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
,
537 REQ_F_TIMEOUT_NOSEQ_BIT
,
538 REQ_F_COMP_LOCKED_BIT
,
539 REQ_F_NEED_CLEANUP_BIT
,
542 REQ_F_BUFFER_SELECTED_BIT
,
543 REQ_F_NO_FILE_TABLE_BIT
,
544 REQ_F_QUEUE_TIMEOUT_BIT
,
545 REQ_F_WORK_INITIALIZED_BIT
,
546 REQ_F_TASK_PINNED_BIT
,
548 /* not a real bit, just to check we're not overflowing the space */
554 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
555 /* drain existing IO first */
556 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
558 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
559 /* doesn't sever on completion < 0 */
560 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
562 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
563 /* IOSQE_BUFFER_SELECT */
564 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
567 REQ_F_LINK_HEAD
= BIT(REQ_F_LINK_HEAD_BIT
),
568 /* already grabbed next link */
569 REQ_F_LINK_NEXT
= BIT(REQ_F_LINK_NEXT_BIT
),
570 /* fail rest of links */
571 REQ_F_FAIL_LINK
= BIT(REQ_F_FAIL_LINK_BIT
),
572 /* on inflight list */
573 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
574 /* read/write uses file position */
575 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
576 /* must not punt to workers */
577 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
578 /* has linked timeout */
579 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
580 /* timeout request */
581 REQ_F_TIMEOUT
= BIT(REQ_F_TIMEOUT_BIT
),
583 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
584 /* no timeout sequence */
585 REQ_F_TIMEOUT_NOSEQ
= BIT(REQ_F_TIMEOUT_NOSEQ_BIT
),
586 /* completion under lock */
587 REQ_F_COMP_LOCKED
= BIT(REQ_F_COMP_LOCKED_BIT
),
589 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
590 /* in overflow list */
591 REQ_F_OVERFLOW
= BIT(REQ_F_OVERFLOW_BIT
),
592 /* already went through poll handler */
593 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
594 /* buffer already selected */
595 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
596 /* doesn't need file table for this request */
597 REQ_F_NO_FILE_TABLE
= BIT(REQ_F_NO_FILE_TABLE_BIT
),
598 /* needs to queue linked timeout */
599 REQ_F_QUEUE_TIMEOUT
= BIT(REQ_F_QUEUE_TIMEOUT_BIT
),
600 /* io_wq_work is initialized */
601 REQ_F_WORK_INITIALIZED
= BIT(REQ_F_WORK_INITIALIZED_BIT
),
602 /* req->task is refcounted */
603 REQ_F_TASK_PINNED
= BIT(REQ_F_TASK_PINNED_BIT
),
607 struct io_poll_iocb poll
;
608 struct io_wq_work work
;
612 * NOTE! Each of the iocb union members has the file pointer
613 * as the first entry in their struct definition. So you can
614 * access the file pointer through any of the sub-structs,
615 * or directly as just 'ki_filp' in this struct.
621 struct io_poll_iocb poll
;
622 struct io_accept accept
;
624 struct io_cancel cancel
;
625 struct io_timeout timeout
;
626 struct io_connect connect
;
627 struct io_sr_msg sr_msg
;
629 struct io_close close
;
630 struct io_files_update files_update
;
631 struct io_fadvise fadvise
;
632 struct io_madvise madvise
;
633 struct io_epoll epoll
;
634 struct io_splice splice
;
635 struct io_provide_buf pbuf
;
636 struct io_statx statx
;
639 struct io_async_ctx
*io
;
642 /* polled IO has completed */
647 struct io_ring_ctx
*ctx
;
648 struct list_head list
;
651 struct task_struct
*task
;
657 struct list_head link_list
;
659 struct list_head inflight_entry
;
661 struct percpu_ref
*fixed_file_refs
;
665 * Only commands that never go async can use the below fields,
666 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
667 * async armed poll handlers for regular commands. The latter
668 * restore the work, if needed.
671 struct callback_head task_work
;
672 struct hlist_node hash_node
;
673 struct async_poll
*apoll
;
675 struct io_wq_work work
;
679 #define IO_IOPOLL_BATCH 8
681 struct io_submit_state
{
682 struct blk_plug plug
;
685 * io_kiocb alloc cache
687 void *reqs
[IO_IOPOLL_BATCH
];
688 unsigned int free_reqs
;
691 * File reference cache
695 unsigned int has_refs
;
696 unsigned int used_refs
;
697 unsigned int ios_left
;
701 /* needs req->io allocated for deferral/async */
702 unsigned async_ctx
: 1;
703 /* needs current->mm setup, does mm access */
704 unsigned needs_mm
: 1;
705 /* needs req->file assigned */
706 unsigned needs_file
: 1;
707 /* don't fail if file grab fails */
708 unsigned needs_file_no_error
: 1;
709 /* hash wq insertion if file is a regular file */
710 unsigned hash_reg_file
: 1;
711 /* unbound wq insertion if file is a non-regular file */
712 unsigned unbound_nonreg_file
: 1;
713 /* opcode is not supported by this kernel */
714 unsigned not_supported
: 1;
715 /* needs file table */
716 unsigned file_table
: 1;
718 unsigned needs_fs
: 1;
719 /* set if opcode supports polled "wait" */
721 unsigned pollout
: 1;
722 /* op supports buffer selection */
723 unsigned buffer_select
: 1;
726 static const struct io_op_def io_op_defs
[] = {
727 [IORING_OP_NOP
] = {},
728 [IORING_OP_READV
] = {
732 .unbound_nonreg_file
= 1,
736 [IORING_OP_WRITEV
] = {
741 .unbound_nonreg_file
= 1,
744 [IORING_OP_FSYNC
] = {
747 [IORING_OP_READ_FIXED
] = {
749 .unbound_nonreg_file
= 1,
752 [IORING_OP_WRITE_FIXED
] = {
755 .unbound_nonreg_file
= 1,
758 [IORING_OP_POLL_ADD
] = {
760 .unbound_nonreg_file
= 1,
762 [IORING_OP_POLL_REMOVE
] = {},
763 [IORING_OP_SYNC_FILE_RANGE
] = {
766 [IORING_OP_SENDMSG
] = {
770 .unbound_nonreg_file
= 1,
774 [IORING_OP_RECVMSG
] = {
778 .unbound_nonreg_file
= 1,
783 [IORING_OP_TIMEOUT
] = {
787 [IORING_OP_TIMEOUT_REMOVE
] = {},
788 [IORING_OP_ACCEPT
] = {
791 .unbound_nonreg_file
= 1,
795 [IORING_OP_ASYNC_CANCEL
] = {},
796 [IORING_OP_LINK_TIMEOUT
] = {
800 [IORING_OP_CONNECT
] = {
804 .unbound_nonreg_file
= 1,
807 [IORING_OP_FALLOCATE
] = {
810 [IORING_OP_OPENAT
] = {
814 [IORING_OP_CLOSE
] = {
816 .needs_file_no_error
= 1,
819 [IORING_OP_FILES_UPDATE
] = {
823 [IORING_OP_STATX
] = {
831 .unbound_nonreg_file
= 1,
835 [IORING_OP_WRITE
] = {
838 .unbound_nonreg_file
= 1,
841 [IORING_OP_FADVISE
] = {
844 [IORING_OP_MADVISE
] = {
850 .unbound_nonreg_file
= 1,
856 .unbound_nonreg_file
= 1,
860 [IORING_OP_OPENAT2
] = {
864 [IORING_OP_EPOLL_CTL
] = {
865 .unbound_nonreg_file
= 1,
868 [IORING_OP_SPLICE
] = {
871 .unbound_nonreg_file
= 1,
873 [IORING_OP_PROVIDE_BUFFERS
] = {},
874 [IORING_OP_REMOVE_BUFFERS
] = {},
878 .unbound_nonreg_file
= 1,
882 enum io_mem_account
{
887 static void io_wq_submit_work(struct io_wq_work
**workptr
);
888 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
);
889 static void io_put_req(struct io_kiocb
*req
);
890 static void __io_double_put_req(struct io_kiocb
*req
);
891 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
);
892 static void io_queue_linked_timeout(struct io_kiocb
*req
);
893 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
894 struct io_uring_files_update
*ip
,
896 static int io_grab_files(struct io_kiocb
*req
);
897 static void io_cleanup_req(struct io_kiocb
*req
);
898 static int io_file_get(struct io_submit_state
*state
, struct io_kiocb
*req
,
899 int fd
, struct file
**out_file
, bool fixed
);
900 static void __io_queue_sqe(struct io_kiocb
*req
,
901 const struct io_uring_sqe
*sqe
);
903 static ssize_t
io_import_iovec(int rw
, struct io_kiocb
*req
,
904 struct iovec
**iovec
, struct iov_iter
*iter
,
906 static int io_setup_async_rw(struct io_kiocb
*req
, ssize_t io_size
,
907 struct iovec
*iovec
, struct iovec
*fast_iov
,
908 struct iov_iter
*iter
);
910 static struct kmem_cache
*req_cachep
;
912 static const struct file_operations io_uring_fops
;
914 struct sock
*io_uring_get_socket(struct file
*file
)
916 #if defined(CONFIG_UNIX)
917 if (file
->f_op
== &io_uring_fops
) {
918 struct io_ring_ctx
*ctx
= file
->private_data
;
920 return ctx
->ring_sock
->sk
;
925 EXPORT_SYMBOL(io_uring_get_socket
);
927 static void io_get_req_task(struct io_kiocb
*req
)
929 if (req
->flags
& REQ_F_TASK_PINNED
)
931 get_task_struct(req
->task
);
932 req
->flags
|= REQ_F_TASK_PINNED
;
935 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
936 static void __io_put_req_task(struct io_kiocb
*req
)
938 if (req
->flags
& REQ_F_TASK_PINNED
)
939 put_task_struct(req
->task
);
942 static void io_file_put_work(struct work_struct
*work
);
945 * Note: must call io_req_init_async() for the first time you
946 * touch any members of io_wq_work.
948 static inline void io_req_init_async(struct io_kiocb
*req
)
950 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
953 memset(&req
->work
, 0, sizeof(req
->work
));
954 req
->flags
|= REQ_F_WORK_INITIALIZED
;
957 static inline bool io_async_submit(struct io_ring_ctx
*ctx
)
959 return ctx
->flags
& IORING_SETUP_SQPOLL
;
962 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
964 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
966 complete(&ctx
->ref_comp
);
969 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
971 struct io_ring_ctx
*ctx
;
974 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
978 ctx
->fallback_req
= kmem_cache_alloc(req_cachep
, GFP_KERNEL
);
979 if (!ctx
->fallback_req
)
983 * Use 5 bits less than the max cq entries, that should give us around
984 * 32 entries per hash list if totally full and uniformly spread.
986 hash_bits
= ilog2(p
->cq_entries
);
990 ctx
->cancel_hash_bits
= hash_bits
;
991 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
993 if (!ctx
->cancel_hash
)
995 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
997 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
998 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1001 ctx
->flags
= p
->flags
;
1002 init_waitqueue_head(&ctx
->sqo_wait
);
1003 init_waitqueue_head(&ctx
->cq_wait
);
1004 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1005 init_completion(&ctx
->ref_comp
);
1006 init_completion(&ctx
->sq_thread_comp
);
1007 idr_init(&ctx
->io_buffer_idr
);
1008 idr_init(&ctx
->personality_idr
);
1009 mutex_init(&ctx
->uring_lock
);
1010 init_waitqueue_head(&ctx
->wait
);
1011 spin_lock_init(&ctx
->completion_lock
);
1012 INIT_LIST_HEAD(&ctx
->poll_list
);
1013 INIT_LIST_HEAD(&ctx
->defer_list
);
1014 INIT_LIST_HEAD(&ctx
->timeout_list
);
1015 init_waitqueue_head(&ctx
->inflight_wait
);
1016 spin_lock_init(&ctx
->inflight_lock
);
1017 INIT_LIST_HEAD(&ctx
->inflight_list
);
1018 INIT_DELAYED_WORK(&ctx
->file_put_work
, io_file_put_work
);
1019 init_llist_head(&ctx
->file_put_llist
);
1022 if (ctx
->fallback_req
)
1023 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
1024 kfree(ctx
->cancel_hash
);
1029 static inline bool __req_need_defer(struct io_kiocb
*req
)
1031 struct io_ring_ctx
*ctx
= req
->ctx
;
1033 return req
->sequence
!= ctx
->cached_cq_tail
1034 + atomic_read(&ctx
->cached_cq_overflow
);
1037 static inline bool req_need_defer(struct io_kiocb
*req
)
1039 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
))
1040 return __req_need_defer(req
);
1045 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
1047 struct io_rings
*rings
= ctx
->rings
;
1049 /* order cqe stores with ring update */
1050 smp_store_release(&rings
->cq
.tail
, ctx
->cached_cq_tail
);
1052 if (wq_has_sleeper(&ctx
->cq_wait
)) {
1053 wake_up_interruptible(&ctx
->cq_wait
);
1054 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
1058 static inline void io_req_work_grab_env(struct io_kiocb
*req
,
1059 const struct io_op_def
*def
)
1061 if (!req
->work
.mm
&& def
->needs_mm
) {
1062 mmgrab(current
->mm
);
1063 req
->work
.mm
= current
->mm
;
1065 if (!req
->work
.creds
)
1066 req
->work
.creds
= get_current_cred();
1067 if (!req
->work
.fs
&& def
->needs_fs
) {
1068 spin_lock(¤t
->fs
->lock
);
1069 if (!current
->fs
->in_exec
) {
1070 req
->work
.fs
= current
->fs
;
1071 req
->work
.fs
->users
++;
1073 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1075 spin_unlock(¤t
->fs
->lock
);
1079 static inline void io_req_work_drop_env(struct io_kiocb
*req
)
1081 if (!(req
->flags
& REQ_F_WORK_INITIALIZED
))
1085 mmdrop(req
->work
.mm
);
1086 req
->work
.mm
= NULL
;
1088 if (req
->work
.creds
) {
1089 put_cred(req
->work
.creds
);
1090 req
->work
.creds
= NULL
;
1093 struct fs_struct
*fs
= req
->work
.fs
;
1095 spin_lock(&req
->work
.fs
->lock
);
1098 spin_unlock(&req
->work
.fs
->lock
);
1104 static inline void io_prep_async_work(struct io_kiocb
*req
,
1105 struct io_kiocb
**link
)
1107 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1109 if (req
->flags
& REQ_F_ISREG
) {
1110 if (def
->hash_reg_file
)
1111 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1113 if (def
->unbound_nonreg_file
)
1114 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1117 io_req_init_async(req
);
1118 io_req_work_grab_env(req
, def
);
1120 *link
= io_prep_linked_timeout(req
);
1123 static inline void io_queue_async_work(struct io_kiocb
*req
)
1125 struct io_ring_ctx
*ctx
= req
->ctx
;
1126 struct io_kiocb
*link
;
1128 io_prep_async_work(req
, &link
);
1130 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1131 &req
->work
, req
->flags
);
1132 io_wq_enqueue(ctx
->io_wq
, &req
->work
);
1135 io_queue_linked_timeout(link
);
1138 static void io_kill_timeout(struct io_kiocb
*req
)
1142 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
1144 atomic_inc(&req
->ctx
->cq_timeouts
);
1145 list_del_init(&req
->list
);
1146 req
->flags
|= REQ_F_COMP_LOCKED
;
1147 io_cqring_fill_event(req
, 0);
1152 static void io_kill_timeouts(struct io_ring_ctx
*ctx
)
1154 struct io_kiocb
*req
, *tmp
;
1156 spin_lock_irq(&ctx
->completion_lock
);
1157 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, list
)
1158 io_kill_timeout(req
);
1159 spin_unlock_irq(&ctx
->completion_lock
);
1162 static void __io_queue_deferred(struct io_ring_ctx
*ctx
)
1165 struct io_kiocb
*req
= list_first_entry(&ctx
->defer_list
,
1166 struct io_kiocb
, list
);
1168 if (req_need_defer(req
))
1170 list_del_init(&req
->list
);
1171 io_queue_async_work(req
);
1172 } while (!list_empty(&ctx
->defer_list
));
1175 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1177 while (!list_empty(&ctx
->timeout_list
)) {
1178 struct io_kiocb
*req
= list_first_entry(&ctx
->timeout_list
,
1179 struct io_kiocb
, list
);
1181 if (req
->flags
& REQ_F_TIMEOUT_NOSEQ
)
1183 if (req
->timeout
.target_seq
!= ctx
->cached_cq_tail
1184 - atomic_read(&ctx
->cq_timeouts
))
1187 list_del_init(&req
->list
);
1188 io_kill_timeout(req
);
1192 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
1194 io_flush_timeouts(ctx
);
1195 __io_commit_cqring(ctx
);
1197 if (unlikely(!list_empty(&ctx
->defer_list
)))
1198 __io_queue_deferred(ctx
);
1201 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
1203 struct io_rings
*rings
= ctx
->rings
;
1206 tail
= ctx
->cached_cq_tail
;
1208 * writes to the cq entry need to come after reading head; the
1209 * control dependency is enough as we're using WRITE_ONCE to
1212 if (tail
- READ_ONCE(rings
->cq
.head
) == rings
->cq_ring_entries
)
1215 ctx
->cached_cq_tail
++;
1216 return &rings
->cqes
[tail
& ctx
->cq_mask
];
1219 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1223 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1225 if (!ctx
->eventfd_async
)
1227 return io_wq_current_is_worker();
1230 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1232 if (waitqueue_active(&ctx
->wait
))
1233 wake_up(&ctx
->wait
);
1234 if (waitqueue_active(&ctx
->sqo_wait
))
1235 wake_up(&ctx
->sqo_wait
);
1236 if (io_should_trigger_evfd(ctx
))
1237 eventfd_signal(ctx
->cq_ev_fd
, 1);
1240 /* Returns true if there are no backlogged entries after the flush */
1241 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1243 struct io_rings
*rings
= ctx
->rings
;
1244 struct io_uring_cqe
*cqe
;
1245 struct io_kiocb
*req
;
1246 unsigned long flags
;
1250 if (list_empty_careful(&ctx
->cq_overflow_list
))
1252 if ((ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
) ==
1253 rings
->cq_ring_entries
))
1257 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1259 /* if force is set, the ring is going away. always drop after that */
1261 ctx
->cq_overflow_flushed
= 1;
1264 while (!list_empty(&ctx
->cq_overflow_list
)) {
1265 cqe
= io_get_cqring(ctx
);
1269 req
= list_first_entry(&ctx
->cq_overflow_list
, struct io_kiocb
,
1271 list_move(&req
->list
, &list
);
1272 req
->flags
&= ~REQ_F_OVERFLOW
;
1274 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1275 WRITE_ONCE(cqe
->res
, req
->result
);
1276 WRITE_ONCE(cqe
->flags
, req
->cflags
);
1278 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1279 atomic_inc_return(&ctx
->cached_cq_overflow
));
1283 io_commit_cqring(ctx
);
1285 clear_bit(0, &ctx
->sq_check_overflow
);
1286 clear_bit(0, &ctx
->cq_check_overflow
);
1288 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1289 io_cqring_ev_posted(ctx
);
1291 while (!list_empty(&list
)) {
1292 req
= list_first_entry(&list
, struct io_kiocb
, list
);
1293 list_del(&req
->list
);
1300 static void __io_cqring_fill_event(struct io_kiocb
*req
, long res
, long cflags
)
1302 struct io_ring_ctx
*ctx
= req
->ctx
;
1303 struct io_uring_cqe
*cqe
;
1305 trace_io_uring_complete(ctx
, req
->user_data
, res
);
1308 * If we can't get a cq entry, userspace overflowed the
1309 * submission (by quite a lot). Increment the overflow count in
1312 cqe
= io_get_cqring(ctx
);
1314 WRITE_ONCE(cqe
->user_data
, req
->user_data
);
1315 WRITE_ONCE(cqe
->res
, res
);
1316 WRITE_ONCE(cqe
->flags
, cflags
);
1317 } else if (ctx
->cq_overflow_flushed
) {
1318 WRITE_ONCE(ctx
->rings
->cq_overflow
,
1319 atomic_inc_return(&ctx
->cached_cq_overflow
));
1321 if (list_empty(&ctx
->cq_overflow_list
)) {
1322 set_bit(0, &ctx
->sq_check_overflow
);
1323 set_bit(0, &ctx
->cq_check_overflow
);
1325 req
->flags
|= REQ_F_OVERFLOW
;
1326 refcount_inc(&req
->refs
);
1328 req
->cflags
= cflags
;
1329 list_add_tail(&req
->list
, &ctx
->cq_overflow_list
);
1333 static void io_cqring_fill_event(struct io_kiocb
*req
, long res
)
1335 __io_cqring_fill_event(req
, res
, 0);
1338 static void __io_cqring_add_event(struct io_kiocb
*req
, long res
, long cflags
)
1340 struct io_ring_ctx
*ctx
= req
->ctx
;
1341 unsigned long flags
;
1343 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1344 __io_cqring_fill_event(req
, res
, cflags
);
1345 io_commit_cqring(ctx
);
1346 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1348 io_cqring_ev_posted(ctx
);
1351 static void io_cqring_add_event(struct io_kiocb
*req
, long res
)
1353 __io_cqring_add_event(req
, res
, 0);
1356 static inline bool io_is_fallback_req(struct io_kiocb
*req
)
1358 return req
== (struct io_kiocb
*)
1359 ((unsigned long) req
->ctx
->fallback_req
& ~1UL);
1362 static struct io_kiocb
*io_get_fallback_req(struct io_ring_ctx
*ctx
)
1364 struct io_kiocb
*req
;
1366 req
= ctx
->fallback_req
;
1367 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx
->fallback_req
))
1373 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
,
1374 struct io_submit_state
*state
)
1376 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1377 struct io_kiocb
*req
;
1379 if (!state
->free_reqs
) {
1383 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
1384 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
1387 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1388 * retry single alloc to be on the safe side.
