1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/tracehook.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
93 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
96 #define IORING_MAX_FIXED_FILES (1U << 15)
97 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
98 IORING_REGISTER_LAST + IORING_OP_LAST)
100 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
101 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
102 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
104 #define IORING_MAX_REG_BUFFERS (1U << 14)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
112 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
115 u32 head ____cacheline_aligned_in_smp
;
116 u32 tail ____cacheline_aligned_in_smp
;
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
128 * Head and tail offsets into the ring; the offsets need to be
129 * masked to get valid indices.
131 * The kernel controls head of the sq ring and the tail of the cq ring,
132 * and the application controls tail of the sq ring and the head of the
135 struct io_uring sq
, cq
;
137 * Bitmasks to apply to head and tail offsets (constant, equals
140 u32 sq_ring_mask
, cq_ring_mask
;
141 /* Ring sizes (constant, power of 2) */
142 u32 sq_ring_entries
, cq_ring_entries
;
144 * Number of invalid entries dropped by the kernel due to
145 * invalid index stored in array
147 * Written by the kernel, shouldn't be modified by the
148 * application (i.e. get number of "new events" by comparing to
151 * After a new SQ head value was read by the application this
152 * counter includes all submissions that were dropped reaching
153 * the new SQ head (and possibly more).
159 * Written by the kernel, shouldn't be modified by the
162 * The application needs a full memory barrier before checking
163 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
169 * Written by the application, shouldn't be modified by the
174 * Number of completion events lost because the queue was full;
175 * this should be avoided by the application by making sure
176 * there are not more requests pending than there is space in
177 * the completion queue.
179 * Written by the kernel, shouldn't be modified by the
180 * application (i.e. get number of "new events" by comparing to
183 * As completion events come in out of order this counter is not
184 * ordered with any other data.
188 * Ring buffer of completion events.
190 * The kernel writes completion events fresh every time they are
191 * produced, so the application is allowed to modify pending
194 struct io_uring_cqe cqes
[] ____cacheline_aligned_in_smp
;
197 enum io_uring_cmd_flags
{
198 IO_URING_F_NONBLOCK
= 1,
199 IO_URING_F_COMPLETE_DEFER
= 2,
202 struct io_mapped_ubuf
{
205 unsigned int nr_bvecs
;
206 unsigned long acct_pages
;
207 struct bio_vec bvec
[];
212 struct io_overflow_cqe
{
213 struct io_uring_cqe cqe
;
214 struct list_head list
;
217 struct io_fixed_file
{
218 /* file * with additional FFS_* flags */
219 unsigned long file_ptr
;
223 struct list_head list
;
228 struct io_mapped_ubuf
*buf
;
232 struct io_file_table
{
233 struct io_fixed_file
*files
;
236 struct io_rsrc_node
{
237 struct percpu_ref refs
;
238 struct list_head node
;
239 struct list_head rsrc_list
;
240 struct io_rsrc_data
*rsrc_data
;
241 struct llist_node llist
;
245 typedef void (rsrc_put_fn
)(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
);
247 struct io_rsrc_data
{
248 struct io_ring_ctx
*ctx
;
254 struct completion done
;
259 struct list_head list
;
265 struct io_restriction
{
266 DECLARE_BITMAP(register_op
, IORING_REGISTER_LAST
);
267 DECLARE_BITMAP(sqe_op
, IORING_OP_LAST
);
268 u8 sqe_flags_allowed
;
269 u8 sqe_flags_required
;
274 IO_SQ_THREAD_SHOULD_STOP
= 0,
275 IO_SQ_THREAD_SHOULD_PARK
,
280 atomic_t park_pending
;
283 /* ctx's that are using this sqd */
284 struct list_head ctx_list
;
286 struct task_struct
*thread
;
287 struct wait_queue_head wait
;
289 unsigned sq_thread_idle
;
295 struct completion exited
;
298 #define IO_COMPL_BATCH 32
299 #define IO_REQ_CACHE_SIZE 32
300 #define IO_REQ_ALLOC_BATCH 8
302 struct io_submit_link
{
303 struct io_kiocb
*head
;
304 struct io_kiocb
*last
;
307 struct io_submit_state
{
308 struct blk_plug plug
;
309 struct io_submit_link link
;
312 * io_kiocb alloc cache
314 void *reqs
[IO_REQ_CACHE_SIZE
];
315 unsigned int free_reqs
;
320 * Batch completion logic
322 struct io_kiocb
*compl_reqs
[IO_COMPL_BATCH
];
323 unsigned int compl_nr
;
324 /* inline/task_work completion list, under ->uring_lock */
325 struct list_head free_list
;
327 unsigned int ios_left
;
331 /* const or read-mostly hot data */
333 struct percpu_ref refs
;
335 struct io_rings
*rings
;
337 unsigned int compat
: 1;
338 unsigned int drain_next
: 1;
339 unsigned int eventfd_async
: 1;
340 unsigned int restricted
: 1;
341 unsigned int off_timeout_used
: 1;
342 unsigned int drain_active
: 1;
343 } ____cacheline_aligned_in_smp
;
345 /* submission data */
347 struct mutex uring_lock
;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 struct io_uring_sqe
*sq_sqes
;
362 unsigned cached_sq_head
;
364 struct list_head defer_list
;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node
*rsrc_node
;
371 struct io_file_table file_table
;
372 unsigned nr_user_files
;
373 unsigned nr_user_bufs
;
374 struct io_mapped_ubuf
**user_bufs
;
376 struct io_submit_state submit_state
;
377 struct list_head timeout_list
;
378 struct list_head ltimeout_list
;
379 struct list_head cq_overflow_list
;
380 struct xarray io_buffers
;
381 struct xarray personalities
;
383 unsigned sq_thread_idle
;
384 } ____cacheline_aligned_in_smp
;
386 /* IRQ completion list, under ->completion_lock */
387 struct list_head locked_free_list
;
388 unsigned int locked_free_nr
;
390 const struct cred
*sq_creds
; /* cred used for __io_sq_thread() */
391 struct io_sq_data
*sq_data
; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait
;
394 struct list_head sqd_list
;
396 unsigned long check_cq_overflow
;
399 unsigned cached_cq_tail
;
401 struct eventfd_ctx
*cq_ev_fd
;
402 struct wait_queue_head poll_wait
;
403 struct wait_queue_head cq_wait
;
405 atomic_t cq_timeouts
;
406 unsigned cq_last_tm_flush
;
407 } ____cacheline_aligned_in_smp
;
410 spinlock_t completion_lock
;
412 spinlock_t timeout_lock
;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list
;
421 struct hlist_head
*cancel_hash
;
422 unsigned cancel_hash_bits
;
423 bool poll_multi_queue
;
424 } ____cacheline_aligned_in_smp
;
426 struct io_restriction restrictions
;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node
*rsrc_backup_node
;
431 struct io_mapped_ubuf
*dummy_ubuf
;
432 struct io_rsrc_data
*file_data
;
433 struct io_rsrc_data
*buf_data
;
435 struct delayed_work rsrc_put_work
;
436 struct llist_head rsrc_put_llist
;
437 struct list_head rsrc_ref_list
;
438 spinlock_t rsrc_ref_lock
;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket
*ring_sock
;
446 /* hashed buffered write serialization */
447 struct io_wq_hash
*hash_map
;
449 /* Only used for accounting purposes */
450 struct user_struct
*user
;
451 struct mm_struct
*mm_account
;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist
;
455 struct delayed_work fallback_work
;
456 struct work_struct exit_work
;
457 struct list_head tctx_list
;
458 struct completion ref_comp
;
460 bool iowq_limits_set
;
464 struct io_uring_task
{
465 /* submission side */
468 struct wait_queue_head wait
;
469 const struct io_ring_ctx
*last
;
471 struct percpu_counter inflight
;
472 atomic_t inflight_tracked
;
475 spinlock_t task_lock
;
476 struct io_wq_work_list task_list
;
477 struct callback_head task_work
;
482 * First field must be the file pointer in all the
483 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
485 struct io_poll_iocb
{
487 struct wait_queue_head
*head
;
489 struct wait_queue_entry wait
;
492 struct io_poll_update
{
498 bool update_user_data
;
507 struct io_timeout_data
{
508 struct io_kiocb
*req
;
509 struct hrtimer timer
;
510 struct timespec64 ts
;
511 enum hrtimer_mode mode
;
517 struct sockaddr __user
*addr
;
518 int __user
*addr_len
;
521 unsigned long nofile
;
541 struct list_head list
;
542 /* head of the link, used by linked timeouts only */
543 struct io_kiocb
*head
;
544 /* for linked completions */
545 struct io_kiocb
*prev
;
548 struct io_timeout_rem
{
553 struct timespec64 ts
;
559 /* NOTE: kiocb has the file as the first member, so don't do it here */
567 struct sockaddr __user
*addr
;
574 struct compat_msghdr __user
*umsg_compat
;
575 struct user_msghdr __user
*umsg
;
581 struct io_buffer
*kbuf
;
588 struct filename
*filename
;
590 unsigned long nofile
;
593 struct io_rsrc_update
{
619 struct epoll_event event
;
623 struct file
*file_out
;
631 struct io_provide_buf
{
645 const char __user
*filename
;
646 struct statx __user
*buffer
;
658 struct filename
*oldpath
;
659 struct filename
*newpath
;
667 struct filename
*filename
;
674 struct filename
*filename
;
680 struct filename
*oldpath
;
681 struct filename
*newpath
;
688 struct filename
*oldpath
;
689 struct filename
*newpath
;
693 struct io_completion
{
698 struct io_async_connect
{
699 struct sockaddr_storage address
;
702 struct io_async_msghdr
{
703 struct iovec fast_iov
[UIO_FASTIOV
];
704 /* points to an allocated iov, if NULL we use fast_iov instead */
705 struct iovec
*free_iov
;
706 struct sockaddr __user
*uaddr
;
708 struct sockaddr_storage addr
;
712 struct iovec fast_iov
[UIO_FASTIOV
];
713 const struct iovec
*free_iovec
;
714 struct iov_iter iter
;
715 struct iov_iter_state iter_state
;
717 struct wait_page_queue wpq
;
721 REQ_F_FIXED_FILE_BIT
= IOSQE_FIXED_FILE_BIT
,
722 REQ_F_IO_DRAIN_BIT
= IOSQE_IO_DRAIN_BIT
,
723 REQ_F_LINK_BIT
= IOSQE_IO_LINK_BIT
,
724 REQ_F_HARDLINK_BIT
= IOSQE_IO_HARDLINK_BIT
,
725 REQ_F_FORCE_ASYNC_BIT
= IOSQE_ASYNC_BIT
,
726 REQ_F_BUFFER_SELECT_BIT
= IOSQE_BUFFER_SELECT_BIT
,
728 /* first byte is taken by user flags, shift it to not overlap */
733 REQ_F_LINK_TIMEOUT_BIT
,
734 REQ_F_NEED_CLEANUP_BIT
,
736 REQ_F_BUFFER_SELECTED_BIT
,
737 REQ_F_COMPLETE_INLINE_BIT
,
741 REQ_F_ARM_LTIMEOUT_BIT
,
742 /* keep async read/write and isreg together and in order */
743 REQ_F_NOWAIT_READ_BIT
,
744 REQ_F_NOWAIT_WRITE_BIT
,
747 /* not a real bit, just to check we're not overflowing the space */
753 REQ_F_FIXED_FILE
= BIT(REQ_F_FIXED_FILE_BIT
),
754 /* drain existing IO first */
755 REQ_F_IO_DRAIN
= BIT(REQ_F_IO_DRAIN_BIT
),
757 REQ_F_LINK
= BIT(REQ_F_LINK_BIT
),
758 /* doesn't sever on completion < 0 */
759 REQ_F_HARDLINK
= BIT(REQ_F_HARDLINK_BIT
),
761 REQ_F_FORCE_ASYNC
= BIT(REQ_F_FORCE_ASYNC_BIT
),
762 /* IOSQE_BUFFER_SELECT */
763 REQ_F_BUFFER_SELECT
= BIT(REQ_F_BUFFER_SELECT_BIT
),
765 /* fail rest of links */
766 REQ_F_FAIL
= BIT(REQ_F_FAIL_BIT
),
767 /* on inflight list, should be cancelled and waited on exit reliably */
768 REQ_F_INFLIGHT
= BIT(REQ_F_INFLIGHT_BIT
),
769 /* read/write uses file position */
770 REQ_F_CUR_POS
= BIT(REQ_F_CUR_POS_BIT
),
771 /* must not punt to workers */
772 REQ_F_NOWAIT
= BIT(REQ_F_NOWAIT_BIT
),
773 /* has or had linked timeout */
774 REQ_F_LINK_TIMEOUT
= BIT(REQ_F_LINK_TIMEOUT_BIT
),
776 REQ_F_NEED_CLEANUP
= BIT(REQ_F_NEED_CLEANUP_BIT
),
777 /* already went through poll handler */
778 REQ_F_POLLED
= BIT(REQ_F_POLLED_BIT
),
779 /* buffer already selected */
780 REQ_F_BUFFER_SELECTED
= BIT(REQ_F_BUFFER_SELECTED_BIT
),
781 /* completion is deferred through io_comp_state */
782 REQ_F_COMPLETE_INLINE
= BIT(REQ_F_COMPLETE_INLINE_BIT
),
783 /* caller should reissue async */
784 REQ_F_REISSUE
= BIT(REQ_F_REISSUE_BIT
),
785 /* supports async reads */
786 REQ_F_NOWAIT_READ
= BIT(REQ_F_NOWAIT_READ_BIT
),
787 /* supports async writes */
788 REQ_F_NOWAIT_WRITE
= BIT(REQ_F_NOWAIT_WRITE_BIT
),
790 REQ_F_ISREG
= BIT(REQ_F_ISREG_BIT
),
791 /* has creds assigned */
792 REQ_F_CREDS
= BIT(REQ_F_CREDS_BIT
),
793 /* skip refcounting if not set */
794 REQ_F_REFCOUNT
= BIT(REQ_F_REFCOUNT_BIT
),
795 /* there is a linked timeout that has to be armed */
796 REQ_F_ARM_LTIMEOUT
= BIT(REQ_F_ARM_LTIMEOUT_BIT
),
800 struct io_poll_iocb poll
;
801 struct io_poll_iocb
*double_poll
;
804 typedef void (*io_req_tw_func_t
)(struct io_kiocb
*req
, bool *locked
);
806 struct io_task_work
{
808 struct io_wq_work_node node
;
809 struct llist_node fallback_node
;
811 io_req_tw_func_t func
;
815 IORING_RSRC_FILE
= 0,
816 IORING_RSRC_BUFFER
= 1,
820 * NOTE! Each of the iocb union members has the file pointer
821 * as the first entry in their struct definition. So you can
822 * access the file pointer through any of the sub-structs,
823 * or directly as just 'ki_filp' in this struct.
829 struct io_poll_iocb poll
;
830 struct io_poll_update poll_update
;
831 struct io_accept accept
;
833 struct io_cancel cancel
;
834 struct io_timeout timeout
;
835 struct io_timeout_rem timeout_rem
;
836 struct io_connect connect
;
837 struct io_sr_msg sr_msg
;
839 struct io_close close
;
840 struct io_rsrc_update rsrc_update
;
841 struct io_fadvise fadvise
;
842 struct io_madvise madvise
;
843 struct io_epoll epoll
;
844 struct io_splice splice
;
845 struct io_provide_buf pbuf
;
846 struct io_statx statx
;
847 struct io_shutdown shutdown
;
848 struct io_rename rename
;
849 struct io_unlink unlink
;
850 struct io_mkdir mkdir
;
851 struct io_symlink symlink
;
852 struct io_hardlink hardlink
;
853 /* use only after cleaning per-op data, see io_clean_op() */
854 struct io_completion
compl;
857 /* opcode allocated if it needs to store data for async defer */
860 /* polled IO has completed */
866 struct io_ring_ctx
*ctx
;
869 struct task_struct
*task
;
872 struct io_kiocb
*link
;
873 struct percpu_ref
*fixed_rsrc_refs
;
875 /* used with ctx->iopoll_list with reads/writes */
876 struct list_head inflight_entry
;
877 struct io_task_work io_task_work
;
878 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
879 struct hlist_node hash_node
;
880 struct async_poll
*apoll
;
881 struct io_wq_work work
;
882 const struct cred
*creds
;
884 /* store used ubuf, so we can prevent reloading */
885 struct io_mapped_ubuf
*imu
;
886 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
887 struct io_buffer
*kbuf
;
891 struct io_tctx_node
{
892 struct list_head ctx_node
;
893 struct task_struct
*task
;
894 struct io_ring_ctx
*ctx
;
897 struct io_defer_entry
{
898 struct list_head list
;
899 struct io_kiocb
*req
;
904 /* needs req->file assigned */
905 unsigned needs_file
: 1;
906 /* hash wq insertion if file is a regular file */
907 unsigned hash_reg_file
: 1;
908 /* unbound wq insertion if file is a non-regular file */
909 unsigned unbound_nonreg_file
: 1;
910 /* opcode is not supported by this kernel */
911 unsigned not_supported
: 1;
912 /* set if opcode supports polled "wait" */
914 unsigned pollout
: 1;
915 /* op supports buffer selection */
916 unsigned buffer_select
: 1;
917 /* do prep async if is going to be punted */
918 unsigned needs_async_setup
: 1;
919 /* should block plug */
921 /* size of async data needed, if any */
922 unsigned short async_size
;
925 static const struct io_op_def io_op_defs
[] = {
926 [IORING_OP_NOP
] = {},
927 [IORING_OP_READV
] = {
929 .unbound_nonreg_file
= 1,
932 .needs_async_setup
= 1,
934 .async_size
= sizeof(struct io_async_rw
),
936 [IORING_OP_WRITEV
] = {
939 .unbound_nonreg_file
= 1,
941 .needs_async_setup
= 1,
943 .async_size
= sizeof(struct io_async_rw
),
945 [IORING_OP_FSYNC
] = {
948 [IORING_OP_READ_FIXED
] = {
950 .unbound_nonreg_file
= 1,
953 .async_size
= sizeof(struct io_async_rw
),
955 [IORING_OP_WRITE_FIXED
] = {
958 .unbound_nonreg_file
= 1,
961 .async_size
= sizeof(struct io_async_rw
),
963 [IORING_OP_POLL_ADD
] = {
965 .unbound_nonreg_file
= 1,
967 [IORING_OP_POLL_REMOVE
] = {},
968 [IORING_OP_SYNC_FILE_RANGE
] = {
971 [IORING_OP_SENDMSG
] = {
973 .unbound_nonreg_file
= 1,
975 .needs_async_setup
= 1,
976 .async_size
= sizeof(struct io_async_msghdr
),
978 [IORING_OP_RECVMSG
] = {
980 .unbound_nonreg_file
= 1,
983 .needs_async_setup
= 1,
984 .async_size
= sizeof(struct io_async_msghdr
),
986 [IORING_OP_TIMEOUT
] = {
987 .async_size
= sizeof(struct io_timeout_data
),
989 [IORING_OP_TIMEOUT_REMOVE
] = {
990 /* used by timeout updates' prep() */
992 [IORING_OP_ACCEPT
] = {
994 .unbound_nonreg_file
= 1,
997 [IORING_OP_ASYNC_CANCEL
] = {},
998 [IORING_OP_LINK_TIMEOUT
] = {
999 .async_size
= sizeof(struct io_timeout_data
),
1001 [IORING_OP_CONNECT
] = {
1003 .unbound_nonreg_file
= 1,
1005 .needs_async_setup
= 1,
1006 .async_size
= sizeof(struct io_async_connect
),
1008 [IORING_OP_FALLOCATE
] = {
1011 [IORING_OP_OPENAT
] = {},
1012 [IORING_OP_CLOSE
] = {},
1013 [IORING_OP_FILES_UPDATE
] = {},
1014 [IORING_OP_STATX
] = {},
1015 [IORING_OP_READ
] = {
1017 .unbound_nonreg_file
= 1,
1021 .async_size
= sizeof(struct io_async_rw
),
1023 [IORING_OP_WRITE
] = {
1026 .unbound_nonreg_file
= 1,
1029 .async_size
= sizeof(struct io_async_rw
),
1031 [IORING_OP_FADVISE
] = {
1034 [IORING_OP_MADVISE
] = {},
1035 [IORING_OP_SEND
] = {
1037 .unbound_nonreg_file
= 1,
1040 [IORING_OP_RECV
] = {
1042 .unbound_nonreg_file
= 1,
1046 [IORING_OP_OPENAT2
] = {
1048 [IORING_OP_EPOLL_CTL
] = {
1049 .unbound_nonreg_file
= 1,
1051 [IORING_OP_SPLICE
] = {
1054 .unbound_nonreg_file
= 1,
1056 [IORING_OP_PROVIDE_BUFFERS
] = {},
1057 [IORING_OP_REMOVE_BUFFERS
] = {},
1061 .unbound_nonreg_file
= 1,
1063 [IORING_OP_SHUTDOWN
] = {
1066 [IORING_OP_RENAMEAT
] = {},
1067 [IORING_OP_UNLINKAT
] = {},
1068 [IORING_OP_MKDIRAT
] = {},
1069 [IORING_OP_SYMLINKAT
] = {},
1070 [IORING_OP_LINKAT
] = {},
1073 /* requests with any of those set should undergo io_disarm_next() */
1074 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1076 static bool io_disarm_next(struct io_kiocb
*req
);
1077 static void io_uring_del_tctx_node(unsigned long index
);
1078 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
1079 struct task_struct
*task
,
1081 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
);
1083 static void io_fill_cqe_req(struct io_kiocb
*req
, s32 res
, u32 cflags
);
1085 static void io_put_req(struct io_kiocb
*req
);
1086 static void io_put_req_deferred(struct io_kiocb
*req
);
1087 static void io_dismantle_req(struct io_kiocb
*req
);
1088 static void io_queue_linked_timeout(struct io_kiocb
*req
);
1089 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
1090 struct io_uring_rsrc_update2
*up
,
1092 static void io_clean_op(struct io_kiocb
*req
);
1093 static struct file
*io_file_get(struct io_ring_ctx
*ctx
,
1094 struct io_kiocb
*req
, int fd
, bool fixed
);
1095 static void __io_queue_sqe(struct io_kiocb
*req
);
1096 static void io_rsrc_put_work(struct work_struct
*work
);
1098 static void io_req_task_queue(struct io_kiocb
*req
);
1099 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
);
1100 static int io_req_prep_async(struct io_kiocb
*req
);
1102 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
1103 unsigned int issue_flags
, u32 slot_index
);
1104 static int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
);
1106 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
);
1108 static struct kmem_cache
*req_cachep
;
1110 static const struct file_operations io_uring_fops
;
1112 struct sock
*io_uring_get_socket(struct file
*file
)
1114 #if defined(CONFIG_UNIX)
1115 if (file
->f_op
== &io_uring_fops
) {
1116 struct io_ring_ctx
*ctx
= file
->private_data
;
1118 return ctx
->ring_sock
->sk
;
1123 EXPORT_SYMBOL(io_uring_get_socket
);
1125 static inline void io_tw_lock(struct io_ring_ctx
*ctx
, bool *locked
)
1128 mutex_lock(&ctx
->uring_lock
);
1133 #define io_for_each_link(pos, head) \
1134 for (pos = (head); pos; pos = pos->link)
1137 * Shamelessly stolen from the mm implementation of page reference checking,
1138 * see commit f958d7b528b1 for details.
