wq_work
*work
)
7102 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7104 /* Consider that io_steal_work() relies on this ref */
7108 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
7109 struct io_uring_params
*p
)
7111 struct io_wq_data data
;
7113 struct io_ring_ctx
*ctx_attach
;
7114 unsigned int concurrency
;
7117 data
.user
= ctx
->user
;
7118 data
.free_work
= io_free_work
;
7119 data
.do_work
= io_wq_submit_work
;
7121 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
7122 /* Do QD, or 4 * CPUS, whatever is smallest */
7123 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
7125 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
7126 if (IS_ERR(ctx
->io_wq
)) {
7127 ret
= PTR_ERR(ctx
->io_wq
);
7133 f
= fdget(p
->wq_fd
);
7137 if (f
.file
->f_op
!= &io_uring_fops
) {
7142 ctx_attach
= f
.file
->private_data
;
7143 /* @io_wq is protected by holding the fd */
7144 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
7149 ctx
->io_wq
= ctx_attach
->io_wq
;
7155 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
7156 struct io_uring_params
*p
)
7160 mmgrab(current
->mm
);
7161 ctx
->sqo_mm
= current
->mm
;
7163 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7165 if (!capable(CAP_SYS_ADMIN
))
7168 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
7169 if (!ctx
->sq_thread_idle
)
7170 ctx
->sq_thread_idle
= HZ
;
7172 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7173 int cpu
= p
->sq_thread_cpu
;
7176 if (cpu
>= nr_cpu_ids
)
7178 if (!cpu_online(cpu
))
7181 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
7185 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
7188 if (IS_ERR(ctx
->sqo_thread
)) {
7189 ret
= PTR_ERR(ctx
->sqo_thread
);
7190 ctx
->sqo_thread
= NULL
;
7193 wake_up_process(ctx
->sqo_thread
);
7194 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7195 /* Can't have SQ_AFF without SQPOLL */
7200 ret
= io_init_wq_offload(ctx
, p
);
7206 io_finish_async(ctx
);
7207 mmdrop(ctx
->sqo_mm
);
7212 static inline void __io_unaccount_mem(struct user_struct
*user
,
7213 unsigned long nr_pages
)
7215 atomic_long_sub(nr_pages
, &user
->locked_vm
);
7218 static inline int __io_account_mem(struct user_struct
*user
,
7219 unsigned long nr_pages
)
7221 unsigned long page_limit
, cur_pages
, new_pages
;
7223 /* Don't allow more pages than we can safely lock */
7224 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
7227 cur_pages
= atomic_long_read(&user
->locked_vm
);
7228 new_pages
= cur_pages
+ nr_pages
;
7229 if (new_pages
> page_limit
)
7231 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
7232 new_pages
) != cur_pages
);
7237 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
7238 enum io_mem_account acct
)
7241 __io_unaccount_mem(ctx
->user
, nr_pages
);
7244 if (acct
== ACCT_LOCKED
)
7245 ctx
->sqo_mm
->locked_vm
-= nr_pages
;
7246 else if (acct
== ACCT_PINNED
)
7247 atomic64_sub(nr_pages
, &ctx
->sqo_mm
->pinned_vm
);
7251 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
7252 enum io_mem_account acct
)
7256 if (ctx
->limit_mem
) {
7257 ret
= __io_account_mem(ctx
->user
, nr_pages
);
7263 if (acct
== ACCT_LOCKED
)
7264 ctx
->sqo_mm
->locked_vm
+= nr_pages
;
7265 else if (acct
== ACCT_PINNED
)
7266 atomic64_add(nr_pages
, &ctx
->sqo_mm
->pinned_vm
);
7272 static void io_mem_free(void *ptr
)
7279 page
= virt_to_head_page(ptr
);
7280 if (put_page_testzero(page
))
7281 free_compound_page(page
);
7284 static void *io_mem_alloc(size_t size
)
7286 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
7289 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
7292 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
7295 struct io_rings
*rings
;
7296 size_t off
, sq_array_size
;
7298 off
= struct_size(rings
, cqes
, cq_entries
);
7299 if (off
== SIZE_MAX
)
7303 off
= ALIGN(off
, SMP_CACHE_BYTES
);
7308 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
7309 if (sq_array_size
== SIZE_MAX
)
7312 if (check_add_overflow(off
, sq_array_size
, &off
))
7321 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
7325 pages
= (size_t)1 << get_order(
7326 rings_size(sq_entries
, cq_entries
, NULL
));