1390 if (unlikely(ret
<= 0)) {
1391 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1392 if (!state
->reqs
[0])
1396 state
->free_reqs
= ret
- 1;
1397 req
= state
->reqs
[ret
- 1];
1400 req
= state
->reqs
[state
->free_reqs
];
1405 return io_get_fallback_req(ctx
);
1408 static inline void io_put_file(struct io_kiocb
*req
, struct file
*file
,
1412 percpu_ref_put(req
->fixed_file_refs
);
1417 static void __io_req_aux_free(struct io_kiocb
*req
)
1419 if (req
->flags
& REQ_F_NEED_CLEANUP
)
1420 io_cleanup_req(req
);
1424 io_put_file(req
, req
->file
, (req
->flags
& REQ_F_FIXED_FILE
));
1425 __io_put_req_task(req
);
1426 io_req_work_drop_env(req
);
1429 static void __io_free_req(struct io_kiocb
*req
)
1431 __io_req_aux_free(req
);
1433 if (req
->flags
& REQ_F_INFLIGHT
) {
1434 struct io_ring_ctx
*ctx
= req
->ctx
;
1435 unsigned long flags
;
1437 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1438 list_del(&req
->inflight_entry
);
1439 if (waitqueue_active(&ctx
->inflight_wait
))
1440 wake_up(&ctx
->inflight_wait
);
1441 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1444 percpu_ref_put(&req
->ctx
->refs
);
1445 if (likely(!io_is_fallback_req(req
)))
1446 kmem_cache_free(req_cachep
, req
);
1448 clear_bit_unlock(0, (unsigned long *) &req
->ctx
->fallback_req
);
1452 void *reqs
[IO_IOPOLL_BATCH
];
1457 static void io_free_req_many(struct io_ring_ctx
*ctx
, struct req_batch
*rb
)
1461 if (rb
->need_iter
) {
1462 int i
, inflight
= 0;
1463 unsigned long flags
;
1465 for (i
= 0; i
< rb
->to_free
; i
++) {
1466 struct io_kiocb
*req
= rb
->reqs
[i
];
1468 if (req
->flags
& REQ_F_INFLIGHT
)
1470 __io_req_aux_free(req
);
1475 spin_lock_irqsave(&ctx
->inflight_lock
, flags
);
1476 for (i
= 0; i
< rb
->to_free
; i
++) {
1477 struct io_kiocb
*req
= rb
->reqs
[i
];
1479 if (req
->flags
& REQ_F_INFLIGHT
) {
1480 list_del(&req
->inflight_entry
);
1485 spin_unlock_irqrestore(&ctx
->inflight_lock
, flags
);
1487 if (waitqueue_active(&ctx
->inflight_wait
))
1488 wake_up(&ctx
->inflight_wait
);
1491 kmem_cache_free_bulk(req_cachep
, rb
->to_free
, rb
->reqs
);
1492 percpu_ref_put_many(&ctx
->refs
, rb
->to_free
);
1493 rb
->to_free
= rb
->need_iter
= 0;
1496 static bool io_link_cancel_timeout(struct io_kiocb
*req
)
1498 struct io_ring_ctx
*ctx
= req
->ctx
;
1501 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
1503 io_cqring_fill_event(req
, -ECANCELED
);
1504 io_commit_cqring(ctx
);
1505 req
->flags
&= ~REQ_F_LINK_HEAD
;
1513 static void io_req_link_next(struct io_kiocb
*req
, struct io_kiocb
**nxtptr
)
1515 struct io_ring_ctx
*ctx
= req
->ctx
;
1516 bool wake_ev
= false;
1518 /* Already got next link */
1519 if (req
->flags
& REQ_F_LINK_NEXT
)
1523 * The list should never be empty when we are called here. But could
1524 * potentially happen if the chain is messed up, check to be on the
1527 while (!list_empty(&req
->link_list
)) {
1528 struct io_kiocb
*nxt
= list_first_entry(&req
->link_list
,
1529 struct io_kiocb
, link_list
);
1531 if (unlikely((req
->flags
& REQ_F_LINK_TIMEOUT
) &&
1532 (nxt
->flags
& REQ_F_TIMEOUT
))) {
1533 list_del_init(&nxt
->link_list
);
1534 wake_ev
|= io_link_cancel_timeout(nxt
);
1535 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1539 list_del_init(&req
->link_list
);
1540 if (!list_empty(&nxt
->link_list
))
1541 nxt
->flags
|= REQ_F_LINK_HEAD
;
1546 req
->flags
|= REQ_F_LINK_NEXT
;
1548 io_cqring_ev_posted(ctx
);
1552 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1554 static void io_fail_links(struct io_kiocb
*req
)
1556 struct io_ring_ctx
*ctx
= req
->ctx
;
1557 unsigned long flags
;
1559 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1561 while (!list_empty(&req
->link_list
)) {
1562 struct io_kiocb
*link
= list_first_entry(&req
->link_list
,
1563 struct io_kiocb
, link_list
);
1565 list_del_init(&link
->link_list
);
1566 trace_io_uring_fail_link(req
, link
);
1568 if ((req
->flags
& REQ_F_LINK_TIMEOUT
) &&
1569 link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
1570 io_link_cancel_timeout(link
);
1572 io_cqring_fill_event(link
, -ECANCELED
);
1573 __io_double_put_req(link
);
1575 req
->flags
&= ~REQ_F_LINK_TIMEOUT
;
1578 io_commit_cqring(ctx
);
1579 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1580 io_cqring_ev_posted(ctx
);
1583 static void io_req_find_next(struct io_kiocb
*req
, struct io_kiocb
**nxt
)
1585 if (likely(!(req
->flags
& REQ_F_LINK_HEAD
)))
1589 * If LINK is set, we have dependent requests in this chain. If we
1590 * didn't fail this request, queue the first one up, moving any other
1591 * dependencies to the next request. In case of failure, fail the rest
1594 if (req
->flags
& REQ_F_FAIL_LINK
) {
1596 } else if ((req
->flags
& (REQ_F_LINK_TIMEOUT
| REQ_F_COMP_LOCKED
)) ==
1597 REQ_F_LINK_TIMEOUT
) {
1598 struct io_ring_ctx
*ctx
= req
->ctx
;
1599 unsigned long flags
;
1602 * If this is a timeout link, we could be racing with the
1603 * timeout timer. Grab the completion lock for this case to
1604 * protect against that.
1606 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1607 io_req_link_next(req
, nxt
);
1608 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1610 io_req_link_next(req
, nxt
);
1614 static void io_free_req(struct io_kiocb
*req
)
1616 struct io_kiocb
*nxt
= NULL
;
1618 io_req_find_next(req
, &nxt
);
1622 io_queue_async_work(nxt
);
1625 static void io_wq_assign_next(struct io_wq_work
**workptr
, struct io_kiocb
*nxt
)
1627 struct io_kiocb
*link
;
1628 const struct io_op_def
*def
= &io_op_defs
[nxt
->opcode
];
1630 if ((nxt
->flags
& REQ_F_ISREG
) && def
->hash_reg_file
)
1631 io_wq_hash_work(&nxt
->work
, file_inode(nxt
->file
));
1633 *workptr
= &nxt
->work
;
1634 link
= io_prep_linked_timeout(nxt
);
1636 nxt
->flags
|= REQ_F_QUEUE_TIMEOUT
;
1640 * Drop reference to request, return next in chain (if there is one) if this
1641 * was the last reference to this request.
1643 __attribute__((nonnull
))
1644 static void io_put_req_find_next(struct io_kiocb
*req
, struct io_kiocb
**nxtptr
)
1646 if (refcount_dec_and_test(&req
->refs
)) {
1647 io_req_find_next(req
, nxtptr
);
1652 static void io_put_req(struct io_kiocb
*req
)
1654 if (refcount_dec_and_test(&req
->refs
))
1658 static void io_steal_work(struct io_kiocb
*req
,
1659 struct io_wq_work
**workptr
)
1662 * It's in an io-wq worker, so there always should be at least
1663 * one reference, which will be dropped in io_put_work() just
1664 * after the current handler returns.
1666 * It also means, that if the counter dropped to 1, then there is
1667 * no asynchronous users left, so it's safe to steal the next work.
1669 if (refcount_read(&req
->refs
) == 1) {
1670 struct io_kiocb
*nxt
= NULL
;
1672 io_req_find_next(req
, &nxt
);
1674 io_wq_assign_next(workptr
, nxt
);
1679 * Must only be used if we don't need to care about links, usually from
1680 * within the completion handling itself.
1682 static void __io_double_put_req(struct io_kiocb
*req
)
1684 /* drop both submit and complete references */
1685 if (refcount_sub_and_test(2, &req
->refs
))
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 unsigned io_cqring_events(struct io_ring_ctx
*ctx
, bool noflush
)
1698 struct io_rings
*rings
= ctx
->rings
;
1700 if (test_bit(0, &ctx
->cq_check_overflow
)) {
1702 * noflush == true is from the waitqueue handler, just ensure
1703 * we wake up the task, and the next invocation will flush the
1704 * entries. We cannot safely to it from here.
1706 if (noflush
&& !list_empty(&ctx
->cq_overflow_list
))
1709 io_cqring_overflow_flush(ctx
, false);
1712 /* See comment at the top of this file */
1714 return ctx
->cached_cq_tail
- READ_ONCE(rings
->cq
.head
);
1717 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
1719 struct io_rings
*rings
= ctx
->rings
;
1721 /* make sure SQ entry isn't read before tail */
1722 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
1725 static inline bool io_req_multi_free(struct req_batch
*rb
, struct io_kiocb
*req
)
1727 if ((req
->flags
& REQ_F_LINK_HEAD
) || io_is_fallback_req(req
))
1730 if (req
->file
|| req
->io
)
1733 rb
->reqs
[rb
->to_free
++] = req
;
1734 if (unlikely(rb
->to_free
== ARRAY_SIZE(rb
->reqs
)))
1735 io_free_req_many(req
->ctx
, rb
);
1739 static int io_put_kbuf(struct io_kiocb
*req
)
1741 struct io_buffer
*kbuf
;
1744 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
1745 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
1746 cflags
|= IORING_CQE_F_BUFFER
;
1752 static void io_iopoll_queue(struct list_head
*again
)
1754 struct io_kiocb
*req
;
1757 req
= list_first_entry(again
, struct io_kiocb
, list
);
1758 list_del(&req
->list
);
1759 refcount_inc(&req
->refs
);
1760 io_queue_async_work(req
);
1761 } while (!list_empty(again
));
1765 * Find and free completed poll iocbs
1767 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1768 struct list_head
*done
)
1770 struct req_batch rb
;
1771 struct io_kiocb
*req
;
1774 /* order with ->result store in io_complete_rw_iopoll() */
1777 rb
.to_free
= rb
.need_iter
= 0;
1778 while (!list_empty(done
)) {
1781 req
= list_first_entry(done
, struct io_kiocb
, list
);
1782 if (READ_ONCE(req
->result
) == -EAGAIN
) {
1783 req
->iopoll_completed
= 0;
1784 list_move_tail(&req
->list
, &again
);
1787 list_del(&req
->list
);
1789 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
1790 cflags
= io_put_kbuf(req
);
1792 __io_cqring_fill_event(req
, req
->result
, cflags
);
1795 if (refcount_dec_and_test(&req
->refs
) &&
1796 !io_req_multi_free(&rb
, req
))
1800 io_commit_cqring(ctx
);
1801 if (ctx
->flags
& IORING_SETUP_SQPOLL
)
1802 io_cqring_ev_posted(ctx
);
1803 io_free_req_many(ctx
, &rb
);
1805 if (!list_empty(&again
))
1806 io_iopoll_queue(&again
);
1809 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1812 struct io_kiocb
*req
, *tmp
;
1818 * Only spin for completions if we don't have multiple devices hanging
1819 * off our complete list, and we're under the requested amount.
1821 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
1824 list_for_each_entry_safe(req
, tmp
, &ctx
->poll_list
, list
) {
1825 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
1828 * Move completed and retryable entries to our local lists.
1829 * If we find a request that requires polling, break out
1830 * and complete those lists first, if we have entries there.
1832 if (READ_ONCE(req
->iopoll_completed
)) {
1833 list_move_tail(&req
->list
, &done
);
1836 if (!list_empty(&done
))
1839 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
1848 if (!list_empty(&done
))
1849 io_iopoll_complete(ctx
, nr_events
, &done
);
1855 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1856 * non-spinning poll check - we'll still enter the driver poll loop, but only
1857 * as a non-spinning completion check.
1859 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
1862 while (!list_empty(&ctx
->poll_list
) && !need_resched()) {
1865 ret
= io_do_iopoll(ctx
, nr_events
, min
);
1868 if (!min
|| *nr_events
>= min
)
1876 * We can't just wait for polled events to come to us, we have to actively
1877 * find and complete them.
1879 static void io_iopoll_reap_events(struct io_ring_ctx
*ctx
)
1881 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1884 mutex_lock(&ctx
->uring_lock
);
1885 while (!list_empty(&ctx
->poll_list
)) {
1886 unsigned int nr_events
= 0;
1888 io_iopoll_getevents(ctx
, &nr_events
, 1);
1891 * Ensure we allow local-to-the-cpu processing to take place,
1892 * in this case we need to ensure that we reap all events.
1896 mutex_unlock(&ctx
->uring_lock
);
1899 static int io_iopoll_check(struct io_ring_ctx
*ctx
, unsigned *nr_events
,
1902 int iters
= 0, ret
= 0;
1905 * We disallow the app entering submit/complete with polling, but we
1906 * still need to lock the ring to prevent racing with polled issue
1907 * that got punted to a workqueue.
1909 mutex_lock(&ctx
->uring_lock
);
1914 * Don't enter poll loop if we already have events pending.
1915 * If we do, we can potentially be spinning for commands that
1916 * already triggered a CQE (eg in error).
1918 if (io_cqring_events(ctx
, false))
1922 * If a submit got punted to a workqueue, we can have the
1923 * application entering polling for a command before it gets
1924 * issued. That app will hold the uring_lock for the duration
1925 * of the poll right here, so we need to take a breather every
1926 * now and then to ensure that the issue has a chance to add
1927 * the poll to the issued list. Otherwise we can spin here
1928 * forever, while the workqueue is stuck trying to acquire the
1931 if (!(++iters
& 7)) {
1932 mutex_unlock(&ctx
->uring_lock
);
1933 mutex_lock(&ctx
->uring_lock
);
1936 if (*nr_events
< min
)
1937 tmin
= min
- *nr_events
;
1939 ret
= io_iopoll_getevents(ctx
, nr_events
, tmin
);
1943 } while (min
&& !*nr_events
&& !need_resched());
1945 mutex_unlock(&ctx
->uring_lock
);
1949 static void kiocb_end_write(struct io_kiocb
*req
)
1952 * Tell lockdep we inherited freeze protection from submission
1955 if (req
->flags
& REQ_F_ISREG
) {
1956 struct inode
*inode
= file_inode(req
->file
);
1958 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
1960 file_end_write(req
->file
);
1963 static inline void req_set_fail_links(struct io_kiocb
*req
)
1965 if ((req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) == REQ_F_LINK
)
1966 req
->flags
|= REQ_F_FAIL_LINK
;
1969 static void io_complete_rw_common(struct kiocb
*kiocb
, long res
)
1971 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
1974 if (kiocb
->ki_flags
& IOCB_WRITE
)
1975 kiocb_end_write(req
);
1977 if (res
!= req
->result
)
1978 req_set_fail_links(req
);
1979 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
1980 cflags
= io_put_kbuf(req
);
1981 __io_cqring_add_event(req
, res
, cflags
);
1984 static void io_sq_thread_drop_mm(struct io_ring_ctx
*ctx
)
1986 struct mm_struct
*mm
= current
->mm
;
1989 kthread_unuse_mm(mm
);
1994 static int __io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
)
1997 if (unlikely(!mmget_not_zero(ctx
->sqo_mm
)))
1999 kthread_use_mm(ctx
->sqo_mm
);
2005 static int io_sq_thread_acquire_mm(struct io_ring_ctx
*ctx
,
2006 struct io_kiocb
*req
)
2008 if (!io_op_defs
[req
->opcode
].needs_mm
)
2010 return __io_sq_thread_acquire_mm(ctx
);
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
)) {
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
;
2897 /* If the file doesn't support async, just async punt */
2898 if (force_nonblock
&& !io_file_supports_async(req
->file
, READ
))
2901 iov_count
= iov_iter_count(&iter
);
2902 ret
= rw_verify_area(READ
, req
->file
, &kiocb
->ki_pos
, iov_count
);
2904 unsigned long nr_segs
= iter
.nr_segs
;
2907 ret2
= io_iter_do_read(req
, &iter
);
2909 /* Catch -EAGAIN return for forced non-blocking submission */
2910 if (!force_nonblock
|| (ret2
!= -EAGAIN
&& ret2
!= -EIO
)) {
2911 kiocb_done(kiocb
, ret2
);
2913 iter
.count
= iov_count
;
2914 iter
.nr_segs
= nr_segs
;
2916 ret
= io_setup_async_rw(req
, io_size
, iovec
,
2917 inline_vecs
, &iter
);
2920 /* if we can retry, do so with the callbacks armed */
2921 if (io_rw_should_retry(req
)) {
2922 ret2
= io_iter_do_read(req
, &iter
);
2923 if (ret2
== -EIOCBQUEUED
) {
2925 } else if (ret2
!= -EAGAIN
) {
2926 kiocb_done(kiocb
, ret2
);
2930 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
2935 if (!(req
->flags
& REQ_F_NEED_CLEANUP
))
2940 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2941 bool force_nonblock
)
2943 struct io_async_ctx
*io
;
2944 struct iov_iter iter
;
2947 ret
= io_prep_rw(req
, sqe
, force_nonblock
);
2951 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
2954 req
->fsize
= rlimit(RLIMIT_FSIZE
);
2956 /* either don't need iovec imported or already have it */
2957 if (!req
->io
|| req
->flags
& REQ_F_NEED_CLEANUP
)
2961 io
->rw
.iov
= io
->rw
.fast_iov
;
2963 ret
= io_import_iovec(WRITE
, req
, &io
->rw
.iov
, &iter
, !force_nonblock
);
2968 io_req_map_rw(req
, ret
, io
->rw
.iov
, io
->rw
.fast_iov
, &iter
);
2972 static int io_write(struct io_kiocb
*req
, bool force_nonblock
)
2974 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
2975 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2976 struct iov_iter iter
;
2978 ssize_t ret
, io_size
;
2980 ret
= io_import_iovec(WRITE
, req
, &iovec
, &iter
, !force_nonblock
);
2984 /* Ensure we clear previously set non-block flag */
2985 if (!force_nonblock
)
2986 req
->rw
.kiocb
.ki_flags
&= ~IOCB_NOWAIT
;
2990 if (req
->flags
& REQ_F_LINK_HEAD
)
2991 req
->result
= io_size
;
2993 /* If the file doesn't support async, just async punt */
2994 if (force_nonblock
&& !io_file_supports_async(req
->file
, WRITE
))
2997 /* file path doesn't support NOWAIT for non-direct_IO */
2998 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
2999 (req
->flags
& REQ_F_ISREG
))
3002 iov_count
= iov_iter_count(&iter
);
3003 ret
= rw_verify_area(WRITE
, req
->file
, &kiocb
->ki_pos
, iov_count
);
3005 unsigned long nr_segs
= iter
.nr_segs
;
3009 * Open-code file_start_write here to grab freeze protection,
3010 * which will be released by another thread in
3011 * io_complete_rw(). Fool lockdep by telling it the lock got
3012 * released so that it doesn't complain about the held lock when
3013 * we return to userspace.
3015 if (req
->flags
& REQ_F_ISREG
) {
3016 __sb_start_write(file_inode(req
->file
)->i_sb
,
3017 SB_FREEZE_WRITE
, true);
3018 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3021 kiocb
->ki_flags
|= IOCB_WRITE
;
3023 if (!force_nonblock
)
3024 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= req
->fsize
;
3026 if (req
->file
->f_op
->write_iter
)
3027 ret2
= call_write_iter(req
->file
, kiocb
, &iter
);
3029 ret2
= loop_rw_iter(WRITE
, req
->file
, kiocb
, &iter
);
3031 if (!force_nonblock
)
3032 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= RLIM_INFINITY
;
3035 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3036 * retry them without IOCB_NOWAIT.