1140 #define req_ref_zero_or_close_to_overflow(req) \
1141 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1143 static inline bool req_ref_inc_not_zero(struct io_kiocb
*req
)
1145 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1146 return atomic_inc_not_zero(&req
->refs
);
1149 static inline bool req_ref_put_and_test(struct io_kiocb
*req
)
1151 if (likely(!(req
->flags
& REQ_F_REFCOUNT
)))
1154 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1155 return atomic_dec_and_test(&req
->refs
);
1158 static inline void req_ref_put(struct io_kiocb
*req
)
1160 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1161 WARN_ON_ONCE(req_ref_put_and_test(req
));
1164 static inline void req_ref_get(struct io_kiocb
*req
)
1166 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1167 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1168 atomic_inc(&req
->refs
);
1171 static inline void __io_req_set_refcount(struct io_kiocb
*req
, int nr
)
1173 if (!(req
->flags
& REQ_F_REFCOUNT
)) {
1174 req
->flags
|= REQ_F_REFCOUNT
;
1175 atomic_set(&req
->refs
, nr
);
1179 static inline void io_req_set_refcount(struct io_kiocb
*req
)
1181 __io_req_set_refcount(req
, 1);
1184 static inline void io_req_set_rsrc_node(struct io_kiocb
*req
)
1186 struct io_ring_ctx
*ctx
= req
->ctx
;
1188 if (!req
->fixed_rsrc_refs
) {
1189 req
->fixed_rsrc_refs
= &ctx
->rsrc_node
->refs
;
1190 percpu_ref_get(req
->fixed_rsrc_refs
);
1194 static void io_refs_resurrect(struct percpu_ref
*ref
, struct completion
*compl)
1196 bool got
= percpu_ref_tryget(ref
);
1198 /* already at zero, wait for ->release() */
1200 wait_for_completion(compl);
1201 percpu_ref_resurrect(ref
);
1203 percpu_ref_put(ref
);
1206 static bool io_match_task(struct io_kiocb
*head
, struct task_struct
*task
,
1208 __must_hold(&req
->ctx
->timeout_lock
)
1210 struct io_kiocb
*req
;
1212 if (task
&& head
->task
!= task
)
1217 io_for_each_link(req
, head
) {
1218 if (req
->flags
& REQ_F_INFLIGHT
)
1224 static bool io_match_linked(struct io_kiocb
*head
)
1226 struct io_kiocb
*req
;
1228 io_for_each_link(req
, head
) {
1229 if (req
->flags
& REQ_F_INFLIGHT
)
1236 * As io_match_task() but protected against racing with linked timeouts.
1237 * User must not hold timeout_lock.
1239 static bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
1244 if (task
&& head
->task
!= task
)
1249 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
1250 struct io_ring_ctx
*ctx
= head
->ctx
;
1252 /* protect against races with linked timeouts */
1253 spin_lock_irq(&ctx
->timeout_lock
);
1254 matched
= io_match_linked(head
);
1255 spin_unlock_irq(&ctx
->timeout_lock
);
1257 matched
= io_match_linked(head
);
1262 static inline void req_set_fail(struct io_kiocb
*req
)
1264 req
->flags
|= REQ_F_FAIL
;
1267 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
1273 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1275 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1277 complete(&ctx
->ref_comp
);
1280 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1282 return !req
->timeout
.off
;
1285 static void io_fallback_req_func(struct work_struct
*work
)
1287 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
1288 fallback_work
.work
);
1289 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
1290 struct io_kiocb
*req
, *tmp
;
1291 bool locked
= false;
1293 percpu_ref_get(&ctx
->refs
);
1294 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.fallback_node
)
1295 req
->io_task_work
.func(req
, &locked
);
1298 if (ctx
->submit_state
.compl_nr
)
1299 io_submit_flush_completions(ctx
);
1300 mutex_unlock(&ctx
->uring_lock
);
1302 percpu_ref_put(&ctx
->refs
);
1306 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1308 struct io_ring_ctx
*ctx
;
1311 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1316 * Use 5 bits less than the max cq entries, that should give us around
1317 * 32 entries per hash list if totally full and uniformly spread.
1319 hash_bits
= ilog2(p
->cq_entries
);
1323 ctx
->cancel_hash_bits
= hash_bits
;
1324 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1326 if (!ctx
->cancel_hash
)
1328 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1330 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
1331 if (!ctx
->dummy_ubuf
)
1333 /* set invalid range, so io_import_fixed() fails meeting it */
1334 ctx
->dummy_ubuf
->ubuf
= -1UL;
1336 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1337 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1340 ctx
->flags
= p
->flags
;
1341 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1342 INIT_LIST_HEAD(&ctx
->sqd_list
);
1343 init_waitqueue_head(&ctx
->poll_wait
);
1344 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1345 init_completion(&ctx
->ref_comp
);
1346 xa_init_flags(&ctx
->io_buffers
, XA_FLAGS_ALLOC1
);
1347 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1348 mutex_init(&ctx
->uring_lock
);
1349 init_waitqueue_head(&ctx
->cq_wait
);
1350 spin_lock_init(&ctx
->completion_lock
);
1351 spin_lock_init(&ctx
->timeout_lock
);
1352 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1353 INIT_LIST_HEAD(&ctx
->defer_list
);
1354 INIT_LIST_HEAD(&ctx
->timeout_list
);
1355 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
1356 spin_lock_init(&ctx
->rsrc_ref_lock
);
1357 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1358 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1359 init_llist_head(&ctx
->rsrc_put_llist
);
1360 INIT_LIST_HEAD(&ctx
->tctx_list
);
1361 INIT_LIST_HEAD(&ctx
->submit_state
.free_list
);
1362 INIT_LIST_HEAD(&ctx
->locked_free_list
);
1363 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
1366 kfree(ctx
->dummy_ubuf
);
1367 kfree(ctx
->cancel_hash
);
1372 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
1374 struct io_rings
*r
= ctx
->rings
;
1376 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
1380 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1382 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1383 struct io_ring_ctx
*ctx
= req
->ctx
;
1385 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
1391 #define FFS_ASYNC_READ 0x1UL
1392 #define FFS_ASYNC_WRITE 0x2UL
1394 #define FFS_ISREG 0x4UL
1396 #define FFS_ISREG 0x0UL
1398 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1400 static inline bool io_req_ffs_set(struct io_kiocb
*req
)
1402 return IS_ENABLED(CONFIG_64BIT
) && (req
->flags
& REQ_F_FIXED_FILE
);
1405 static void io_req_track_inflight(struct io_kiocb
*req
)
1407 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1408 req
->flags
|= REQ_F_INFLIGHT
;
1409 atomic_inc(¤t
->io_uring
->inflight_tracked
);
1413 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
1415 if (WARN_ON_ONCE(!req
->link
))
1418 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
1419 req
->flags
|= REQ_F_LINK_TIMEOUT
;
1421 /* linked timeouts should have two refs once prep'ed */
1422 io_req_set_refcount(req
);
1423 __io_req_set_refcount(req
->link
, 2);
1427 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
1429 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
1431 return __io_prep_linked_timeout(req
);
1434 static void io_prep_async_work(struct io_kiocb
*req
)
1436 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1437 struct io_ring_ctx
*ctx
= req
->ctx
;
1439 if (!(req
->flags
& REQ_F_CREDS
)) {
1440 req
->flags
|= REQ_F_CREDS
;
1441 req
->creds
= get_current_cred();
1444 req
->work
.list
.next
= NULL
;
1445 req
->work
.flags
= 0;
1446 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1447 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1449 if (req
->flags
& REQ_F_ISREG
) {
1450 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1451 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1452 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
1453 if (def
->unbound_nonreg_file
)
1454 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1458 static void io_prep_async_link(struct io_kiocb
*req
)
1460 struct io_kiocb
*cur
;
1462 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
1463 struct io_ring_ctx
*ctx
= req
->ctx
;
1465 spin_lock_irq(&ctx
->timeout_lock
);
1466 io_for_each_link(cur
, req
)
1467 io_prep_async_work(cur
);
1468 spin_unlock_irq(&ctx
->timeout_lock
);
1470 io_for_each_link(cur
, req
)
1471 io_prep_async_work(cur
);
1475 static void io_queue_async_work(struct io_kiocb
*req
, bool *locked
)
1477 struct io_ring_ctx
*ctx
= req
->ctx
;
1478 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1479 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1481 /* must not take the lock, NULL it as a precaution */
1485 BUG_ON(!tctx
->io_wq
);
1487 /* init ->work of the whole link before punting */
1488 io_prep_async_link(req
);
1491 * Not expected to happen, but if we do have a bug where this _can_
1492 * happen, catch it here and ensure the request is marked as
1493 * canceled. That will make io-wq go through the usual work cancel
1494 * procedure rather than attempt to run this request (or create a new
1497 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
1498 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1500 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1501 &req
->work
, req
->flags
);
1502 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
1504 io_queue_linked_timeout(link
);
1507 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1508 __must_hold(&req
->ctx
->completion_lock
)
1509 __must_hold(&req
->ctx
->timeout_lock
)
1511 struct io_timeout_data
*io
= req
->async_data
;
1513 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1516 atomic_set(&req
->ctx
->cq_timeouts
,
1517 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1518 list_del_init(&req
->timeout
.list
);
1519 io_fill_cqe_req(req
, status
, 0);
1520 io_put_req_deferred(req
);
1524 static void io_queue_deferred(struct io_ring_ctx
*ctx
)
1526 while (!list_empty(&ctx
->defer_list
)) {
1527 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1528 struct io_defer_entry
, list
);
1530 if (req_need_defer(de
->req
, de
->seq
))
1532 list_del_init(&de
->list
);
1533 io_req_task_queue(de
->req
);
1538 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1539 __must_hold(&ctx
->completion_lock
)
1541 u32 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1542 struct io_kiocb
*req
, *tmp
;
1544 spin_lock_irq(&ctx
->timeout_lock
);
1545 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1546 u32 events_needed
, events_got
;
1548 if (io_is_timeout_noseq(req
))
1552 * Since seq can easily wrap around over time, subtract
1553 * the last seq at which timeouts were flushed before comparing.
1554 * Assuming not more than 2^31-1 events have happened since,
1555 * these subtractions won't have wrapped, so we can check if
1556 * target is in [last_seq, current_seq] by comparing the two.
1558 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1559 events_got
= seq
- ctx
->cq_last_tm_flush
;
1560 if (events_got
< events_needed
)
1563 io_kill_timeout(req
, 0);
1565 ctx
->cq_last_tm_flush
= seq
;
1566 spin_unlock_irq(&ctx
->timeout_lock
);
1569 static void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
1571 if (ctx
->off_timeout_used
)
1572 io_flush_timeouts(ctx
);
1573 if (ctx
->drain_active
)
1574 io_queue_deferred(ctx
);
1577 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
1579 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
))
1580 __io_commit_cqring_flush(ctx
);
1581 /* order cqe stores with ring update */
1582 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1585 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1587 struct io_rings
*r
= ctx
->rings
;
1589 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
1592 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1594 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1597 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
1599 struct io_rings
*rings
= ctx
->rings
;
1600 unsigned tail
, mask
= ctx
->cq_entries
- 1;
1603 * writes to the cq entry need to come after reading head; the
1604 * control dependency is enough as we're using WRITE_ONCE to
1607 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
1610 tail
= ctx
->cached_cq_tail
++;
1611 return &rings
->cqes
[tail
& mask
];
1614 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1616 if (likely(!ctx
->cq_ev_fd
))
1618 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1620 return !ctx
->eventfd_async
|| io_wq_current_is_worker();
1624 * This should only get called when at least one event has been posted.
1625 * Some applications rely on the eventfd notification count only changing
1626 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1627 * 1:1 relationship between how many times this function is called (and
1628 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1630 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1633 * wake_up_all() may seem excessive, but io_wake_function() and
1634 * io_should_wake() handle the termination of the loop and only
1635 * wake as many waiters as we need to.
1637 if (wq_has_sleeper(&ctx
->cq_wait
))
1638 wake_up_all(&ctx
->cq_wait
);
1639 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1640 wake_up(&ctx
->sq_data
->wait
);
1641 if (io_should_trigger_evfd(ctx
))
1642 eventfd_signal(ctx
->cq_ev_fd
, 1);
1643 if (waitqueue_active(&ctx
->poll_wait
))
1644 wake_up_interruptible(&ctx
->poll_wait
);
1647 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1649 /* see waitqueue_active() comment */
1652 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1653 if (waitqueue_active(&ctx
->cq_wait
))
1654 wake_up_all(&ctx
->cq_wait
);
1656 if (io_should_trigger_evfd(ctx
))
1657 eventfd_signal(ctx
->cq_ev_fd
, 1);
1658 if (waitqueue_active(&ctx
->poll_wait
))
1659 wake_up_interruptible(&ctx
->poll_wait
);
1662 /* Returns true if there are no backlogged entries after the flush */
1663 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1665 bool all_flushed
, posted
;
1667 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
1671 spin_lock(&ctx
->completion_lock
);
1672 while (!list_empty(&ctx
->cq_overflow_list
)) {
1673 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
1674 struct io_overflow_cqe
*ocqe
;
1678 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
1679 struct io_overflow_cqe
, list
);
1681 memcpy(cqe
, &ocqe
->cqe
, sizeof(*cqe
));
1683 io_account_cq_overflow(ctx
);
1686 list_del(&ocqe
->list
);
1690 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1692 clear_bit(0, &ctx
->check_cq_overflow
);
1693 WRITE_ONCE(ctx
->rings
->sq_flags
,
1694 ctx
->rings
->sq_flags
& ~IORING_SQ_CQ_OVERFLOW
);
1698 io_commit_cqring(ctx
);
1699 spin_unlock(&ctx
->completion_lock
);
1701 io_cqring_ev_posted(ctx
);
1705 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
1709 if (test_bit(0, &ctx
->check_cq_overflow
)) {
1710 /* iopoll syncs against uring_lock, not completion_lock */
1711 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1712 mutex_lock(&ctx
->uring_lock
);
1713 ret
= __io_cqring_overflow_flush(ctx
, false);
1714 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1715 mutex_unlock(&ctx
->uring_lock
);
1721 /* must to be called somewhat shortly after putting a request */
1722 static inline void io_put_task(struct task_struct
*task
, int nr
)
1724 struct io_uring_task
*tctx
= task
->io_uring
;
1726 if (likely(task
== current
)) {
1727 tctx
->cached_refs
+= nr
;
1729 percpu_counter_sub(&tctx
->inflight
, nr
);
1730 if (unlikely(atomic_read(&tctx
->in_idle
)))
1731 wake_up(&tctx
->wait
);
1732 put_task_struct_many(task
, nr
);
1736 static void io_task_refs_refill(struct io_uring_task
*tctx
)
1738 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
1740 percpu_counter_add(&tctx
->inflight
, refill
);
1741 refcount_add(refill
, ¤t
->usage
);
1742 tctx
->cached_refs
+= refill
;
1745 static inline void io_get_task_refs(int nr
)
1747 struct io_uring_task
*tctx
= current
->io_uring
;
1749 tctx
->cached_refs
-= nr
;
1750 if (unlikely(tctx
->cached_refs
< 0))
1751 io_task_refs_refill(tctx
);
1754 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
1756 struct io_uring_task
*tctx
= task
->io_uring
;
1757 unsigned int refs
= tctx
->cached_refs
;
1760 tctx
->cached_refs
= 0;
1761 percpu_counter_sub(&tctx
->inflight
, refs
);
1762 put_task_struct_many(task
, refs
);
1766 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
1767 s32 res
, u32 cflags
)
1769 struct io_overflow_cqe
*ocqe
;
1771 ocqe
= kmalloc(sizeof(*ocqe
), GFP_ATOMIC
| __GFP_ACCOUNT
);
1774 * If we're in ring overflow flush mode, or in task cancel mode,
1775 * or cannot allocate an overflow entry, then we need to drop it
1778 io_account_cq_overflow(ctx
);
1781 if (list_empty(&ctx
->cq_overflow_list
)) {
1782 set_bit(0, &ctx
->check_cq_overflow
);
1783 WRITE_ONCE(ctx
->rings
->sq_flags
,
1784 ctx
->rings
->sq_flags
| IORING_SQ_CQ_OVERFLOW
);
1787 ocqe
->cqe
.user_data
= user_data
;
1788 ocqe
->cqe
.res
= res
;
1789 ocqe
->cqe
.flags
= cflags
;
1790 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
1794 static inline bool __io_fill_cqe(struct io_ring_ctx
*ctx
, u64 user_data
,
1795 s32 res
, u32 cflags
)
1797 struct io_uring_cqe
*cqe
;
1799 trace_io_uring_complete(ctx
, user_data
, res
, cflags
);
1802 * If we can't get a cq entry, userspace overflowed the
1803 * submission (by quite a lot). Increment the overflow count in
1806 cqe
= io_get_cqe(ctx
);
1808 WRITE_ONCE(cqe
->user_data
, user_data
);
1809 WRITE_ONCE(cqe
->res
, res
);
1810 WRITE_ONCE(cqe
->flags
, cflags
);
1813 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
);
1816 static noinline
void io_fill_cqe_req(struct io_kiocb
*req
, s32 res
, u32 cflags
)
1818 __io_fill_cqe(req
->ctx
, req
->user_data
, res
, cflags
);
1821 static noinline
bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
,
1822 s32 res
, u32 cflags
)
1825 return __io_fill_cqe(ctx
, user_data
, res
, cflags
);
1828 static void io_req_complete_post(struct io_kiocb
*req
, s32 res
,
1831 struct io_ring_ctx
*ctx
= req
->ctx
;
1833 spin_lock(&ctx
->completion_lock
);
1834 __io_fill_cqe(ctx
, req
->user_data
, res
, cflags
);
1836 * If we're the last reference to this request, add to our locked
1839 if (req_ref_put_and_test(req
)) {
1840 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1841 if (req
->flags
& IO_DISARM_MASK
)
1842 io_disarm_next(req
);
1844 io_req_task_queue(req
->link
);
1848 io_dismantle_req(req
);
1849 io_put_task(req
->task
, 1);
1850 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
1851 ctx
->locked_free_nr
++;
1853 if (!percpu_ref_tryget(&ctx
->refs
))
1856 io_commit_cqring(ctx
);
1857 spin_unlock(&ctx
->completion_lock
);
1860 io_cqring_ev_posted(ctx
);
1861 percpu_ref_put(&ctx
->refs
);
1865 static inline bool io_req_needs_clean(struct io_kiocb
*req
)
1867 return req
->flags
& IO_REQ_CLEAN_FLAGS
;
1870 static inline void io_req_complete_state(struct io_kiocb
*req
, s32 res
,
1873 if (io_req_needs_clean(req
))
1876 req
->compl.cflags
= cflags
;
1877 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1880 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1881 s32 res
, u32 cflags
)
1883 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1884 io_req_complete_state(req
, res
, cflags
);
1886 io_req_complete_post(req
, res
, cflags
);
1889 static inline void io_req_complete(struct io_kiocb
*req
, s32 res
)
1891 __io_req_complete(req
, 0, res
, 0);
1894 static void io_req_complete_failed(struct io_kiocb
*req
, s32 res
)
1897 io_req_complete_post(req
, res
, 0);
1900 static void io_req_complete_fail_submit(struct io_kiocb
*req
)
1903 * We don't submit, fail them all, for that replace hardlinks with
1904 * normal links. Extra REQ_F_LINK is tolerated.