7327 pages
+= (size_t)1 << get_order(
7328 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
7333 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
7337 if (!ctx
->user_bufs
)
7340 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
7341 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
7343 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
7344 unpin_user_page(imu
->bvec
[j
].bv_page
);
7346 io_unaccount_mem(ctx
, imu
->nr_bvecs
, ACCT_PINNED
);
7351 kfree(ctx
->user_bufs
);
7352 ctx
->user_bufs
= NULL
;
7353 ctx
->nr_user_bufs
= 0;
7357 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
7358 void __user
*arg
, unsigned index
)
7360 struct iovec __user
*src
;
7362 #ifdef CONFIG_COMPAT
7364 struct compat_iovec __user
*ciovs
;
7365 struct compat_iovec ciov
;
7367 ciovs
= (struct compat_iovec __user
*) arg
;
7368 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
7371 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
7372 dst
->iov_len
= ciov
.iov_len
;
7376 src
= (struct iovec __user
*) arg
;
7377 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
7382 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7385 struct vm_area_struct
**vmas
= NULL
;
7386 struct page
**pages
= NULL
;
7387 int i
, j
, got_pages
= 0;
7392 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
7395 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
7397 if (!ctx
->user_bufs
)
7400 for (i
= 0; i
< nr_args
; i
++) {
7401 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
7402 unsigned long off
, start
, end
, ubuf
;
7407 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
7412 * Don't impose further limits on the size and buffer
7413 * constraints here, we'll -EINVAL later when IO is
7414 * submitted if they are wrong.
7417 if (!iov
.iov_base
|| !iov
.iov_len
)
7420 /* arbitrary limit, but we need something */
7421 if (iov
.iov_len
> SZ_1G
)
7424 ubuf
= (unsigned long) iov
.iov_base
;
7425 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
7426 start
= ubuf
>> PAGE_SHIFT
;
7427 nr_pages
= end
- start
;
7429 ret
= io_account_mem(ctx
, nr_pages
, ACCT_PINNED
);
7434 if (!pages
|| nr_pages
> got_pages
) {
7437 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
7439 vmas
= kvmalloc_array(nr_pages
,
7440 sizeof(struct vm_area_struct
*),
7442 if (!pages
|| !vmas
) {
7444 io_unaccount_mem(ctx
, nr_pages
, ACCT_PINNED
);
7447 got_pages
= nr_pages
;
7450 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
7454 io_unaccount_mem(ctx
, nr_pages
, ACCT_PINNED
);
7459 mmap_read_lock(current
->mm
);
7460 pret
= pin_user_pages(ubuf
, nr_pages
,
7461 FOLL_WRITE
| FOLL_LONGTERM
,
7463 if (pret
== nr_pages
) {
7464 /* don't support file backed memory */
7465 for (j
= 0; j
< nr_pages
; j
++) {
7466 struct vm_area_struct
*vma
= vmas
[j
];
7469 !is_file_hugepages(vma
->vm_file
)) {
7475 ret
= pret
< 0 ? pret
: -EFAULT
;
7477 mmap_read_unlock(current
->mm
);
7480 * if we did partial map, or found file backed vmas,
7481 * release any pages we did get
7484 unpin_user_pages(pages
, pret
);
7485 io_unaccount_mem(ctx
, nr_pages
, ACCT_PINNED
);
7490 off
= ubuf
& ~PAGE_MASK
;
7492 for (j
= 0; j
< nr_pages
; j
++) {
7495 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
7496 imu
->bvec
[j
].bv_page
= pages
[j
];
7497 imu
->bvec
[j
].bv_len
= vec_len
;
7498 imu
->bvec
[j
].bv_offset
= off
;
7502 /* store original address for later verification */
7504 imu
->len
= iov
.iov_len
;
7505 imu
->nr_bvecs
= nr_pages
;
7507 ctx
->nr_user_bufs
++;
7515 io_sqe_buffer_unregister(ctx
);
7519 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
7521 __s32 __user
*fds
= arg
;
7527 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
7530 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
7531 if (IS_ERR(ctx
->cq_ev_fd
)) {
7532 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
7533 ctx
->cq_ev_fd
= NULL
;
7540 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
7542 if (ctx
->cq_ev_fd
) {
7543 eventfd_ctx_put(ctx
->cq_ev_fd
);
7544 ctx
->cq_ev_fd
= NULL
;
7551 static int __io_destroy_buffers(int id
, void *p
, void *data
)
7553 struct io_ring_ctx
*ctx
= data
;