3038 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3040 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3041 kiocb_done(kiocb
, ret2
);
3043 iter
.count
= iov_count
;
3044 iter
.nr_segs
= nr_segs
;
3046 ret
= io_setup_async_rw(req
, io_size
, iovec
,
3047 inline_vecs
, &iter
);
3054 if (!(req
->flags
& REQ_F_NEED_CLEANUP
))
3059 static int __io_splice_prep(struct io_kiocb
*req
,
3060 const struct io_uring_sqe
*sqe
)
3062 struct io_splice
* sp
= &req
->splice
;
3063 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
3066 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3068 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3072 sp
->len
= READ_ONCE(sqe
->len
);
3073 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
3075 if (unlikely(sp
->flags
& ~valid_flags
))
3078 ret
= io_file_get(NULL
, req
, READ_ONCE(sqe
->splice_fd_in
), &sp
->file_in
,
3079 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3082 req
->flags
|= REQ_F_NEED_CLEANUP
;
3084 if (!S_ISREG(file_inode(sp
->file_in
)->i_mode
)) {
3086 * Splice operation will be punted aync, and here need to
3087 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3089 io_req_init_async(req
);
3090 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
3096 static int io_tee_prep(struct io_kiocb
*req
,
3097 const struct io_uring_sqe
*sqe
)
3099 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
3101 return __io_splice_prep(req
, sqe
);
3104 static int io_tee(struct io_kiocb
*req
, bool force_nonblock
)
3106 struct io_splice
*sp
= &req
->splice
;
3107 struct file
*in
= sp
->file_in
;
3108 struct file
*out
= sp
->file_out
;
3109 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3115 ret
= do_tee(in
, out
, sp
->len
, flags
);
3117 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3118 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3120 io_cqring_add_event(req
, ret
);
3122 req_set_fail_links(req
);
3127 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3129 struct io_splice
* sp
= &req
->splice
;
3131 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
3132 sp
->off_out
= READ_ONCE(sqe
->off
);
3133 return __io_splice_prep(req
, sqe
);
3136 static int io_splice(struct io_kiocb
*req
, bool force_nonblock
)
3138 struct io_splice
*sp
= &req
->splice
;
3139 struct file
*in
= sp
->file_in
;
3140 struct file
*out
= sp
->file_out
;
3141 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
3142 loff_t
*poff_in
, *poff_out
;
3148 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
3149 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
3152 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
3154 io_put_file(req
, in
, (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
3155 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3157 io_cqring_add_event(req
, ret
);
3159 req_set_fail_links(req
);
3165 * IORING_OP_NOP just posts a completion event, nothing else.
3167 static int io_nop(struct io_kiocb
*req
)
3169 struct io_ring_ctx
*ctx
= req
->ctx
;
3171 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3174 io_cqring_add_event(req
, 0);
3179 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3181 struct io_ring_ctx
*ctx
= req
->ctx
;
3186 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3188 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3191 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
3192 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
3195 req
->sync
.off
= READ_ONCE(sqe
->off
);
3196 req
->sync
.len
= READ_ONCE(sqe
->len
);
3200 static int io_fsync(struct io_kiocb
*req
, bool force_nonblock
)
3202 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
3205 /* fsync always requires a blocking context */
3209 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
3210 end
> 0 ? end
: LLONG_MAX
,
3211 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
3213 req_set_fail_links(req
);
3214 io_cqring_add_event(req
, ret
);
3219 static int io_fallocate_prep(struct io_kiocb
*req
,
3220 const struct io_uring_sqe
*sqe
)
3222 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
)
3224 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3227 req
->sync
.off
= READ_ONCE(sqe
->off
);
3228 req
->sync
.len
= READ_ONCE(sqe
->addr
);
3229 req
->sync
.mode
= READ_ONCE(sqe
->len
);
3230 req
->fsize
= rlimit(RLIMIT_FSIZE
);
3234 static int io_fallocate(struct io_kiocb
*req
, bool force_nonblock
)
3238 /* fallocate always requiring blocking context */
3242 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= req
->fsize
;
3243 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
3245 current
->signal
->rlim
[RLIMIT_FSIZE
].rlim_cur
= RLIM_INFINITY
;
3247 req_set_fail_links(req
);
3248 io_cqring_add_event(req
, ret
);
3253 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3255 const char __user
*fname
;
3258 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3260 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
3262 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3265 /* open.how should be already initialised */
3266 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
3267 req
->open
.how
.flags
|= O_LARGEFILE
;
3269 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
3270 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3271 req
->open
.filename
= getname(fname
);
3272 if (IS_ERR(req
->open
.filename
)) {
3273 ret
= PTR_ERR(req
->open
.filename
);
3274 req
->open
.filename
= NULL
;
3277 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
3278 req
->flags
|= REQ_F_NEED_CLEANUP
;
3282 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3286 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3288 mode
= READ_ONCE(sqe
->len
);
3289 flags
= READ_ONCE(sqe
->open_flags
);
3290 req
->open
.how
= build_open_how(flags
, mode
);
3291 return __io_openat_prep(req
, sqe
);
3294 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3296 struct open_how __user
*how
;
3300 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3302 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3303 len
= READ_ONCE(sqe
->len
);
3304 if (len
< OPEN_HOW_SIZE_VER0
)
3307 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
3312 return __io_openat_prep(req
, sqe
);
3315 static int io_openat2(struct io_kiocb
*req
, bool force_nonblock
)
3317 struct open_flags op
;
3324 ret
= build_open_flags(&req
->open
.how
, &op
);
3328 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
3332 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
3335 ret
= PTR_ERR(file
);
3337 fsnotify_open(file
);
3338 fd_install(ret
, file
);
3341 putname(req
->open
.filename
);
3342 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3344 req_set_fail_links(req
);
3345 io_cqring_add_event(req
, ret
);
3350 static int io_openat(struct io_kiocb
*req
, bool force_nonblock
)
3352 return io_openat2(req
, force_nonblock
);
3355 static int io_remove_buffers_prep(struct io_kiocb
*req
,
3356 const struct io_uring_sqe
*sqe
)
3358 struct io_provide_buf
*p
= &req
->pbuf
;
3361 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
)
3364 tmp
= READ_ONCE(sqe
->fd
);
3365 if (!tmp
|| tmp
> USHRT_MAX
)
3368 memset(p
, 0, sizeof(*p
));
3370 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3374 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
3375 int bgid
, unsigned nbufs
)
3379 /* shouldn't happen */
3383 /* the head kbuf is the list itself */
3384 while (!list_empty(&buf
->list
)) {
3385 struct io_buffer
*nxt
;
3387 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
3388 list_del(&nxt
->list
);
3395 idr_remove(&ctx
->io_buffer_idr
, bgid
);
3400 static int io_remove_buffers(struct io_kiocb
*req
, bool force_nonblock
)
3402 struct io_provide_buf
*p
= &req
->pbuf
;
3403 struct io_ring_ctx
*ctx
= req
->ctx
;
3404 struct io_buffer
*head
;
3407 io_ring_submit_lock(ctx
, !force_nonblock
);
3409 lockdep_assert_held(&ctx
->uring_lock
);
3412 head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3414 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
3416 io_ring_submit_lock(ctx
, !force_nonblock
);
3418 req_set_fail_links(req
);
3419 io_cqring_add_event(req
, ret
);
3424 static int io_provide_buffers_prep(struct io_kiocb
*req
,
3425 const struct io_uring_sqe
*sqe
)
3427 struct io_provide_buf
*p
= &req
->pbuf
;
3430 if (sqe
->ioprio
|| sqe
->rw_flags
)
3433 tmp
= READ_ONCE(sqe
->fd
);
3434 if (!tmp
|| tmp
> USHRT_MAX
)
3437 p
->addr
= READ_ONCE(sqe
->addr
);
3438 p
->len
= READ_ONCE(sqe
->len
);
3440 if (!access_ok(u64_to_user_ptr(p
->addr
), (p
->len
* p
->nbufs
)))
3443 p
->bgid
= READ_ONCE(sqe
->buf_group
);
3444 tmp
= READ_ONCE(sqe
->off
);
3445 if (tmp
> USHRT_MAX
)
3451 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
3453 struct io_buffer
*buf
;
3454 u64 addr
= pbuf
->addr
;
3455 int i
, bid
= pbuf
->bid
;
3457 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
3458 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL
);
3463 buf
->len
= pbuf
->len
;
3468 INIT_LIST_HEAD(&buf
->list
);
3471 list_add_tail(&buf
->list
, &(*head
)->list
);
3475 return i
? i
: -ENOMEM
;
3478 static int io_provide_buffers(struct io_kiocb
*req
, bool force_nonblock
)
3480 struct io_provide_buf
*p
= &req
->pbuf
;
3481 struct io_ring_ctx
*ctx
= req
->ctx
;
3482 struct io_buffer
*head
, *list
;
3485 io_ring_submit_lock(ctx
, !force_nonblock
);
3487 lockdep_assert_held(&ctx
->uring_lock
);
3489 list
= head
= idr_find(&ctx
->io_buffer_idr
, p
->bgid
);
3491 ret
= io_add_buffers(p
, &head
);
3496 ret
= idr_alloc(&ctx
->io_buffer_idr
, head
, p
->bgid
, p
->bgid
+ 1,
3499 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
3504 io_ring_submit_unlock(ctx
, !force_nonblock
);
3506 req_set_fail_links(req
);
3507 io_cqring_add_event(req
, ret
);
3512 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
3513 const struct io_uring_sqe
*sqe
)
3515 #if defined(CONFIG_EPOLL)
3516 if (sqe
->ioprio
|| sqe
->buf_index
)
3518 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3521 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
3522 req
->epoll
.op
= READ_ONCE(sqe
->len
);
3523 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
3525 if (ep_op_has_event(req
->epoll
.op
)) {
3526 struct epoll_event __user
*ev
;
3528 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3529 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
3539 static int io_epoll_ctl(struct io_kiocb
*req
, bool force_nonblock
)
3541 #if defined(CONFIG_EPOLL)
3542 struct io_epoll
*ie
= &req
->epoll
;
3545 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
3546 if (force_nonblock
&& ret
== -EAGAIN
)
3550 req_set_fail_links(req
);
3551 io_cqring_add_event(req
, ret
);
3559 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3561 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3562 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
)
3564 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3567 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
3568 req
->madvise
.len
= READ_ONCE(sqe
->len
);
3569 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
3576 static int io_madvise(struct io_kiocb
*req
, bool force_nonblock
)
3578 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3579 struct io_madvise
*ma
= &req
->madvise
;
3585 ret
= do_madvise(ma
->addr
, ma
->len
, ma
->advice
);
3587 req_set_fail_links(req
);
3588 io_cqring_add_event(req
, ret
);
3596 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3598 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
)
3600 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3603 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
3604 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
3605 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
3609 static int io_fadvise(struct io_kiocb
*req
, bool force_nonblock
)
3611 struct io_fadvise
*fa
= &req
->fadvise
;
3614 if (force_nonblock
) {
3615 switch (fa
->advice
) {
3616 case POSIX_FADV_NORMAL
:
3617 case POSIX_FADV_RANDOM
:
3618 case POSIX_FADV_SEQUENTIAL
:
3625 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
3627 req_set_fail_links(req
);
3628 io_cqring_add_event(req
, ret
);
3633 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3635 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3637 if (sqe
->ioprio
|| sqe
->buf_index
)
3639 if (req
->flags
& REQ_F_FIXED_FILE
)
3642 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
3643 req
->statx
.mask
= READ_ONCE(sqe
->len
);
3644 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3645 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3646 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
3651 static int io_statx(struct io_kiocb
*req
, bool force_nonblock
)
3653 struct io_statx
*ctx
= &req
->statx
;
3656 if (force_nonblock
) {
3657 /* only need file table for an actual valid fd */
3658 if (ctx
->dfd
== -1 || ctx
->dfd
== AT_FDCWD
)
3659 req
->flags
|= REQ_F_NO_FILE_TABLE
;
3663 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
3667 req_set_fail_links(req
);
3668 io_cqring_add_event(req
, ret
);
3673 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3676 * If we queue this for async, it must not be cancellable. That would
3677 * leave the 'file' in an undeterminate state, and here need to modify
3678 * io_wq_work.flags, so initialize io_wq_work firstly.
3680 io_req_init_async(req
);
3681 req
->work
.flags
|= IO_WQ_WORK_NO_CANCEL
;
3683 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
3685 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
3686 sqe
->rw_flags
|| sqe
->buf_index
)
3688 if (req
->flags
& REQ_F_FIXED_FILE
)
3691 req
->close
.fd
= READ_ONCE(sqe
->fd
);
3692 if ((req
->file
&& req
->file
->f_op
== &io_uring_fops
) ||
3693 req
->close
.fd
== req
->ctx
->ring_fd
)
3696 req
->close
.put_file
= NULL
;
3700 static int io_close(struct io_kiocb
*req
, bool force_nonblock
)
3702 struct io_close
*close
= &req
->close
;
3705 /* might be already done during nonblock submission */
3706 if (!close
->put_file
) {
3707 ret
= __close_fd_get_file(close
->fd
, &close
->put_file
);
3709 return (ret
== -ENOENT
) ? -EBADF
: ret
;
3712 /* if the file has a flush method, be safe and punt to async */
3713 if (close
->put_file
->f_op
->flush
&& force_nonblock
) {
3714 /* was never set, but play safe */
3715 req
->flags
&= ~REQ_F_NOWAIT
;
3716 /* avoid grabbing files - we don't need the files */
3717 req
->flags
|= REQ_F_NO_FILE_TABLE
;
3721 /* No ->flush() or already async, safely close from here */
3722 ret
= filp_close(close
->put_file
, req
->work
.files
);
3724 req_set_fail_links(req
);
3725 io_cqring_add_event(req
, ret
);
3726 fput(close
->put_file
);
3727 close
->put_file
= NULL
;
3732 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3734 struct io_ring_ctx
*ctx
= req
->ctx
;
3739 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
3741 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
3744 req
->sync
.off
= READ_ONCE(sqe
->off
);
3745 req
->sync
.len
= READ_ONCE(sqe
->len
);
3746 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
3750 static int io_sync_file_range(struct io_kiocb
*req
, bool force_nonblock
)
3754 /* sync_file_range always requires a blocking context */
3758 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
3761 req_set_fail_links(req
);
3762 io_cqring_add_event(req
, ret
);
3767 #if defined(CONFIG_NET)
3768 static int io_setup_async_msg(struct io_kiocb
*req
,
3769 struct io_async_msghdr
*kmsg
)
3773 if (io_alloc_async_ctx(req
)) {
3774 if (kmsg
->iov
!= kmsg
->fast_iov
)
3778 req
->flags
|= REQ_F_NEED_CLEANUP
;
3779 memcpy(&req
->io
->msg
, kmsg
, sizeof(*kmsg
));
3783 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3785 struct io_sr_msg
*sr
= &req
->sr_msg
;
3786 struct io_async_ctx
*io
= req
->io
;
3789 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3792 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
3793 sr
->msg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3794 sr
->len
= READ_ONCE(sqe
->len
);
3796 #ifdef CONFIG_COMPAT
3797 if (req
->ctx
->compat
)
3798 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
3801 if (!io
|| req
->opcode
== IORING_OP_SEND
)
3803 /* iovec is already imported */
3804 if (req
->flags
& REQ_F_NEED_CLEANUP
)
3807 io
->msg
.iov
= io
->msg
.fast_iov
;
3808 ret
= sendmsg_copy_msghdr(&io
->msg
.msg
, sr
->msg
, sr
->msg_flags
,
3811 req
->flags
|= REQ_F_NEED_CLEANUP
;
3815 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
)
3817 struct io_async_msghdr
*kmsg
= NULL
;
3818 struct socket
*sock
;
3821 sock
= sock_from_file(req
->file
, &ret
);
3823 struct io_async_ctx io
;
3827 kmsg
= &req
->io
->msg
;
3828 kmsg
->msg
.msg_name
= &req
->io
->msg
.addr
;
3829 /* if iov is set, it's allocated already */
3831 kmsg
->iov
= kmsg
->fast_iov
;
3832 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
3834 struct io_sr_msg
*sr
= &req
->sr_msg
;
3837 kmsg
->msg
.msg_name
= &io
.msg
.addr
;
3839 io
.msg
.iov
= io
.msg
.fast_iov
;
3840 ret
= sendmsg_copy_msghdr(&io
.msg
.msg
, sr
->msg
,
3841 sr
->msg_flags
, &io
.msg
.iov
);
3846 flags
= req
->sr_msg
.msg_flags
;
3847 if (flags
& MSG_DONTWAIT
)
3848 req
->flags
|= REQ_F_NOWAIT
;
3849 else if (force_nonblock
)
3850 flags
|= MSG_DONTWAIT
;
3852 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
3853 if (force_nonblock
&& ret
== -EAGAIN
)
3854 return io_setup_async_msg(req
, kmsg
);
3855 if (ret
== -ERESTARTSYS
)
3859 if (kmsg
&& kmsg
->iov
!= kmsg
->fast_iov
)
3861 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3862 io_cqring_add_event(req
, ret
);
3864 req_set_fail_links(req
);
3869 static int io_send(struct io_kiocb
*req
, bool force_nonblock
)
3871 struct socket
*sock
;
3874 sock
= sock_from_file(req
->file
, &ret
);
3876 struct io_sr_msg
*sr
= &req
->sr_msg
;
3881 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
,
3886 msg
.msg_name
= NULL
;
3887 msg
.msg_control
= NULL
;
3888 msg
.msg_controllen
= 0;
3889 msg
.msg_namelen
= 0;
3891 flags
= req
->sr_msg
.msg_flags
;
3892 if (flags
& MSG_DONTWAIT
)
3893 req
->flags
|= REQ_F_NOWAIT
;
3894 else if (force_nonblock
)
3895 flags
|= MSG_DONTWAIT
;
3897 msg
.msg_flags
= flags
;
3898 ret
= sock_sendmsg(sock
, &msg
);
3899 if (force_nonblock
&& ret
== -EAGAIN
)
3901 if (ret
== -ERESTARTSYS
)
3905 io_cqring_add_event(req
, ret
);
3907 req_set_fail_links(req
);
3912 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
, struct io_async_ctx
*io
)
3914 struct io_sr_msg
*sr
= &req
->sr_msg
;
3915 struct iovec __user
*uiov
;
3919 ret
= __copy_msghdr_from_user(&io
->msg
.msg
, sr
->msg
, &io
->msg
.uaddr
,
3924 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3927 if (copy_from_user(io
->msg
.iov
, uiov
, sizeof(*uiov
)))
3929 sr
->len
= io
->msg
.iov
[0].iov_len
;
3930 iov_iter_init(&io
->msg
.msg
.msg_iter
, READ
, io
->msg
.iov
, 1,
3934 ret
= import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
3935 &io
->msg
.iov
, &io
->msg
.msg
.msg_iter
);
3943 #ifdef CONFIG_COMPAT
3944 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
3945 struct io_async_ctx
*io
)
3947 struct compat_msghdr __user
*msg_compat
;
3948 struct io_sr_msg
*sr
= &req
->sr_msg
;
3949 struct compat_iovec __user
*uiov
;
3954 msg_compat
= (struct compat_msghdr __user
*) sr
->msg
;
3955 ret
= __get_compat_msghdr(&io
->msg
.msg
, msg_compat
, &io
->msg
.uaddr
,
3960 uiov
= compat_ptr(ptr
);
3961 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3962 compat_ssize_t clen
;
3966 if (!access_ok(uiov
, sizeof(*uiov
)))
3968 if (__get_user(clen
, &uiov
->iov_len
))
3972 sr
->len
= io
->msg
.iov
[0].iov_len
;
3975 ret
= compat_import_iovec(READ
, uiov
, len
, UIO_FASTIOV
,
3977 &io
->msg
.msg
.msg_iter
);
3986 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
, struct io_async_ctx
*io
)
3988 io
->msg
.iov
= io
->msg
.fast_iov
;
3990 #ifdef CONFIG_COMPAT
3991 if (req
->ctx
->compat
)
3992 return __io_compat_recvmsg_copy_hdr(req
, io
);
3995 return __io_recvmsg_copy_hdr(req
, io
);
3998 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
3999 int *cflags
, bool needs_lock
)
4001 struct io_sr_msg
*sr
= &req
->sr_msg
;
4002 struct io_buffer
*kbuf
;
4004 if (!(req
->flags
& REQ_F_BUFFER_SELECT
))
4007 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4012 req
->flags
|= REQ_F_BUFFER_SELECTED
;
4014 *cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
4015 *cflags
|= IORING_CQE_F_BUFFER
;
4019 static int io_recvmsg_prep(struct io_kiocb
*req
,
4020 const struct io_uring_sqe
*sqe
)
4022 struct io_sr_msg
*sr
= &req
->sr_msg
;
4023 struct io_async_ctx
*io
= req
->io
;
4026 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4029 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
);
4030 sr
->msg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4031 sr
->len
= READ_ONCE(sqe
->len
);
4032 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
4034 #ifdef CONFIG_COMPAT
4035 if (req
->ctx
->compat
)
4036 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4039 if (!io
|| req
->opcode
== IORING_OP_RECV
)
4041 /* iovec is already imported */
4042 if (req
->flags
& REQ_F_NEED_CLEANUP
)
4045 ret
= io_recvmsg_copy_hdr(req
, io
);
4047 req
->flags
|= REQ_F_NEED_CLEANUP
;
4051 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
)
4053 struct io_async_msghdr
*kmsg
= NULL
;
4054 struct socket
*sock
;
4055 int ret
, cflags
= 0;
4057 sock
= sock_from_file(req
->file
, &ret
);
4059 struct io_buffer
*kbuf
;
4060 struct io_async_ctx io
;
4064 kmsg
= &req
->io
->msg
;
4065 kmsg
->msg
.msg_name
= &req
->io
->msg
.addr
;
4066 /* if iov is set, it's allocated already */
4068 kmsg
->iov
= kmsg
->fast_iov
;
4069 kmsg
->msg
.msg_iter
.iov
= kmsg
->iov
;
4072 kmsg
->msg
.msg_name
= &io
.msg
.addr
;
4074 ret
= io_recvmsg_copy_hdr(req
, &io
);
4079 kbuf
= io_recv_buffer_select(req
, &cflags
, !force_nonblock
);
4081 return PTR_ERR(kbuf
);
4083 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
4084 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->iov
,
4085 1, req
->sr_msg
.len
);
4088 flags
= req
->sr_msg
.msg_flags
;
4089 if (flags
& MSG_DONTWAIT
)
4090 req
->flags
|= REQ_F_NOWAIT
;
4091 else if (force_nonblock
)
4092 flags
|= MSG_DONTWAIT
;
4094 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.msg
,
4095 kmsg
->uaddr
, flags
);
4096 if (force_nonblock
&& ret
== -EAGAIN
)
4097 return io_setup_async_msg(req
, kmsg
);
4098 if (ret
== -ERESTARTSYS
)
4102 if (kmsg
&& kmsg
->iov
!= kmsg
->fast_iov
)
4104 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4105 __io_cqring_add_event(req
, ret
, cflags
);
4107 req_set_fail_links(req
);
4112 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
)
4114 struct io_buffer
*kbuf
= NULL
;
4115 struct socket
*sock
;
4116 int ret
, cflags
= 0;
4118 sock
= sock_from_file(req
->file
, &ret
);
4120 struct io_sr_msg
*sr
= &req
->sr_msg
;
4121 void __user
*buf
= sr
->buf
;
4126 kbuf
= io_recv_buffer_select(req
, &cflags
, !force_nonblock
);
4128 return PTR_ERR(kbuf
);
4130 buf
= u64_to_user_ptr(kbuf
->addr
);
4132 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
,
4139 req
->flags
|= REQ_F_NEED_CLEANUP
;
4140 msg
.msg_name
= NULL
;
4141 msg
.msg_control
= NULL
;
4142 msg
.msg_controllen
= 0;
4143 msg
.msg_namelen
= 0;
4144 msg
.msg_iocb
= NULL
;
4147 flags
= req
->sr_msg
.msg_flags
;
4148 if (flags
& MSG_DONTWAIT
)
4149 req
->flags
|= REQ_F_NOWAIT
;
4150 else if (force_nonblock
)
4151 flags
|= MSG_DONTWAIT
;
4153 ret
= sock_recvmsg(sock
, &msg
, flags
);
4154 if (force_nonblock
&& ret
== -EAGAIN
)
4156 if (ret
== -ERESTARTSYS
)
4161 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4162 __io_cqring_add_event(req
, ret
, cflags
);
4164 req_set_fail_links(req
);
4169 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4171 struct io_accept
*accept
= &req
->accept
;
4173 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
4175 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
4178 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4179 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4180 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
4181 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
4185 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
)
4187 struct io_accept
*accept
= &req
->accept
;
4188 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4191 if (req
->file
->f_flags
& O_NONBLOCK
)
4192 req
->flags
|= REQ_F_NOWAIT
;
4194 ret
= __sys_accept4_file(req
->file
, file_flags
, accept
->addr
,
4195 accept
->addr_len
, accept
->flags
,
4197 if (ret
== -EAGAIN
&& force_nonblock
)
4200 if (ret
== -ERESTARTSYS
)
4202 req_set_fail_links(req
);
4204 io_cqring_add_event(req
, ret
);
4209 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4211 struct io_connect
*conn
= &req
->connect
;
4212 struct io_async_ctx
*io
= req
->io
;
4214 if (unlikely(req
->ctx
->flags
& (IORING_SETUP_IOPOLL
|IORING_SETUP_SQPOLL
)))
4216 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
)
4219 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4220 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
4225 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
,
4226 &io
->connect
.address
);
4229 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
)
4231 struct io_async_ctx __io
, *io
;
4232 unsigned file_flags
;
4238 ret
= move_addr_to_kernel(req
->connect
.addr
,
4239 req
->connect
.addr_len
,
4240 &__io
.connect
.address
);
4246 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
4248 ret
= __sys_connect_file(req
->file
, &io
->connect
.address
,
4249 req
->connect
.addr_len
, file_flags
);
4250 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
4253 if (io_alloc_async_ctx(req
)) {
4257 memcpy(&req
->io
->connect
, &__io
.connect
, sizeof(__io
.connect
));
4260 if (ret
== -ERESTARTSYS
)
4264 req_set_fail_links(req
);
4265 io_cqring_add_event(req
, ret
);
4269 #else /* !CONFIG_NET */
4270 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4275 static int io_sendmsg(struct io_kiocb
*req
, bool force_nonblock
)
4280 static int io_send(struct io_kiocb
*req
, bool force_nonblock
)
4285 static int io_recvmsg_prep(struct io_kiocb
*req
,
4286 const struct io_uring_sqe
*sqe
)
4291 static int io_recvmsg(struct io_kiocb
*req
, bool force_nonblock
)
4296 static int io_recv(struct io_kiocb
*req
, bool force_nonblock
)
4301 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4306 static int io_accept(struct io_kiocb
*req
, bool force_nonblock
)
4311 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4316 static int io_connect(struct io_kiocb
*req
, bool force_nonblock
)
4320 #endif /* CONFIG_NET */
4322 struct io_poll_table
{
4323 struct poll_table_struct pt
;
4324 struct io_kiocb
*req
;
4328 static int __io_async_wake(struct io_kiocb
*req
, struct io_poll_iocb
*poll
,
4329 __poll_t mask
, task_work_func_t func
)
4331 struct task_struct
*tsk
;
4334 /* for instances that support it check for an event match first: */
4335 if (mask
&& !(mask
& poll
->events
))
4338 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
4340 list_del_init(&poll
->wait
.entry
);
4344 init_task_work(&req
->task_work
, func
);
4346 * If this fails, then the task is exiting. When a task exits, the
4347 * work gets canceled, so just cancel this request as well instead
4348 * of executing it. We can't safely execute it anyway, as we may not
4349 * have the needed state needed for it anyway.