1906 req
->flags
&= ~REQ_F_HARDLINK
;
1907 req
->flags
|= REQ_F_LINK
;
1908 io_req_complete_failed(req
, req
->result
);
1912 * Don't initialise the fields below on every allocation, but do that in
1913 * advance and keep them valid across allocations.
1915 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1919 req
->async_data
= NULL
;
1920 /* not necessary, but safer to zero */
1924 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1925 struct io_submit_state
*state
)
1927 spin_lock(&ctx
->completion_lock
);
1928 list_splice_init(&ctx
->locked_free_list
, &state
->free_list
);
1929 ctx
->locked_free_nr
= 0;
1930 spin_unlock(&ctx
->completion_lock
);
1933 /* Returns true IFF there are requests in the cache */
1934 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1936 struct io_submit_state
*state
= &ctx
->submit_state
;
1940 * If we have more than a batch's worth of requests in our IRQ side
1941 * locked cache, grab the lock and move them over to our submission
1944 if (READ_ONCE(ctx
->locked_free_nr
) > IO_COMPL_BATCH
)
1945 io_flush_cached_locked_reqs(ctx
, state
);
1947 nr
= state
->free_reqs
;
1948 while (!list_empty(&state
->free_list
)) {
1949 struct io_kiocb
*req
= list_first_entry(&state
->free_list
,
1950 struct io_kiocb
, inflight_entry
);
1952 list_del(&req
->inflight_entry
);
1953 state
->reqs
[nr
++] = req
;
1954 if (nr
== ARRAY_SIZE(state
->reqs
))
1958 state
->free_reqs
= nr
;
1963 * A request might get retired back into the request caches even before opcode
1964 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1965 * Because of that, io_alloc_req() should be called only under ->uring_lock
1966 * and with extra caution to not get a request that is still worked on.
1968 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1969 __must_hold(&ctx
->uring_lock
)
1971 struct io_submit_state
*state
= &ctx
->submit_state
;
1972 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1975 BUILD_BUG_ON(ARRAY_SIZE(state
->reqs
) < IO_REQ_ALLOC_BATCH
);
1977 if (likely(state
->free_reqs
|| io_flush_cached_reqs(ctx
)))
1980 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1984 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1985 * retry single alloc to be on the safe side.
1987 if (unlikely(ret
<= 0)) {
1988 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1989 if (!state
->reqs
[0])
1994 for (i
= 0; i
< ret
; i
++)
1995 io_preinit_req(state
->reqs
[i
], ctx
);
1996 state
->free_reqs
= ret
;
1999 return state
->reqs
[state
->free_reqs
];
2002 static inline void io_put_file(struct file
*file
)
2008 static void io_dismantle_req(struct io_kiocb
*req
)
2010 unsigned int flags
= req
->flags
;
2012 if (io_req_needs_clean(req
))
2014 if (!(flags
& REQ_F_FIXED_FILE
))
2015 io_put_file(req
->file
);
2016 if (req
->fixed_rsrc_refs
)
2017 percpu_ref_put(req
->fixed_rsrc_refs
);
2018 if (req
->async_data
) {
2019 kfree(req
->async_data
);
2020 req
->async_data
= NULL
;
2024 static void __io_free_req(struct io_kiocb
*req
)
2026 struct io_ring_ctx
*ctx
= req
->ctx
;
2028 io_dismantle_req(req
);
2029 io_put_task(req
->task
, 1);
2031 spin_lock(&ctx
->completion_lock
);
2032 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
2033 ctx
->locked_free_nr
++;
2034 spin_unlock(&ctx
->completion_lock
);
2036 percpu_ref_put(&ctx
->refs
);
2039 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2041 struct io_kiocb
*nxt
= req
->link
;
2043 req
->link
= nxt
->link
;
2047 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
2048 __must_hold(&req
->ctx
->completion_lock
)
2049 __must_hold(&req
->ctx
->timeout_lock
)
2051 struct io_kiocb
*link
= req
->link
;
2053 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2054 struct io_timeout_data
*io
= link
->async_data
;
2056 io_remove_next_linked(req
);
2057 link
->timeout
.head
= NULL
;
2058 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
2059 list_del(&link
->timeout
.list
);
2060 io_fill_cqe_req(link
, -ECANCELED
, 0);
2061 io_put_req_deferred(link
);
2068 static void io_fail_links(struct io_kiocb
*req
)
2069 __must_hold(&req
->ctx
->completion_lock
)
2071 struct io_kiocb
*nxt
, *link
= req
->link
;
2075 long res
= -ECANCELED
;
2077 if (link
->flags
& REQ_F_FAIL
)
2083 trace_io_uring_fail_link(req
, link
);
2084 io_fill_cqe_req(link
, res
, 0);
2085 io_put_req_deferred(link
);
2090 static bool io_disarm_next(struct io_kiocb
*req
)
2091 __must_hold(&req
->ctx
->completion_lock
)
2093 bool posted
= false;
2095 if (req
->flags
& REQ_F_ARM_LTIMEOUT
) {
2096 struct io_kiocb
*link
= req
->link
;
2098 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
2099 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2100 io_remove_next_linked(req
);
2101 io_fill_cqe_req(link
, -ECANCELED
, 0);
2102 io_put_req_deferred(link
);
2105 } else if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
2106 struct io_ring_ctx
*ctx
= req
->ctx
;
2108 spin_lock_irq(&ctx
->timeout_lock
);
2109 posted
= io_kill_linked_timeout(req
);
2110 spin_unlock_irq(&ctx
->timeout_lock
);
2112 if (unlikely((req
->flags
& REQ_F_FAIL
) &&
2113 !(req
->flags
& REQ_F_HARDLINK
))) {
2114 posted
|= (req
->link
!= NULL
);
2120 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2122 struct io_kiocb
*nxt
;
2125 * If LINK is set, we have dependent requests in this chain. If we
2126 * didn't fail this request, queue the first one up, moving any other
2127 * dependencies to the next request. In case of failure, fail the rest
2130 if (req
->flags
& IO_DISARM_MASK
) {
2131 struct io_ring_ctx
*ctx
= req
->ctx
;
2134 spin_lock(&ctx
->completion_lock
);
2135 posted
= io_disarm_next(req
);
2137 io_commit_cqring(req
->ctx
);
2138 spin_unlock(&ctx
->completion_lock
);
2140 io_cqring_ev_posted(ctx
);
2147 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2149 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
2151 return __io_req_find_next(req
);
2154 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, bool *locked
)
2159 if (ctx
->submit_state
.compl_nr
)
2160 io_submit_flush_completions(ctx
);
2161 mutex_unlock(&ctx
->uring_lock
);
2164 percpu_ref_put(&ctx
->refs
);
2167 static void tctx_task_work(struct callback_head
*cb
)
2169 bool locked
= false;
2170 struct io_ring_ctx
*ctx
= NULL
;
2171 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
2175 struct io_wq_work_node
*node
;
2177 if (!tctx
->task_list
.first
&& locked
&& ctx
->submit_state
.compl_nr
)
2178 io_submit_flush_completions(ctx
);
2180 spin_lock_irq(&tctx
->task_lock
);
2181 node
= tctx
->task_list
.first
;
2182 INIT_WQ_LIST(&tctx
->task_list
);
2184 tctx
->task_running
= false;
2185 spin_unlock_irq(&tctx
->task_lock
);
2190 struct io_wq_work_node
*next
= node
->next
;
2191 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
2194 if (req
->ctx
!= ctx
) {
2195 ctx_flush_and_put(ctx
, &locked
);
2197 /* if not contended, grab and improve batching */
2198 locked
= mutex_trylock(&ctx
->uring_lock
);
2199 percpu_ref_get(&ctx
->refs
);
2201 req
->io_task_work
.func(req
, &locked
);
2208 ctx_flush_and_put(ctx
, &locked
);
2210 /* relaxed read is enough as only the task itself sets ->in_idle */
2211 if (unlikely(atomic_read(&tctx
->in_idle
)))
2212 io_uring_drop_tctx_refs(current
);
2215 static void io_req_task_work_add(struct io_kiocb
*req
)
2217 struct task_struct
*tsk
= req
->task
;
2218 struct io_uring_task
*tctx
= tsk
->io_uring
;
2219 enum task_work_notify_mode notify
;
2220 struct io_wq_work_node
*node
;
2221 unsigned long flags
;
2224 WARN_ON_ONCE(!tctx
);
2226 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2227 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
2228 running
= tctx
->task_running
;
2230 tctx
->task_running
= true;
2231 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2233 /* task_work already pending, we're done */
2238 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2239 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2240 * processing task_work. There's no reliable way to tell if TWA_RESUME
2243 notify
= (req
->ctx
->flags
& IORING_SETUP_SQPOLL
) ? TWA_NONE
: TWA_SIGNAL
;
2244 if (!task_work_add(tsk
, &tctx
->task_work
, notify
)) {
2245 wake_up_process(tsk
);
2249 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2250 tctx
->task_running
= false;
2251 node
= tctx
->task_list
.first
;
2252 INIT_WQ_LIST(&tctx
->task_list
);
2253 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2256 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
2258 if (llist_add(&req
->io_task_work
.fallback_node
,
2259 &req
->ctx
->fallback_llist
))
2260 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
2264 static void io_req_task_cancel(struct io_kiocb
*req
, bool *locked
)
2266 struct io_ring_ctx
*ctx
= req
->ctx
;
2268 /* not needed for normal modes, but SQPOLL depends on it */
2269 io_tw_lock(ctx
, locked
);
2270 io_req_complete_failed(req
, req
->result
);
2273 static void io_req_task_submit(struct io_kiocb
*req
, bool *locked
)
2275 struct io_ring_ctx
*ctx
= req
->ctx
;
2277 io_tw_lock(ctx
, locked
);
2278 /* req->task == current here, checking PF_EXITING is safe */
2279 if (likely(!(req
->task
->flags
& PF_EXITING
)))
2280 __io_queue_sqe(req
);
2282 io_req_complete_failed(req
, -EFAULT
);
2285 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2288 req
->io_task_work
.func
= io_req_task_cancel
;
2289 io_req_task_work_add(req
);
2292 static void io_req_task_queue(struct io_kiocb
*req
)
2294 req
->io_task_work
.func
= io_req_task_submit
;
2295 io_req_task_work_add(req
);
2298 static void io_req_task_queue_reissue(struct io_kiocb
*req
)
2300 req
->io_task_work
.func
= io_queue_async_work
;
2301 io_req_task_work_add(req
);
2304 static inline void io_queue_next(struct io_kiocb
*req
)
2306 struct io_kiocb
*nxt
= io_req_find_next(req
);
2309 io_req_task_queue(nxt
);
2312 static void io_free_req(struct io_kiocb
*req
)
2318 static void io_free_req_work(struct io_kiocb
*req
, bool *locked
)
2324 struct task_struct
*task
;
2329 static inline void io_init_req_batch(struct req_batch
*rb
)
2336 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2337 struct req_batch
*rb
)
2340 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2342 io_put_task(rb
->task
, rb
->task_refs
);
2345 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2346 struct io_submit_state
*state
)
2349 io_dismantle_req(req
);
2351 if (req
->task
!= rb
->task
) {
2353 io_put_task(rb
->task
, rb
->task_refs
);
2354 rb
->task
= req
->task
;
2360 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2361 state
->reqs
[state
->free_reqs
++] = req
;
2363 list_add(&req
->inflight_entry
, &state
->free_list
);
2366 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
2367 __must_hold(&ctx
->uring_lock
)
2369 struct io_submit_state
*state
= &ctx
->submit_state
;
2370 int i
, nr
= state
->compl_nr
;
2371 struct req_batch rb
;
2373 spin_lock(&ctx
->completion_lock
);
2374 for (i
= 0; i
< nr
; i
++) {
2375 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2377 __io_fill_cqe(ctx
, req
->user_data
, req
->result
,
2380 io_commit_cqring(ctx
);
2381 spin_unlock(&ctx
->completion_lock
);
2382 io_cqring_ev_posted(ctx
);
2384 io_init_req_batch(&rb
);
2385 for (i
= 0; i
< nr
; i
++) {
2386 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2388 if (req_ref_put_and_test(req
))
2389 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2392 io_req_free_batch_finish(ctx
, &rb
);
2393 state
->compl_nr
= 0;
2397 * Drop reference to request, return next in chain (if there is one) if this
2398 * was the last reference to this request.
2400 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2402 struct io_kiocb
*nxt
= NULL
;
2404 if (req_ref_put_and_test(req
)) {
2405 nxt
= io_req_find_next(req
);
2411 static inline void io_put_req(struct io_kiocb
*req
)
2413 if (req_ref_put_and_test(req
))
2417 static inline void io_put_req_deferred(struct io_kiocb
*req
)
2419 if (req_ref_put_and_test(req
)) {
2420 req
->io_task_work
.func
= io_free_req_work
;
2421 io_req_task_work_add(req
);
2425 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2427 /* See comment at the top of this file */
2429 return __io_cqring_events(ctx
);
2432 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2434 struct io_rings
*rings
= ctx
->rings
;
2436 /* make sure SQ entry isn't read before tail */
2437 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2440 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2442 unsigned int cflags
;
2444 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2445 cflags
|= IORING_CQE_F_BUFFER
;
2446 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2451 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2453 struct io_buffer
*kbuf
;
2455 if (likely(!(req
->flags
& REQ_F_BUFFER_SELECTED
)))
2457 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2458 return io_put_kbuf(req
, kbuf
);
2461 static inline bool io_run_task_work(void)
2463 if (test_thread_flag(TIF_NOTIFY_SIGNAL
) || current
->task_works
) {
2464 __set_current_state(TASK_RUNNING
);
2465 tracehook_notify_signal();
2473 * Find and free completed poll iocbs
2475 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2476 struct list_head
*done
)
2478 struct req_batch rb
;
2479 struct io_kiocb
*req
;
2481 /* order with ->result store in io_complete_rw_iopoll() */
2484 io_init_req_batch(&rb
);
2485 while (!list_empty(done
)) {
2486 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2487 list_del(&req
->inflight_entry
);
2489 io_fill_cqe_req(req
, req
->result
, io_put_rw_kbuf(req
));
2492 if (req_ref_put_and_test(req
))
2493 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2496 io_commit_cqring(ctx
);
2497 io_cqring_ev_posted_iopoll(ctx
);
2498 io_req_free_batch_finish(ctx
, &rb
);
2501 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2504 struct io_kiocb
*req
, *tmp
;
2509 * Only spin for completions if we don't have multiple devices hanging
2510 * off our complete list, and we're under the requested amount.
2512 spin
= !ctx
->poll_multi_queue
&& *nr_events
< min
;
2514 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2515 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2519 * Move completed and retryable entries to our local lists.
2520 * If we find a request that requires polling, break out
2521 * and complete those lists first, if we have entries there.
2523 if (READ_ONCE(req
->iopoll_completed
)) {
2524 list_move_tail(&req
->inflight_entry
, &done
);
2527 if (!list_empty(&done
))
2530 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2531 if (unlikely(ret
< 0))
2536 /* iopoll may have completed current req */
2537 if (READ_ONCE(req
->iopoll_completed
))
2538 list_move_tail(&req
->inflight_entry
, &done
);
2541 if (!list_empty(&done
))
2542 io_iopoll_complete(ctx
, nr_events
, &done
);
2548 * We can't just wait for polled events to come to us, we have to actively
2549 * find and complete them.
2551 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2553 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2556 mutex_lock(&ctx
->uring_lock
);
2557 while (!list_empty(&ctx
->iopoll_list
)) {
2558 unsigned int nr_events
= 0;
2560 io_do_iopoll(ctx
, &nr_events
, 0);
2562 /* let it sleep and repeat later if can't complete a request */
2566 * Ensure we allow local-to-the-cpu processing to take place,
2567 * in this case we need to ensure that we reap all events.
2568 * Also let task_work, etc. to progress by releasing the mutex
2570 if (need_resched()) {
2571 mutex_unlock(&ctx
->uring_lock
);
2573 mutex_lock(&ctx
->uring_lock
);
2576 mutex_unlock(&ctx
->uring_lock
);
2579 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2581 unsigned int nr_events
= 0;
2585 * We disallow the app entering submit/complete with polling, but we
2586 * still need to lock the ring to prevent racing with polled issue
2587 * that got punted to a workqueue.
2589 mutex_lock(&ctx
->uring_lock
);
2591 * Don't enter poll loop if we already have events pending.
2592 * If we do, we can potentially be spinning for commands that
2593 * already triggered a CQE (eg in error).
2595 if (test_bit(0, &ctx
->check_cq_overflow
))
2596 __io_cqring_overflow_flush(ctx
, false);
2597 if (io_cqring_events(ctx
))
2601 * If a submit got punted to a workqueue, we can have the
2602 * application entering polling for a command before it gets
2603 * issued. That app will hold the uring_lock for the duration
2604 * of the poll right here, so we need to take a breather every
2605 * now and then to ensure that the issue has a chance to add
2606 * the poll to the issued list. Otherwise we can spin here
2607 * forever, while the workqueue is stuck trying to acquire the
2610 if (list_empty(&ctx
->iopoll_list
)) {
2611 u32 tail
= ctx
->cached_cq_tail
;
2613 mutex_unlock(&ctx
->uring_lock
);
2615 mutex_lock(&ctx
->uring_lock
);
2617 /* some requests don't go through iopoll_list */
2618 if (tail
!= ctx
->cached_cq_tail
||
2619 list_empty(&ctx
->iopoll_list
))
2622 ret
= io_do_iopoll(ctx
, &nr_events
, min
);
2623 } while (!ret
&& nr_events
< min
&& !need_resched());
2625 mutex_unlock(&ctx
->uring_lock
);
2629 static void kiocb_end_write(struct io_kiocb
*req
)
2632 * Tell lockdep we inherited freeze protection from submission
2635 if (req
->flags
& REQ_F_ISREG
) {
2636 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
2638 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
2644 static bool io_resubmit_prep(struct io_kiocb
*req
)
2646 struct io_async_rw
*rw
= req
->async_data
;
2649 return !io_req_prep_async(req
);
2650 iov_iter_restore(&rw
->iter
, &rw
->iter_state
);
2654 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2656 umode_t mode
= file_inode(req
->file
)->i_mode
;
2657 struct io_ring_ctx
*ctx
= req
->ctx
;
2659 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2661 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2662 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2665 * If ref is dying, we might be running poll reap from the exit work.
2666 * Don't attempt to reissue from that path, just let it fail with
2669 if (percpu_ref_is_dying(&ctx
->refs
))
2672 * Play it safe and assume not safe to re-import and reissue if we're
2673 * not in the original thread group (or in task context).
2675 if (!same_thread_group(req
->task
, current
) || !in_task())
2680 static bool io_resubmit_prep(struct io_kiocb
*req
)
2684 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2690 static bool __io_complete_rw_common(struct io_kiocb
*req
, long res
)
2692 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
) {
2693 kiocb_end_write(req
);
2694 fsnotify_modify(req
->file
);
2696 fsnotify_access(req
->file
);
2698 if (res
!= req
->result
) {
2699 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
2700 io_rw_should_reissue(req
)) {
2701 req
->flags
|= REQ_F_REISSUE
;
2710 static void io_req_task_complete(struct io_kiocb
*req
, bool *locked
)
2712 unsigned int cflags
= io_put_rw_kbuf(req
);
2713 int res
= req
->result
;
2716 struct io_ring_ctx
*ctx
= req
->ctx
;
2717 struct io_submit_state
*state
= &ctx
->submit_state
;
2719 io_req_complete_state(req
, res
, cflags
);
2720 state
->compl_reqs
[state
->compl_nr
++] = req
;
2721 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
2722 io_submit_flush_completions(ctx
);
2724 io_req_complete_post(req
, res
, cflags
);
2728 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2729 unsigned int issue_flags
)
2731 if (__io_complete_rw_common(req
, res
))
2733 __io_req_complete(req
, issue_flags
, req
->result
, io_put_rw_kbuf(req
));
2736 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2738 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2740 if (__io_complete_rw_common(req
, res
))
2743 req
->io_task_work
.func
= io_req_task_complete
;
2744 io_req_task_work_add(req
);
2747 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2749 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2751 if (kiocb
->ki_flags
& IOCB_WRITE
)
2752 kiocb_end_write(req
);
2753 if (unlikely(res
!= req
->result
)) {
2754 if (res
== -EAGAIN
&& io_rw_should_reissue(req
)) {
2755 req
->flags
|= REQ_F_REISSUE
;
2760 WRITE_ONCE(req
->result
, res
);
2761 /* order with io_iopoll_complete() checking ->result */
2763 WRITE_ONCE(req
->iopoll_completed
, 1);
2767 * After the iocb has been issued, it's safe to be found on the poll list.