7554 struct io_buffer
*buf
= p
;
7556 __io_remove_buffers(ctx
, buf
, id
, -1U);
7560 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
7562 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
7563 idr_destroy(&ctx
->io_buffer_idr
);
7566 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
7568 io_finish_async(ctx
);
7570 mmdrop(ctx
->sqo_mm
);
7574 io_iopoll_reap_events(ctx
);
7575 io_sqe_buffer_unregister(ctx
);
7576 io_sqe_files_unregister(ctx
);
7577 io_eventfd_unregister(ctx
);
7578 io_destroy_buffers(ctx
);
7579 idr_destroy(&ctx
->personality_idr
);
7581 #if defined(CONFIG_UNIX)
7582 if (ctx
->ring_sock
) {
7583 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
7584 sock_release(ctx
->ring_sock
);
7588 io_mem_free(ctx
->rings
);
7589 io_mem_free(ctx
->sq_sqes
);
7591 percpu_ref_exit(&ctx
->refs
);
7592 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
7594 free_uid(ctx
->user
);
7595 put_cred(ctx
->creds
);
7596 kfree(ctx
->cancel_hash
);
7597 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
7601 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
7603 struct io_ring_ctx
*ctx
= file
->private_data
;
7606 poll_wait(file
, &ctx
->cq_wait
, wait
);
7608 * synchronizes with barrier from wq_has_sleeper call in
7612 if (READ_ONCE(ctx
->rings
->sq
.tail
) - ctx
->cached_sq_head
!=
7613 ctx
->rings
->sq_ring_entries
)
7614 mask
|= EPOLLOUT
| EPOLLWRNORM
;
7615 if (io_cqring_events(ctx
, false))
7616 mask
|= EPOLLIN
| EPOLLRDNORM
;
7621 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
7623 struct io_ring_ctx
*ctx
= file
->private_data
;
7625 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
7628 static int io_remove_personalities(int id
, void *p
, void *data
)
7630 struct io_ring_ctx
*ctx
= data
;
7631 const struct cred
*cred
;
7633 cred
= idr_remove(&ctx
->personality_idr
, id
);
7639 static void io_ring_exit_work(struct work_struct
*work
)
7641 struct io_ring_ctx
*ctx
;
7643 ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
7645 io_cqring_overflow_flush(ctx
, true);
7648 * If we're doing polled IO and end up having requests being
7649 * submitted async (out-of-line), then completions can come in while
7650 * we're waiting for refs to drop. We need to reap these manually,
7651 * as nobody else will be looking for them.
7653 while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20)) {
7654 io_iopoll_reap_events(ctx
);
7656 io_cqring_overflow_flush(ctx
, true);
7658 io_ring_ctx_free(ctx
);
7661 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
7663 mutex_lock(&ctx
->uring_lock
);
7664 percpu_ref_kill(&ctx
->refs
);
7665 mutex_unlock(&ctx
->uring_lock
);
7667 io_kill_timeouts(ctx
);
7668 io_poll_remove_all(ctx
);
7671 io_wq_cancel_all(ctx
->io_wq
);
7673 io_iopoll_reap_events(ctx
);
7674 /* if we failed setting up the ctx, we might not have any rings */
7676 io_cqring_overflow_flush(ctx
, true);
7677 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
7678 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
7679 queue_work(system_wq
, &ctx
->exit_work
);
7682 static int io_uring_release(struct inode
*inode
, struct file
*file
)
7684 struct io_ring_ctx
*ctx
= file
->private_data
;
7686 file
->private_data
= NULL
;
7687 io_ring_ctx_wait_and_kill(ctx
);
7691 static bool io_wq_files_match(struct io_wq_work
*work
, void *data
)
7693 struct files_struct
*files
= data
;
7695 return work
->files
== files
;
7698 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
7699 struct files_struct
*files
)
7701 if (list_empty_careful(&ctx
->inflight_list
))
7704 /* cancel all at once, should be faster than doing it one by one*/
7705 io_wq_cancel_cb(ctx
->io_wq
, io_wq_files_match
, files
, true);
7707 while (!list_empty_careful(&ctx
->inflight_list
)) {
7708 struct io_kiocb
*cancel_req
= NULL
, *req
;
7711 spin_lock_irq(&ctx
->inflight_lock
);
7712 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
) {
7713 if (req
->work
.files
!= files
)
7715 /* req is being completed, ignore */
7716 if (!refcount_inc_not_zero(&req
->refs
))
7722 prepare_to_wait(&ctx
->inflight_wait
, &wait
,
7723 TASK_UNINTERRUPTIBLE
);
7724 spin_unlock_irq(&ctx