4351 ret
= task_work_add(tsk
, &req
->task_work
, true);
4352 if (unlikely(ret
)) {
4353 WRITE_ONCE(poll
->canceled
, true);
4354 tsk
= io_wq_get_task(req
->ctx
->io_wq
);
4355 task_work_add(tsk
, &req
->task_work
, true);
4357 wake_up_process(tsk
);
4361 static bool io_poll_rewait(struct io_kiocb
*req
, struct io_poll_iocb
*poll
)
4362 __acquires(&req
->ctx
->completion_lock
)
4364 struct io_ring_ctx
*ctx
= req
->ctx
;
4366 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4367 struct poll_table_struct pt
= { ._key
= poll
->events
};
4369 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
4372 spin_lock_irq(&ctx
->completion_lock
);
4373 if (!req
->result
&& !READ_ONCE(poll
->canceled
)) {
4374 add_wait_queue(poll
->head
, &poll
->wait
);
4381 static void io_poll_remove_double(struct io_kiocb
*req
)
4383 struct io_poll_iocb
*poll
= (struct io_poll_iocb
*) req
->io
;
4385 lockdep_assert_held(&req
->ctx
->completion_lock
);
4387 if (poll
&& poll
->head
) {
4388 struct wait_queue_head
*head
= poll
->head
;
4390 spin_lock(&head
->lock
);
4391 list_del_init(&poll
->wait
.entry
);
4392 if (poll
->wait
.private)
4393 refcount_dec(&req
->refs
);
4395 spin_unlock(&head
->lock
);
4399 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
, int error
)
4401 struct io_ring_ctx
*ctx
= req
->ctx
;
4403 io_poll_remove_double(req
);
4404 req
->poll
.done
= true;
4405 io_cqring_fill_event(req
, error
? error
: mangle_poll(mask
));
4406 io_commit_cqring(ctx
);
4409 static void io_poll_task_handler(struct io_kiocb
*req
, struct io_kiocb
**nxt
)
4411 struct io_ring_ctx
*ctx
= req
->ctx
;
4413 if (io_poll_rewait(req
, &req
->poll
)) {
4414 spin_unlock_irq(&ctx
->completion_lock
);
4418 hash_del(&req
->hash_node
);
4419 io_poll_complete(req
, req
->result
, 0);
4420 req
->flags
|= REQ_F_COMP_LOCKED
;
4421 io_put_req_find_next(req
, nxt
);
4422 spin_unlock_irq(&ctx
->completion_lock
);
4424 io_cqring_ev_posted(ctx
);
4427 static void io_poll_task_func(struct callback_head
*cb
)
4429 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4430 struct io_kiocb
*nxt
= NULL
;
4432 io_poll_task_handler(req
, &nxt
);
4434 struct io_ring_ctx
*ctx
= nxt
->ctx
;
4436 mutex_lock(&ctx
->uring_lock
);
4437 __io_queue_sqe(nxt
, NULL
);
4438 mutex_unlock(&ctx
->uring_lock
);
4442 static int io_poll_double_wake(struct wait_queue_entry
*wait
, unsigned mode
,
4443 int sync
, void *key
)
4445 struct io_kiocb
*req
= wait
->private;
4446 struct io_poll_iocb
*poll
= (struct io_poll_iocb
*) req
->io
;
4447 __poll_t mask
= key_to_poll(key
);
4449 /* for instances that support it check for an event match first: */
4450 if (mask
&& !(mask
& poll
->events
))
4453 if (req
->poll
.head
) {
4456 spin_lock(&req
->poll
.head
->lock
);
4457 done
= list_empty(&req
->poll
.wait
.entry
);
4459 list_del_init(&req
->poll
.wait
.entry
);
4460 spin_unlock(&req
->poll
.head
->lock
);
4462 __io_async_wake(req
, poll
, mask
, io_poll_task_func
);
4464 refcount_dec(&req
->refs
);
4468 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
4469 wait_queue_func_t wake_func
)
4473 poll
->canceled
= false;
4474 poll
->events
= events
;
4475 INIT_LIST_HEAD(&poll
->wait
.entry
);
4476 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
4479 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
4480 struct wait_queue_head
*head
)
4482 struct io_kiocb
*req
= pt
->req
;
4485 * If poll->head is already set, it's because the file being polled
4486 * uses multiple waitqueues for poll handling (eg one for read, one
4487 * for write). Setup a separate io_poll_iocb if this happens.
4489 if (unlikely(poll
->head
)) {
4490 /* already have a 2nd entry, fail a third attempt */
4492 pt
->error
= -EINVAL
;
4495 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
4497 pt
->error
= -ENOMEM
;
4500 io_init_poll_iocb(poll
, req
->poll
.events
, io_poll_double_wake
);
4501 refcount_inc(&req
->refs
);
4502 poll
->wait
.private = req
;
4503 req
->io
= (void *) poll
;
4509 if (poll
->events
& EPOLLEXCLUSIVE
)
4510 add_wait_queue_exclusive(head
, &poll
->wait
);
4512 add_wait_queue(head
, &poll
->wait
);
4515 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
4516 struct poll_table_struct
*p
)
4518 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
4520 __io_queue_proc(&pt
->req
->apoll
->poll
, pt
, head
);
4523 static void io_async_task_func(struct callback_head
*cb
)
4525 struct io_kiocb
*req
= container_of(cb
, struct io_kiocb
, task_work
);
4526 struct async_poll
*apoll
= req
->apoll
;
4527 struct io_ring_ctx
*ctx
= req
->ctx
;
4528 bool canceled
= false;
4530 trace_io_uring_task_run(req
->ctx
, req
->opcode
, req
->user_data
);
4532 if (io_poll_rewait(req
, &apoll
->poll
)) {
4533 spin_unlock_irq(&ctx
->completion_lock
);
4537 /* If req is still hashed, it cannot have been canceled. Don't check. */
4538 if (hash_hashed(&req
->hash_node
)) {
4539 hash_del(&req
->hash_node
);
4541 canceled
= READ_ONCE(apoll
->poll
.canceled
);
4543 io_cqring_fill_event(req
, -ECANCELED
);
4544 io_commit_cqring(ctx
);
4548 spin_unlock_irq(&ctx
->completion_lock
);
4550 /* restore ->work in case we need to retry again */
4551 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4552 memcpy(&req
->work
, &apoll
->work
, sizeof(req
->work
));
4556 __set_current_state(TASK_RUNNING
);
4557 if (io_sq_thread_acquire_mm(ctx
, req
)) {
4558 io_cqring_add_event(req
, -EFAULT
);
4561 mutex_lock(&ctx
->uring_lock
);
4562 __io_queue_sqe(req
, NULL
);
4563 mutex_unlock(&ctx
->uring_lock
);
4565 io_cqring_ev_posted(ctx
);
4567 req_set_fail_links(req
);
4568 io_double_put_req(req
);
4572 static int io_async_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
4575 struct io_kiocb
*req
= wait
->private;
4576 struct io_poll_iocb
*poll
= &req
->apoll
->poll
;
4578 trace_io_uring_poll_wake(req
->ctx
, req
->opcode
, req
->user_data
,
4581 return __io_async_wake(req
, poll
, key_to_poll(key
), io_async_task_func
);
4584 static void io_poll_req_insert(struct io_kiocb
*req
)
4586 struct io_ring_ctx
*ctx
= req
->ctx
;
4587 struct hlist_head
*list
;
4589 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
4590 hlist_add_head(&req
->hash_node
, list
);
4593 static __poll_t
__io_arm_poll_handler(struct io_kiocb
*req
,
4594 struct io_poll_iocb
*poll
,
4595 struct io_poll_table
*ipt
, __poll_t mask
,
4596 wait_queue_func_t wake_func
)
4597 __acquires(&ctx
->completion_lock
)
4599 struct io_ring_ctx
*ctx
= req
->ctx
;
4600 bool cancel
= false;
4602 io_init_poll_iocb(poll
, mask
, wake_func
);
4603 poll
->file
= req
->file
;
4604 poll
->wait
.private = req
;
4606 ipt
->pt
._key
= mask
;
4608 ipt
->error
= -EINVAL
;
4610 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
4612 spin_lock_irq(&ctx
->completion_lock
);
4613 if (likely(poll
->head
)) {
4614 spin_lock(&poll
->head
->lock
);
4615 if (unlikely(list_empty(&poll
->wait
.entry
))) {
4621 if (mask
|| ipt
->error
)
4622 list_del_init(&poll
->wait
.entry
);
4624 WRITE_ONCE(poll
->canceled
, true);
4625 else if (!poll
->done
) /* actually waiting for an event */
4626 io_poll_req_insert(req
);
4627 spin_unlock(&poll
->head
->lock
);
4633 static bool io_arm_poll_handler(struct io_kiocb
*req
)
4635 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
4636 struct io_ring_ctx
*ctx
= req
->ctx
;
4637 struct async_poll
*apoll
;
4638 struct io_poll_table ipt
;
4642 if (!req
->file
|| !file_can_poll(req
->file
))
4644 if (req
->flags
& REQ_F_POLLED
)
4646 if (!def
->pollin
&& !def
->pollout
)
4649 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
4650 if (unlikely(!apoll
))
4653 req
->flags
|= REQ_F_POLLED
;
4654 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4655 memcpy(&apoll
->work
, &req
->work
, sizeof(req
->work
));
4656 had_io
= req
->io
!= NULL
;
4658 io_get_req_task(req
);
4660 INIT_HLIST_NODE(&req
->hash_node
);
4664 mask
|= POLLIN
| POLLRDNORM
;
4666 mask
|= POLLOUT
| POLLWRNORM
;
4667 mask
|= POLLERR
| POLLPRI
;
4669 ipt
.pt
._qproc
= io_async_queue_proc
;
4671 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
,
4675 /* only remove double add if we did it here */
4677 io_poll_remove_double(req
);
4678 spin_unlock_irq(&ctx
->completion_lock
);
4679 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4680 memcpy(&req
->work
, &apoll
->work
, sizeof(req
->work
));
4684 spin_unlock_irq(&ctx
->completion_lock
);
4685 trace_io_uring_poll_arm(ctx
, req
->opcode
, req
->user_data
, mask
,
4686 apoll
->poll
.events
);
4690 static bool __io_poll_remove_one(struct io_kiocb
*req
,
4691 struct io_poll_iocb
*poll
)
4693 bool do_complete
= false;
4695 spin_lock(&poll
->head
->lock
);
4696 WRITE_ONCE(poll
->canceled
, true);
4697 if (!list_empty(&poll
->wait
.entry
)) {
4698 list_del_init(&poll
->wait
.entry
);
4701 spin_unlock(&poll
->head
->lock
);
4702 hash_del(&req
->hash_node
);
4706 static bool io_poll_remove_one(struct io_kiocb
*req
)
4710 if (req
->opcode
== IORING_OP_POLL_ADD
) {
4711 io_poll_remove_double(req
);
4712 do_complete
= __io_poll_remove_one(req
, &req
->poll
);
4714 struct async_poll
*apoll
= req
->apoll
;
4716 /* non-poll requests have submit ref still */
4717 do_complete
= __io_poll_remove_one(req
, &apoll
->poll
);
4721 * restore ->work because we will call
4722 * io_req_work_drop_env below when dropping the
4725 if (req
->flags
& REQ_F_WORK_INITIALIZED
)
4726 memcpy(&req
->work
, &apoll
->work
,
4733 io_cqring_fill_event(req
, -ECANCELED
);
4734 io_commit_cqring(req
->ctx
);
4735 req
->flags
|= REQ_F_COMP_LOCKED
;
4742 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
4744 struct hlist_node
*tmp
;
4745 struct io_kiocb
*req
;
4748 spin_lock_irq(&ctx
->completion_lock
);
4749 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
4750 struct hlist_head
*list
;
4752 list
= &ctx
->cancel_hash
[i
];
4753 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
)
4754 posted
+= io_poll_remove_one(req
);
4756 spin_unlock_irq(&ctx
->completion_lock
);
4759 io_cqring_ev_posted(ctx
);
4762 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
)
4764 struct hlist_head
*list
;
4765 struct io_kiocb
*req
;
4767 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
4768 hlist_for_each_entry(req
, list
, hash_node
) {
4769 if (sqe_addr
!= req
->user_data
)
4771 if (io_poll_remove_one(req
))
4779 static int io_poll_remove_prep(struct io_kiocb
*req
,
4780 const struct io_uring_sqe
*sqe
)
4782 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4784 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
4788 req
->poll
.addr
= READ_ONCE(sqe
->addr
);
4793 * Find a running poll command that matches one specified in sqe->addr,
4794 * and remove it if found.
4796 static int io_poll_remove(struct io_kiocb
*req
)
4798 struct io_ring_ctx
*ctx
= req
->ctx
;
4802 addr
= req
->poll
.addr
;
4803 spin_lock_irq(&ctx
->completion_lock
);
4804 ret
= io_poll_cancel(ctx
, addr
);
4805 spin_unlock_irq(&ctx
->completion_lock
);
4807 io_cqring_add_event(req
, ret
);
4809 req_set_fail_links(req
);
4814 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
4817 struct io_kiocb
*req
= wait
->private;
4818 struct io_poll_iocb
*poll
= &req
->poll
;
4820 return __io_async_wake(req
, poll
, key_to_poll(key
), io_poll_task_func
);
4823 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
4824 struct poll_table_struct
*p
)
4826 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
4828 __io_queue_proc(&pt
->req
->poll
, pt
, head
);
4831 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4833 struct io_poll_iocb
*poll
= &req
->poll
;
4836 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4838 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
4843 events
= READ_ONCE(sqe
->poll32_events
);
4845 events
= swahw32(events
);
4847 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
|
4848 (events
& EPOLLEXCLUSIVE
);
4850 io_get_req_task(req
);
4854 static int io_poll_add(struct io_kiocb
*req
)
4856 struct io_poll_iocb
*poll
= &req
->poll
;
4857 struct io_ring_ctx
*ctx
= req
->ctx
;
4858 struct io_poll_table ipt
;
4861 INIT_HLIST_NODE(&req
->hash_node
);
4862 INIT_LIST_HEAD(&req
->list
);
4863 ipt
.pt
._qproc
= io_poll_queue_proc
;
4865 mask
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
,
4868 if (mask
) { /* no async, we'd stolen it */
4870 io_poll_complete(req
, mask
, 0);
4872 spin_unlock_irq(&ctx
->completion_lock
);
4875 io_cqring_ev_posted(ctx
);
4881 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
4883 struct io_timeout_data
*data
= container_of(timer
,
4884 struct io_timeout_data
, timer
);
4885 struct io_kiocb
*req
= data
->req
;
4886 struct io_ring_ctx
*ctx
= req
->ctx
;
4887 unsigned long flags
;
4889 atomic_inc(&ctx
->cq_timeouts
);
4891 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
4893 * We could be racing with timeout deletion. If the list is empty,
4894 * then timeout lookup already found it and will be handling it.