2768 * Adding the kiocb to the list AFTER submission ensures that we don't
2769 * find it from a io_do_iopoll() thread before the issuer is done
2770 * accessing the kiocb cookie.
2772 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2774 struct io_ring_ctx
*ctx
= req
->ctx
;
2775 const bool in_async
= io_wq_current_is_worker();
2777 /* workqueue context doesn't hold uring_lock, grab it now */
2778 if (unlikely(in_async
))
2779 mutex_lock(&ctx
->uring_lock
);
2782 * Track whether we have multiple files in our lists. This will impact
2783 * how we do polling eventually, not spinning if we're on potentially
2784 * different devices.
2786 if (list_empty(&ctx
->iopoll_list
)) {
2787 ctx
->poll_multi_queue
= false;
2788 } else if (!ctx
->poll_multi_queue
) {
2789 struct io_kiocb
*list_req
;
2790 unsigned int queue_num0
, queue_num1
;
2792 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2795 if (list_req
->file
!= req
->file
) {
2796 ctx
->poll_multi_queue
= true;
2798 queue_num0
= blk_qc_t_to_queue_num(list_req
->rw
.kiocb
.ki_cookie
);
2799 queue_num1
= blk_qc_t_to_queue_num(req
->rw
.kiocb
.ki_cookie
);
2800 if (queue_num0
!= queue_num1
)
2801 ctx
->poll_multi_queue
= true;
2806 * For fast devices, IO may have already completed. If it has, add
2807 * it to the front so we find it first.
2809 if (READ_ONCE(req
->iopoll_completed
))
2810 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2812 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2814 if (unlikely(in_async
)) {
2816 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2817 * in sq thread task context or in io worker task context. If
2818 * current task context is sq thread, we don't need to check
2819 * whether should wake up sq thread.
2821 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2822 wq_has_sleeper(&ctx
->sq_data
->wait
))
2823 wake_up(&ctx
->sq_data
->wait
);
2825 mutex_unlock(&ctx
->uring_lock
);
2829 static bool io_bdev_nowait(struct block_device
*bdev
)
2831 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2835 * If we tracked the file through the SCM inflight mechanism, we could support
2836 * any file. For now, just ensure that anything potentially problematic is done
2839 static bool __io_file_supports_nowait(struct file
*file
, int rw
)
2841 umode_t mode
= file_inode(file
)->i_mode
;
2843 if (S_ISBLK(mode
)) {
2844 if (IS_ENABLED(CONFIG_BLOCK
) &&
2845 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2851 if (S_ISREG(mode
)) {
2852 if (IS_ENABLED(CONFIG_BLOCK
) &&
2853 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2854 file
->f_op
!= &io_uring_fops
)
2859 /* any ->read/write should understand O_NONBLOCK */
2860 if (file
->f_flags
& O_NONBLOCK
)
2863 if (!(file
->f_mode
& FMODE_NOWAIT
))
2867 return file
->f_op
->read_iter
!= NULL
;
2869 return file
->f_op
->write_iter
!= NULL
;
2872 static bool io_file_supports_nowait(struct io_kiocb
*req
, int rw
)
2874 if (rw
== READ
&& (req
->flags
& REQ_F_NOWAIT_READ
))
2876 else if (rw
== WRITE
&& (req
->flags
& REQ_F_NOWAIT_WRITE
))
2879 return __io_file_supports_nowait(req
->file
, rw
);
2882 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2885 struct io_ring_ctx
*ctx
= req
->ctx
;
2886 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2887 struct file
*file
= req
->file
;
2891 if (!io_req_ffs_set(req
) && S_ISREG(file_inode(file
)->i_mode
))
2892 req
->flags
|= REQ_F_ISREG
;
2894 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2895 if (kiocb
->ki_pos
== -1) {
2896 if (!(file
->f_mode
& FMODE_STREAM
)) {
2897 req
->flags
|= REQ_F_CUR_POS
;
2898 kiocb
->ki_pos
= file
->f_pos
;
2903 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2904 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2905 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2910 * If the file is marked O_NONBLOCK, still allow retry for it if it
2911 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2912 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2914 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) ||
2915 ((file
->f_flags
& O_NONBLOCK
) && !io_file_supports_nowait(req
, rw
)))
2916 req
->flags
|= REQ_F_NOWAIT
;
2918 ioprio
= READ_ONCE(sqe
->ioprio
);
2920 ret
= ioprio_check_cap(ioprio
);
2924 kiocb
->ki_ioprio
= ioprio
;
2926 kiocb
->ki_ioprio
= get_current_ioprio();
2928 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2929 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2930 !kiocb
->ki_filp
->f_op
->iopoll
)
2933 kiocb
->ki_flags
|= IOCB_HIPRI
| IOCB_ALLOC_CACHE
;
2934 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2935 req
->iopoll_completed
= 0;
2937 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2939 kiocb
->ki_complete
= io_complete_rw
;
2942 /* used for fixed read/write too - just read unconditionally */
2943 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2946 if (req
->opcode
== IORING_OP_READ_FIXED
||
2947 req
->opcode
== IORING_OP_WRITE_FIXED
) {
2948 struct io_ring_ctx
*ctx
= req
->ctx
;
2951 if (unlikely(req
->buf_index
>= ctx
->nr_user_bufs
))
2953 index
= array_index_nospec(req
->buf_index
, ctx
->nr_user_bufs
);
2954 req
->imu
= ctx
->user_bufs
[index
];
2955 io_req_set_rsrc_node(req
);
2958 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2959 req
->rw
.len
= READ_ONCE(sqe
->len
);
2963 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2969 case -ERESTARTNOINTR
:
2970 case -ERESTARTNOHAND
:
2971 case -ERESTART_RESTARTBLOCK
:
2973 * We can't just restart the syscall, since previously
2974 * submitted sqes may already be in progress. Just fail this
2980 kiocb
->ki_complete(kiocb
, ret
, 0);
2984 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2985 unsigned int issue_flags
)
2987 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2988 struct io_async_rw
*io
= req
->async_data
;
2990 /* add previously done IO, if any */
2991 if (io
&& io
->bytes_done
> 0) {
2993 ret
= io
->bytes_done
;
2995 ret
+= io
->bytes_done
;
2998 if (req
->flags
& REQ_F_CUR_POS
)
2999 req
->file
->f_pos
= kiocb
->ki_pos
;
3000 if (ret
>= 0 && (kiocb
->ki_complete
== io_complete_rw
))
3001 __io_complete_rw(req
, ret
, 0, issue_flags
);
3003 io_rw_done(kiocb
, ret
);
3005 if (req
->flags
& REQ_F_REISSUE
) {
3006 req
->flags
&= ~REQ_F_REISSUE
;
3007 if (io_resubmit_prep(req
)) {
3008 io_req_task_queue_reissue(req
);
3010 unsigned int cflags
= io_put_rw_kbuf(req
);
3011 struct io_ring_ctx
*ctx
= req
->ctx
;
3014 if (!(issue_flags
& IO_URING_F_NONBLOCK
)) {
3015 mutex_lock(&ctx
->uring_lock
);
3016 __io_req_complete(req
, issue_flags
, ret
, cflags
);
3017 mutex_unlock(&ctx
->uring_lock
);
3019 __io_req_complete(req
, issue_flags
, ret
, cflags
);
3025 static int __io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
,
3026 struct io_mapped_ubuf
*imu
)
3028 size_t len
= req
->rw
.len
;
3029 u64 buf_end
, buf_addr
= req
->rw
.addr
;
3032 if (unlikely(check_add_overflow(buf_addr
, (u64
)len
, &buf_end
)))
3034 /* not inside the mapped region */
3035 if (unlikely(buf_addr
< imu
->ubuf
|| buf_end
> imu
->ubuf_end
))
3039 * May not be a start of buffer, set size appropriately
3040 * and advance us to the beginning.
3042 offset
= buf_addr
- imu
->ubuf
;
3043 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
3047 * Don't use iov_iter_advance() here, as it's really slow for
3048 * using the latter parts of a big fixed buffer - it iterates
3049 * over each segment manually. We can cheat a bit here, because
3052 * 1) it's a BVEC iter, we set it up
3053 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3054 * first and last bvec
3056 * So just find our index, and adjust the iterator afterwards.
3057 * If the offset is within the first bvec (or the whole first
3058 * bvec, just use iov_iter_advance(). This makes it easier
3059 * since we can just skip the first segment, which may not
3060 * be PAGE_SIZE aligned.
3062 const struct bio_vec
*bvec
= imu
->bvec
;
3064 if (offset
<= bvec
->bv_len
) {
3065 iov_iter_advance(iter
, offset
);
3067 unsigned long seg_skip
;
3069 /* skip first vec */
3070 offset
-= bvec
->bv_len
;
3071 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
3073 iter
->bvec
= bvec
+ seg_skip
;
3074 iter
->nr_segs
-= seg_skip
;
3075 iter
->count
-= bvec
->bv_len
+ offset
;
3076 iter
->iov_offset
= offset
& ~PAGE_MASK
;
3083 static int io_import_fixed(struct io_kiocb
*req
, int rw
, struct iov_iter
*iter
)
3085 if (WARN_ON_ONCE(!req
->imu
))
3087 return __io_import_fixed(req
, rw
, iter
, req
->imu
);
3090 static void io_ring_submit_unlock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3093 mutex_unlock(&ctx
->uring_lock
);
3096 static void io_ring_submit_lock(struct io_ring_ctx
*ctx
, bool needs_lock
)
3099 * "Normal" inline submissions always hold the uring_lock, since we
3100 * grab it from the system call. Same is true for the SQPOLL offload.
3101 * The only exception is when we've detached the request and issue it
3102 * from an async worker thread, grab the lock for that case.
3105 mutex_lock(&ctx
->uring_lock
);
3108 static struct io_buffer
*io_buffer_select(struct io_kiocb
*req
, size_t *len
,
3109 int bgid
, struct io_buffer
*kbuf
,
3112 struct io_buffer
*head
;
3114 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
3117 io_ring_submit_lock(req
->ctx
, needs_lock
);
3119 lockdep_assert_held(&req
->ctx
->uring_lock
);
3121 head
= xa_load(&req
->ctx
->io_buffers
, bgid
);
3123 if (!list_empty(&head
->list
)) {
3124 kbuf
= list_last_entry(&head
->list
, struct io_buffer
,
3126 list_del(&kbuf
->list
);
3129 xa_erase(&req
->ctx
->io_buffers
, bgid
);
3131 if (*len
> kbuf
->len
)
3134 kbuf
= ERR_PTR(-ENOBUFS
);
3137 io_ring_submit_unlock(req
->ctx
, needs_lock
);
3142 static void __user
*io_rw_buffer_select(struct io_kiocb
*req
, size_t *len
,
3145 struct io_buffer
*kbuf
;
3148 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3149 bgid
= req
->buf_index
;
3150 kbuf
= io_buffer_select(req
, len
, bgid
, kbuf
, needs_lock
);
3153 req
->rw
.addr
= (u64
) (unsigned long) kbuf
;
3154 req
->flags
|= REQ_F_BUFFER_SELECTED
;
3155 return u64_to_user_ptr(kbuf
->addr
);
3158 #ifdef CONFIG_COMPAT
3159 static ssize_t
io_compat_import(struct io_kiocb
*req
, struct iovec
*iov
,
3162 struct compat_iovec __user
*uiov
;
3163 compat_ssize_t clen
;
3167 uiov
= u64_to_user_ptr(req
->rw
.addr
);
3168 if (!access_ok(uiov
, sizeof(*uiov
)))
3170 if (__get_user(clen
, &uiov
->iov_len
))
3176 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3178 return PTR_ERR(buf
);
3179 iov
[0].iov_base
= buf
;
3180 iov
[0].iov_len
= (compat_size_t
) len
;
3185 static ssize_t
__io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3188 struct iovec __user
*uiov
= u64_to_user_ptr(req
->rw
.addr
);
3192 if (copy_from_user(iov
, uiov
, sizeof(*uiov
)))
3195 len
= iov
[0].iov_len
;
3198 buf
= io_rw_buffer_select(req
, &len
, needs_lock
);
3200 return PTR_ERR(buf
);
3201 iov
[0].iov_base
= buf
;
3202 iov
[0].iov_len
= len
;
3206 static ssize_t
io_iov_buffer_select(struct io_kiocb
*req
, struct iovec
*iov
,
3209 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
3210 struct io_buffer
*kbuf
;
3212 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
3213 iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
3214 iov
[0].iov_len
= kbuf
->len
;
3217 if (req
->rw
.len
!= 1)
3220 #ifdef CONFIG_COMPAT
3221 if (req
->ctx
->compat
)
3222 return io_compat_import(req
, iov
, needs_lock
);
3225 return __io_iov_buffer_select(req
, iov
, needs_lock
);
3228 static int io_import_iovec(int rw
, struct io_kiocb
*req
, struct iovec
**iovec
,
3229 struct iov_iter
*iter
, bool needs_lock
)
3231 void __user
*buf
= u64_to_user_ptr(req
->rw
.addr
);
3232 size_t sqe_len
= req
->rw
.len
;
3233 u8 opcode
= req
->opcode
;
3236 if (opcode
== IORING_OP_READ_FIXED
|| opcode
== IORING_OP_WRITE_FIXED
) {
3238 return io_import_fixed(req
, rw
, iter
);
3241 /* buffer index only valid with fixed read/write, or buffer select */
3242 if (req
->buf_index
&& !(req
->flags
& REQ_F_BUFFER_SELECT
))
3245 if (opcode
== IORING_OP_READ
|| opcode
== IORING_OP_WRITE
) {
3246 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3247 buf
= io_rw_buffer_select(req
, &sqe_len
, needs_lock
);
3249 return PTR_ERR(buf
);
3250 req
->rw
.len
= sqe_len
;
3253 ret
= import_single_range(rw
, buf
, sqe_len
, *iovec
, iter
);
3258 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
3259 ret
= io_iov_buffer_select(req
, *iovec
, needs_lock
);
3261 iov_iter_init(iter
, rw
, *iovec
, 1, (*iovec
)->iov_len
);
3266 return __import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
,
3270 static inline loff_t
*io_kiocb_ppos(struct kiocb
*kiocb
)
3272 return (kiocb
->ki_filp
->f_mode
& FMODE_STREAM
) ? NULL
: &kiocb
->ki_pos
;
3276 * For files that don't have ->read_iter() and ->write_iter(), handle them
3277 * by looping over ->read() or ->write() manually.
3279 static ssize_t
loop_rw_iter(int rw
, struct io_kiocb
*req
, struct iov_iter
*iter
)
3281 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3282 struct file
*file
= req
->file
;
3286 * Don't support polled IO through this interface, and we can't
3287 * support non-blocking either. For the latter, this just causes
3288 * the kiocb to be handled from an async context.
3290 if (kiocb
->ki_flags
& IOCB_HIPRI
)
3292 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
3295 while (iov_iter_count(iter
)) {
3299 if (!iov_iter_is_bvec(iter
)) {
3300 iovec
= iov_iter_iovec(iter
);
3302 iovec
.iov_base
= u64_to_user_ptr(req
->rw
.addr
);
3303 iovec
.iov_len
= req
->rw
.len
;
3307 nr
= file
->f_op
->read(file
, iovec
.iov_base
,
3308 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3310 nr
= file
->f_op
->write(file
, iovec
.iov_base
,
3311 iovec
.iov_len
, io_kiocb_ppos(kiocb
));
3320 if (!iov_iter_is_bvec(iter
)) {
3321 iov_iter_advance(iter
, nr
);
3328 if (nr
!= iovec
.iov_len
)
3335 static void io_req_map_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3336 const struct iovec
*fast_iov
, struct iov_iter
*iter
)
3338 struct io_async_rw
*rw
= req
->async_data
;
3340 memcpy(&rw
->iter
, iter
, sizeof(*iter
));
3341 rw
->free_iovec
= iovec
;
3343 /* can only be fixed buffers, no need to do anything */
3344 if (iov_iter_is_bvec(iter
))
3347 unsigned iov_off
= 0;
3349 rw
->iter
.iov
= rw
->fast_iov
;
3350 if (iter
->iov
!= fast_iov
) {
3351 iov_off
= iter
->iov
- fast_iov
;
3352 rw
->iter
.iov
+= iov_off
;
3354 if (rw
->fast_iov
!= fast_iov
)
3355 memcpy(rw
->fast_iov
+ iov_off
, fast_iov
+ iov_off
,
3356 sizeof(struct iovec
) * iter
->nr_segs
);
3358 req
->flags
|= REQ_F_NEED_CLEANUP
;
3362 static inline int io_alloc_async_data(struct io_kiocb
*req
)
3364 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
3365 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
3366 return req
->async_data
== NULL
;
3369 static int io_setup_async_rw(struct io_kiocb
*req
, const struct iovec
*iovec
,
3370 const struct iovec
*fast_iov
,
3371 struct iov_iter
*iter
, bool force
)
3373 if (!force
&& !io_op_defs
[req
->opcode
].needs_async_setup
)
3375 if (!req
->async_data
) {
3376 struct io_async_rw
*iorw
;
3378 if (io_alloc_async_data(req
)) {
3383 io_req_map_rw(req
, iovec
, fast_iov
, iter
);
3384 iorw
= req
->async_data
;
3385 /* we've copied and mapped the iter, ensure state is saved */
3386 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3391 static inline int io_rw_prep_async(struct io_kiocb
*req
, int rw
)
3393 struct io_async_rw
*iorw
= req
->async_data
;
3394 struct iovec
*iov
= iorw
->fast_iov
;
3397 ret
= io_import_iovec(rw
, req
, &iov
, &iorw
->iter
, false);
3398 if (unlikely(ret
< 0))
3401 iorw
->bytes_done
= 0;
3402 iorw
->free_iovec
= iov
;
3404 req
->flags
|= REQ_F_NEED_CLEANUP
;
3405 iov_iter_save_state(&iorw
->iter
, &iorw
->iter_state
);
3409 static int io_read_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3411 if (unlikely(!(req
->file
->f_mode
& FMODE_READ
)))
3413 return io_prep_rw(req
, sqe
, READ
);
3417 * This is our waitqueue callback handler, registered through lock_page_async()
3418 * when we initially tried to do the IO with the iocb armed our waitqueue.
3419 * This gets called when the page is unlocked, and we generally expect that to
3420 * happen when the page IO is completed and the page is now uptodate. This will
3421 * queue a task_work based retry of the operation, attempting to copy the data
3422 * again. If the latter fails because the page was NOT uptodate, then we will
3423 * do a thread based blocking retry of the operation. That's the unexpected
3426 static int io_async_buf_func(struct wait_queue_entry
*wait
, unsigned mode
,
3427 int sync
, void *arg
)
3429 struct wait_page_queue
*wpq
;
3430 struct io_kiocb
*req
= wait
->private;
3431 struct wait_page_key
*key
= arg
;
3433 wpq
= container_of(wait
, struct wait_page_queue
, wait
);
3435 if (!wake_page_match(wpq
, key
))
3438 req
->rw
.kiocb
.ki_flags
&= ~IOCB_WAITQ
;
3439 list_del_init(&wait
->entry
);
3440 io_req_task_queue(req
);
3445 * This controls whether a given IO request should be armed for async page
3446 * based retry. If we return false here, the request is handed to the async
3447 * worker threads for retry. If we're doing buffered reads on a regular file,
3448 * we prepare a private wait_page_queue entry and retry the operation. This
3449 * will either succeed because the page is now uptodate and unlocked, or it
3450 * will register a callback when the page is unlocked at IO completion. Through
3451 * that callback, io_uring uses task_work to setup a retry of the operation.
3452 * That retry will attempt the buffered read again. The retry will generally
3453 * succeed, or in rare cases where it fails, we then fall back to using the
3454 * async worker threads for a blocking retry.
3456 static bool io_rw_should_retry(struct io_kiocb
*req
)
3458 struct io_async_rw
*rw
= req
->async_data
;
3459 struct wait_page_queue
*wait
= &rw
->wpq
;
3460 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3462 /* never retry for NOWAIT, we just complete with -EAGAIN */
3463 if (req
->flags
& REQ_F_NOWAIT
)
3466 /* Only for buffered IO */
3467 if (kiocb
->ki_flags
& (IOCB_DIRECT
| IOCB_HIPRI
))
3471 * just use poll if we can, and don't attempt if the fs doesn't
3472 * support callback based unlocks
3474 if (file_can_poll(req
->file
) || !(req
->file
->f_mode
& FMODE_BUF_RASYNC
))
3477 wait
->wait
.func
= io_async_buf_func
;
3478 wait
->wait
.private = req
;
3479 wait
->wait
.flags
= 0;
3480 INIT_LIST_HEAD(&wait
->wait
.entry
);
3481 kiocb
->ki_flags
|= IOCB_WAITQ
;
3482 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3483 kiocb
->ki_waitq
= wait
;
3487 static inline int io_iter_do_read(struct io_kiocb
*req
, struct iov_iter
*iter
)
3489 if (req
->file
->f_op
->read_iter
)
3490 return call_read_iter(req
->file
, &req
->rw
.kiocb
, iter
);
3491 else if (req
->file
->f_op
->read
)
3492 return loop_rw_iter(READ
, req
, iter
);
3497 static bool need_read_all(struct io_kiocb
*req
)
3499 return req
->flags
& REQ_F_ISREG
||
3500 S_ISBLK(file_inode(req
->file
)->i_mode
);
3503 static int io_read(struct io_kiocb
*req
, unsigned int issue_flags
)
3505 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3506 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3507 struct iov_iter __iter
, *iter
= &__iter
;
3508 struct io_async_rw
*rw
= req
->async_data
;
3509 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3510 struct iov_iter_state __state
, *state
;
3515 state
= &rw
->iter_state
;
3517 * We come here from an earlier attempt, restore our state to
3518 * match in case it doesn't. It's cheap enough that we don't
3519 * need to make this conditional.