->inflight_lock
);
7726 /* We need to keep going until we don't find a matching req */
7730 if (cancel_req
->flags
& REQ_F_OVERFLOW
) {
7731 spin_lock_irq(&ctx
->completion_lock
);
7732 list_del(&cancel_req
->list
);
7733 cancel_req
->flags
&= ~REQ_F_OVERFLOW
;
7734 if (list_empty(&ctx
->cq_overflow_list
)) {
7735 clear_bit(0, &ctx
->sq_check_overflow
);
7736 clear_bit(0, &ctx
->cq_check_overflow
);
7738 spin_unlock_irq(&ctx
->completion_lock
);
7740 WRITE_ONCE(ctx
->rings
->cq_overflow
,
7741 atomic_inc_return(&ctx
->cached_cq_overflow
));
7744 * Put inflight ref and overflow ref. If that's
7745 * all we had, then we're done with this request.
7747 if (refcount_sub_and_test(2, &cancel_req
->refs
)) {
7748 io_free_req(cancel_req
);
7749 finish_wait(&ctx
->inflight_wait
, &wait
);
7753 io_wq_cancel_work(ctx
->io_wq
, &cancel_req
->work
);
7754 io_put_req(cancel_req
);
7758 finish_wait(&ctx
->inflight_wait
, &wait
);
7762 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
7764 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7765 struct task_struct
*task
= data
;
7767 return req
->task
== task
;
7770 static int io_uring_flush(struct file
*file
, void *data
)
7772 struct io_ring_ctx
*ctx
= file
->private_data
;
7774 io_uring_cancel_files(ctx
, data
);
7777 * If the task is going away, cancel work it may have pending
7779 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
))
7780 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
, current
, true);
7785 static void *io_uring_validate_mmap_request(struct file
*file
,
7786 loff_t pgoff
, size_t sz
)
7788 struct io_ring_ctx
*ctx
= file
->private_data
;
7789 loff_t offset
= pgoff
<< PAGE_SHIFT
;
7794 case IORING_OFF_SQ_RING
:
7795 case IORING_OFF_CQ_RING
:
7798 case IORING_OFF_SQES
:
7802 return ERR_PTR(-EINVAL
);
7805 page
= virt_to_head_page(ptr
);
7806 if (sz
> page_size(page
))
7807 return ERR_PTR(-EINVAL
);
7814 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
7816 size_t sz
= vma
->vm_end
- vma
->vm_start
;
7820 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
7822 return PTR_ERR(ptr
);
7824 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
7825 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
7828 #else /* !CONFIG_MMU */
7830 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
7832 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
7835 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
7837 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
7840 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
7841 unsigned long addr
, unsigned long len
,
7842 unsigned long pgoff
, unsigned long flags
)
7846 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
7848 return PTR_ERR(ptr
);
7850 return (unsigned long) ptr
;
7853 #endif /* !CONFIG_MMU */
7855 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
7856 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
7859 struct io_ring_ctx
*ctx
;
7864 if (current
->task_works
)
7867 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
7875 if (f
.file
->f_op
!= &io_uring_fops
)
7879 ctx
= f
.file
->private_data
;
7880 if (!percpu_ref_tryget(&ctx
->refs
))
7884 * For SQ polling, the thread will do all submissions and completions.
7885 * Just return the requested submit count, and wake the thread if
7889 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7890 if (!list_empty_careful(&ctx
->cq_overflow_list
))
7891 io_cqring_overflow_flush(ctx
, false);
7892 if (flags
& IORING_ENTER_SQ_WAKEUP
)
7893 wake_up(&ctx
->sqo_wait
);
7894 submitted
= to_submit
;
7895 } else if (to_submit
) {
7896 mutex_lock(&ctx
->uring_lock
);
7897 submitted
= io_submit_sqes(ctx
, to_submit
, f
.file
, fd
);
7898 mutex_unlock(&ctx
->uring_lock
);
7900 if (submitted
!= to_submit
)
7903 if (flags
& IORING_ENTER_GETEVENTS
) {
7904 unsigned nr_events
= 0;
7906 min_complete
= min(min_complete
, ctx
->cq_entries
);
7909 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7910 * space applications don't need to do io completion events
7911 * polling again, they can rely on io_sq_thread to do polling
7912 * work, which can reduce cpu usage and uring_lock contention.