4896 if (!list_empty(&req
->list
))
4897 list_del_init(&req
->list
);
4899 io_cqring_fill_event(req
, -ETIME
);
4900 io_commit_cqring(ctx
);
4901 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
4903 io_cqring_ev_posted(ctx
);
4904 req_set_fail_links(req
);
4906 return HRTIMER_NORESTART
;
4909 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
4911 struct io_kiocb
*req
;
4914 list_for_each_entry(req
, &ctx
->timeout_list
, list
) {
4915 if (user_data
== req
->user_data
) {
4916 list_del_init(&req
->list
);
4925 ret
= hrtimer_try_to_cancel(&req
->io
->timeout
.timer
);
4929 req_set_fail_links(req
);
4930 io_cqring_fill_event(req
, -ECANCELED
);
4935 static int io_timeout_remove_prep(struct io_kiocb
*req
,
4936 const struct io_uring_sqe
*sqe
)
4938 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4940 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
)
4943 req
->timeout
.addr
= READ_ONCE(sqe
->addr
);
4944 req
->timeout
.flags
= READ_ONCE(sqe
->timeout_flags
);
4945 if (req
->timeout
.flags
)
4952 * Remove or update an existing timeout command
4954 static int io_timeout_remove(struct io_kiocb
*req
)
4956 struct io_ring_ctx
*ctx
= req
->ctx
;
4959 spin_lock_irq(&ctx
->completion_lock
);
4960 ret
= io_timeout_cancel(ctx
, req
->timeout
.addr
);
4962 io_cqring_fill_event(req
, ret
);
4963 io_commit_cqring(ctx
);
4964 spin_unlock_irq(&ctx
->completion_lock
);
4965 io_cqring_ev_posted(ctx
);
4967 req_set_fail_links(req
);
4972 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
4973 bool is_timeout_link
)
4975 struct io_timeout_data
*data
;
4977 u32 off
= READ_ONCE(sqe
->off
);
4979 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4981 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1)
4983 if (off
&& is_timeout_link
)
4985 flags
= READ_ONCE(sqe
->timeout_flags
);
4986 if (flags
& ~IORING_TIMEOUT_ABS
)
4989 req
->timeout
.off
= off
;
4991 if (!req
->io
&& io_alloc_async_ctx(req
))
4994 data
= &req
->io
->timeout
;
4996 req
->flags
|= REQ_F_TIMEOUT
;
4998 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
5001 if (flags
& IORING_TIMEOUT_ABS
)
5002 data
->mode
= HRTIMER_MODE_ABS
;
5004 data
->mode
= HRTIMER_MODE_REL
;
5006 hrtimer_init(&data
->timer
, CLOCK_MONOTONIC
, data
->mode
);
5010 static int io_timeout(struct io_kiocb
*req
)
5012 struct io_ring_ctx
*ctx
= req
->ctx
;
5013 struct io_timeout_data
*data
= &req
->io
->timeout
;
5014 struct list_head
*entry
;
5015 u32 tail
, off
= req
->timeout
.off
;
5017 spin_lock_irq(&ctx
->completion_lock
);
5020 * sqe->off holds how many events that need to occur for this
5021 * timeout event to be satisfied. If it isn't set, then this is
5022 * a pure timeout request, sequence isn't used.
5025 req
->flags
|= REQ_F_TIMEOUT_NOSEQ
;
5026 entry
= ctx
->timeout_list
.prev
;
5030 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
5031 req
->timeout
.target_seq
= tail
+ off
;
5034 * Insertion sort, ensuring the first entry in the list is always
5035 * the one we need first.
5037 list_for_each_prev(entry
, &ctx
->timeout_list
) {
5038 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
, list
);
5040 if (nxt
->flags
& REQ_F_TIMEOUT_NOSEQ
)
5042 /* nxt.seq is behind @tail, otherwise would've been completed */
5043 if (off
>= nxt
->timeout
.target_seq
- tail
)
5047 list_add(&req
->list
, entry
);
5048 data
->timer
.function
= io_timeout_fn
;
5049 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
5050 spin_unlock_irq(&ctx
->completion_lock
);
5054 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
5056 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5058 return req
->user_data
== (unsigned long) data
;
5061 static int io_async_cancel_one(struct io_ring_ctx
*ctx
, void *sqe_addr
)
5063 enum io_wq_cancel cancel_ret
;
5066 cancel_ret
= io_wq_cancel_cb(ctx
->io_wq
, io_cancel_cb
, sqe_addr
, false);
5067 switch (cancel_ret
) {
5068 case IO_WQ_CANCEL_OK
:
5071 case IO_WQ_CANCEL_RUNNING
:
5074 case IO_WQ_CANCEL_NOTFOUND
:
5082 static void io_async_find_and_cancel(struct io_ring_ctx
*ctx
,
5083 struct io_kiocb
*req
, __u64 sqe_addr
,
5086 unsigned long flags
;
5089 ret
= io_async_cancel_one(ctx
, (void *) (unsigned long) sqe_addr
);
5090 if (ret
!= -ENOENT
) {
5091 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5095 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5096 ret
= io_timeout_cancel(ctx
, sqe_addr
);
5099 ret
= io_poll_cancel(ctx
, sqe_addr
);
5103 io_cqring_fill_event(req
, ret
);
5104 io_commit_cqring(ctx
);
5105 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5106 io_cqring_ev_posted(ctx
);
5109 req_set_fail_links(req
);
5113 static int io_async_cancel_prep(struct io_kiocb
*req
,
5114 const struct io_uring_sqe
*sqe
)
5116 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5118 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
||
5122 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
5126 static int io_async_cancel(struct io_kiocb
*req
)
5128 struct io_ring_ctx
*ctx
= req
->ctx
;
5130 io_async_find_and_cancel(ctx
, req
, req
->cancel
.addr
, 0);
5134 static int io_files_update_prep(struct io_kiocb
*req
,
5135 const struct io_uring_sqe
*sqe
)
5137 if (sqe
->flags
|| sqe
->ioprio
|| sqe
->rw_flags
)
5140 req
->files_update
.offset
= READ_ONCE(sqe
->off
);
5141 req
->files_update
.nr_args
= READ_ONCE(sqe
->len
);
5142 if (!req
->files_update
.nr_args
)
5144 req
->files_update
.arg
= READ_ONCE(sqe
->addr
);
5148 static int io_files_update(struct io_kiocb
*req
, bool force_nonblock
)
5150 struct io_ring_ctx
*ctx
= req
->ctx
;
5151 struct io_uring_files_update up
;
5157 up
.offset
= req
->files_update
.offset
;
5158 up
.fds
= req
->files_update
.arg
;
5160 mutex_lock(&ctx
->uring_lock
);
5161 ret
= __io_sqe_files_update(ctx
, &up
, req
->files_update
.nr_args
);
5162 mutex_unlock(&ctx
->uring_lock
);
5165 req_set_fail_links(req
);
5166 io_cqring_add_event(req
, ret
);
5171 static int io_req_defer_prep(struct io_kiocb
*req
,
5172 const struct io_uring_sqe
*sqe
)
5179 io_req_init_async(req
);
5181 if (io_op_defs
[req
->opcode
].file_table
) {
5182 ret
= io_grab_files(req
);
5187 io_req_work_grab_env(req
, &io_op_defs
[req
->opcode
]);
5189 switch (req
->opcode
) {
5192 case IORING_OP_READV
:
5193 case IORING_OP_READ_FIXED
:
5194 case IORING_OP_READ
:
5195 ret
= io_read_prep(req
, sqe
, true);
5197 case IORING_OP_WRITEV
:
5198 case IORING_OP_WRITE_FIXED
:
5199 case IORING_OP_WRITE
:
5200 ret
= io_write_prep(req
, sqe
, true);
5202 case IORING_OP_POLL_ADD
:
5203 ret
= io_poll_add_prep(req
, sqe
);
5205 case IORING_OP_POLL_REMOVE
:
5206 ret
= io_poll_remove_prep(req
, sqe
);
5208 case IORING_OP_FSYNC
:
5209 ret
= io_prep_fsync(req
, sqe
);
5211 case IORING_OP_SYNC_FILE_RANGE
:
5212 ret
= io_prep_sfr(req
, sqe
);
5214 case IORING_OP_SENDMSG
:
5215 case IORING_OP_SEND
:
5216 ret
= io_sendmsg_prep(req
, sqe
);
5218 case IORING_OP_RECVMSG
:
5219 case IORING_OP_RECV
:
5220 ret
= io_recvmsg_prep(req
, sqe
);
5222 case IORING_OP_CONNECT
:
5223 ret
= io_connect_prep(req
, sqe
);
5225 case IORING_OP_TIMEOUT
:
5226 ret
= io_timeout_prep(req
, sqe
, false);
5228 case IORING_OP_TIMEOUT_REMOVE
:
5229 ret
= io_timeout_remove_prep(req
, sqe
);
5231 case IORING_OP_ASYNC_CANCEL
:
5232 ret
= io_async_cancel_prep(req
, sqe
);
5234 case IORING_OP_LINK_TIMEOUT
:
5235 ret
= io_timeout_prep(req
, sqe
, true);
5237 case IORING_OP_ACCEPT
:
5238 ret
= io_accept_prep(req
, sqe
);
5240 case IORING_OP_FALLOCATE
:
5241 ret
= io_fallocate_prep(req
, sqe
);
5243 case IORING_OP_OPENAT
:
5244 ret
= io_openat_prep(req
, sqe
);
5246 case IORING_OP_CLOSE
:
5247 ret
= io_close_prep(req
, sqe
);
5249 case IORING_OP_FILES_UPDATE
:
5250 ret
= io_files_update_prep(req
, sqe
);
5252 case IORING_OP_STATX
:
5253 ret
= io_statx_prep(req
, sqe
);
5255 case IORING_OP_FADVISE
:
5256 ret
= io_fadvise_prep(req
, sqe
);
5258 case IORING_OP_MADVISE
:
5259 ret
= io_madvise_prep(req
, sqe
);
5261 case IORING_OP_OPENAT2
:
5262 ret
= io_openat2_prep(req
, sqe
);
5264 case IORING_OP_EPOLL_CTL
:
5265 ret
= io_epoll_ctl_prep(req
, sqe
);
5267 case IORING_OP_SPLICE
:
5268 ret
= io_splice_prep(req
, sqe
);
5270 case IORING_OP_PROVIDE_BUFFERS
:
5271 ret
= io_provide_buffers_prep(req
, sqe
);
5273 case IORING_OP_REMOVE_BUFFERS
:
5274 ret
= io_remove_buffers_prep(req
, sqe
);
5277 ret
= io_tee_prep(req
, sqe
);
5280 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
5289 static int io_req_defer(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5291 struct io_ring_ctx
*ctx
= req
->ctx
;
5294 /* Still need defer if there is pending req in defer list. */
5295 if (!req_need_defer(req
) && list_empty_careful(&ctx
->defer_list
))
5299 if (io_alloc_async_ctx(req
))
5301 ret
= io_req_defer_prep(req
, sqe
);
5306 spin_lock_irq(&ctx
->completion_lock
);
5307 if (!req_need_defer(req
) && list_empty(&ctx
->defer_list
)) {
5308 spin_unlock_irq(&ctx
->completion_lock
);
5312 trace_io_uring_defer(ctx
, req
, req
->user_data
);
5313 list_add_tail(&req
->list
, &ctx
->defer_list
);
5314 spin_unlock_irq(&ctx
->completion_lock
);
5315 return -EIOCBQUEUED
;
5318 static void io_cleanup_req(struct io_kiocb
*req
)
5320 struct io_async_ctx
*io
= req
->io
;
5322 switch (req
->opcode
) {
5323 case IORING_OP_READV
:
5324 case IORING_OP_READ_FIXED
:
5325 case IORING_OP_READ
:
5326 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5327 kfree((void *)(unsigned long)req
->rw
.addr
);
5329 case IORING_OP_WRITEV
:
5330 case IORING_OP_WRITE_FIXED
:
5331 case IORING_OP_WRITE
:
5332 if (io
->rw
.iov
!= io
->rw
.fast_iov
)
5335 case IORING_OP_RECVMSG
:
5336 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5337 kfree(req
->sr_msg
.kbuf
);
5339 case IORING_OP_SENDMSG
:
5340 if (io
->msg
.iov
!= io
->msg
.fast_iov
)
5343 case IORING_OP_RECV
:
5344 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5345 kfree(req
->sr_msg
.kbuf
);
5347 case IORING_OP_OPENAT
:
5348 case IORING_OP_OPENAT2
:
5350 case IORING_OP_SPLICE
:
5352 io_put_file(req
, req
->splice
.file_in
,
5353 (req
->splice
.flags
& SPLICE_F_FD_IN_FIXED
));
5357 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5360 static int io_issue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5361 bool force_nonblock
)
5363 struct io_ring_ctx
*ctx
= req
->ctx
;
5366 switch (req
->opcode
) {
5370 case IORING_OP_READV
:
5371 case IORING_OP_READ_FIXED
:
5372 case IORING_OP_READ
:
5374 ret
= io_read_prep(req
, sqe
, force_nonblock
);
5378 ret
= io_read(req
, force_nonblock
);
5380 case IORING_OP_WRITEV
:
5381 case IORING_OP_WRITE_FIXED
:
5382 case IORING_OP_WRITE
:
5384 ret
= io_write_prep(req
, sqe
, force_nonblock
);
5388 ret
= io_write(req
, force_nonblock
);
5390 case IORING_OP_FSYNC
:
5392 ret
= io_prep_fsync(req
, sqe
);
5396 ret
= io_fsync(req
, force_nonblock
);
5398 case IORING_OP_POLL_ADD
:
5400 ret
= io_poll_add_prep(req
, sqe
);
5404 ret
= io_poll_add(req
);
5406 case IORING_OP_POLL_REMOVE
:
5408 ret
= io_poll_remove_prep(req
, sqe
);
5412 ret
= io_poll_remove(req
);
5414 case IORING_OP_SYNC_FILE_RANGE
:
5416 ret
= io_prep_sfr(req
, sqe
);
5420 ret
= io_sync_file_range(req
, force_nonblock
);
5422 case IORING_OP_SENDMSG
:
5423 case IORING_OP_SEND
:
5425 ret
= io_sendmsg_prep(req
, sqe
);
5429 if (req
->opcode
== IORING_OP_SENDMSG
)
5430 ret
= io_sendmsg(req
, force_nonblock
);
5432 ret
= io_send(req
, force_nonblock
);
5434 case IORING_OP_RECVMSG
:
5435 case IORING_OP_RECV
:
5437 ret
= io_recvmsg_prep(req
, sqe
);
5441 if (req
->opcode
== IORING_OP_RECVMSG
)
5442 ret
= io_recvmsg(req
, force_nonblock
);
5444 ret
= io_recv(req
, force_nonblock
);
5446 case IORING_OP_TIMEOUT
:
5448 ret
= io_timeout_prep(req
, sqe
, false);
5452 ret
= io_timeout(req
);
5454 case IORING_OP_TIMEOUT_REMOVE
:
5456 ret
= io_timeout_remove_prep(req
, sqe
);
5460 ret
= io_timeout_remove(req
);
5462 case IORING_OP_ACCEPT
:
5464 ret
= io_accept_prep(req
, sqe
);
5468 ret
= io_accept(req
, force_nonblock
);
5470 case IORING_OP_CONNECT
:
5472 ret
= io_connect_prep(req
, sqe
);
5476 ret
= io_connect(req
, force_nonblock
);
5478 case IORING_OP_ASYNC_CANCEL
:
5480 ret
= io_async_cancel_prep(req
, sqe
);
5484 ret
= io_async_cancel(req
);
5486 case IORING_OP_FALLOCATE
:
5488 ret
= io_fallocate_prep(req
, sqe
);
5492 ret
= io_fallocate(req
, force_nonblock
);
5494 case IORING_OP_OPENAT
:
5496 ret
= io_openat_prep(req
, sqe
);
5500 ret
= io_openat(req
, force_nonblock
);
5502 case IORING_OP_CLOSE
:
5504 ret
= io_close_prep(req
, sqe
);
5508 ret
= io_close(req
, force_nonblock
);
5510 case IORING_OP_FILES_UPDATE
:
5512 ret
= io_files_update_prep(req
, sqe
);
5516 ret
= io_files_update(req
, force_nonblock
);
5518 case IORING_OP_STATX
:
5520 ret
= io_statx_prep(req
, sqe
);
5524 ret
= io_statx(req
, force_nonblock
);
5526 case IORING_OP_FADVISE
:
5528 ret
= io_fadvise_prep(req
, sqe
);
5532 ret
= io_fadvise(req
, force_nonblock
);
5534 case IORING_OP_MADVISE
:
5536 ret
= io_madvise_prep(req
, sqe
);
5540 ret
= io_madvise(req
, force_nonblock
);
5542 case IORING_OP_OPENAT2
:
5544 ret
= io_openat2_prep(req
, sqe
);
5548 ret
= io_openat2(req
, force_nonblock
);
5550 case IORING_OP_EPOLL_CTL
:
5552 ret
= io_epoll_ctl_prep(req
, sqe
);
5556 ret
= io_epoll_ctl(req
, force_nonblock
);
5558 case IORING_OP_SPLICE
:
5560 ret
= io_splice_prep(req
, sqe
);
5564 ret
= io_splice(req
, force_nonblock
);
5566 case IORING_OP_PROVIDE_BUFFERS
:
5568 ret
= io_provide_buffers_prep(req
, sqe
);
5572 ret
= io_provide_buffers(req
, force_nonblock
);
5574 case IORING_OP_REMOVE_BUFFERS
:
5576 ret
= io_remove_buffers_prep(req
, sqe
);
5580 ret
= io_remove_buffers(req
, force_nonblock
);
5584 ret
= io_tee_prep(req
, sqe
);
5588 ret
= io_tee(req
, force_nonblock
);
5598 /* If the op doesn't have a file, we're not polling for it */
5599 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
) {
5600 const bool in_async
= io_wq_current_is_worker();
5602 if (req
->result
== -EAGAIN
)
5605 /* workqueue context doesn't hold uring_lock, grab it now */
5607 mutex_lock(&ctx
->uring_lock
);
5609 io_iopoll_req_issued(req
);
5612 mutex_unlock(&ctx
->uring_lock
);
5618 static void io_arm_async_linked_timeout(struct io_kiocb
*req
)
5620 struct io_kiocb
*link
;
5622 /* link head's timeout is queued in io_queue_async_work() */
5623 if (!(req
->flags
& REQ_F_QUEUE_TIMEOUT
))
5626 link
= list_first_entry(&req
->link_list
, struct io_kiocb
, link_list
);
5627 io_queue_linked_timeout(link
);
5630 static void io_wq_submit_work(struct io_wq_work
**workptr
)
5632 struct io_wq_work
*work
= *workptr
;
5633 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
5636 io_arm_async_linked_timeout(req
);
5638 /* if NO_CANCEL is set, we must still run the work */
5639 if ((work
->flags
& (IO_WQ_WORK_CANCEL
|IO_WQ_WORK_NO_CANCEL
)) ==
5640 IO_WQ_WORK_CANCEL
) {
5646 ret
= io_issue_sqe(req
, NULL
, false);
5648 * We can get EAGAIN for polled IO even though we're
5649 * forcing a sync submission from here, since we can't
5650 * wait for request slots on the block side.
5659 req_set_fail_links(req
);
5660 io_cqring_add_event(req
, ret
);
5664 io_steal_work(req
, workptr
);
5667 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
5670 struct fixed_file_table
*table
;
5672 table
= &ctx
->file_data
->table
[index
>> IORING_FILE_TABLE_SHIFT
];
5673 return table
->files
[index
& IORING_FILE_TABLE_MASK
];
5676 static int io_file_get(struct io_submit_state
*state
, struct io_kiocb
*req
,
5677 int fd
, struct file
**out_file
, bool fixed
)
5679 struct io_ring_ctx
*ctx
= req
->ctx
;
5683 if (unlikely(!ctx
->file_data
||
5684 (unsigned) fd
>= ctx
->nr_user_files
))
5686 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
5687 file
= io_file_from_index(ctx
, fd
);
5689 req
->fixed_file_refs
= ctx
->file_data
->cur_refs
;
5690 percpu_ref_get(req
->fixed_file_refs
);
5693 trace_io_uring_file_get(ctx
, fd
);
5694 file
= __io_file_get(state
, fd
);
5697 if (file
|| io_op_defs
[req
->opcode
].needs_file_no_error
) {
5704 static int io_req_set_file(struct io_submit_state
*state
, struct io_kiocb
*req
,
5709 fixed
= (req
->flags
& REQ_F_FIXED_FILE
) != 0;
5710 if (unlikely(!fixed
&& io_async_submit(req
->ctx
)))
5713 return io_file_get(state
, req
, fd
, &req
->file
, fixed
);
5716 static int io_grab_files(struct io_kiocb
*req
)
5719 struct io_ring_ctx
*ctx
= req
->ctx
;
5721 if (req
->work
.files
|| (req
->flags
& REQ_F_NO_FILE_TABLE
))
5723 if (!ctx
->ring_file
)
5727 spin_lock_irq(&ctx
->inflight_lock
);
5729 * We use the f_ops->flush() handler to ensure that we can flush
5730 * out work accessing these files if the fd is closed. Check if
5731 * the fd has changed since we started down this path, and disallow
5732 * this operation if it has.
5734 if (fcheck(ctx
->ring_fd
) == ctx
->ring_file
) {
5735 list_add(&req
->inflight_entry
, &ctx
->inflight_list
);
5736 req
->flags
|= REQ_F_INFLIGHT
;
5737 req
->work
.files
= current
->files
;
5740 spin_unlock_irq(&ctx
->inflight_lock
);
5746 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
5748 struct io_timeout_data
*data
= container_of(timer
,
5749 struct io_timeout_data
, timer
);
5750 struct io_kiocb
*req
= data
->req
;
5751 struct io_ring_ctx
*ctx
= req
->ctx
;
5752 struct io_kiocb
*prev
= NULL
;
5753 unsigned long flags
;
5755 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
5758 * We don't expect the list to be empty, that will only happen if we
5759 * race with the completion of the linked work.