3521 iov_iter_restore(iter
, state
);
3524 ret
= io_import_iovec(READ
, req
, &iovec
, iter
, !force_nonblock
);
3528 iov_iter_save_state(iter
, state
);
3530 req
->result
= iov_iter_count(iter
);
3532 /* Ensure we clear previously set non-block flag */
3533 if (!force_nonblock
)
3534 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3536 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3538 /* If the file doesn't support async, just async punt */
3539 if (force_nonblock
&& !io_file_supports_nowait(req
, READ
)) {
3540 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3541 return ret
?: -EAGAIN
;
3544 ret
= rw_verify_area(READ
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3545 if (unlikely(ret
)) {
3550 ret
= io_iter_do_read(req
, iter
);
3552 if (ret
== -EAGAIN
|| (req
->flags
& REQ_F_REISSUE
)) {
3553 req
->flags
&= ~REQ_F_REISSUE
;
3554 /* IOPOLL retry should happen for io-wq threads */
3555 if (!force_nonblock
&& !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3557 /* no retry on NONBLOCK nor RWF_NOWAIT */
3558 if (req
->flags
& REQ_F_NOWAIT
)
3561 } else if (ret
== -EIOCBQUEUED
) {
3563 } else if (ret
<= 0 || ret
== req
->result
|| !force_nonblock
||
3564 (req
->flags
& REQ_F_NOWAIT
) || !need_read_all(req
)) {
3565 /* read all, failed, already did sync or don't want to retry */
3570 * Don't depend on the iter state matching what was consumed, or being
3571 * untouched in case of error. Restore it and we'll advance it
3572 * manually if we need to.
3574 iov_iter_restore(iter
, state
);
3576 ret2
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, true);
3581 rw
= req
->async_data
;
3583 * Now use our persistent iterator and state, if we aren't already.
3584 * We've restored and mapped the iter to match.
3586 if (iter
!= &rw
->iter
) {
3588 state
= &rw
->iter_state
;
3593 * We end up here because of a partial read, either from
3594 * above or inside this loop. Advance the iter by the bytes
3595 * that were consumed.
3597 iov_iter_advance(iter
, ret
);
3598 if (!iov_iter_count(iter
))
3600 rw
->bytes_done
+= ret
;
3601 iov_iter_save_state(iter
, state
);
3603 /* if we can retry, do so with the callbacks armed */
3604 if (!io_rw_should_retry(req
)) {
3605 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3610 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3611 * we get -EIOCBQUEUED, then we'll get a notification when the
3612 * desired page gets unlocked. We can also get a partial read
3613 * here, and if we do, then just retry at the new offset.
3615 ret
= io_iter_do_read(req
, iter
);
3616 if (ret
== -EIOCBQUEUED
)
3618 /* we got some bytes, but not all. retry. */
3619 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3620 iov_iter_restore(iter
, state
);
3623 kiocb_done(kiocb
, ret
, issue_flags
);
3625 /* it's faster to check here then delegate to kfree */
3631 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3633 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3635 return io_prep_rw(req
, sqe
, WRITE
);
3638 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3640 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3641 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3642 struct iov_iter __iter
, *iter
= &__iter
;
3643 struct io_async_rw
*rw
= req
->async_data
;
3644 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3645 struct iov_iter_state __state
, *state
;
3650 state
= &rw
->iter_state
;
3651 iov_iter_restore(iter
, state
);
3654 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3658 iov_iter_save_state(iter
, state
);
3660 req
->result
= iov_iter_count(iter
);
3662 /* Ensure we clear previously set non-block flag */
3663 if (!force_nonblock
)
3664 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3666 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3668 /* If the file doesn't support async, just async punt */
3669 if (force_nonblock
&& !io_file_supports_nowait(req
, WRITE
))
3672 /* file path doesn't support NOWAIT for non-direct_IO */
3673 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3674 (req
->flags
& REQ_F_ISREG
))
3677 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3682 * Open-code file_start_write here to grab freeze protection,
3683 * which will be released by another thread in
3684 * io_complete_rw(). Fool lockdep by telling it the lock got
3685 * released so that it doesn't complain about the held lock when
3686 * we return to userspace.
3688 if (req
->flags
& REQ_F_ISREG
) {
3689 sb_start_write(file_inode(req
->file
)->i_sb
);
3690 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3693 kiocb
->ki_flags
|= IOCB_WRITE
;
3695 if (req
->file
->f_op
->write_iter
)
3696 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3697 else if (req
->file
->f_op
->write
)
3698 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3702 if (req
->flags
& REQ_F_REISSUE
) {
3703 req
->flags
&= ~REQ_F_REISSUE
;
3708 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3709 * retry them without IOCB_NOWAIT.
3711 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3713 /* no retry on NONBLOCK nor RWF_NOWAIT */
3714 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3716 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3717 /* IOPOLL retry should happen for io-wq threads */
3718 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3721 kiocb_done(kiocb
, ret2
, issue_flags
);
3724 iov_iter_restore(iter
, state
);
3725 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3726 return ret
?: -EAGAIN
;
3729 /* it's reportedly faster than delegating the null check to kfree() */
3735 static int io_renameat_prep(struct io_kiocb
*req
,
3736 const struct io_uring_sqe
*sqe
)
3738 struct io_rename
*ren
= &req
->rename
;
3739 const char __user
*oldf
, *newf
;
3741 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3743 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3745 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3748 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3749 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3750 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3751 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3752 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3754 ren
->oldpath
= getname(oldf
);
3755 if (IS_ERR(ren
->oldpath
))
3756 return PTR_ERR(ren
->oldpath
);
3758 ren
->newpath
= getname(newf
);
3759 if (IS_ERR(ren
->newpath
)) {
3760 putname(ren
->oldpath
);
3761 return PTR_ERR(ren
->newpath
);
3764 req
->flags
|= REQ_F_NEED_CLEANUP
;
3768 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3770 struct io_rename
*ren
= &req
->rename
;
3773 if (issue_flags
& IO_URING_F_NONBLOCK
)
3776 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3777 ren
->newpath
, ren
->flags
);
3779 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3782 io_req_complete(req
, ret
);
3786 static int io_unlinkat_prep(struct io_kiocb
*req
,
3787 const struct io_uring_sqe
*sqe
)
3789 struct io_unlink
*un
= &req
->unlink
;
3790 const char __user
*fname
;
3792 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3794 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
3797 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3800 un
->dfd
= READ_ONCE(sqe
->fd
);
3802 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3803 if (un
->flags
& ~AT_REMOVEDIR
)
3806 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3807 un
->filename
= getname(fname
);
3808 if (IS_ERR(un
->filename
))
3809 return PTR_ERR(un
->filename
);
3811 req
->flags
|= REQ_F_NEED_CLEANUP
;
3815 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3817 struct io_unlink
*un
= &req
->unlink
;
3820 if (issue_flags
& IO_URING_F_NONBLOCK
)
3823 if (un
->flags
& AT_REMOVEDIR
)
3824 ret
= do_rmdir(un
->dfd
, un
->filename
);
3826 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3828 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3831 io_req_complete(req
, ret
);
3835 static int io_mkdirat_prep(struct io_kiocb
*req
,
3836 const struct io_uring_sqe
*sqe
)
3838 struct io_mkdir
*mkd
= &req
->mkdir
;
3839 const char __user
*fname
;
3841 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3843 if (sqe
->ioprio
|| sqe
->off
|| sqe
->rw_flags
|| sqe
->buf_index
||
3846 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3849 mkd
->dfd
= READ_ONCE(sqe
->fd
);
3850 mkd
->mode
= READ_ONCE(sqe
->len
);
3852 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3853 mkd
->filename
= getname(fname
);
3854 if (IS_ERR(mkd
->filename
))
3855 return PTR_ERR(mkd
->filename
);
3857 req
->flags
|= REQ_F_NEED_CLEANUP
;
3861 static int io_mkdirat(struct io_kiocb
*req
, int issue_flags
)
3863 struct io_mkdir
*mkd
= &req
->mkdir
;
3866 if (issue_flags
& IO_URING_F_NONBLOCK
)
3869 ret
= do_mkdirat(mkd
->dfd
, mkd
->filename
, mkd
->mode
);
3871 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3874 io_req_complete(req
, ret
);
3878 static int io_symlinkat_prep(struct io_kiocb
*req
,
3879 const struct io_uring_sqe
*sqe
)
3881 struct io_symlink
*sl
= &req
->symlink
;
3882 const char __user
*oldpath
, *newpath
;
3884 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3886 if (sqe
->ioprio
|| sqe
->len
|| sqe
->rw_flags
|| sqe
->buf_index
||
3889 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3892 sl
->new_dfd
= READ_ONCE(sqe
->fd
);
3893 oldpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3894 newpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3896 sl
->oldpath
= getname(oldpath
);
3897 if (IS_ERR(sl
->oldpath
))
3898 return PTR_ERR(sl
->oldpath
);
3900 sl
->newpath
= getname(newpath
);
3901 if (IS_ERR(sl
->newpath
)) {
3902 putname(sl
->oldpath
);
3903 return PTR_ERR(sl
->newpath
);
3906 req
->flags
|= REQ_F_NEED_CLEANUP
;
3910 static int io_symlinkat(struct io_kiocb
*req
, int issue_flags
)
3912 struct io_symlink
*sl
= &req
->symlink
;
3915 if (issue_flags
& IO_URING_F_NONBLOCK
)
3918 ret
= do_symlinkat(sl
->oldpath
, sl
->new_dfd
, sl
->newpath
);
3920 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3923 io_req_complete(req
, ret
);
3927 static int io_linkat_prep(struct io_kiocb
*req
,
3928 const struct io_uring_sqe
*sqe
)
3930 struct io_hardlink
*lnk
= &req
->hardlink
;
3931 const char __user
*oldf
, *newf
;
3933 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3935 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3937 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3940 lnk
->old_dfd
= READ_ONCE(sqe
->fd
);
3941 lnk
->new_dfd
= READ_ONCE(sqe
->len
);
3942 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3943 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3944 lnk
->flags
= READ_ONCE(sqe
->hardlink_flags
);
3946 lnk
->oldpath
= getname(oldf
);
3947 if (IS_ERR(lnk
->oldpath
))
3948 return PTR_ERR(lnk
->oldpath
);
3950 lnk
->newpath
= getname(newf
);
3951 if (IS_ERR(lnk
->newpath
)) {
3952 putname(lnk
->oldpath
);
3953 return PTR_ERR(lnk
->newpath
);
3956 req
->flags
|= REQ_F_NEED_CLEANUP
;
3960 static int io_linkat(struct io_kiocb
*req
, int issue_flags
)
3962 struct io_hardlink
*lnk
= &req
->hardlink
;
3965 if (issue_flags
& IO_URING_F_NONBLOCK
)
3968 ret
= do_linkat(lnk
->old_dfd
, lnk
->oldpath
, lnk
->new_dfd
,
3969 lnk
->newpath
, lnk
->flags
);
3971 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3974 io_req_complete(req
, ret
);
3978 static int io_shutdown_prep(struct io_kiocb
*req
,
3979 const struct io_uring_sqe
*sqe
)
3981 #if defined(CONFIG_NET)
3982 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3984 if (unlikely(sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3985 sqe
->buf_index
|| sqe
->splice_fd_in
))
3988 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
3995 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
3997 #if defined(CONFIG_NET)
3998 struct socket
*sock
;
4001 if (issue_flags
& IO_URING_F_NONBLOCK
)
4004 sock
= sock_from_file(req
->file
);
4005 if (unlikely(!sock
))
4008 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
4011 io_req_complete(req
, ret
);
4018 static int __io_splice_prep(struct io_kiocb
*req
,
4019 const struct io_uring_sqe
*sqe
)
4021 struct io_splice
*sp
= &req
->splice
;
4022 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
4024 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4027 sp
->len
= READ_ONCE(sqe
->len
);
4028 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
4029 if (unlikely(sp
->flags
& ~valid_flags
))
4031 sp
->splice_fd_in
= READ_ONCE(sqe
->splice_fd_in
);
4035 static int io_tee_prep(struct io_kiocb
*req
,
4036 const struct io_uring_sqe
*sqe
)
4038 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
4040 return __io_splice_prep(req
, sqe
);
4043 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
4045 struct io_splice
*sp
= &req
->splice
;
4046 struct file
*out
= sp
->file_out
;
4047 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4051 if (issue_flags
& IO_URING_F_NONBLOCK
)
4054 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4055 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4062 ret
= do_tee(in
, out
, sp
->len
, flags
);
4064 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4069 io_req_complete(req
, ret
);
4073 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4075 struct io_splice
*sp
= &req
->splice
;
4077 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
4078 sp
->off_out
= READ_ONCE(sqe
->off
);
4079 return __io_splice_prep(req
, sqe
);
4082 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
4084 struct io_splice
*sp
= &req
->splice
;
4085 struct file
*out
= sp
->file_out
;
4086 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4087 loff_t
*poff_in
, *poff_out
;
4091 if (issue_flags
& IO_URING_F_NONBLOCK
)
4094 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4095 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4101 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
4102 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
4105 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
4107 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4112 io_req_complete(req
, ret
);
4117 * IORING_OP_NOP just posts a completion event, nothing else.
4119 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
4121 struct io_ring_ctx
*ctx
= req
->ctx
;
4123 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4126 __io_req_complete(req
, issue_flags
, 0, 0);
4130 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4132 struct io_ring_ctx
*ctx
= req
->ctx
;
4134 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4136 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4140 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
4141 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
4144 req
->sync
.off
= READ_ONCE(sqe
->off
);
4145 req
->sync
.len
= READ_ONCE(sqe
->len
);
4149 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
4151 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
4154 /* fsync always requires a blocking context */
4155 if (issue_flags
& IO_URING_F_NONBLOCK
)
4158 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
4159 end
> 0 ? end
: LLONG_MAX
,
4160 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
4163 io_req_complete(req
, ret
);
4167 static int io_fallocate_prep(struct io_kiocb
*req
,
4168 const struct io_uring_sqe
*sqe
)
4170 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
||
4173 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4176 req
->sync
.off
= READ_ONCE(sqe
->off
);
4177 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4178 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4182 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
4186 /* fallocate always requiring blocking context */
4187 if (issue_flags
& IO_URING_F_NONBLOCK
)
4189 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4194 fsnotify_modify(req
->file
);
4195 io_req_complete(req
, ret
);
4199 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4201 const char __user
*fname
;
4204 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4206 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4208 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4211 /* open.how should be already initialised */
4212 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4213 req
->open
.how
.flags
|= O_LARGEFILE
;
4215 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4216 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4217 req
->open
.filename
= getname(fname
);
4218 if (IS_ERR(req
->open
.filename
)) {
4219 ret
= PTR_ERR(req
->open
.filename
);
4220 req
->open
.filename
= NULL
;
4224 req
->open
.file_slot
= READ_ONCE(sqe
->file_index
);
4225 if (req
->open
.file_slot
&& (req
->open
.how
.flags
& O_CLOEXEC
))
4228 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4229 req
->flags
|= REQ_F_NEED_CLEANUP
;
4233 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4235 u64 mode
= READ_ONCE(sqe
->len
);
4236 u64 flags
= READ_ONCE(sqe
->open_flags
);
4238 req
->open
.how
= build_open_how(flags
, mode
);
4239 return __io_openat_prep(req
, sqe
);
4242 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4244 struct open_how __user
*how
;
4248 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4249 len
= READ_ONCE(sqe
->len
);
4250 if (len
< OPEN_HOW_SIZE_VER0
)
4253 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4258 return __io_openat_prep(req
, sqe
);
4261 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
4263 struct open_flags op
;
4265 bool resolve_nonblock
, nonblock_set
;
4266 bool fixed
= !!req
->open
.file_slot
;
4269 ret
= build_open_flags(&req
->open
.how
, &op
);
4272 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
4273 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
4274 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4276 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4277 * it'll always -EAGAIN
4279 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
4281 op
.lookup_flags
|= LOOKUP_CACHED
;
4282 op
.open_flag
|= O_NONBLOCK
;
4286 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4291 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4294 * We could hang on to this 'fd' on retrying, but seems like
4295 * marginal gain for something that is now known to be a slower
4296 * path. So just put it, and we'll get a new one when we retry.