7914 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
7915 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
7916 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
7918 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
7923 percpu_ref_put(&ctx
->refs
);
7926 return submitted
? submitted
: ret
;
7929 #ifdef CONFIG_PROC_FS
7930 static int io_uring_show_cred(int id
, void *p
, void *data
)
7932 const struct cred
*cred
= p
;
7933 struct seq_file
*m
= data
;
7934 struct user_namespace
*uns
= seq_user_ns(m
);
7935 struct group_info
*gi
;
7940 seq_printf(m
, "%5d\n", id
);
7941 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
7942 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
7943 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
7944 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
7945 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
7946 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
7947 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
7948 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
7949 seq_puts(m
, "\n\tGroups:\t");
7950 gi
= cred
->group_info
;
7951 for (g
= 0; g
< gi
->ngroups
; g
++) {
7952 seq_put_decimal_ull(m
, g
? " " : "",
7953 from_kgid_munged(uns
, gi
->gid
[g
]));
7955 seq_puts(m
, "\n\tCapEff:\t");
7956 cap
= cred
->cap_effective
;
7957 CAP_FOR_EACH_U32(__capi
)
7958 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
7963 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
7967 mutex_lock(&ctx
->uring_lock
);
7968 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
7969 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7970 struct fixed_file_table
*table
;
7973 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7974 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
7976 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
7978 seq_printf(m
, "%5u: <none>\n", i
);
7980 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
7981 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
7982 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
7984 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
7985 (unsigned int) buf
->len
);
7987 if (!idr_is_empty(&ctx
->personality_idr
)) {
7988 seq_printf(m
, "Personalities:\n");
7989 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
7991 seq_printf(m
, "PollList:\n");
7992 spin_lock_irq(&ctx
->completion_lock
);
7993 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
7994 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
7995 struct io_kiocb
*req
;
7997 hlist_for_each_entry(req
, list
, hash_node
)
7998 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
7999 req
->task
->task_works
!= NULL
);
8001 spin_unlock_irq(&ctx
->completion_lock
);
8002 mutex_unlock(&ctx
->uring_lock
);
8005 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
8007 struct io_ring_ctx
*ctx
= f
->private_data
;
8009 if (percpu_ref_tryget(&ctx
->refs
)) {
8010 __io_uring_show_fdinfo(ctx
, m
);
8011 percpu_ref_put(&ctx
->refs
);
8016 static const struct file_operations io_uring_fops
= {
8017 .release
= io_uring_release
,
8018 .flush
= io_uring_flush
,
8019 .mmap
= io_uring_mmap
,
8021 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
8022 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
8024 .poll
= io_uring_poll
,
8025 .fasync
= io_uring_fasync
,
8026 #ifdef CONFIG_PROC_FS
8027 .show_fdinfo
= io_uring_show_fdinfo
,
8031 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
8032 struct io_uring_params
*p
)
8034 struct io_rings
*rings
;
8035 size_t size
, sq_array_offset
;
8037 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
8038 if (size
== SIZE_MAX
)
8041 rings
= io_mem_alloc(size
);
8046 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
8047 rings
->sq_ring_mask
= p
->sq_entries
- 1;
8048 rings
->cq_ring_mask
= p
->cq_entries
- 1;
8049 rings
->sq_ring_entries
= p
->sq_entries
;
8050 rings
->cq_ring_entries
= p
->cq_entries
;
8051 ctx
->sq_mask
= rings
->sq_ring_mask
;
8052 ctx
->cq_mask
= rings
->cq_ring_mask
;
8053 ctx
->sq_entries
= rings
->sq_ring_entries
;
8054 ctx
->cq_entries
= rings
->cq_ring_entries
;
8056 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
8057 if (size
== SIZE_MAX
) {
8058 io_mem_free(ctx
->rings
);
8063 ctx
->sq_sqes
= io_mem_alloc(size
);
8064 if (!ctx
->sq_sqes
) {
8065 io_mem_free(ctx
->rings
);
8074 * Allocate an anonymous fd, this is what constitutes the application
8075 * visible backing of an io_uring instance. The application mmaps this
8076 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8077 * we have to tie this fd to a socket for file garbage collection purposes.