5761 if (!list_empty(&req
->link_list
)) {
5762 prev
= list_entry(req
->link_list
.prev
, struct io_kiocb
,
5764 if (refcount_inc_not_zero(&prev
->refs
)) {
5765 list_del_init(&req
->link_list
);
5766 prev
->flags
&= ~REQ_F_LINK_TIMEOUT
;
5771 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
5774 req_set_fail_links(prev
);
5775 io_async_find_and_cancel(ctx
, req
, prev
->user_data
, -ETIME
);
5778 io_cqring_add_event(req
, -ETIME
);
5781 return HRTIMER_NORESTART
;
5784 static void io_queue_linked_timeout(struct io_kiocb
*req
)
5786 struct io_ring_ctx
*ctx
= req
->ctx
;
5789 * If the list is now empty, then our linked request finished before
5790 * we got a chance to setup the timer
5792 spin_lock_irq(&ctx
->completion_lock
);
5793 if (!list_empty(&req
->link_list
)) {
5794 struct io_timeout_data
*data
= &req
->io
->timeout
;
5796 data
->timer
.function
= io_link_timeout_fn
;
5797 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
5800 spin_unlock_irq(&ctx
->completion_lock
);
5802 /* drop submission reference */
5806 static struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
5808 struct io_kiocb
*nxt
;
5810 if (!(req
->flags
& REQ_F_LINK_HEAD
))
5812 /* for polled retry, if flag is set, we already went through here */
5813 if (req
->flags
& REQ_F_POLLED
)
5816 nxt
= list_first_entry_or_null(&req
->link_list
, struct io_kiocb
,
5818 if (!nxt
|| nxt
->opcode
!= IORING_OP_LINK_TIMEOUT
)
5821 req
->flags
|= REQ_F_LINK_TIMEOUT
;
5825 static void __io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5827 struct io_kiocb
*linked_timeout
;
5828 struct io_kiocb
*nxt
;
5829 const struct cred
*old_creds
= NULL
;
5833 linked_timeout
= io_prep_linked_timeout(req
);
5835 if ((req
->flags
& REQ_F_WORK_INITIALIZED
) && req
->work
.creds
&&
5836 req
->work
.creds
!= current_cred()) {
5838 revert_creds(old_creds
);
5839 if (old_creds
== req
->work
.creds
)
5840 old_creds
= NULL
; /* restored original creds */
5842 old_creds
= override_creds(req
->work
.creds
);
5845 ret
= io_issue_sqe(req
, sqe
, true);
5848 * We async punt it if the file wasn't marked NOWAIT, or if the file
5849 * doesn't support non-blocking read/write attempts
5851 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
5852 if (io_arm_poll_handler(req
)) {
5854 io_queue_linked_timeout(linked_timeout
);
5858 io_req_init_async(req
);
5860 if (io_op_defs
[req
->opcode
].file_table
) {
5861 ret
= io_grab_files(req
);
5867 * Queued up for async execution, worker will release
5868 * submit reference when the iocb is actually submitted.
5870 io_queue_async_work(req
);
5876 /* drop submission reference */
5877 io_put_req_find_next(req
, &nxt
);
5879 if (linked_timeout
) {
5881 io_queue_linked_timeout(linked_timeout
);
5883 io_put_req(linked_timeout
);
5886 /* and drop final reference, if we failed */
5888 io_cqring_add_event(req
, ret
);
5889 req_set_fail_links(req
);
5895 if (req
->flags
& REQ_F_FORCE_ASYNC
)
5901 revert_creds(old_creds
);
5904 static void io_queue_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5908 ret
= io_req_defer(req
, sqe
);
5910 if (ret
!= -EIOCBQUEUED
) {
5912 io_cqring_add_event(req
, ret
);
5913 req_set_fail_links(req
);
5914 io_double_put_req(req
);
5916 } else if (req
->flags
& REQ_F_FORCE_ASYNC
) {
5919 if (io_alloc_async_ctx(req
))
5921 ret
= io_req_defer_prep(req
, sqe
);
5922 if (unlikely(ret
< 0))
5927 * Never try inline submit of IOSQE_ASYNC is set, go straight
5928 * to async execution.
5930 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
5931 io_queue_async_work(req
);
5933 __io_queue_sqe(req
, sqe
);
5937 static inline void io_queue_link_head(struct io_kiocb
*req
)
5939 if (unlikely(req
->flags
& REQ_F_FAIL_LINK
)) {
5940 io_cqring_add_event(req
, -ECANCELED
);
5941 io_double_put_req(req
);
5943 io_queue_sqe(req
, NULL
);
5946 static int io_submit_sqe(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
5947 struct io_kiocb
**link
)
5949 struct io_ring_ctx
*ctx
= req
->ctx
;
5953 * If we already have a head request, queue this one for async
5954 * submittal once the head completes. If we don't have a head but
5955 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5956 * submitted sync once the chain is complete. If none of those
5957 * conditions are true (normal request), then just queue it.
5960 struct io_kiocb
*head
= *link
;
5963 * Taking sequential execution of a link, draining both sides
5964 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5965 * requests in the link. So, it drains the head and the
5966 * next after the link request. The last one is done via
5967 * drain_next flag to persist the effect across calls.
5969 if (req
->flags
& REQ_F_IO_DRAIN
) {
5970 head
->flags
|= REQ_F_IO_DRAIN
;
5971 ctx
->drain_next
= 1;
5973 if (io_alloc_async_ctx(req
))
5976 ret
= io_req_defer_prep(req
, sqe
);
5978 /* fail even hard links since we don't submit */
5979 head
->flags
|= REQ_F_FAIL_LINK
;
5982 trace_io_uring_link(ctx
, req
, head
);
5983 list_add_tail(&req
->link_list
, &head
->link_list
);
5985 /* last request of a link, enqueue the link */
5986 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
5987 io_queue_link_head(head
);
5991 if (unlikely(ctx
->drain_next
)) {
5992 req
->flags
|= REQ_F_IO_DRAIN
;
5993 ctx
->drain_next
= 0;
5995 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
5996 req
->flags
|= REQ_F_LINK_HEAD
;
5997 INIT_LIST_HEAD(&req
->link_list
);
5999 if (io_alloc_async_ctx(req
))
6002 ret
= io_req_defer_prep(req
, sqe
);
6004 req
->flags
|= REQ_F_FAIL_LINK
;
6007 io_queue_sqe(req
, sqe
);
6015 * Batched submission is done, ensure local IO is flushed out.
6017 static void io_submit_state_end(struct io_submit_state
*state
)
6019 blk_finish_plug(&state
->plug
);
6020 io_state_file_put(state
);
6021 if (state
->free_reqs
)
6022 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
6026 * Start submission side cache.
6028 static void io_submit_state_start(struct io_submit_state
*state
,
6029 unsigned int max_ios
)
6031 blk_start_plug(&state
->plug
);
6033 state
->plug
.nowait
= true;
6035 state
->free_reqs
= 0;
6037 state
->ios_left
= max_ios
;
6040 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
6042 struct io_rings
*rings
= ctx
->rings
;
6045 * Ensure any loads from the SQEs are done at this point,
6046 * since once we write the new head, the application could
6047 * write new data to them.
6049 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
6053 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6054 * that is mapped by userspace. This means that care needs to be taken to
6055 * ensure that reads are stable, as we cannot rely on userspace always
6056 * being a good citizen. If members of the sqe are validated and then later
6057 * used, it's important that those reads are done through READ_ONCE() to
6058 * prevent a re-load down the line.
6060 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
6062 u32
*sq_array
= ctx
->sq_array
;
6066 * The cached sq head (or cq tail) serves two purposes:
6068 * 1) allows us to batch the cost of updating the user visible
6070 * 2) allows the kernel side to track the head on its own, even
6071 * though the application is the one updating it.
6073 head
= READ_ONCE(sq_array
[ctx
->cached_sq_head
& ctx
->sq_mask
]);
6074 if (likely(head
< ctx
->sq_entries
))
6075 return &ctx
->sq_sqes
[head
];
6077 /* drop invalid entries */
6078 ctx
->cached_sq_dropped
++;
6079 WRITE_ONCE(ctx
->rings
->sq_dropped
, ctx
->cached_sq_dropped
);
6083 static inline void io_consume_sqe(struct io_ring_ctx
*ctx
)
6085 ctx
->cached_sq_head
++;
6088 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6089 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6090 IOSQE_BUFFER_SELECT)
6092 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
6093 const struct io_uring_sqe
*sqe
,
6094 struct io_submit_state
*state
)
6096 unsigned int sqe_flags
;
6100 * All io need record the previous position, if LINK vs DARIN,
6101 * it can be used to mark the position of the first IO in the
6104 req
->sequence
= ctx
->cached_sq_head
- ctx
->cached_sq_dropped
;
6105 req
->opcode
= READ_ONCE(sqe
->opcode
);
6106 req
->user_data
= READ_ONCE(sqe
->user_data
);
6111 /* one is dropped after submission, the other at completion */
6112 refcount_set(&req
->refs
, 2);
6113 req
->task
= current
;
6116 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
6119 if (unlikely(io_sq_thread_acquire_mm(ctx
, req
)))
6122 sqe_flags
= READ_ONCE(sqe
->flags
);
6123 /* enforce forwards compatibility on users */
6124 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
6127 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
6128 !io_op_defs
[req
->opcode
].buffer_select
)
6131 id
= READ_ONCE(sqe
->personality
);
6133 io_req_init_async(req
);
6134 req
->work
.creds
= idr_find(&ctx
->personality_idr
, id
);
6135 if (unlikely(!req
->work
.creds
))
6137 get_cred(req
->work
.creds
);
6140 /* same numerical values with corresponding REQ_F_*, safe to copy */
6141 req
->flags
|= sqe_flags
;
6143 if (!io_op_defs
[req
->opcode
].needs_file
)
6146 return io_req_set_file(state
, req
, READ_ONCE(sqe
->fd
));
6149 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
,
6150 struct file
*ring_file
, int ring_fd
)
6152 struct io_submit_state state
;
6153 struct io_kiocb
*link
= NULL
;
6154 int i
, submitted
= 0;
6156 /* if we have a backlog and couldn't flush it all, return BUSY */
6157 if (test_bit(0, &ctx
->sq_check_overflow
)) {
6158 if (!list_empty(&ctx
->cq_overflow_list
) &&
6159 !io_cqring_overflow_flush(ctx
, false))
6163 /* make sure SQ entry isn't read before tail */
6164 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
6166 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
6169 io_submit_state_start(&state
, nr
);
6171 ctx
->ring_fd
= ring_fd
;
6172 ctx
->ring_file
= ring_file
;
6174 for (i
= 0; i
< nr
; i
++) {
6175 const struct io_uring_sqe
*sqe
;
6176 struct io_kiocb
*req
;
6179 sqe
= io_get_sqe(ctx
);
6180 if (unlikely(!sqe
)) {
6181 io_consume_sqe(ctx
);
6184 req
= io_alloc_req(ctx
, &state
);
6185 if (unlikely(!req
)) {
6187 submitted
= -EAGAIN
;
6191 err
= io_init_req(ctx
, req
, sqe
, &state
);
6192 io_consume_sqe(ctx
);
6193 /* will complete beyond this point, count as submitted */
6196 if (unlikely(err
)) {
6198 io_cqring_add_event(req
, err
);
6199 io_double_put_req(req
);
6203 trace_io_uring_submit_sqe(ctx
, req
->opcode
, req
->user_data
,
6204 true, io_async_submit(ctx
));
6205 err
= io_submit_sqe(req
, sqe
, &link
);
6210 if (unlikely(submitted
!= nr
)) {
6211 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
6213 percpu_ref_put_many(&ctx
->refs
, nr
- ref_used
);
6216 io_queue_link_head(link
);
6217 io_submit_state_end(&state
);
6219 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6220 io_commit_sqring(ctx
);
6225 static int io_sq_thread(void *data
)
6227 struct io_ring_ctx
*ctx
= data
;
6228 const struct cred
*old_cred
;
6230 unsigned long timeout
;
6233 complete(&ctx
->sq_thread_comp
);
6235 old_cred
= override_creds(ctx
->creds
);
6237 timeout
= jiffies
+ ctx
->sq_thread_idle
;
6238 while (!kthread_should_park()) {
6239 unsigned int to_submit
;
6241 if (!list_empty(&ctx
->poll_list
)) {
6242 unsigned nr_events
= 0;
6244 mutex_lock(&ctx
->uring_lock
);
6245 if (!list_empty(&ctx
->poll_list
))
6246 io_iopoll_getevents(ctx
, &nr_events
, 0);
6248 timeout
= jiffies
+ ctx
->sq_thread_idle
;
6249 mutex_unlock(&ctx
->uring_lock
);
6252 to_submit
= io_sqring_entries(ctx
);
6255 * If submit got -EBUSY, flag us as needing the application
6256 * to enter the kernel to reap and flush events.
6258 if (!to_submit
|| ret
== -EBUSY
) {
6260 * Drop cur_mm before scheduling, we can't hold it for
6261 * long periods (or over schedule()). Do this before
6262 * adding ourselves to the waitqueue, as the unuse/drop
6265 io_sq_thread_drop_mm(ctx
);
6268 * We're polling. If we're within the defined idle
6269 * period, then let us spin without work before going
6270 * to sleep. The exception is if we got EBUSY doing
6271 * more IO, we should wait for the application to
6272 * reap events and wake us up.
6274 if (!list_empty(&ctx
->poll_list
) ||
6275 (!time_after(jiffies
, timeout
) && ret
!= -EBUSY
&&
6276 !percpu_ref_is_dying(&ctx
->refs
))) {
6277 if (current
->task_works
)
6283 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
6284 TASK_INTERRUPTIBLE
);
6287 * While doing polled IO, before going to sleep, we need
6288 * to check if there are new reqs added to poll_list, it
6289 * is because reqs may have been punted to io worker and
6290 * will be added to poll_list later, hence check the
6293 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
6294 !list_empty_careful(&ctx
->poll_list
)) {
6295 finish_wait(&ctx
->sqo_wait
, &wait
);
6299 /* Tell userspace we may need a wakeup call */
6300 ctx
->rings
->sq_flags
|= IORING_SQ_NEED_WAKEUP
;
6301 /* make sure to read SQ tail after writing flags */
6304 to_submit
= io_sqring_entries(ctx
);
6305 if (!to_submit
|| ret
== -EBUSY
) {
6306 if (kthread_should_park()) {
6307 finish_wait(&ctx
->sqo_wait
, &wait
);
6310 if (current
->task_works
) {
6312 finish_wait(&ctx
->sqo_wait
, &wait
);
6315 if (signal_pending(current
))
6316 flush_signals(current
);
6318 finish_wait(&ctx
->sqo_wait
, &wait
);
6320 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6324 finish_wait(&ctx
->sqo_wait
, &wait
);
6326 ctx
->rings
->sq_flags
&= ~IORING_SQ_NEED_WAKEUP
;
6329 mutex_lock(&ctx
->uring_lock
);
6330 if (likely(!percpu_ref_is_dying(&ctx
->refs
)))
6331 ret
= io_submit_sqes(ctx
, to_submit
, NULL
, -1);
6332 mutex_unlock(&ctx
->uring_lock
);
6333 timeout
= jiffies
+ ctx
->sq_thread_idle
;
6336 if (current
->task_works
)
6339 io_sq_thread_drop_mm(ctx
);
6340 revert_creds(old_cred
);
6347 struct io_wait_queue
{
6348 struct wait_queue_entry wq
;
6349 struct io_ring_ctx
*ctx
;
6351 unsigned nr_timeouts
;
6354 static inline bool io_should_wake(struct io_wait_queue
*iowq
, bool noflush
)
6356 struct io_ring_ctx
*ctx
= iowq
->ctx
;
6359 * Wake up if we have enough events, or if a timeout occurred since we
6360 * started waiting. For timeouts, we always want to return to userspace,
6361 * regardless of event count.
6363 return io_cqring_events(ctx
, noflush
) >= iowq
->to_wait
||
6364 atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
6367 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
6368 int wake_flags
, void *key
)
6370 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
6373 /* use noflush == true, as we can't safely rely on locking context */
6374 if (!io_should_wake(iowq
, true))
6377 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
6381 * Wait until events become available, if we don't already have some. The
6382 * application must reap them itself, as they reside on the shared cq ring.
6384 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
6385 const sigset_t __user
*sig
, size_t sigsz
)
6387 struct io_wait_queue iowq
= {
6390 .func
= io_wake_function
,
6391 .entry
= LIST_HEAD_INIT(iowq
.wq
.entry
),
6394 .to_wait
= min_events
,
6396 struct io_rings
*rings
= ctx
->rings
;
6400 if (io_cqring_events(ctx
, false) >= min_events
)
6402 if (!current
->task_works
)
6408 #ifdef CONFIG_COMPAT
6409 if (in_compat_syscall())
6410 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
6414 ret
= set_user_sigmask(sig
, sigsz
);
6420 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
6421 trace_io_uring_cqring_wait(ctx
, min_events
);
6423 prepare_to_wait_exclusive(&ctx
->wait
, &iowq
.wq
,
6424 TASK_INTERRUPTIBLE
);
6425 if (current
->task_works
)
6427 if (io_should_wake(&iowq
, false))
6430 if (signal_pending(current
)) {
6435 finish_wait(&ctx
->wait
, &iowq
.wq
);
6437 restore_saved_sigmask_unless(ret
== -EINTR
);
6439 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
6442 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6444 #if defined(CONFIG_UNIX)
6445 if (ctx
->ring_sock
) {
6446 struct sock
*sock
= ctx
->ring_sock
->sk
;
6447 struct sk_buff
*skb
;
6449 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
6455 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6458 file
= io_file_from_index(ctx
, i
);
6465 static void io_file_ref_kill(struct percpu_ref
*ref
)
6467 struct fixed_file_data
*data
;
6469 data
= container_of(ref
, struct fixed_file_data
, refs
);
6470 complete(&data
->done
);
6473 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
6475 struct fixed_file_data
*data
= ctx
->file_data
;
6476 struct fixed_file_ref_node
*ref_node
= NULL
;
6477 unsigned nr_tables
, i
;
6482 spin_lock(&data
->lock
);
6483 if (!list_empty(&data
->ref_list
))
6484 ref_node
= list_first_entry(&data
->ref_list
,
6485 struct fixed_file_ref_node
, node
);
6486 spin_unlock(&data
->lock
);
6488 percpu_ref_kill(&ref_node
->refs
);
6490 percpu_ref_kill(&data
->refs
);
6492 /* wait for all refs nodes to complete */
6493 flush_delayed_work(&ctx
->file_put_work
);
6494 wait_for_completion(&data
->done
);
6496 __io_sqe_files_unregister(ctx
);
6497 nr_tables
= DIV_ROUND_UP(ctx
->nr_user_files
, IORING_MAX_FILES_TABLE
);
6498 for (i
= 0; i
< nr_tables
; i
++)
6499 kfree(data
->table
[i
].files
);
6501 percpu_ref_exit(&data
->refs
);
6503 ctx
->file_data
= NULL
;
6504 ctx
->nr_user_files
= 0;
6508 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
6510 if (ctx
->sqo_thread
) {
6511 wait_for_completion(&ctx
->sq_thread_comp
);
6513 * The park is a bit of a work-around, without it we get
6514 * warning spews on shutdown with SQPOLL set and affinity
6515 * set to a single CPU.
6517 kthread_park(ctx
->sqo_thread
);
6518 kthread_stop(ctx
->sqo_thread
);
6519 ctx
->sqo_thread
= NULL
;
6523 static void io_finish_async(struct io_ring_ctx
*ctx
)
6525 io_sq_thread_stop(ctx
);
6528 io_wq_destroy(ctx
->io_wq
);
6533 #if defined(CONFIG_UNIX)
6535 * Ensure the UNIX gc is aware of our file set, so we are certain that
6536 * the io_uring can be safely unregistered on process exit, even if we have
6537 * loops in the file referencing.
6539 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
6541 struct sock
*sk
= ctx
->ring_sock
->sk
;
6542 struct scm_fp_list
*fpl
;
6543 struct sk_buff
*skb
;
6546 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
6550 skb
= alloc_skb(0, GFP_KERNEL
);
6559 fpl
->user
= get_uid(ctx
->user
);
6560 for (i
= 0; i
< nr
; i
++) {
6561 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
6565 fpl
->fp
[nr_files
] = get_file(file
);
6566 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
6571 fpl
->max
= SCM_MAX_FD
;
6572 fpl
->count
= nr_files
;
6573 UNIXCB(skb
).fp
= fpl
;
6574 skb
->destructor
= unix_destruct_scm
;
6575 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
6576 skb_queue_head(&sk
->sk_receive_queue
, skb
);
6578 for (i
= 0; i
< nr_files
; i
++)
6589 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6590 * causes regular reference counting to break down. We rely on the UNIX
6591 * garbage collection to take care of this problem for us.