4301 ret
= PTR_ERR(file
);
4302 /* only retry if RESOLVE_CACHED wasn't already set by application */
4303 if (ret
== -EAGAIN
&&
4304 (!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)))
4309 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
4310 file
->f_flags
&= ~O_NONBLOCK
;
4311 fsnotify_open(file
);
4314 fd_install(ret
, file
);
4316 ret
= io_install_fixed_file(req
, file
, issue_flags
,
4317 req
->open
.file_slot
- 1);
4319 putname(req
->open
.filename
);
4320 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4323 __io_req_complete(req
, issue_flags
, ret
, 0);
4327 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
4329 return io_openat2(req
, issue_flags
);
4332 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4333 const struct io_uring_sqe
*sqe
)
4335 struct io_provide_buf
*p
= &req
->pbuf
;
4338 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
4342 tmp
= READ_ONCE(sqe
->fd
);
4343 if (!tmp
|| tmp
> USHRT_MAX
)
4346 memset(p
, 0, sizeof(*p
));
4348 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4352 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4353 int bgid
, unsigned nbufs
)
4357 /* shouldn't happen */
4361 /* the head kbuf is the list itself */
4362 while (!list_empty(&buf
->list
)) {
4363 struct io_buffer
*nxt
;
4365 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4366 list_del(&nxt
->list
);
4374 xa_erase(&ctx
->io_buffers
, bgid
);
4379 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4381 struct io_provide_buf
*p
= &req
->pbuf
;
4382 struct io_ring_ctx
*ctx
= req
->ctx
;
4383 struct io_buffer
*head
;
4385 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4387 io_ring_submit_lock(ctx
, !force_nonblock
);
4389 lockdep_assert_held(&ctx
->uring_lock
);
4392 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4394 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4398 /* complete before unlock, IOPOLL may need the lock */
4399 __io_req_complete(req
, issue_flags
, ret
, 0);
4400 io_ring_submit_unlock(ctx
, !force_nonblock
);
4404 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4405 const struct io_uring_sqe
*sqe
)
4407 unsigned long size
, tmp_check
;
4408 struct io_provide_buf
*p
= &req
->pbuf
;
4411 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
4414 tmp
= READ_ONCE(sqe
->fd
);
4415 if (!tmp
|| tmp
> USHRT_MAX
)
4418 p
->addr
= READ_ONCE(sqe
->addr
);
4419 p
->len
= READ_ONCE(sqe
->len
);
4421 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
4424 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
4427 size
= (unsigned long)p
->len
* p
->nbufs
;
4428 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4431 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4432 tmp
= READ_ONCE(sqe
->off
);
4433 if (tmp
> USHRT_MAX
)
4439 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4441 struct io_buffer
*buf
;
4442 u64 addr
= pbuf
->addr
;
4443 int i
, bid
= pbuf
->bid
;
4445 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4446 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL_ACCOUNT
);
4451 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
4456 INIT_LIST_HEAD(&buf
->list
);
4459 list_add_tail(&buf
->list
, &(*head
)->list
);
4464 return i
? i
: -ENOMEM
;
4467 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4469 struct io_provide_buf
*p
= &req
->pbuf
;
4470 struct io_ring_ctx
*ctx
= req
->ctx
;
4471 struct io_buffer
*head
, *list
;
4473 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4475 io_ring_submit_lock(ctx
, !force_nonblock
);
4477 lockdep_assert_held(&ctx
->uring_lock
);
4479 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4481 ret
= io_add_buffers(p
, &head
);
4482 if (ret
>= 0 && !list
) {
4483 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
, GFP_KERNEL
);
4485 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4489 /* complete before unlock, IOPOLL may need the lock */
4490 __io_req_complete(req
, issue_flags
, ret
, 0);
4491 io_ring_submit_unlock(ctx
, !force_nonblock
);
4495 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4496 const struct io_uring_sqe
*sqe
)
4498 #if defined(CONFIG_EPOLL)
4499 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4501 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4504 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4505 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4506 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4508 if (ep_op_has_event(req
->epoll
.op
)) {
4509 struct epoll_event __user
*ev
;
4511 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4512 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4522 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4524 #if defined(CONFIG_EPOLL)
4525 struct io_epoll
*ie
= &req
->epoll
;
4527 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4529 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4530 if (force_nonblock
&& ret
== -EAGAIN
)
4535 __io_req_complete(req
, issue_flags
, ret
, 0);
4542 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4544 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4545 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->splice_fd_in
)
4547 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4550 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4551 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4552 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4559 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4561 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4562 struct io_madvise
*ma
= &req
->madvise
;
4565 if (issue_flags
& IO_URING_F_NONBLOCK
)
4568 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4571 io_req_complete(req
, ret
);
4578 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4580 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
|| sqe
->splice_fd_in
)
4582 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4585 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4586 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4587 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4591 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4593 struct io_fadvise
*fa
= &req
->fadvise
;
4596 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4597 switch (fa
->advice
) {
4598 case POSIX_FADV_NORMAL
:
4599 case POSIX_FADV_RANDOM
:
4600 case POSIX_FADV_SEQUENTIAL
:
4607 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4610 __io_req_complete(req
, issue_flags
, ret
, 0);
4614 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4616 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4618 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4620 if (req
->flags
& REQ_F_FIXED_FILE
)
4623 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4624 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4625 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4626 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4627 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4632 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4634 struct io_statx
*ctx
= &req
->statx
;
4637 if (issue_flags
& IO_URING_F_NONBLOCK
)
4640 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4645 io_req_complete(req
, ret
);
4649 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4651 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4653 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4654 sqe
->rw_flags
|| sqe
->buf_index
)
4656 if (req
->flags
& REQ_F_FIXED_FILE
)
4659 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4660 req
->close
.file_slot
= READ_ONCE(sqe
->file_index
);
4661 if (req
->close
.file_slot
&& req
->close
.fd
)
4667 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4669 struct files_struct
*files
= current
->files
;
4670 struct io_close
*close
= &req
->close
;
4671 struct fdtable
*fdt
;
4672 struct file
*file
= NULL
;
4675 if (req
->close
.file_slot
) {
4676 ret
= io_close_fixed(req
, issue_flags
);
4680 spin_lock(&files
->file_lock
);
4681 fdt
= files_fdtable(files
);
4682 if (close
->fd
>= fdt
->max_fds
) {
4683 spin_unlock(&files
->file_lock
);
4686 file
= fdt
->fd
[close
->fd
];
4687 if (!file
|| file
->f_op
== &io_uring_fops
) {
4688 spin_unlock(&files
->file_lock
);
4693 /* if the file has a flush method, be safe and punt to async */
4694 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4695 spin_unlock(&files
->file_lock
);
4699 ret
= __close_fd_get_file(close
->fd
, &file
);
4700 spin_unlock(&files
->file_lock
);
4707 /* No ->flush() or already async, safely close from here */
4708 ret
= filp_close(file
, current
->files
);
4714 __io_req_complete(req
, issue_flags
, ret
, 0);
4718 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4720 struct io_ring_ctx
*ctx
= req
->ctx
;
4722 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4724 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4728 req
->sync
.off
= READ_ONCE(sqe
->off
);
4729 req
->sync
.len
= READ_ONCE(sqe
->len
);
4730 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4734 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4738 /* sync_file_range always requires a blocking context */
4739 if (issue_flags
& IO_URING_F_NONBLOCK
)
4742 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4746 io_req_complete(req
, ret
);
4750 #if defined(CONFIG_NET)
4751 static int io_setup_async_msg(struct io_kiocb
*req
,
4752 struct io_async_msghdr
*kmsg
)
4754 struct io_async_msghdr
*async_msg
= req
->async_data
;
4758 if (io_alloc_async_data(req
)) {
4759 kfree(kmsg
->free_iov
);
4762 async_msg
= req
->async_data
;
4763 req
->flags
|= REQ_F_NEED_CLEANUP
;
4764 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4765 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4766 /* if were using fast_iov, set it to the new one */
4767 if (!async_msg
->free_iov
)
4768 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4773 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4774 struct io_async_msghdr
*iomsg
)
4776 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4777 iomsg
->free_iov
= iomsg
->fast_iov
;
4778 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4779 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4782 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4786 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4788 req
->flags
|= REQ_F_NEED_CLEANUP
;
4792 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4794 struct io_sr_msg
*sr
= &req
->sr_msg
;
4796 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4798 if (unlikely(sqe
->addr2
|| sqe
->file_index
))
4800 if (unlikely(sqe
->addr2
|| sqe
->file_index
|| sqe
->ioprio
))
4803 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4804 sr
->len
= READ_ONCE(sqe
->len
);
4805 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4806 if (sr
->msg_flags
& MSG_DONTWAIT
)
4807 req
->flags
|= REQ_F_NOWAIT
;
4809 #ifdef CONFIG_COMPAT
4810 if (req
->ctx
->compat
)
4811 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4816 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4818 struct io_async_msghdr iomsg
, *kmsg
;
4819 struct socket
*sock
;
4824 sock
= sock_from_file(req
->file
);
4825 if (unlikely(!sock
))
4828 kmsg
= req
->async_data
;
4830 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4836 flags
= req
->sr_msg
.msg_flags
;
4837 if (issue_flags
& IO_URING_F_NONBLOCK
)
4838 flags
|= MSG_DONTWAIT
;
4839 if (flags
& MSG_WAITALL
)
4840 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4842 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4843 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4844 return io_setup_async_msg(req
, kmsg
);
4845 if (ret
== -ERESTARTSYS
)
4848 /* fast path, check for non-NULL to avoid function call */
4850 kfree(kmsg
->free_iov
);
4851 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4854 __io_req_complete(req
, issue_flags
, ret
, 0);
4858 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4860 struct io_sr_msg
*sr
= &req
->sr_msg
;
4863 struct socket
*sock
;
4868 sock
= sock_from_file(req
->file
);
4869 if (unlikely(!sock
))
4872 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4876 msg
.msg_name
= NULL
;
4877 msg
.msg_control
= NULL
;
4878 msg
.msg_controllen
= 0;
4879 msg
.msg_namelen
= 0;
4881 flags
= req
->sr_msg
.msg_flags
;
4882 if (issue_flags
& IO_URING_F_NONBLOCK
)
4883 flags
|= MSG_DONTWAIT
;
4884 if (flags
& MSG_WAITALL
)
4885 min_ret
= iov_iter_count(&msg
.msg_iter
);
4887 msg
.msg_flags
= flags
;
4888 ret
= sock_sendmsg(sock
, &msg
);
4889 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4891 if (ret
== -ERESTARTSYS
)
4896 __io_req_complete(req
, issue_flags
, ret
, 0);
4900 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4901 struct io_async_msghdr
*iomsg
)
4903 struct io_sr_msg
*sr
= &req
->sr_msg
;
4904 struct iovec __user
*uiov
;
4908 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4909 &iomsg
->uaddr
, &uiov
, &iov_len
);
4913 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4916 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4918 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4919 iomsg
->free_iov
= NULL
;
4921 iomsg
->free_iov
= iomsg
->fast_iov
;
4922 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4923 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4932 #ifdef CONFIG_COMPAT
4933 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4934 struct io_async_msghdr
*iomsg
)
4936 struct io_sr_msg
*sr
= &req
->sr_msg
;
4937 struct compat_iovec __user
*uiov
;
4942 ret
= __get_compat_msghdr(&iomsg
->msg
, sr
->umsg_compat
, &iomsg
->uaddr
,
4947 uiov
= compat_ptr(ptr
);
4948 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4949 compat_ssize_t clen
;
4953 if (!access_ok(uiov
, sizeof(*uiov
)))
4955 if (__get_user(clen
, &uiov
->iov_len
))
4960 iomsg
->free_iov
= NULL
;
4962 iomsg
->free_iov
= iomsg
->fast_iov
;
4963 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4964 UIO_FASTIOV
, &iomsg
->free_iov
,
4965 &iomsg
->msg
.msg_iter
, true);
4974 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4975 struct io_async_msghdr
*iomsg
)
4977 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4979 #ifdef CONFIG_COMPAT
4980 if (req
->ctx
->compat
)
4981 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4984 return __io_recvmsg_copy_hdr(req
, iomsg
);
4987 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4990 struct io_sr_msg
*sr
= &req
->sr_msg
;
4991 struct io_buffer
*kbuf
;
4993 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
4998 req
->flags
|= REQ_F_BUFFER_SELECTED
;
5002 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
5004 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
5007 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
5011 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
5013 req
->flags
|= REQ_F_NEED_CLEANUP
;
5017 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5019 struct io_sr_msg
*sr
= &req
->sr_msg
;
5021 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5023 if (unlikely(sqe
->addr2
|| sqe
->file_index
))
5025 if (unlikely(sqe
->addr2
|| sqe
->file_index
|| sqe
->ioprio
))
5028 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5029 sr
->len
= READ_ONCE(sqe
->len
);
5030 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
5031 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
5032 if (sr
->msg_flags
& MSG_DONTWAIT
)
5033 req
->flags
|= REQ_F_NOWAIT
;
5035 #ifdef CONFIG_COMPAT
5036 if (req
->ctx
->compat
)
5037 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
5042 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
5044 struct io_async_msghdr iomsg
, *kmsg
;
5045 struct socket
*sock
;
5046 struct io_buffer
*kbuf
;
5049 int ret
, cflags
= 0;
5050 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5052 sock
= sock_from_file(req
->file
);
5053 if (unlikely(!sock
))
5056 kmsg
= req
->async_data
;
5058 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
5064 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5065 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5067 return PTR_ERR(kbuf
);
5068 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
5069 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
5070 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
5071 1, req
->sr_msg
.len
);
5074 flags
= req
->sr_msg
.msg_flags
;
5076 flags
|= MSG_DONTWAIT
;
5077 if (flags
& MSG_WAITALL
)
5078 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
5080 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
5081 kmsg
->uaddr
, flags
);
5082 if (force_nonblock
&& ret
== -EAGAIN
)
5083 return io_setup_async_msg(req
, kmsg
);
5084 if (ret
== -ERESTARTSYS
)
5087 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5088 cflags
= io_put_recv_kbuf(req
);
5089 /* fast path, check for non-NULL to avoid function call */
5091 kfree(kmsg
->free_iov
);
5092 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5093 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5095 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5099 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
5101 struct io_buffer
*kbuf
;
5102 struct io_sr_msg
*sr
= &req
->sr_msg
;
5104 void __user
*buf
= sr
->buf
;
5105 struct socket
*sock
;
5109 int ret
, cflags
= 0;
5110 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5112 sock
= sock_from_file(req
->file
);
5113 if (unlikely(!sock
))
5116 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5117 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5119 return PTR_ERR(kbuf
);
5120 buf
= u64_to_user_ptr(kbuf
->addr
);
5123 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
5127 msg
.msg_name
= NULL
;
5128 msg
.msg_control
= NULL
;
5129 msg
.msg_controllen
= 0;
5130 msg
.msg_namelen
= 0;
5131 msg
.msg_iocb
= NULL
;
5134 flags
= req
->sr_msg
.msg_flags
;
5136 flags
|= MSG_DONTWAIT
;
5137 if (flags
& MSG_WAITALL
)
5138 min_ret
= iov_iter_count(&msg
.msg_iter
);
5140 ret
= sock_recvmsg(sock
, &msg
, flags
);
5141 if (force_nonblock
&& ret
== -EAGAIN
)
5143 if (ret
== -ERESTARTSYS
)
5146 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5147 cflags
= io_put_recv_kbuf(req
);
5148 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5150 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5154 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5156 struct io_accept
*accept
= &req
->accept
;
5158 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5160 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
5163 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5164 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5165 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
5166 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
5168 accept
->file_slot
= READ_ONCE(sqe
->file_index
);
5169 if (accept
->file_slot
&& (accept
->flags
& SOCK_CLOEXEC
))
5171 if (accept
->flags
& ~(SOCK_CLOEXEC
| SOCK_NONBLOCK
))
5173 if (SOCK_NONBLOCK
!= O_NONBLOCK
&& (accept
->flags
& SOCK_NONBLOCK
))
5174 accept
->flags
= (accept
->flags
& ~SOCK_NONBLOCK
) | O_NONBLOCK
;
5178 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
5180 struct io_accept
*accept
= &req
->accept
;
5181 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5182 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5183 bool fixed
= !!accept
->file_slot
;
5187 if (req
->file
->f_flags
& O_NONBLOCK
)
5188 req
->flags
|= REQ_F_NOWAIT
;
5191 fd
= __get_unused_fd_flags(accept
->flags
, accept
->nofile
);
5192 if (unlikely(fd
< 0))
5195 file
= do_accept(req
->file
, file_flags
, accept
->addr
, accept
->addr_len
,
5200 ret
= PTR_ERR(file
);
5201 if (ret
== -EAGAIN
&& force_nonblock
)
5203 if (ret
== -ERESTARTSYS
)
5206 } else if (!fixed
) {
5207 fd_install(fd
, file
);
5210 ret
= io_install_fixed_file(req
, file
, issue_flags
,
5211 accept
->file_slot
- 1);
5213 __io_req_complete(req
, issue_flags
, ret
, 0);
5217 static int io_connect_prep_async(struct io_kiocb
*req
)
5219 struct io_async_connect
*io
= req
->async_data
;
5220 struct io_connect
*conn
= &req
->connect
;
5222 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
5225 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5227 struct io_connect
*conn
= &req
->connect
;
5229 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5231 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
||
5235 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5236 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5240 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
5242 struct io_async_connect __io
, *io
;
5243 unsigned file_flags
;
5245 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5247 if (req
->async_data
) {
5248 io
= req
->async_data
;
5250 ret
= move_addr_to_kernel(req
->connect
.addr
,
5251 req
->connect
.addr_len
,
5258 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5260 ret
= __sys_connect_file(req
->file
, &io
->address
,
5261 req
->connect
.addr_len
, file_flags
);
5262 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5263 if (req
->async_data
)
5265 if (io_alloc_async_data(req
)) {
5269 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5272 if (ret
== -ERESTARTSYS
)
5277 __io_req_complete(req
, issue_flags
, ret
, 0);
5280 #else /* !CONFIG_NET */
5281 #define IO_NETOP_FN(op) \
5282 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5284 return -EOPNOTSUPP; \
5287 #define IO_NETOP_PREP(op) \
5289 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5291 return -EOPNOTSUPP; \
5294 #define IO_NETOP_PREP_ASYNC(op) \
5296 static int io_##op##_prep_async(struct io_kiocb *req) \
5298 return -EOPNOTSUPP; \
5301 IO_NETOP_PREP_ASYNC(sendmsg
);
5302 IO_NETOP_PREP_ASYNC(recvmsg
);
5303 IO_NETOP_PREP_ASYNC(connect
);
5304 IO_NETOP_PREP(accept
);
5307 #endif /* CONFIG_NET */
5309 struct io_poll_table
{
5310 struct poll_table_struct pt
;
5311 struct io_kiocb
*req
;
5316 #define IO_POLL_CANCEL_FLAG BIT(31)
5317 #define IO_POLL_REF_MASK GENMASK(30, 0)
5320 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
5321 * bump it and acquire ownership. It's disallowed to modify requests while not
5322 * owning it, that prevents from races for enqueueing task_work's and b/w
5323 * arming poll and wakeups.
5325 static inline bool io_poll_get_ownership(struct io_kiocb
*req
)
5327 return !(atomic_fetch_inc(&req
->poll_refs
) & IO_POLL_REF_MASK
);
5330 static void io_poll_mark_cancelled(struct io_kiocb
*req
)
5332 atomic_or(IO_POLL_CANCEL_FLAG
, &req
->poll_refs
);
5335 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5337 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5338 if (req
->opcode
== IORING_OP_POLL_ADD
)
5339 return req
->async_data
;
5340 return req
->apoll
->double_poll
;
5343 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5345 if (req
->opcode
== IORING_OP_POLL_ADD
)
5347 return &req
->apoll
->poll
;
5350 static void io_poll_req_insert(struct io_kiocb
*req
)
5352 struct io_ring_ctx
*ctx
= req
->ctx
;
5353 struct hlist_head
*list
;
5355 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5356 hlist_add_head(&req
->hash_node
, list
);
5359 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5360 wait_queue_func_t wake_func
)
5363 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5364 /* mask in events that we always want/need */
5365 poll
->events
= events
| IO_POLL_UNMASK
;
5366 INIT_LIST_HEAD(&poll
->wait
.entry
);
5367 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5370 static inline void io_poll_remove_entry(struct io_poll_iocb
*poll
)
5372 struct wait_queue_head
*head
= poll
->head
;
5374 spin_lock_irq(&head
->lock
);
5375 list_del_init(&poll
->wait
.entry
);
5377 spin_unlock_irq(&head
->lock
);
5380 static void io_poll_remove_entries(struct io_kiocb
*req
)
5382 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5383 struct io_poll_iocb
*poll_double
= io_poll_get_double(req
);
5386 io_poll_remove_entry(poll
);
5387 if (poll_double
&& poll_double
->head
)
5388 io_poll_remove_entry(poll_double
);
5392 * All poll tw should go through this. Checks for poll events, manages
5393 * references, does rewait, etc.
5395 * Returns a negative error on failure. >0 when no action require, which is
5396 * either spurious wakeup or multishot CQE is served. 0 when it's done with
5397 * the request, then the mask is stored in req->result.
5399 static int io_poll_check_events(struct io_kiocb
*req
)
5401 struct io_ring_ctx
*ctx
= req
->ctx
;
5402 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5405 /* req->task == current here, checking PF_EXITING is safe */
5406 if (unlikely(req
->task
->flags
& PF_EXITING
))
5407 io_poll_mark_cancelled(req
);
5410 v
= atomic_read(&req
->poll_refs
);
5412 /* tw handler should be the owner, and so have some references */
5413 if (WARN_ON_ONCE(!(v
& IO_POLL_REF_MASK
)))
5415 if (v
& IO_POLL_CANCEL_FLAG
)
5419 struct poll_table_struct pt
= { ._key
= poll
->events
};
5421 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5424 /* multishot, just fill an CQE and proceed */
5425 if (req
->result
&& !(poll
->events
& EPOLLONESHOT
)) {
5426 __poll_t mask
= mangle_poll(req
->result
& poll
->events
);
5429 spin_lock(&ctx
->completion_lock
);
5430 filled
= io_fill_cqe_aux(ctx
, req
->user_data
, mask
,
5432 io_commit_cqring(ctx
);
5433 spin_unlock(&ctx
->completion_lock
);
5434 if (unlikely(!filled
))
5436 io_cqring_ev_posted(ctx
);
5437 } else if (req
->result
) {
5442 * Release all references, retry if someone tried to restart
5443 * task_work while we were executing it.
5445 } while (atomic_sub_return(v
& IO_POLL_REF_MASK
, &req
->poll_refs
));
5450 static void io_poll_task_func(struct io_kiocb
*req
, bool *locked
)
5452 struct io_ring_ctx
*ctx
= req
->ctx
;
5455 ret
= io_poll_check_events(req
);
5460 req
->result
= mangle_poll(req
->result
& req
->poll
.events
);
5466 io_poll_remove_entries(req
);
5467 spin_lock(&ctx
->completion_lock
);
5468 hash_del(&req
->hash_node
);
5469 spin_unlock(&ctx
->completion_lock
);
5470 io_req_complete_post(req
, req
->result
, 0);
5473 static void io_apoll_task_func(struct io_kiocb
*req
, bool *locked
)
5475 struct io_ring_ctx
*ctx
= req
->ctx
;
5478 ret
= io_poll_check_events(req
);
5482 io_poll_remove_entries(req
);
5483 spin_lock(&ctx
->completion_lock
);
5484 hash_del(&req
->hash_node
);
5485 spin_unlock(&ctx
->completion_lock
);
5488 io_req_task_submit(req
, locked
);
5490 io_req_complete_failed(req
, ret
);
5493 static void __io_poll_execute(struct io_kiocb
*req
, int mask
)
5496 if (req
->opcode
== IORING_OP_POLL_ADD
)
5497 req
->io_task_work
.func
= io_poll_task_func
;
5499 req
->io_task_work
.func
= io_apoll_task_func
;
5501 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5502 io_req_task_work_add(req
);
5505 static inline void io_poll_execute(struct io_kiocb
*req
, int res
)
5507 if (io_poll_get_ownership(req
))
5508 __io_poll_execute(req
, res
);
5511 static void io_poll_cancel_req(struct io_kiocb
*req
)
5513 io_poll_mark_cancelled(req
);
5514 /* kick tw, which should complete the request */
5515 io_poll_execute(req
, 0);
5518 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5521 struct io_kiocb
*req
= wait
->private;
5522 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
5524 __poll_t mask
= key_to_poll(key
);
5526 /* for instances that support it check for an event match first */
5527 if (mask
&& !(mask
& poll
->events
))
5530 if (io_poll_get_ownership(req
))
5531 __io_poll_execute(req
, mask
);
5535 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5536 struct wait_queue_head
*head
,
5537 struct io_poll_iocb
**poll_ptr
)
5539 struct io_kiocb
*req
= pt
->req
;
5542 * The file being polled uses multiple waitqueues for poll handling
5543 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5546 if (unlikely(pt
->nr_entries
)) {
5547 struct io_poll_iocb
*first
= poll
;
5549 /* double add on the same waitqueue head, ignore */
5550 if (first
->head
== head
)
5552 /* already have a 2nd entry, fail a third attempt */
5554 if ((*poll_ptr
)->head
== head
)
5556 pt
->error
= -EINVAL
;
5560 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5562 pt
->error
= -ENOMEM
;
5565 io_init_poll_iocb(poll
, first
->events
, first
->wait
.func
);
5571 poll
->wait
.private = req
;
5573 if (poll
->events
& EPOLLEXCLUSIVE
)
5574 add_wait_queue_exclusive(head
, &poll
->wait
);
5576 add_wait_queue(head
, &poll
->wait
);
5579 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5580 struct poll_table_struct
*p
)
5582 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5584 __io_queue_proc(&pt
->req
->poll
, pt
, head
,
5585 (struct io_poll_iocb
**) &pt
->req
->async_data
);
5588 static int __io_arm_poll_handler(struct io_kiocb
*req
,
5589 struct io_poll_iocb
*poll
,
5590 struct io_poll_table
*ipt
, __poll_t mask
)
5592 struct io_ring_ctx
*ctx
= req
->ctx
;
5595 INIT_HLIST_NODE(&req
->hash_node
);
5596 io_init_poll_iocb(poll
, mask
, io_poll_wake
);
5597 poll
->file
= req
->file
;
5598 poll
->wait
.private = req
;
5600 ipt
->pt
._key
= mask
;
5603 ipt
->nr_entries
= 0;
5606 * Take the ownership to delay any tw execution up until we're done
5607 * with poll arming. see io_poll_get_ownership().