8079 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
8084 #if defined(CONFIG_UNIX)
8085 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
8091 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
8095 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
8096 O_RDWR
| O_CLOEXEC
);
8099 ret
= PTR_ERR(file
);
8103 #if defined(CONFIG_UNIX)
8104 ctx
->ring_sock
->file
= file
;
8106 fd_install(ret
, file
);
8109 #if defined(CONFIG_UNIX)
8110 sock_release(ctx
->ring_sock
);
8111 ctx
->ring_sock
= NULL
;
8116 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
8117 struct io_uring_params __user
*params
)
8119 struct user_struct
*user
= NULL
;
8120 struct io_ring_ctx
*ctx
;
8126 if (entries
> IORING_MAX_ENTRIES
) {
8127 if (!(p
->flags
& IORING_SETUP_CLAMP
))
8129 entries
= IORING_MAX_ENTRIES
;
8133 * Use twice as many entries for the CQ ring. It's possible for the
8134 * application to drive a higher depth than the size of the SQ ring,
8135 * since the sqes are only used at submission time. This allows for
8136 * some flexibility in overcommitting a bit. If the application has
8137 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8138 * of CQ ring entries manually.
8140 p
->sq_entries
= roundup_pow_of_two(entries
);
8141 if (p
->flags
& IORING_SETUP_CQSIZE
) {
8143 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8144 * to a power-of-two, if it isn't already. We do NOT impose
8145 * any cq vs sq ring sizing.
8147 if (p
->cq_entries
< p
->sq_entries
)
8149 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
8150 if (!(p
->flags
& IORING_SETUP_CLAMP
))
8152 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
8154 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
8156 p
->cq_entries
= 2 * p
->sq_entries
;
8159 user
= get_uid(current_user());
8160 limit_mem
= !capable(CAP_IPC_LOCK
);
8163 ret
= __io_account_mem(user
,
8164 ring_pages(p
->sq_entries
, p
->cq_entries
));
8171 ctx
= io_ring_ctx_alloc(p
);
8174 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
8179 ctx
->compat
= in_compat_syscall();
8181 ctx
->creds
= get_current_cred();
8183 ret
= io_allocate_scq_urings(ctx
, p
);
8187 ret
= io_sq_offload_start(ctx
, p
);
8191 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
8192 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
8193 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
8194 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
8195 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
8196 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
8197 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
8198 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
8200 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
8201 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
8202 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
8203 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
8204 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
8205 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
8206 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
8207 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
8209 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
8210 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
8211 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
8212 IORING_FEAT_POLL_32BITS
;
8214 if (copy_to_user(params
, p
, sizeof(*p
))) {
8219 * Install ring fd as the very last thing, so we don't risk someone
8220 * having closed it before we finish setup
8222 ret
= io_uring_get_fd(ctx
);
8226 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
8227 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
8229 ctx
->limit_mem
= limit_mem
;
8232 io_ring_ctx_wait_and_kill(ctx
);
8237 * Sets up an aio uring context, and returns the fd. Applications asks for a
8238 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8239 * params structure passed in.
8241 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
8243 struct io_uring_params p
;
8246 if (copy_from_user(&p
, params
, sizeof(p
)))
8248 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
8253 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
8254 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
8255 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
))
8258 return io_uring_create(entries
, &p
, params
);
8261 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
8262 struct io_uring_params __user
*, params
)
8264 return io_uring_setup(entries
, params
);
8267 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
8269 struct io_uring_probe
*p
;
8273 size
= struct_size(p
, ops
, nr_args
);
8274 if (size
== SIZE_MAX
)
8276 p
= kzalloc(size
, GFP_KERNEL
);
8281 if (copy_from_user(p
, arg
, size
))
8284 if (memchr_inv(p
, 0, size
))
8287 p
->last_op
= IORING_OP_LAST
- 1;
8288 if (nr_args
> IORING_OP_LAST
)
8289 nr_args
= IORING_OP_LAST
;
8291 for (i
= 0; i
< nr_args
; i
++) {
8293 if (!io_op_defs
[i
].not_supported
)
8294 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
8299 if (copy_to_user(arg
, p
, size
))
8306 static int io_register_personality(struct io_ring_ctx
*ctx
)
8308 const struct cred
*creds
= get_current_cred();
8311 id
= idr_alloc_cyclic(&ctx
->personality_idr
, (void *) creds
, 1,
8312 USHRT_MAX
, GFP_KERNEL
);
8318 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8320 const struct cred
*old_creds
;
8322 old_creds
= idr_remove(&ctx
->personality_idr
, id
);
8324 put_cred(old_creds
);
8331 static bool io_register_op_must_quiesce(int op
)
8334 case IORING_UNREGISTER_FILES
:
8335 case IORING_REGISTER_FILES_UPDATE
:
8336 case IORING_REGISTER_PROBE
:
8337 case IORING_REGISTER_PERSONALITY
:
8338 case IORING_UNREGISTER_PERSONALITY
:
8345 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
8346 void __user
*arg
, unsigned nr_args
)
8347 __releases(ctx
->uring_lock
)
8348 __acquires(ctx
->uring_lock
)
8353 * We're inside the ring mutex, if the ref is already dying, then
8354 * someone else killed the ctx or is already going through
8355 * io_uring_register().