6593 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
6595 unsigned left
, total
;
6599 left
= ctx
->nr_user_files
;
6601 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
6603 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
6607 total
+= this_files
;
6613 while (total
< ctx
->nr_user_files
) {
6614 struct file
*file
= io_file_from_index(ctx
, total
);
6624 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
6630 static int io_sqe_alloc_file_tables(struct io_ring_ctx
*ctx
, unsigned nr_tables
,
6635 for (i
= 0; i
< nr_tables
; i
++) {
6636 struct fixed_file_table
*table
= &ctx
->file_data
->table
[i
];
6637 unsigned this_files
;
6639 this_files
= min(nr_files
, IORING_MAX_FILES_TABLE
);
6640 table
->files
= kcalloc(this_files
, sizeof(struct file
*),
6644 nr_files
-= this_files
;
6650 for (i
= 0; i
< nr_tables
; i
++) {
6651 struct fixed_file_table
*table
= &ctx
->file_data
->table
[i
];
6652 kfree(table
->files
);
6657 static void io_ring_file_put(struct io_ring_ctx
*ctx
, struct file
*file
)
6659 #if defined(CONFIG_UNIX)
6660 struct sock
*sock
= ctx
->ring_sock
->sk
;
6661 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
6662 struct sk_buff
*skb
;
6665 __skb_queue_head_init(&list
);
6668 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6669 * remove this entry and rearrange the file array.
6671 skb
= skb_dequeue(head
);
6673 struct scm_fp_list
*fp
;
6675 fp
= UNIXCB(skb
).fp
;
6676 for (i
= 0; i
< fp
->count
; i
++) {
6679 if (fp
->fp
[i
] != file
)
6682 unix_notinflight(fp
->user
, fp
->fp
[i
]);
6683 left
= fp
->count
- 1 - i
;
6685 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
6686 left
* sizeof(struct file
*));
6693 __skb_queue_tail(&list
, skb
);
6703 __skb_queue_tail(&list
, skb
);
6705 skb
= skb_dequeue(head
);
6708 if (skb_peek(&list
)) {
6709 spin_lock_irq(&head
->lock
);
6710 while ((skb
= __skb_dequeue(&list
)) != NULL
)
6711 __skb_queue_tail(head
, skb
);
6712 spin_unlock_irq(&head
->lock
);
6719 struct io_file_put
{
6720 struct list_head list
;
6724 static void __io_file_put_work(struct fixed_file_ref_node
*ref_node
)
6726 struct fixed_file_data
*file_data
= ref_node
->file_data
;
6727 struct io_ring_ctx
*ctx
= file_data
->ctx
;
6728 struct io_file_put
*pfile
, *tmp
;
6730 list_for_each_entry_safe(pfile
, tmp
, &ref_node
->file_list
, list
) {
6731 list_del(&pfile
->list
);
6732 io_ring_file_put(ctx
, pfile
->file
);
6736 spin_lock(&file_data
->lock
);
6737 list_del(&ref_node
->node
);
6738 spin_unlock(&file_data
->lock
);
6740 percpu_ref_exit(&ref_node
->refs
);
6742 percpu_ref_put(&file_data
->refs
);
6745 static void io_file_put_work(struct work_struct
*work
)
6747 struct io_ring_ctx
*ctx
;
6748 struct llist_node
*node
;
6750 ctx
= container_of(work
, struct io_ring_ctx
, file_put_work
.work
);
6751 node
= llist_del_all(&ctx
->file_put_llist
);
6754 struct fixed_file_ref_node
*ref_node
;
6755 struct llist_node
*next
= node
->next
;
6757 ref_node
= llist_entry(node
, struct fixed_file_ref_node
, llist
);
6758 __io_file_put_work(ref_node
);
6763 static void io_file_data_ref_zero(struct percpu_ref
*ref
)
6765 struct fixed_file_ref_node
*ref_node
;
6766 struct io_ring_ctx
*ctx
;
6770 ref_node
= container_of(ref
, struct fixed_file_ref_node
, refs
);
6771 ctx
= ref_node
->file_data
->ctx
;
6773 if (percpu_ref_is_dying(&ctx
->file_data
->refs
))
6776 first_add
= llist_add(&ref_node
->llist
, &ctx
->file_put_llist
);
6778 mod_delayed_work(system_wq
, &ctx
->file_put_work
, 0);
6780 queue_delayed_work(system_wq
, &ctx
->file_put_work
, delay
);
6783 static struct fixed_file_ref_node
*alloc_fixed_file_ref_node(
6784 struct io_ring_ctx
*ctx
)
6786 struct fixed_file_ref_node
*ref_node
;
6788 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
6790 return ERR_PTR(-ENOMEM
);
6792 if (percpu_ref_init(&ref_node
->refs
, io_file_data_ref_zero
,
6795 return ERR_PTR(-ENOMEM
);
6797 INIT_LIST_HEAD(&ref_node
->node
);
6798 INIT_LIST_HEAD(&ref_node
->file_list
);
6799 ref_node
->file_data
= ctx
->file_data
;
6803 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node
*ref_node
)
6805 percpu_ref_exit(&ref_node
->refs
);
6809 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
6812 __s32 __user
*fds
= (__s32 __user
*) arg
;
6817 struct fixed_file_ref_node
*ref_node
;
6823 if (nr_args
> IORING_MAX_FIXED_FILES
)
6826 ctx
->file_data
= kzalloc(sizeof(*ctx
->file_data
), GFP_KERNEL
);
6827 if (!ctx
->file_data
)
6829 ctx
->file_data
->ctx
= ctx
;
6830 init_completion(&ctx
->file_data
->done
);
6831 INIT_LIST_HEAD(&ctx
->file_data
->ref_list
);
6832 spin_lock_init(&ctx
->file_data
->lock
);
6834 nr_tables
= DIV_ROUND_UP(nr_args
, IORING_MAX_FILES_TABLE
);
6835 ctx
->file_data
->table
= kcalloc(nr_tables
,
6836 sizeof(struct fixed_file_table
),
6838 if (!ctx
->file_data
->table
) {
6839 kfree(ctx
->file_data
);
6840 ctx
->file_data
= NULL
;
6844 if (percpu_ref_init(&ctx
->file_data
->refs
, io_file_ref_kill
,
6845 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
)) {
6846 kfree(ctx
->file_data
->table
);
6847 kfree(ctx
->file_data
);
6848 ctx
->file_data
= NULL
;
6852 if (io_sqe_alloc_file_tables(ctx
, nr_tables
, nr_args
)) {
6853 percpu_ref_exit(&ctx
->file_data
->refs
);
6854 kfree(ctx
->file_data
->table
);
6855 kfree(ctx
->file_data
);
6856 ctx
->file_data
= NULL
;
6860 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
6861 struct fixed_file_table
*table
;
6865 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
6867 /* allow sparse sets */
6873 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
6874 index
= i
& IORING_FILE_TABLE_MASK
;
6882 * Don't allow io_uring instances to be registered. If UNIX
6883 * isn't enabled, then this causes a reference cycle and this
6884 * instance can never get freed. If UNIX is enabled we'll
6885 * handle it just fine, but there's still no point in allowing
6886 * a ring fd as it doesn't support regular read/write anyway.
6888 if (file
->f_op
== &io_uring_fops
) {
6893 table
->files
[index
] = file
;
6897 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
6898 file
= io_file_from_index(ctx
, i
);
6902 for (i
= 0; i
< nr_tables
; i
++)
6903 kfree(ctx
->file_data
->table
[i
].files
);
6905 kfree(ctx
->file_data
->table
);
6906 kfree(ctx
->file_data
);
6907 ctx
->file_data
= NULL
;
6908 ctx
->nr_user_files
= 0;
6912 ret
= io_sqe_files_scm(ctx
);
6914 io_sqe_files_unregister(ctx
);
6918 ref_node
= alloc_fixed_file_ref_node(ctx
);
6919 if (IS_ERR(ref_node
)) {
6920 io_sqe_files_unregister(ctx
);
6921 return PTR_ERR(ref_node
);
6924 ctx
->file_data
->cur_refs
= &ref_node
->refs
;
6925 spin_lock(&ctx
->file_data
->lock
);
6926 list_add(&ref_node
->node
, &ctx
->file_data
->ref_list
);
6927 spin_unlock(&ctx
->file_data
->lock
);
6928 percpu_ref_get(&ctx
->file_data
->refs
);
6932 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
6935 #if defined(CONFIG_UNIX)
6936 struct sock
*sock
= ctx
->ring_sock
->sk
;
6937 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
6938 struct sk_buff
*skb
;
6941 * See if we can merge this file into an existing skb SCM_RIGHTS
6942 * file set. If there's no room, fall back to allocating a new skb
6943 * and filling it in.
6945 spin_lock_irq(&head
->lock
);
6946 skb
= skb_peek(head
);
6948 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
6950 if (fpl
->count
< SCM_MAX_FD
) {
6951 __skb_unlink(skb
, head
);
6952 spin_unlock_irq(&head
->lock
);
6953 fpl
->fp
[fpl
->count
] = get_file(file
);
6954 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
6956 spin_lock_irq(&head
->lock
);
6957 __skb_queue_head(head
, skb
);
6962 spin_unlock_irq(&head
->lock
);
6969 return __io_sqe_files_scm(ctx
, 1, index
);
6975 static int io_queue_file_removal(struct fixed_file_data
*data
,
6978 struct io_file_put
*pfile
;
6979 struct percpu_ref
*refs
= data
->cur_refs
;
6980 struct fixed_file_ref_node
*ref_node
;
6982 pfile
= kzalloc(sizeof(*pfile
), GFP_KERNEL
);
6986 ref_node
= container_of(refs
, struct fixed_file_ref_node
, refs
);
6988 list_add(&pfile
->list
, &ref_node
->file_list
);
6993 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
6994 struct io_uring_files_update
*up
,
6997 struct fixed_file_data
*data
= ctx
->file_data
;
6998 struct fixed_file_ref_node
*ref_node
;
7003 bool needs_switch
= false;
7005 if (check_add_overflow(up
->offset
, nr_args
, &done
))
7007 if (done
> ctx
->nr_user_files
)
7010 ref_node
= alloc_fixed_file_ref_node(ctx
);
7011 if (IS_ERR(ref_node
))
7012 return PTR_ERR(ref_node
);
7015 fds
= u64_to_user_ptr(up
->fds
);
7017 struct fixed_file_table
*table
;
7021 if (copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
7025 i
= array_index_nospec(up
->offset
, ctx
->nr_user_files
);
7026 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7027 index
= i
& IORING_FILE_TABLE_MASK
;
7028 if (table
->files
[index
]) {
7029 file
= io_file_from_index(ctx
, index
);
7030 err
= io_queue_file_removal(data
, file
);
7033 table
->files
[index
] = NULL
;
7034 needs_switch
= true;
7043 * Don't allow io_uring instances to be registered. If
7044 * UNIX isn't enabled, then this causes a reference
7045 * cycle and this instance can never get freed. If UNIX
7046 * is enabled we'll handle it just fine, but there's
7047 * still no point in allowing a ring fd as it doesn't
7048 * support regular read/write anyway.
7050 if (file
->f_op
== &io_uring_fops
) {
7055 table
->files
[index
] = file
;
7056 err
= io_sqe_file_register(ctx
, file
, i
);
7066 percpu_ref_kill(data
->cur_refs
);
7067 spin_lock(&data
->lock
);
7068 list_add(&ref_node
->node
, &data
->ref_list
);
7069 data
->cur_refs
= &ref_node
->refs
;
7070 spin_unlock(&data
->lock
);
7071 percpu_ref_get(&ctx
->file_data
->refs
);
7073 destroy_fixed_file_ref_node(ref_node
);
7075 return done
? done
: err
;
7078 static int io_sqe_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
7081 struct io_uring_files_update up
;
7083 if (!ctx
->file_data
)
7087 if (copy_from_user(&up
, arg
, sizeof(up
)))
7092 return __io_sqe_files_update(ctx
, &up
, nr_args
);
7095 static void io_free_work(struct io_wq_work
*work
)
7097 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7099 /* Consider that io_steal_work() relies on this ref */
7103 static int io_init_wq_offload(struct io_ring_ctx
*ctx
,
7104 struct io_uring_params
*p
)
7106 struct io_wq_data data
;
7108 struct io_ring_ctx
*ctx_attach
;
7109 unsigned int concurrency
;
7112 data
.user
= ctx
->user
;
7113 data
.free_work
= io_free_work
;
7114 data
.do_work
= io_wq_submit_work
;
7116 if (!(p
->flags
& IORING_SETUP_ATTACH_WQ
)) {
7117 /* Do QD, or 4 * CPUS, whatever is smallest */
7118 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
7120 ctx
->io_wq
= io_wq_create(concurrency
, &data
);
7121 if (IS_ERR(ctx
->io_wq
)) {
7122 ret
= PTR_ERR(ctx
->io_wq
);
7128 f
= fdget(p
->wq_fd
);
7132 if (f
.file
->f_op
!= &io_uring_fops
) {
7137 ctx_attach
= f
.file
->private_data
;
7138 /* @io_wq is protected by holding the fd */
7139 if (!io_wq_get(ctx_attach
->io_wq
, &data
)) {
7144 ctx
->io_wq
= ctx_attach
->io_wq
;
7150 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
7151 struct io_uring_params
*p
)
7155 mmgrab(current
->mm
);
7156 ctx
->sqo_mm
= current
->mm
;
7158 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7160 if (!capable(CAP_SYS_ADMIN
))
7163 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
7164 if (!ctx
->sq_thread_idle
)
7165 ctx
->sq_thread_idle
= HZ
;
7167 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7168 int cpu
= p
->sq_thread_cpu
;
7171 if (cpu
>= nr_cpu_ids
)
7173 if (!cpu_online(cpu
))
7176 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
7180 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
7183 if (IS_ERR(ctx
->sqo_thread
)) {
7184 ret
= PTR_ERR(ctx
->sqo_thread
);
7185 ctx
->sqo_thread
= NULL
;
7188 wake_up_process(ctx
->sqo_thread
);
7189 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
7190 /* Can't have SQ_AFF without SQPOLL */
7195 ret
= io_init_wq_offload(ctx
, p
);
7201 io_finish_async(ctx
);
7202 mmdrop(ctx
->sqo_mm
);
7207 static inline void __io_unaccount_mem(struct user_struct
*user
,
7208 unsigned long nr_pages
)
7210 atomic_long_sub(nr_pages
, &user
->locked_vm
);
7213 static inline int __io_account_mem(struct user_struct
*user
,
7214 unsigned long nr_pages
)
7216 unsigned long page_limit
, cur_pages
, new_pages
;
7218 /* Don't allow more pages than we can safely lock */
7219 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
7222 cur_pages
= atomic_long_read(&user
->locked_vm
);
7223 new_pages
= cur_pages
+ nr_pages
;
7224 if (new_pages
> page_limit
)
7226 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
7227 new_pages
) != cur_pages
);
7232 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
7233 enum io_mem_account acct
)
7236 __io_unaccount_mem(ctx
->user
, nr_pages
);
7239 if (acct
== ACCT_LOCKED
)
7240 ctx
->sqo_mm
->locked_vm
-= nr_pages
;
7241 else if (acct
== ACCT_PINNED
)
7242 atomic64_sub(nr_pages
, &ctx
->sqo_mm
->pinned_vm
);
7246 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
,
7247 enum io_mem_account acct
)
7251 if (ctx
->limit_mem
) {
7252 ret
= __io_account_mem(ctx
->user
, nr_pages
);
7258 if (acct
== ACCT_LOCKED
)
7259 ctx
->sqo_mm
->locked_vm
+= nr_pages
;
7260 else if (acct
== ACCT_PINNED
)
7261 atomic64_add(nr_pages
, &ctx
->sqo_mm
->pinned_vm
);
7267 static void io_mem_free(void *ptr
)
7274 page
= virt_to_head_page(ptr
);
7275 if (put_page_testzero(page
))
7276 free_compound_page(page
);
7279 static void *io_mem_alloc(size_t size
)
7281 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
7284 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
7287 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
7290 struct io_rings
*rings
;
7291 size_t off
, sq_array_size
;
7293 off
= struct_size(rings
, cqes
, cq_entries
);
7294 if (off
== SIZE_MAX
)
7298 off
= ALIGN(off
, SMP_CACHE_BYTES
);
7303 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
7304 if (sq_array_size
== SIZE_MAX
)
7307 if (check_add_overflow(off
, sq_array_size
, &off
))
7316 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
7320 pages
= (size_t)1 << get_order(
7321 rings_size(sq_entries
, cq_entries
, NULL
));
7322 pages
+= (size_t)1 << get_order(
7323 array_size(sizeof(struct io_uring_sqe
), sq_entries
));
7328 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
7332 if (!ctx
->user_bufs
)
7335 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
7336 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
7338 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
7339 unpin_user_page(imu
->bvec
[j
].bv_page
);
7341 io_unaccount_mem(ctx
, imu
->nr_bvecs
, ACCT_PINNED
);
7346 kfree(ctx
->user_bufs
);
7347 ctx
->user_bufs
= NULL
;
7348 ctx
->nr_user_bufs
= 0;
7352 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
7353 void __user
*arg
, unsigned index
)
7355 struct iovec __user
*src
;
7357 #ifdef CONFIG_COMPAT
7359 struct compat_iovec __user
*ciovs
;
7360 struct compat_iovec ciov
;
7362 ciovs
= (struct compat_iovec __user
*) arg
;
7363 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
7366 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
7367 dst
->iov_len
= ciov
.iov_len
;
7371 src
= (struct iovec __user
*) arg
;
7372 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
7377 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
7380 struct vm_area_struct
**vmas
= NULL
;
7381 struct page
**pages
= NULL
;
7382 int i
, j
, got_pages
= 0;
7387 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
7390 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
7392 if (!ctx
->user_bufs
)
7395 for (i
= 0; i
< nr_args
; i
++) {
7396 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
7397 unsigned long off
, start
, end
, ubuf
;
7402 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
7407 * Don't impose further limits on the size and buffer
7408 * constraints here, we'll -EINVAL later when IO is
7409 * submitted if they are wrong.
7412 if (!iov
.iov_base
|| !iov
.iov_len
)
7415 /* arbitrary limit, but we need something */
7416 if (iov
.iov_len
> SZ_1G
)
7419 ubuf
= (unsigned long) iov
.iov_base
;
7420 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
7421 start
= ubuf
>> PAGE_SHIFT
;
7422 nr_pages
= end
- start
;
7424 ret
= io_account_mem(ctx
, nr_pages
, ACCT_PINNED
);
7429 if (!pages
|| nr_pages
> got_pages
) {
7432 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
7434 vmas
= kvmalloc_array(nr_pages
,
7435 sizeof(struct vm_area_struct
*),
7437 if (!pages
|| !vmas
) {
7439 io_unaccount_mem(ctx
, nr_pages
, ACCT_PINNED
);
7442 got_pages
= nr_pages
;
7445 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
7449 io_unaccount_mem(ctx
, nr_pages
, ACCT_PINNED
);
7454 mmap_read_lock(current
->mm
);
7455 pret
= pin_user_pages(ubuf
, nr_pages
,
7456 FOLL_WRITE
| FOLL_LONGTERM
,
7458 if (pret
== nr_pages
) {
7459 /* don't support file backed memory */
7460 for (j
= 0; j
< nr_pages
; j
++) {
7461 struct vm_area_struct
*vma
= vmas
[j
];
7464 !is_file_hugepages(vma
->vm_file
)) {
7470 ret
= pret
< 0 ? pret
: -EFAULT
;
7472 mmap_read_unlock(current
->mm
);
7475 * if we did partial map, or found file backed vmas,
7476 * release any pages we did get
7479 unpin_user_pages(pages
, pret
);
7480 io_unaccount_mem(ctx
, nr_pages
, ACCT_PINNED
);
7485 off
= ubuf
& ~PAGE_MASK
;
7487 for (j
= 0; j
< nr_pages
; j
++) {
7490 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
7491 imu
->bvec
[j
].bv_page
= pages
[j
];
7492 imu
->bvec
[j
].bv_len
= vec_len
;
7493 imu
->bvec
[j
].bv_offset
= off
;
7497 /* store original address for later verification */
7499 imu
->len
= iov
.iov_len
;
7500 imu
->nr_bvecs
= nr_pages
;
7502 ctx
->nr_user_bufs
++;
7510 io_sqe_buffer_unregister(ctx
);
7514 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
7516 __s32 __user
*fds
= arg
;
7522 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
7525 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
7526 if (IS_ERR(ctx
->cq_ev_fd
)) {
7527 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
7528 ctx
->cq_ev_fd
= NULL
;
7535 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
7537 if (ctx
->cq_ev_fd
) {
7538 eventfd_ctx_put(ctx
->cq_ev_fd
);
7539 ctx
->cq_ev_fd
= NULL
;
7546 static int __io_destroy_buffers(int id
, void *p
, void *data
)
7548 struct io_ring_ctx
*ctx
= data
;
7549 struct io_buffer
*buf
= p
;
7551 __io_remove_buffers(ctx
, buf
, id
, -1U);
7555 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
7557 idr_for_each(&ctx
->io_buffer_idr
, __io_destroy_buffers
, ctx
);
7558 idr_destroy(&ctx
->io_buffer_idr
);
7561 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
7563 io_finish_async(ctx
);
7565 mmdrop(ctx
->sqo_mm
);
7569 io_iopoll_reap_events(ctx
);
7570 io_sqe_buffer_unregister(ctx
);
7571 io_sqe_files_unregister(ctx
);
7572 io_eventfd_unregister(ctx
);
7573 io_destroy_buffers(ctx
);
7574 idr_destroy(&ctx
->personality_idr
);
7576 #if defined(CONFIG_UNIX)
7577 if (ctx
->ring_sock
) {
7578 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
7579 sock_release(ctx
->ring_sock
);
7583 io_mem_free(ctx
->rings
);
7584 io_mem_free(ctx
->sq_sqes
);
7586 percpu_ref_exit(&ctx
->refs
);
7587 io_unaccount_mem(ctx
, ring_pages(ctx
->sq_entries
, ctx
->cq_entries
),
7589 free_uid(ctx
->user
);
7590 put_cred(ctx
->creds
);
7591 kfree(ctx
->cancel_hash
);
7592 kmem_cache_free(req_cachep
, ctx
->fallback_req
);
7596 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
7598 struct io_ring_ctx
*ctx
= file
->private_data
;
7601 poll_wait(file
, &ctx
->cq_wait
, wait
);
7603 * synchronizes with barrier from wq_has_sleeper call in
7607 if (READ_ONCE(ctx
->rings
->sq
.tail
) - ctx
->cached_sq_head
!=
7608 ctx
->rings
->sq_ring_entries
)
7609 mask
|= EPOLLOUT
| EPOLLWRNORM
;
7610 if (io_cqring_events(ctx
, false))
7611 mask
|= EPOLLIN
| EPOLLRDNORM
;
7616 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
7618 struct io_ring_ctx
*ctx
= file
->private_data
;
7620 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
7623 static int io_remove_personalities(int id
, void *p
, void *data
)
7625 struct io_ring_ctx
*ctx
= data
;
7626 const struct cred
*cred
;
7628 cred
= idr_remove(&ctx
->personality_idr
, id
);
7634 static void io_ring_exit_work(struct work_struct
*work
)
7636 struct io_ring_ctx
*ctx
;
7638 ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
7640 io_cqring_overflow_flush(ctx
, true);
7643 * If we're doing polled IO and end up having requests being
7644 * submitted async (out-of-line), then completions can come in while
7645 * we're waiting for refs to drop. We need to reap these manually,
7646 * as nobody else will be looking for them.