5609 atomic_set(&req
->poll_refs
, 1);
5610 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5612 if (mask
&& (poll
->events
& EPOLLONESHOT
)) {
5613 io_poll_remove_entries(req
);
5614 /* no one else has access to the req, forget about the ref */
5617 if (!mask
&& unlikely(ipt
->error
|| !ipt
->nr_entries
)) {
5618 io_poll_remove_entries(req
);
5620 ipt
->error
= -EINVAL
;
5624 spin_lock(&ctx
->completion_lock
);
5625 io_poll_req_insert(req
);
5626 spin_unlock(&ctx
->completion_lock
);
5629 /* can't multishot if failed, just queue the event we've got */
5630 if (unlikely(ipt
->error
|| !ipt
->nr_entries
))
5631 poll
->events
|= EPOLLONESHOT
;
5632 __io_poll_execute(req
, mask
);
5637 * Release ownership. If someone tried to queue a tw while it was
5638 * locked, kick it off for them.
5640 v
= atomic_dec_return(&req
->poll_refs
);
5641 if (unlikely(v
& IO_POLL_REF_MASK
))
5642 __io_poll_execute(req
, 0);
5646 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5647 struct poll_table_struct
*p
)
5649 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5650 struct async_poll
*apoll
= pt
->req
->apoll
;
5652 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5661 static int io_arm_poll_handler(struct io_kiocb
*req
)
5663 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5664 struct io_ring_ctx
*ctx
= req
->ctx
;
5665 struct async_poll
*apoll
;
5666 struct io_poll_table ipt
;
5667 __poll_t mask
= EPOLLONESHOT
| POLLERR
| POLLPRI
;
5670 if (!req
->file
|| !file_can_poll(req
->file
))
5671 return IO_APOLL_ABORTED
;
5672 if (req
->flags
& REQ_F_POLLED
)
5673 return IO_APOLL_ABORTED
;
5674 if (!def
->pollin
&& !def
->pollout
)
5675 return IO_APOLL_ABORTED
;
5678 mask
|= POLLIN
| POLLRDNORM
;
5680 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5681 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5682 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5685 mask
|= POLLOUT
| POLLWRNORM
;
5688 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5689 if (unlikely(!apoll
))
5690 return IO_APOLL_ABORTED
;
5691 apoll
->double_poll
= NULL
;
5693 req
->flags
|= REQ_F_POLLED
;
5694 ipt
.pt
._qproc
= io_async_queue_proc
;
5696 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
);
5697 if (ret
|| ipt
.error
)
5698 return ret
? IO_APOLL_READY
: IO_APOLL_ABORTED
;
5700 trace_io_uring_poll_arm(ctx
, req
, req
->opcode
, req
->user_data
,
5701 mask
, apoll
->poll
.events
);
5706 * Returns true if we found and killed one or more poll requests
5708 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5711 struct hlist_node
*tmp
;
5712 struct io_kiocb
*req
;
5716 spin_lock(&ctx
->completion_lock
);
5717 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5718 struct hlist_head
*list
;
5720 list
= &ctx
->cancel_hash
[i
];
5721 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5722 if (io_match_task_safe(req
, tsk
, cancel_all
)) {
5723 hlist_del_init(&req
->hash_node
);
5724 io_poll_cancel_req(req
);
5729 spin_unlock(&ctx
->completion_lock
);
5733 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5735 __must_hold(&ctx
->completion_lock
)
5737 struct hlist_head
*list
;
5738 struct io_kiocb
*req
;
5740 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5741 hlist_for_each_entry(req
, list
, hash_node
) {
5742 if (sqe_addr
!= req
->user_data
)
5744 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
5751 static bool io_poll_disarm(struct io_kiocb
*req
)
5752 __must_hold(&ctx
->completion_lock
)
5754 if (!io_poll_get_ownership(req
))
5756 io_poll_remove_entries(req
);
5757 hash_del(&req
->hash_node
);
5761 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5763 __must_hold(&ctx
->completion_lock
)
5765 struct io_kiocb
*req
= io_poll_find(ctx
, sqe_addr
, poll_only
);
5769 io_poll_cancel_req(req
);
5773 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
5778 events
= READ_ONCE(sqe
->poll32_events
);
5780 events
= swahw32(events
);
5782 if (!(flags
& IORING_POLL_ADD_MULTI
))
5783 events
|= EPOLLONESHOT
;
5784 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
5787 static int io_poll_update_prep(struct io_kiocb
*req
,
5788 const struct io_uring_sqe
*sqe
)
5790 struct io_poll_update
*upd
= &req
->poll_update
;
5793 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5795 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
5797 flags
= READ_ONCE(sqe
->len
);
5798 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
5799 IORING_POLL_ADD_MULTI
))
5801 /* meaningless without update */
5802 if (flags
== IORING_POLL_ADD_MULTI
)
5805 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
5806 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
5807 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
5809 upd
->new_user_data
= READ_ONCE(sqe
->off
);
5810 if (!upd
->update_user_data
&& upd
->new_user_data
)
5812 if (upd
->update_events
)
5813 upd
->events
= io_poll_parse_events(sqe
, flags
);
5814 else if (sqe
->poll32_events
)
5820 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5822 struct io_poll_iocb
*poll
= &req
->poll
;
5825 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5827 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
5829 flags
= READ_ONCE(sqe
->len
);
5830 if (flags
& ~IORING_POLL_ADD_MULTI
)
5833 io_req_set_refcount(req
);
5834 poll
->events
= io_poll_parse_events(sqe
, flags
);
5838 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5840 struct io_poll_iocb
*poll
= &req
->poll
;
5841 struct io_poll_table ipt
;
5844 ipt
.pt
._qproc
= io_poll_queue_proc
;
5846 ret
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
);
5847 ret
= ret
?: ipt
.error
;
5849 __io_req_complete(req
, issue_flags
, ret
, 0);
5853 static int io_poll_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5855 struct io_ring_ctx
*ctx
= req
->ctx
;
5856 struct io_kiocb
*preq
;
5859 spin_lock(&ctx
->completion_lock
);
5860 preq
= io_poll_find(ctx
, req
->poll_update
.old_user_data
, true);
5861 if (!preq
|| !io_poll_disarm(preq
)) {
5862 spin_unlock(&ctx
->completion_lock
);
5863 ret
= preq
? -EALREADY
: -ENOENT
;
5866 spin_unlock(&ctx
->completion_lock
);
5868 if (req
->poll_update
.update_events
|| req
->poll_update
.update_user_data
) {
5869 /* only mask one event flags, keep behavior flags */
5870 if (req
->poll_update
.update_events
) {
5871 preq
->poll
.events
&= ~0xffff;
5872 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
5873 preq
->poll
.events
|= IO_POLL_UNMASK
;
5875 if (req
->poll_update
.update_user_data
)
5876 preq
->user_data
= req
->poll_update
.new_user_data
;
5878 ret2
= io_poll_add(preq
, issue_flags
);
5879 /* successfully updated, don't complete poll request */
5884 io_req_complete(preq
, -ECANCELED
);
5888 /* complete update request, we're done with it */
5889 io_req_complete(req
, ret
);
5893 static void io_req_task_timeout(struct io_kiocb
*req
, bool *locked
)
5896 io_req_complete_post(req
, -ETIME
, 0);
5899 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5901 struct io_timeout_data
*data
= container_of(timer
,
5902 struct io_timeout_data
, timer
);
5903 struct io_kiocb
*req
= data
->req
;
5904 struct io_ring_ctx
*ctx
= req
->ctx
;
5905 unsigned long flags
;
5907 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
5908 list_del_init(&req
->timeout
.list
);
5909 atomic_set(&req
->ctx
->cq_timeouts
,
5910 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
5911 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
5913 req
->io_task_work
.func
= io_req_task_timeout
;
5914 io_req_task_work_add(req
);
5915 return HRTIMER_NORESTART
;
5918 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
5920 __must_hold(&ctx
->timeout_lock
)
5922 struct io_timeout_data
*io
;
5923 struct io_kiocb
*req
;
5926 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
5927 found
= user_data
== req
->user_data
;
5932 return ERR_PTR(-ENOENT
);
5934 io
= req
->async_data
;
5935 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
5936 return ERR_PTR(-EALREADY
);
5937 list_del_init(&req
->timeout
.list
);
5941 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
5942 __must_hold(&ctx
->completion_lock
)
5943 __must_hold(&ctx
->timeout_lock
)
5945 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
5948 return PTR_ERR(req
);
5951 io_fill_cqe_req(req
, -ECANCELED
, 0);
5952 io_put_req_deferred(req
);
5956 static clockid_t
io_timeout_get_clock(struct io_timeout_data
*data
)
5958 switch (data
->flags
& IORING_TIMEOUT_CLOCK_MASK
) {
5959 case IORING_TIMEOUT_BOOTTIME
:
5960 return CLOCK_BOOTTIME
;
5961 case IORING_TIMEOUT_REALTIME
:
5962 return CLOCK_REALTIME
;
5964 /* can't happen, vetted at prep time */
5968 return CLOCK_MONOTONIC
;
5972 static int io_linked_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5973 struct timespec64
*ts
, enum hrtimer_mode mode
)
5974 __must_hold(&ctx
->timeout_lock
)
5976 struct io_timeout_data
*io
;
5977 struct io_kiocb
*req
;
5980 list_for_each_entry(req
, &ctx
->ltimeout_list
, timeout
.list
) {
5981 found
= user_data
== req
->user_data
;
5988 io
= req
->async_data
;
5989 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
5991 hrtimer_init(&io
->timer
, io_timeout_get_clock(io
), mode
);
5992 io
->timer
.function
= io_link_timeout_fn
;
5993 hrtimer_start(&io
->timer
, timespec64_to_ktime(*ts
), mode
);
5997 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
5998 struct timespec64
*ts
, enum hrtimer_mode mode
)
5999 __must_hold(&ctx
->timeout_lock
)
6001 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6002 struct io_timeout_data
*data
;
6005 return PTR_ERR(req
);
6007 req
->timeout
.off
= 0; /* noseq */
6008 data
= req
->async_data
;
6009 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
6010 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), mode
);
6011 data
->timer
.function
= io_timeout_fn
;
6012 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
6016 static int io_timeout_remove_prep(struct io_kiocb
*req
,
6017 const struct io_uring_sqe
*sqe
)
6019 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6021 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6023 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6025 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->splice_fd_in
)
6028 tr
->ltimeout
= false;
6029 tr
->addr
= READ_ONCE(sqe
->addr
);
6030 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
6031 if (tr
->flags
& IORING_TIMEOUT_UPDATE_MASK
) {
6032 if (hweight32(tr
->flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6034 if (tr
->flags
& IORING_LINK_TIMEOUT_UPDATE
)
6035 tr
->ltimeout
= true;
6036 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE_MASK
|IORING_TIMEOUT_ABS
))
6038 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
6040 } else if (tr
->flags
) {
6041 /* timeout removal doesn't support flags */
6048 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
6050 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
6055 * Remove or update an existing timeout command
6057 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
6059 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6060 struct io_ring_ctx
*ctx
= req
->ctx
;
6063 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
)) {
6064 spin_lock(&ctx
->completion_lock
);
6065 spin_lock_irq(&ctx
->timeout_lock
);
6066 ret
= io_timeout_cancel(ctx
, tr
->addr
);
6067 spin_unlock_irq(&ctx
->timeout_lock
);
6068 spin_unlock(&ctx
->completion_lock
);
6070 enum hrtimer_mode mode
= io_translate_timeout_mode(tr
->flags
);
6072 spin_lock_irq(&ctx
->timeout_lock
);
6074 ret
= io_linked_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6076 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6077 spin_unlock_irq(&ctx
->timeout_lock
);
6082 io_req_complete_post(req
, ret
, 0);
6086 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6087 bool is_timeout_link
)
6089 struct io_timeout_data
*data
;
6091 u32 off
= READ_ONCE(sqe
->off
);
6093 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6095 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1 ||
6098 if (off
&& is_timeout_link
)
6100 flags
= READ_ONCE(sqe
->timeout_flags
);
6101 if (flags
& ~(IORING_TIMEOUT_ABS
| IORING_TIMEOUT_CLOCK_MASK
))
6103 /* more than one clock specified is invalid, obviously */
6104 if (hweight32(flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6107 INIT_LIST_HEAD(&req
->timeout
.list
);
6108 req
->timeout
.off
= off
;
6109 if (unlikely(off
&& !req
->ctx
->off_timeout_used
))
6110 req
->ctx
->off_timeout_used
= true;
6112 if (!req
->async_data
&& io_alloc_async_data(req
))
6115 data
= req
->async_data
;
6117 data
->flags
= flags
;
6119 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
6122 INIT_LIST_HEAD(&req
->timeout
.list
);
6123 data
->mode
= io_translate_timeout_mode(flags
);
6124 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), data
->mode
);
6126 if (is_timeout_link
) {
6127 struct io_submit_link
*link
= &req
->ctx
->submit_state
.link
;
6131 if (link
->last
->opcode
== IORING_OP_LINK_TIMEOUT
)
6133 req
->timeout
.head
= link
->last
;
6134 link
->last
->flags
|= REQ_F_ARM_LTIMEOUT
;
6139 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
6141 struct io_ring_ctx
*ctx
= req
->ctx
;
6142 struct io_timeout_data
*data
= req
->async_data
;
6143 struct list_head
*entry
;
6144 u32 tail
, off
= req
->timeout
.off
;
6146 spin_lock_irq(&ctx
->timeout_lock
);
6149 * sqe->off holds how many events that need to occur for this
6150 * timeout event to be satisfied. If it isn't set, then this is
6151 * a pure timeout request, sequence isn't used.
6153 if (io_is_timeout_noseq(req
)) {
6154 entry
= ctx
->timeout_list
.prev
;
6158 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
6159 req
->timeout
.target_seq
= tail
+ off
;
6161 /* Update the last seq here in case io_flush_timeouts() hasn't.
6162 * This is safe because ->completion_lock is held, and submissions
6163 * and completions are never mixed in the same ->completion_lock section.
6165 ctx
->cq_last_tm_flush
= tail
;
6168 * Insertion sort, ensuring the first entry in the list is always
6169 * the one we need first.
6171 list_for_each_prev(entry
, &ctx
->timeout_list
) {
6172 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
6175 if (io_is_timeout_noseq(nxt
))
6177 /* nxt.seq is behind @tail, otherwise would've been completed */
6178 if (off
>= nxt
->timeout
.target_seq
- tail
)
6182 list_add(&req
->timeout
.list
, entry
);
6183 data
->timer
.function
= io_timeout_fn
;
6184 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
6185 spin_unlock_irq(&ctx
->timeout_lock
);
6189 struct io_cancel_data
{
6190 struct io_ring_ctx
*ctx
;
6194 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
6196 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6197 struct io_cancel_data
*cd
= data
;
6199 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
6202 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
6203 struct io_ring_ctx
*ctx
)
6205 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
6206 enum io_wq_cancel cancel_ret
;
6209 if (!tctx
|| !tctx
->io_wq
)
6212 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
6213 switch (cancel_ret
) {
6214 case IO_WQ_CANCEL_OK
:
6217 case IO_WQ_CANCEL_RUNNING
:
6220 case IO_WQ_CANCEL_NOTFOUND
:
6228 static int io_try_cancel_userdata(struct io_kiocb
*req
, u64 sqe_addr
)
6230 struct io_ring_ctx
*ctx
= req
->ctx
;
6233 WARN_ON_ONCE(!io_wq_current_is_worker() && req
->task
!= current
);
6235 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
6239 spin_lock(&ctx
->completion_lock
);
6240 spin_lock_irq(&ctx
->timeout_lock
);
6241 ret
= io_timeout_cancel(ctx
, sqe_addr
);
6242 spin_unlock_irq(&ctx
->timeout_lock
);
6245 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
6247 spin_unlock(&ctx
->completion_lock
);
6251 static int io_async_cancel_prep(struct io_kiocb
*req
,
6252 const struct io_uring_sqe
*sqe
)
6254 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6256 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6258 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
||
6262 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
6266 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
6268 struct io_ring_ctx
*ctx
= req
->ctx
;
6269 u64 sqe_addr
= req
->cancel
.addr
;
6270 struct io_tctx_node
*node
;
6273 ret
= io_try_cancel_userdata(req
, sqe_addr
);
6277 /* slow path, try all io-wq's */
6278 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6280 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
6281 struct io_uring_task
*tctx
= node
->task
->io_uring
;
6283 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
6287 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6291 io_req_complete_post(req
, ret
, 0);
6295 static int io_rsrc_update_prep(struct io_kiocb
*req
,
6296 const struct io_uring_sqe
*sqe
)
6298 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6300 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
6303 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
6304 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
6305 if (!req
->rsrc_update
.nr_args
)
6307 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
6311 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
6313 struct io_ring_ctx
*ctx
= req
->ctx
;
6314 struct io_uring_rsrc_update2 up
;
6317 up
.offset
= req
->rsrc_update
.offset
;
6318 up
.data
= req
->rsrc_update
.arg
;
6324 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6325 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
6326 &up
, req
->rsrc_update
.nr_args
);
6327 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6331 __io_req_complete(req
, issue_flags
, ret
, 0);
6335 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6337 switch (req
->opcode
) {
6340 case IORING_OP_READV
:
6341 case IORING_OP_READ_FIXED
:
6342 case IORING_OP_READ
:
6343 return io_read_prep(req
, sqe
);
6344 case IORING_OP_WRITEV
:
6345 case IORING_OP_WRITE_FIXED
:
6346 case IORING_OP_WRITE
:
6347 return io_write_prep(req
, sqe
);
6348 case IORING_OP_POLL_ADD
:
6349 return io_poll_add_prep(req
, sqe
);
6350 case IORING_OP_POLL_REMOVE
:
6351 return io_poll_update_prep(req
, sqe
);
6352 case IORING_OP_FSYNC
:
6353 return io_fsync_prep(req
, sqe
);
6354 case IORING_OP_SYNC_FILE_RANGE
:
6355 return io_sfr_prep(req
, sqe
);
6356 case IORING_OP_SENDMSG
:
6357 case IORING_OP_SEND
:
6358 return io_sendmsg_prep(req
, sqe
);
6359 case IORING_OP_RECVMSG
:
6360 case IORING_OP_RECV
:
6361 return io_recvmsg_prep(req
, sqe
);
6362 case IORING_OP_CONNECT
:
6363 return io_connect_prep(req
, sqe
);
6364 case IORING_OP_TIMEOUT
:
6365 return io_timeout_prep(req
, sqe
, false);
6366 case IORING_OP_TIMEOUT_REMOVE
:
6367 return io_timeout_remove_prep(req
, sqe
);
6368 case IORING_OP_ASYNC_CANCEL
:
6369 return io_async_cancel_prep(req
, sqe
);
6370 case IORING_OP_LINK_TIMEOUT
:
6371 return io_timeout_prep(req
, sqe
, true);
6372 case IORING_OP_ACCEPT
:
6373 return io_accept_prep(req
, sqe
);
6374 case IORING_OP_FALLOCATE
:
6375 return io_fallocate_prep(req
, sqe
);
6376 case IORING_OP_OPENAT
:
6377 return io_openat_prep(req
, sqe
);
6378 case IORING_OP_CLOSE
:
6379 return io_close_prep(req
, sqe
);
6380 case IORING_OP_FILES_UPDATE
:
6381 return io_rsrc_update_prep(req
, sqe
);
6382 case IORING_OP_STATX
:
6383 return io_statx_prep(req
, sqe
);
6384 case IORING_OP_FADVISE
:
6385 return io_fadvise_prep(req
, sqe
);
6386 case IORING_OP_MADVISE
:
6387 return io_madvise_prep(req
, sqe
);
6388 case IORING_OP_OPENAT2
:
6389 return io_openat2_prep(req
, sqe
);
6390 case IORING_OP_EPOLL_CTL
:
6391 return io_epoll_ctl_prep(req
, sqe
);
6392 case IORING_OP_SPLICE
:
6393 return io_splice_prep(req
, sqe
);
6394 case IORING_OP_PROVIDE_BUFFERS
:
6395 return io_provide_buffers_prep(req
, sqe
);
6396 case IORING_OP_REMOVE_BUFFERS
:
6397 return io_remove_buffers_prep(req
, sqe
);
6399 return io_tee_prep(req
, sqe
);
6400 case IORING_OP_SHUTDOWN
:
6401 return io_shutdown_prep(req
, sqe
);
6402 case IORING_OP_RENAMEAT
:
6403 return io_renameat_prep(req
, sqe
);
6404 case IORING_OP_UNLINKAT
:
6405 return io_unlinkat_prep(req
, sqe
);
6406 case IORING_OP_MKDIRAT
:
6407 return io_mkdirat_prep(req
, sqe
);
6408 case IORING_OP_SYMLINKAT
:
6409 return io_symlinkat_prep(req
, sqe
);
6410 case IORING_OP_LINKAT
:
6411 return io_linkat_prep(req
, sqe
);
6414 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6419 static int io_req_prep_async(struct io_kiocb
*req
)
6421 if (!io_op_defs
[req
->opcode
].needs_async_setup
)
6423 if (WARN_ON_ONCE(req
->async_data
))
6425 if (io_alloc_async_data(req
))
6428 switch (req
->opcode
) {
6429 case IORING_OP_READV
:
6430 return io_rw_prep_async(req
, READ
);
6431 case IORING_OP_WRITEV
:
6432 return io_rw_prep_async(req
, WRITE
);
6433 case IORING_OP_SENDMSG
:
6434 return io_sendmsg_prep_async(req
);
6435 case IORING_OP_RECVMSG
:
6436 return io_recvmsg_prep_async(req
);
6437 case IORING_OP_CONNECT
:
6438 return io_connect_prep_async(req
);
6440 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
6445 static u32
io_get_sequence(struct io_kiocb
*req
)
6447 u32 seq
= req
->ctx
->cached_sq_head
;
6449 /* need original cached_sq_head, but it was increased for each req */
6450 io_for_each_link(req
, req
)
6455 static bool io_drain_req(struct io_kiocb
*req
)
6457 struct io_kiocb
*pos
;
6458 struct io_ring_ctx
*ctx
= req
->ctx
;
6459 struct io_defer_entry
*de
;
6463 if (req
->flags
& REQ_F_FAIL
) {
6464 io_req_complete_fail_submit(req
);
6469 * If we need to drain a request in the middle of a link, drain the
6470 * head request and the next request/link after the current link.