8357 if (percpu_ref_is_dying(&ctx
->refs
))
8360 if (io_register_op_must_quiesce(opcode
)) {
8361 percpu_ref_kill(&ctx
->refs
);
8364 * Drop uring mutex before waiting for references to exit. If
8365 * another thread is currently inside io_uring_enter() it might
8366 * need to grab the uring_lock to make progress. If we hold it
8367 * here across the drain wait, then we can deadlock. It's safe
8368 * to drop the mutex here, since no new references will come in
8369 * after we've killed the percpu ref.
8371 mutex_unlock(&ctx
->uring_lock
);
8372 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
8373 mutex_lock(&ctx
->uring_lock
);
8375 percpu_ref_resurrect(&ctx
->refs
);
8382 case IORING_REGISTER_BUFFERS
:
8383 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
8385 case IORING_UNREGISTER_BUFFERS
:
8389 ret
= io_sqe_buffer_unregister(ctx
);
8391 case IORING_REGISTER_FILES
:
8392 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
8394 case IORING_UNREGISTER_FILES
:
8398 ret
= io_sqe_files_unregister(ctx
);
8400 case IORING_REGISTER_FILES_UPDATE
:
8401 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
8403 case IORING_REGISTER_EVENTFD
:
8404 case IORING_REGISTER_EVENTFD_ASYNC
:
8408 ret
= io_eventfd_register(ctx
, arg
);
8411 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
8412 ctx
->eventfd_async
= 1;
8414 ctx
->eventfd_async
= 0;
8416 case IORING_UNREGISTER_EVENTFD
:
8420 ret
= io_eventfd_unregister(ctx
);
8422 case IORING_REGISTER_PROBE
:
8424 if (!arg
|| nr_args
> 256)
8426 ret
= io_probe(ctx
, arg
, nr_args
);
8428 case IORING_REGISTER_PERSONALITY
:
8432 ret
= io_register_personality(ctx
);
8434 case IORING_UNREGISTER_PERSONALITY
:
8438 ret
= io_unregister_personality(ctx
, nr_args
);
8445 if (io_register_op_must_quiesce(opcode
)) {
8446 /* bring the ctx back to life */
8447 percpu_ref_reinit(&ctx
->refs
);
8449 reinit_completion(&ctx
->ref_comp
);
8454 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
8455 void __user
*, arg
, unsigned int, nr_args
)
8457 struct io_ring_ctx
*ctx
;
8466 if (f
.file
->f_op
!= &io_uring_fops
)
8469 ctx
= f
.file
->private_data
;
8471 mutex_lock(&ctx
->uring_lock
);
8472 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
8473 mutex_unlock(&ctx
->uring_lock
);
8474 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
8475 ctx
->cq_ev_fd
!= NULL
, ret
);
8481 static int __init
io_uring_init(void)
8483 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8484 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8485 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8488 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8489 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8490 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
8491 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
8492 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
8493 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
8494 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
8495 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
8496 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
8497 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
8498 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
8499 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
8500 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
8501 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
8502 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
8503 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
8504 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
8505 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
8506 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
8507 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
8508 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
8509 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
8510 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
8511 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
8512 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
8513 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
8514 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
8515 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
8516 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
8517 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
8518 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
8520 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
8521 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
8522 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
8525 __initcall(io_uring_init
);