7648 while (!wait_for_completion_timeout(&ctx
->ref_comp
, HZ
/20)) {
7649 io_iopoll_reap_events(ctx
);
7651 io_cqring_overflow_flush(ctx
, true);
7653 io_ring_ctx_free(ctx
);
7656 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
7658 mutex_lock(&ctx
->uring_lock
);
7659 percpu_ref_kill(&ctx
->refs
);
7660 mutex_unlock(&ctx
->uring_lock
);
7662 io_kill_timeouts(ctx
);
7663 io_poll_remove_all(ctx
);
7666 io_wq_cancel_all(ctx
->io_wq
);
7668 io_iopoll_reap_events(ctx
);
7669 /* if we failed setting up the ctx, we might not have any rings */
7671 io_cqring_overflow_flush(ctx
, true);
7672 idr_for_each(&ctx
->personality_idr
, io_remove_personalities
, ctx
);
7673 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
7674 queue_work(system_wq
, &ctx
->exit_work
);
7677 static int io_uring_release(struct inode
*inode
, struct file
*file
)
7679 struct io_ring_ctx
*ctx
= file
->private_data
;
7681 file
->private_data
= NULL
;
7682 io_ring_ctx_wait_and_kill(ctx
);
7686 static bool io_wq_files_match(struct io_wq_work
*work
, void *data
)
7688 struct files_struct
*files
= data
;
7690 return work
->files
== files
;
7693 static void io_uring_cancel_files(struct io_ring_ctx
*ctx
,
7694 struct files_struct
*files
)
7696 if (list_empty_careful(&ctx
->inflight_list
))
7699 /* cancel all at once, should be faster than doing it one by one*/
7700 io_wq_cancel_cb(ctx
->io_wq
, io_wq_files_match
, files
, true);
7702 while (!list_empty_careful(&ctx
->inflight_list
)) {
7703 struct io_kiocb
*cancel_req
= NULL
, *req
;
7706 spin_lock_irq(&ctx
->inflight_lock
);
7707 list_for_each_entry(req
, &ctx
->inflight_list
, inflight_entry
) {
7708 if (req
->work
.files
!= files
)
7710 /* req is being completed, ignore */
7711 if (!refcount_inc_not_zero(&req
->refs
))
7717 prepare_to_wait(&ctx
->inflight_wait
, &wait
,
7718 TASK_UNINTERRUPTIBLE
);
7719 spin_unlock_irq(&ctx
->inflight_lock
);
7721 /* We need to keep going until we don't find a matching req */
7725 if (cancel_req
->flags
& REQ_F_OVERFLOW
) {
7726 spin_lock_irq(&ctx
->completion_lock
);
7727 list_del(&cancel_req
->list
);
7728 cancel_req
->flags
&= ~REQ_F_OVERFLOW
;
7729 if (list_empty(&ctx
->cq_overflow_list
)) {
7730 clear_bit(0, &ctx
->sq_check_overflow
);
7731 clear_bit(0, &ctx
->cq_check_overflow
);
7733 spin_unlock_irq(&ctx
->completion_lock
);
7735 WRITE_ONCE(ctx
->rings
->cq_overflow
,
7736 atomic_inc_return(&ctx
->cached_cq_overflow
));
7739 * Put inflight ref and overflow ref. If that's
7740 * all we had, then we're done with this request.
7742 if (refcount_sub_and_test(2, &cancel_req
->refs
)) {
7743 io_free_req(cancel_req
);
7744 finish_wait(&ctx
->inflight_wait
, &wait
);
7748 io_wq_cancel_work(ctx
->io_wq
, &cancel_req
->work
);
7749 io_put_req(cancel_req
);
7753 finish_wait(&ctx
->inflight_wait
, &wait
);
7757 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
7759 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
7760 struct task_struct
*task
= data
;
7762 return req
->task
== task
;
7765 static int io_uring_flush(struct file
*file
, void *data
)
7767 struct io_ring_ctx
*ctx
= file
->private_data
;
7769 io_uring_cancel_files(ctx
, data
);
7772 * If the task is going away, cancel work it may have pending
7774 if (fatal_signal_pending(current
) || (current
->flags
& PF_EXITING
))
7775 io_wq_cancel_cb(ctx
->io_wq
, io_cancel_task_cb
, current
, true);
7780 static void *io_uring_validate_mmap_request(struct file
*file
,
7781 loff_t pgoff
, size_t sz
)
7783 struct io_ring_ctx
*ctx
= file
->private_data
;
7784 loff_t offset
= pgoff
<< PAGE_SHIFT
;
7789 case IORING_OFF_SQ_RING
:
7790 case IORING_OFF_CQ_RING
:
7793 case IORING_OFF_SQES
:
7797 return ERR_PTR(-EINVAL
);
7800 page
= virt_to_head_page(ptr
);
7801 if (sz
> page_size(page
))
7802 return ERR_PTR(-EINVAL
);
7809 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
7811 size_t sz
= vma
->vm_end
- vma
->vm_start
;
7815 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
7817 return PTR_ERR(ptr
);
7819 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
7820 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
7823 #else /* !CONFIG_MMU */
7825 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
7827 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
7830 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
7832 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
7835 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
7836 unsigned long addr
, unsigned long len
,
7837 unsigned long pgoff
, unsigned long flags
)
7841 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
7843 return PTR_ERR(ptr
);
7845 return (unsigned long) ptr
;
7848 #endif /* !CONFIG_MMU */
7850 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
7851 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
7854 struct io_ring_ctx
*ctx
;
7859 if (current
->task_works
)
7862 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
7870 if (f
.file
->f_op
!= &io_uring_fops
)
7874 ctx
= f
.file
->private_data
;
7875 if (!percpu_ref_tryget(&ctx
->refs
))
7879 * For SQ polling, the thread will do all submissions and completions.
7880 * Just return the requested submit count, and wake the thread if
7884 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
7885 if (!list_empty_careful(&ctx
->cq_overflow_list
))
7886 io_cqring_overflow_flush(ctx
, false);
7887 if (flags
& IORING_ENTER_SQ_WAKEUP
)
7888 wake_up(&ctx
->sqo_wait
);
7889 submitted
= to_submit
;
7890 } else if (to_submit
) {
7891 mutex_lock(&ctx
->uring_lock
);
7892 submitted
= io_submit_sqes(ctx
, to_submit
, f
.file
, fd
);
7893 mutex_unlock(&ctx
->uring_lock
);
7895 if (submitted
!= to_submit
)
7898 if (flags
& IORING_ENTER_GETEVENTS
) {
7899 unsigned nr_events
= 0;
7901 min_complete
= min(min_complete
, ctx
->cq_entries
);
7904 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7905 * space applications don't need to do io completion events
7906 * polling again, they can rely on io_sq_thread to do polling
7907 * work, which can reduce cpu usage and uring_lock contention.
7909 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
7910 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
7911 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
7913 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
7918 percpu_ref_put(&ctx
->refs
);
7921 return submitted
? submitted
: ret
;
7924 #ifdef CONFIG_PROC_FS
7925 static int io_uring_show_cred(int id
, void *p
, void *data
)
7927 const struct cred
*cred
= p
;
7928 struct seq_file
*m
= data
;
7929 struct user_namespace
*uns
= seq_user_ns(m
);
7930 struct group_info
*gi
;
7935 seq_printf(m
, "%5d\n", id
);
7936 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
7937 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
7938 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
7939 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
7940 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
7941 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
7942 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
7943 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
7944 seq_puts(m
, "\n\tGroups:\t");
7945 gi
= cred
->group_info
;
7946 for (g
= 0; g
< gi
->ngroups
; g
++) {
7947 seq_put_decimal_ull(m
, g
? " " : "",
7948 from_kgid_munged(uns
, gi
->gid
[g
]));
7950 seq_puts(m
, "\n\tCapEff:\t");
7951 cap
= cred
->cap_effective
;
7952 CAP_FOR_EACH_U32(__capi
)
7953 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
7958 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
7962 mutex_lock(&ctx
->uring_lock
);
7963 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
7964 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7965 struct fixed_file_table
*table
;
7968 table
= &ctx
->file_data
->table
[i
>> IORING_FILE_TABLE_SHIFT
];
7969 f
= table
->files
[i
& IORING_FILE_TABLE_MASK
];
7971 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
7973 seq_printf(m
, "%5u: <none>\n", i
);
7975 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
7976 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
7977 struct io_mapped_ubuf
*buf
= &ctx
->user_bufs
[i
];
7979 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
,
7980 (unsigned int) buf
->len
);
7982 if (!idr_is_empty(&ctx
->personality_idr
)) {
7983 seq_printf(m
, "Personalities:\n");
7984 idr_for_each(&ctx
->personality_idr
, io_uring_show_cred
, m
);
7986 seq_printf(m
, "PollList:\n");
7987 spin_lock_irq(&ctx
->completion_lock
);
7988 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
7989 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
7990 struct io_kiocb
*req
;
7992 hlist_for_each_entry(req
, list
, hash_node
)
7993 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
7994 req
->task
->task_works
!= NULL
);
7996 spin_unlock_irq(&ctx
->completion_lock
);
7997 mutex_unlock(&ctx
->uring_lock
);
8000 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
8002 struct io_ring_ctx
*ctx
= f
->private_data
;
8004 if (percpu_ref_tryget(&ctx
->refs
)) {
8005 __io_uring_show_fdinfo(ctx
, m
);
8006 percpu_ref_put(&ctx
->refs
);
8011 static const struct file_operations io_uring_fops
= {
8012 .release
= io_uring_release
,
8013 .flush
= io_uring_flush
,
8014 .mmap
= io_uring_mmap
,
8016 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
8017 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
8019 .poll
= io_uring_poll
,
8020 .fasync
= io_uring_fasync
,
8021 #ifdef CONFIG_PROC_FS
8022 .show_fdinfo
= io_uring_show_fdinfo
,
8026 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
8027 struct io_uring_params
*p
)
8029 struct io_rings
*rings
;
8030 size_t size
, sq_array_offset
;
8032 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
8033 if (size
== SIZE_MAX
)
8036 rings
= io_mem_alloc(size
);
8041 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
8042 rings
->sq_ring_mask
= p
->sq_entries
- 1;
8043 rings
->cq_ring_mask
= p
->cq_entries
- 1;
8044 rings
->sq_ring_entries
= p
->sq_entries
;
8045 rings
->cq_ring_entries
= p
->cq_entries
;
8046 ctx
->sq_mask
= rings
->sq_ring_mask
;
8047 ctx
->cq_mask
= rings
->cq_ring_mask
;
8048 ctx
->sq_entries
= rings
->sq_ring_entries
;
8049 ctx
->cq_entries
= rings
->cq_ring_entries
;
8051 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
8052 if (size
== SIZE_MAX
) {
8053 io_mem_free(ctx
->rings
);
8058 ctx
->sq_sqes
= io_mem_alloc(size
);
8059 if (!ctx
->sq_sqes
) {
8060 io_mem_free(ctx
->rings
);
8069 * Allocate an anonymous fd, this is what constitutes the application
8070 * visible backing of an io_uring instance. The application mmaps this
8071 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8072 * we have to tie this fd to a socket for file garbage collection purposes.
8074 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
8079 #if defined(CONFIG_UNIX)
8080 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
8086 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
8090 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
8091 O_RDWR
| O_CLOEXEC
);
8094 ret
= PTR_ERR(file
);
8098 #if defined(CONFIG_UNIX)
8099 ctx
->ring_sock
->file
= file
;
8101 fd_install(ret
, file
);
8104 #if defined(CONFIG_UNIX)
8105 sock_release(ctx
->ring_sock
);
8106 ctx
->ring_sock
= NULL
;
8111 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
8112 struct io_uring_params __user
*params
)
8114 struct user_struct
*user
= NULL
;
8115 struct io_ring_ctx
*ctx
;
8121 if (entries
> IORING_MAX_ENTRIES
) {
8122 if (!(p
->flags
& IORING_SETUP_CLAMP
))
8124 entries
= IORING_MAX_ENTRIES
;
8128 * Use twice as many entries for the CQ ring. It's possible for the
8129 * application to drive a higher depth than the size of the SQ ring,
8130 * since the sqes are only used at submission time. This allows for
8131 * some flexibility in overcommitting a bit. If the application has
8132 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8133 * of CQ ring entries manually.
8135 p
->sq_entries
= roundup_pow_of_two(entries
);
8136 if (p
->flags
& IORING_SETUP_CQSIZE
) {
8138 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8139 * to a power-of-two, if it isn't already. We do NOT impose
8140 * any cq vs sq ring sizing.
8142 if (p
->cq_entries
< p
->sq_entries
)
8144 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
8145 if (!(p
->flags
& IORING_SETUP_CLAMP
))
8147 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
8149 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
8151 p
->cq_entries
= 2 * p
->sq_entries
;
8154 user
= get_uid(current_user());
8155 limit_mem
= !capable(CAP_IPC_LOCK
);
8158 ret
= __io_account_mem(user
,
8159 ring_pages(p
->sq_entries
, p
->cq_entries
));
8166 ctx
= io_ring_ctx_alloc(p
);
8169 __io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
8174 ctx
->compat
= in_compat_syscall();
8176 ctx
->creds
= get_current_cred();
8178 ret
= io_allocate_scq_urings(ctx
, p
);
8182 ret
= io_sq_offload_start(ctx
, p
);
8186 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
8187 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
8188 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
8189 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
8190 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
8191 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
8192 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
8193 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
8195 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
8196 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
8197 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
8198 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
8199 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
8200 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
8201 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
8202 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
8204 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
8205 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
8206 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
8207 IORING_FEAT_POLL_32BITS
;
8209 if (copy_to_user(params
, p
, sizeof(*p
))) {
8214 * Install ring fd as the very last thing, so we don't risk someone
8215 * having closed it before we finish setup
8217 ret
= io_uring_get_fd(ctx
);
8221 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
8222 io_account_mem(ctx
, ring_pages(p
->sq_entries
, p
->cq_entries
),
8224 ctx
->limit_mem
= limit_mem
;
8227 io_ring_ctx_wait_and_kill(ctx
);
8232 * Sets up an aio uring context, and returns the fd. Applications asks for a
8233 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8234 * params structure passed in.
8236 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
8238 struct io_uring_params p
;
8241 if (copy_from_user(&p
, params
, sizeof(p
)))
8243 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
8248 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
8249 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
8250 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
))
8253 return io_uring_create(entries
, &p
, params
);
8256 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
8257 struct io_uring_params __user
*, params
)
8259 return io_uring_setup(entries
, params
);
8262 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
8264 struct io_uring_probe
*p
;
8268 size
= struct_size(p
, ops
, nr_args
);
8269 if (size
== SIZE_MAX
)
8271 p
= kzalloc(size
, GFP_KERNEL
);
8276 if (copy_from_user(p
, arg
, size
))
8279 if (memchr_inv(p
, 0, size
))
8282 p
->last_op
= IORING_OP_LAST
- 1;
8283 if (nr_args
> IORING_OP_LAST
)
8284 nr_args
= IORING_OP_LAST
;
8286 for (i
= 0; i
< nr_args
; i
++) {
8288 if (!io_op_defs
[i
].not_supported
)
8289 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
8294 if (copy_to_user(arg
, p
, size
))
8301 static int io_register_personality(struct io_ring_ctx
*ctx
)
8303 const struct cred
*creds
= get_current_cred();
8306 id
= idr_alloc_cyclic(&ctx
->personality_idr
, (void *) creds
, 1,
8307 USHRT_MAX
, GFP_KERNEL
);
8313 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
8315 const struct cred
*old_creds
;
8317 old_creds
= idr_remove(&ctx
->personality_idr
, id
);
8319 put_cred(old_creds
);
8326 static bool io_register_op_must_quiesce(int op
)
8329 case IORING_UNREGISTER_FILES
:
8330 case IORING_REGISTER_FILES_UPDATE
:
8331 case IORING_REGISTER_PROBE
:
8332 case IORING_REGISTER_PERSONALITY
:
8333 case IORING_UNREGISTER_PERSONALITY
:
8340 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
8341 void __user
*arg
, unsigned nr_args
)
8342 __releases(ctx
->uring_lock
)
8343 __acquires(ctx
->uring_lock
)
8348 * We're inside the ring mutex, if the ref is already dying, then
8349 * someone else killed the ctx or is already going through
8350 * io_uring_register().
8352 if (percpu_ref_is_dying(&ctx
->refs
))
8355 if (io_register_op_must_quiesce(opcode
)) {
8356 percpu_ref_kill(&ctx
->refs
);
8359 * Drop uring mutex before waiting for references to exit. If
8360 * another thread is currently inside io_uring_enter() it might
8361 * need to grab the uring_lock to make progress. If we hold it
8362 * here across the drain wait, then we can deadlock. It's safe
8363 * to drop the mutex here, since no new references will come in
8364 * after we've killed the percpu ref.
8366 mutex_unlock(&ctx
->uring_lock
);
8367 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
8368 mutex_lock(&ctx
->uring_lock
);
8370 percpu_ref_resurrect(&ctx
->refs
);
8377 case IORING_REGISTER_BUFFERS
:
8378 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
8380 case IORING_UNREGISTER_BUFFERS
:
8384 ret
= io_sqe_buffer_unregister(ctx
);
8386 case IORING_REGISTER_FILES
:
8387 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
8389 case IORING_UNREGISTER_FILES
:
8393 ret
= io_sqe_files_unregister(ctx
);
8395 case IORING_REGISTER_FILES_UPDATE
:
8396 ret
= io_sqe_files_update(ctx
, arg
, nr_args
);
8398 case IORING_REGISTER_EVENTFD
:
8399 case IORING_REGISTER_EVENTFD_ASYNC
:
8403 ret
= io_eventfd_register(ctx
, arg
);
8406 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
8407 ctx
->eventfd_async
= 1;
8409 ctx
->eventfd_async
= 0;
8411 case IORING_UNREGISTER_EVENTFD
:
8415 ret
= io_eventfd_unregister(ctx
);
8417 case IORING_REGISTER_PROBE
:
8419 if (!arg
|| nr_args
> 256)
8421 ret
= io_probe(ctx
, arg
, nr_args
);
8423 case IORING_REGISTER_PERSONALITY
:
8427 ret
= io_register_personality(ctx
);
8429 case IORING_UNREGISTER_PERSONALITY
:
8433 ret
= io_unregister_personality(ctx
, nr_args
);
8440 if (io_register_op_must_quiesce(opcode
)) {
8441 /* bring the ctx back to life */
8442 percpu_ref_reinit(&ctx
->refs
);
8444 reinit_completion(&ctx
->ref_comp
);
8449 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
8450 void __user
*, arg
, unsigned int, nr_args
)
8452 struct io_ring_ctx
*ctx
;
8461 if (f
.file
->f_op
!= &io_uring_fops
)
8464 ctx
= f
.file
->private_data
;
8466 mutex_lock(&ctx
->uring_lock
);
8467 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
8468 mutex_unlock(&ctx
->uring_lock
);
8469 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
8470 ctx
->cq_ev_fd
!= NULL
, ret
);
8476 static int __init
io_uring_init(void)
8478 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8479 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8480 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8483 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8484 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8485 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
8486 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
8487 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
8488 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
8489 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
8490 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
8491 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
8492 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
8493 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
8494 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
8495 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
8496 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
8497 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
8498 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
8499 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
8500 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
8501 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
8502 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
8503 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
8504 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
8505 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
8506 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
8507 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
8508 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
8509 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
8510 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
8511 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
8512 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
8513 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
8515 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
8516 BUILD_BUG_ON(__REQ_F_LAST_BIT
>= 8 * sizeof(int));
8517 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
8520 __initcall(io_uring_init
);