6471 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6472 * maintained for every request of our link.
6474 if (ctx
->drain_next
) {
6475 req
->flags
|= REQ_F_IO_DRAIN
;
6476 ctx
->drain_next
= false;
6478 /* not interested in head, start from the first linked */
6479 io_for_each_link(pos
, req
->link
) {
6480 if (pos
->flags
& REQ_F_IO_DRAIN
) {
6481 ctx
->drain_next
= true;
6482 req
->flags
|= REQ_F_IO_DRAIN
;
6487 /* Still need defer if there is pending req in defer list. */
6488 spin_lock(&ctx
->completion_lock
);
6489 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6490 !(req
->flags
& REQ_F_IO_DRAIN
))) {
6491 spin_unlock(&ctx
->completion_lock
);
6492 ctx
->drain_active
= false;
6495 spin_unlock(&ctx
->completion_lock
);
6497 seq
= io_get_sequence(req
);
6498 /* Still a chance to pass the sequence check */
6499 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6502 ret
= io_req_prep_async(req
);
6505 io_prep_async_link(req
);
6506 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6510 io_req_complete_failed(req
, ret
);
6514 spin_lock(&ctx
->completion_lock
);
6515 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6516 spin_unlock(&ctx
->completion_lock
);
6518 io_queue_async_work(req
, NULL
);
6522 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6525 list_add_tail(&de
->list
, &ctx
->defer_list
);
6526 spin_unlock(&ctx
->completion_lock
);
6530 static void io_clean_op(struct io_kiocb
*req
)
6532 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6533 switch (req
->opcode
) {
6534 case IORING_OP_READV
:
6535 case IORING_OP_READ_FIXED
:
6536 case IORING_OP_READ
:
6537 kfree((void *)(unsigned long)req
->rw
.addr
);
6539 case IORING_OP_RECVMSG
:
6540 case IORING_OP_RECV
:
6541 kfree(req
->sr_msg
.kbuf
);
6546 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6547 switch (req
->opcode
) {
6548 case IORING_OP_READV
:
6549 case IORING_OP_READ_FIXED
:
6550 case IORING_OP_READ
:
6551 case IORING_OP_WRITEV
:
6552 case IORING_OP_WRITE_FIXED
:
6553 case IORING_OP_WRITE
: {
6554 struct io_async_rw
*io
= req
->async_data
;
6556 kfree(io
->free_iovec
);
6559 case IORING_OP_RECVMSG
:
6560 case IORING_OP_SENDMSG
: {
6561 struct io_async_msghdr
*io
= req
->async_data
;
6563 kfree(io
->free_iov
);
6566 case IORING_OP_OPENAT
:
6567 case IORING_OP_OPENAT2
:
6568 if (req
->open
.filename
)
6569 putname(req
->open
.filename
);
6571 case IORING_OP_RENAMEAT
:
6572 putname(req
->rename
.oldpath
);
6573 putname(req
->rename
.newpath
);
6575 case IORING_OP_UNLINKAT
:
6576 putname(req
->unlink
.filename
);
6578 case IORING_OP_MKDIRAT
:
6579 putname(req
->mkdir
.filename
);
6581 case IORING_OP_SYMLINKAT
:
6582 putname(req
->symlink
.oldpath
);
6583 putname(req
->symlink
.newpath
);
6585 case IORING_OP_LINKAT
:
6586 putname(req
->hardlink
.oldpath
);
6587 putname(req
->hardlink
.newpath
);
6591 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
6592 kfree(req
->apoll
->double_poll
);
6596 if (req
->flags
& REQ_F_INFLIGHT
) {
6597 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6599 atomic_dec(&tctx
->inflight_tracked
);
6601 if (req
->flags
& REQ_F_CREDS
)
6602 put_cred(req
->creds
);
6604 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
6607 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6609 struct io_ring_ctx
*ctx
= req
->ctx
;
6610 const struct cred
*creds
= NULL
;
6613 if ((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred())
6614 creds
= override_creds(req
->creds
);
6616 switch (req
->opcode
) {
6618 ret
= io_nop(req
, issue_flags
);
6620 case IORING_OP_READV
:
6621 case IORING_OP_READ_FIXED
:
6622 case IORING_OP_READ
:
6623 ret
= io_read(req
, issue_flags
);
6625 case IORING_OP_WRITEV
:
6626 case IORING_OP_WRITE_FIXED
:
6627 case IORING_OP_WRITE
:
6628 ret
= io_write(req
, issue_flags
);
6630 case IORING_OP_FSYNC
:
6631 ret
= io_fsync(req
, issue_flags
);
6633 case IORING_OP_POLL_ADD
:
6634 ret
= io_poll_add(req
, issue_flags
);
6636 case IORING_OP_POLL_REMOVE
:
6637 ret
= io_poll_update(req
, issue_flags
);
6639 case IORING_OP_SYNC_FILE_RANGE
:
6640 ret
= io_sync_file_range(req
, issue_flags
);
6642 case IORING_OP_SENDMSG
:
6643 ret
= io_sendmsg(req
, issue_flags
);
6645 case IORING_OP_SEND
:
6646 ret
= io_send(req
, issue_flags
);
6648 case IORING_OP_RECVMSG
:
6649 ret
= io_recvmsg(req
, issue_flags
);
6651 case IORING_OP_RECV
:
6652 ret
= io_recv(req
, issue_flags
);
6654 case IORING_OP_TIMEOUT
:
6655 ret
= io_timeout(req
, issue_flags
);
6657 case IORING_OP_TIMEOUT_REMOVE
:
6658 ret
= io_timeout_remove(req
, issue_flags
);
6660 case IORING_OP_ACCEPT
:
6661 ret
= io_accept(req
, issue_flags
);
6663 case IORING_OP_CONNECT
:
6664 ret
= io_connect(req
, issue_flags
);
6666 case IORING_OP_ASYNC_CANCEL
:
6667 ret
= io_async_cancel(req
, issue_flags
);
6669 case IORING_OP_FALLOCATE
:
6670 ret
= io_fallocate(req
, issue_flags
);
6672 case IORING_OP_OPENAT
:
6673 ret
= io_openat(req
, issue_flags
);
6675 case IORING_OP_CLOSE
:
6676 ret
= io_close(req
, issue_flags
);
6678 case IORING_OP_FILES_UPDATE
:
6679 ret
= io_files_update(req
, issue_flags
);
6681 case IORING_OP_STATX
:
6682 ret
= io_statx(req
, issue_flags
);
6684 case IORING_OP_FADVISE
:
6685 ret
= io_fadvise(req
, issue_flags
);
6687 case IORING_OP_MADVISE
:
6688 ret
= io_madvise(req
, issue_flags
);
6690 case IORING_OP_OPENAT2
:
6691 ret
= io_openat2(req
, issue_flags
);
6693 case IORING_OP_EPOLL_CTL
:
6694 ret
= io_epoll_ctl(req
, issue_flags
);
6696 case IORING_OP_SPLICE
:
6697 ret
= io_splice(req
, issue_flags
);
6699 case IORING_OP_PROVIDE_BUFFERS
:
6700 ret
= io_provide_buffers(req
, issue_flags
);
6702 case IORING_OP_REMOVE_BUFFERS
:
6703 ret
= io_remove_buffers(req
, issue_flags
);
6706 ret
= io_tee(req
, issue_flags
);
6708 case IORING_OP_SHUTDOWN
:
6709 ret
= io_shutdown(req
, issue_flags
);
6711 case IORING_OP_RENAMEAT
:
6712 ret
= io_renameat(req
, issue_flags
);
6714 case IORING_OP_UNLINKAT
:
6715 ret
= io_unlinkat(req
, issue_flags
);
6717 case IORING_OP_MKDIRAT
:
6718 ret
= io_mkdirat(req
, issue_flags
);
6720 case IORING_OP_SYMLINKAT
:
6721 ret
= io_symlinkat(req
, issue_flags
);
6723 case IORING_OP_LINKAT
:
6724 ret
= io_linkat(req
, issue_flags
);
6732 revert_creds(creds
);
6735 /* If the op doesn't have a file, we're not polling for it */
6736 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
6737 io_iopoll_req_issued(req
);
6742 static struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
6744 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6746 req
= io_put_req_find_next(req
);
6747 return req
? &req
->work
: NULL
;
6750 static void io_wq_submit_work(struct io_wq_work
*work
)
6752 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6753 struct io_kiocb
*timeout
;
6756 /* one will be dropped by ->io_free_work() after returning to io-wq */
6757 if (!(req
->flags
& REQ_F_REFCOUNT
))
6758 __io_req_set_refcount(req
, 2);
6762 timeout
= io_prep_linked_timeout(req
);
6764 io_queue_linked_timeout(timeout
);
6766 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6767 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6772 ret
= io_issue_sqe(req
, 0);
6774 * We can get EAGAIN for polled IO even though we're
6775 * forcing a sync submission from here, since we can't
6776 * wait for request slots on the block side.
6778 if (ret
!= -EAGAIN
|| !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6784 /* avoid locking problems by failing it from a clean context */
6786 io_req_task_queue_fail(req
, ret
);
6789 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
6792 return &table
->files
[i
];
6795 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6798 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
6800 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
6803 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
6805 unsigned long file_ptr
= (unsigned long) file
;
6807 if (__io_file_supports_nowait(file
, READ
))
6808 file_ptr
|= FFS_ASYNC_READ
;
6809 if (__io_file_supports_nowait(file
, WRITE
))
6810 file_ptr
|= FFS_ASYNC_WRITE
;
6811 if (S_ISREG(file_inode(file
)->i_mode
))
6812 file_ptr
|= FFS_ISREG
;
6813 file_slot
->file_ptr
= file_ptr
;
6816 static inline struct file
*io_file_get_fixed(struct io_ring_ctx
*ctx
,
6817 struct io_kiocb
*req
, int fd
)
6820 unsigned long file_ptr
;
6822 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6824 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6825 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
6826 file
= (struct file
*) (file_ptr
& FFS_MASK
);
6827 file_ptr
&= ~FFS_MASK
;
6828 /* mask in overlapping REQ_F and FFS bits */
6829 req
->flags
|= (file_ptr
<< REQ_F_NOWAIT_READ_BIT
);
6830 io_req_set_rsrc_node(req
);
6834 static struct file
*io_file_get_normal(struct io_ring_ctx
*ctx
,
6835 struct io_kiocb
*req
, int fd
)
6837 struct file
*file
= fget(fd
);
6839 trace_io_uring_file_get(ctx
, fd
);
6841 /* we don't allow fixed io_uring files */
6842 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6843 io_req_track_inflight(req
);
6847 static inline struct file
*io_file_get(struct io_ring_ctx
*ctx
,
6848 struct io_kiocb
*req
, int fd
, bool fixed
)
6851 return io_file_get_fixed(ctx
, req
, fd
);
6853 return io_file_get_normal(ctx
, req
, fd
);
6856 static void io_req_task_link_timeout(struct io_kiocb
*req
, bool *locked
)
6858 struct io_kiocb
*prev
= req
->timeout
.prev
;
6862 if (!(req
->task
->flags
& PF_EXITING
))
6863 ret
= io_try_cancel_userdata(req
, prev
->user_data
);
6864 io_req_complete_post(req
, ret
?: -ETIME
, 0);
6867 io_req_complete_post(req
, -ETIME
, 0);
6871 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6873 struct io_timeout_data
*data
= container_of(timer
,
6874 struct io_timeout_data
, timer
);
6875 struct io_kiocb
*prev
, *req
= data
->req
;
6876 struct io_ring_ctx
*ctx
= req
->ctx
;
6877 unsigned long flags
;
6879 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
6880 prev
= req
->timeout
.head
;
6881 req
->timeout
.head
= NULL
;
6884 * We don't expect the list to be empty, that will only happen if we
6885 * race with the completion of the linked work.
6888 io_remove_next_linked(prev
);
6889 if (!req_ref_inc_not_zero(prev
))
6892 list_del(&req
->timeout
.list
);
6893 req
->timeout
.prev
= prev
;
6894 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
6896 req
->io_task_work
.func
= io_req_task_link_timeout
;
6897 io_req_task_work_add(req
);
6898 return HRTIMER_NORESTART
;
6901 static void io_queue_linked_timeout(struct io_kiocb
*req
)
6903 struct io_ring_ctx
*ctx
= req
->ctx
;
6905 spin_lock_irq(&ctx
->timeout_lock
);
6907 * If the back reference is NULL, then our linked request finished
6908 * before we got a chance to setup the timer
6910 if (req
->timeout
.head
) {
6911 struct io_timeout_data
*data
= req
->async_data
;
6913 data
->timer
.function
= io_link_timeout_fn
;
6914 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
6916 list_add_tail(&req
->timeout
.list
, &ctx
->ltimeout_list
);
6918 spin_unlock_irq(&ctx
->timeout_lock
);
6919 /* drop submission reference */
6923 static void __io_queue_sqe(struct io_kiocb
*req
)
6924 __must_hold(&req
->ctx
->uring_lock
)
6926 struct io_kiocb
*linked_timeout
;
6930 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
6933 * We async punt it if the file wasn't marked NOWAIT, or if the file
6934 * doesn't support non-blocking read/write attempts
6937 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
6938 struct io_ring_ctx
*ctx
= req
->ctx
;
6939 struct io_submit_state
*state
= &ctx
->submit_state
;
6941 state
->compl_reqs
[state
->compl_nr
++] = req
;
6942 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
6943 io_submit_flush_completions(ctx
);
6947 linked_timeout
= io_prep_linked_timeout(req
);
6949 io_queue_linked_timeout(linked_timeout
);
6950 } else if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
6951 linked_timeout
= io_prep_linked_timeout(req
);
6953 switch (io_arm_poll_handler(req
)) {
6954 case IO_APOLL_READY
:
6956 io_queue_linked_timeout(linked_timeout
);
6958 case IO_APOLL_ABORTED
:
6960 * Queued up for async execution, worker will release
6961 * submit reference when the iocb is actually submitted.
6963 io_queue_async_work(req
, NULL
);
6968 io_queue_linked_timeout(linked_timeout
);
6970 io_req_complete_failed(req
, ret
);
6974 static inline void io_queue_sqe(struct io_kiocb
*req
)
6975 __must_hold(&req
->ctx
->uring_lock
)
6977 if (unlikely(req
->ctx
->drain_active
) && io_drain_req(req
))
6980 if (likely(!(req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
)))) {
6981 __io_queue_sqe(req
);
6982 } else if (req
->flags
& REQ_F_FAIL
) {
6983 io_req_complete_fail_submit(req
);
6985 int ret
= io_req_prep_async(req
);
6988 io_req_complete_failed(req
, ret
);
6990 io_queue_async_work(req
, NULL
);
6995 * Check SQE restrictions (opcode and flags).
6997 * Returns 'true' if SQE is allowed, 'false' otherwise.
6999 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
7000 struct io_kiocb
*req
,
7001 unsigned int sqe_flags
)
7003 if (likely(!ctx
->restricted
))
7006 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
7009 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
7010 ctx
->restrictions
.sqe_flags_required
)
7013 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
7014 ctx
->restrictions
.sqe_flags_required
))
7020 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7021 const struct io_uring_sqe
*sqe
)
7022 __must_hold(&ctx
->uring_lock
)
7024 struct io_submit_state
*state
;
7025 unsigned int sqe_flags
;
7026 int personality
, ret
= 0;
7028 /* req is partially pre-initialised, see io_preinit_req() */
7029 req
->opcode
= READ_ONCE(sqe
->opcode
);
7030 /* same numerical values with corresponding REQ_F_*, safe to copy */
7031 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
7032 req
->user_data
= READ_ONCE(sqe
->user_data
);
7034 req
->fixed_rsrc_refs
= NULL
;
7035 req
->task
= current
;
7037 /* enforce forwards compatibility on users */
7038 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
7040 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
7042 if (!io_check_restriction(ctx
, req
, sqe_flags
))
7045 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
7046 !io_op_defs
[req
->opcode
].buffer_select
)
7048 if (unlikely(sqe_flags
& IOSQE_IO_DRAIN
))
7049 ctx
->drain_active
= true;
7051 personality
= READ_ONCE(sqe
->personality
);
7053 req
->creds
= xa_load(&ctx
->personalities
, personality
);
7056 get_cred(req
->creds
);
7057 req
->flags
|= REQ_F_CREDS
;
7059 state
= &ctx
->submit_state
;
7062 * Plug now if we have more than 1 IO left after this, and the target
7063 * is potentially a read/write to block based storage.
7065 if (!state
->plug_started
&& state
->ios_left
> 1 &&
7066 io_op_defs
[req
->opcode
].plug
) {
7067 blk_start_plug(&state
->plug
);
7068 state
->plug_started
= true;
7071 if (io_op_defs
[req
->opcode
].needs_file
) {
7072 req
->file
= io_file_get(ctx
, req
, READ_ONCE(sqe
->fd
),
7073 (sqe_flags
& IOSQE_FIXED_FILE
));
7074 if (unlikely(!req
->file
))
7082 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7083 const struct io_uring_sqe
*sqe
)
7084 __must_hold(&ctx
->uring_lock
)
7086 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
7089 ret
= io_init_req(ctx
, req
, sqe
);
7090 if (unlikely(ret
)) {
7092 /* fail even hard links since we don't submit */
7095 * we can judge a link req is failed or cancelled by if
7096 * REQ_F_FAIL is set, but the head is an exception since
7097 * it may be set REQ_F_FAIL because of other req's failure
7098 * so let's leverage req->result to distinguish if a head
7099 * is set REQ_F_FAIL because of its failure or other req's
7100 * failure so that we can set the correct ret code for it.
7101 * init result here to avoid affecting the normal path.
7103 if (!(link
->head
->flags
& REQ_F_FAIL
))
7104 req_fail_link_node(link
->head
, -ECANCELED
);
7105 } else if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7107 * the current req is a normal req, we should return
7108 * error and thus break the submittion loop.
7110 io_req_complete_failed(req
, ret
);
7113 req_fail_link_node(req
, ret
);
7115 ret
= io_req_prep(req
, sqe
);
7120 /* don't need @sqe from now on */
7121 trace_io_uring_submit_sqe(ctx
, req
, req
->opcode
, req
->user_data
,
7123 ctx
->flags
& IORING_SETUP_SQPOLL
);
7126 * If we already have a head request, queue this one for async
7127 * submittal once the head completes. If we don't have a head but
7128 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7129 * submitted sync once the chain is complete. If none of those
7130 * conditions are true (normal request), then just queue it.
7133 struct io_kiocb
*head
= link
->head
;
7135 if (!(req
->flags
& REQ_F_FAIL
)) {
7136 ret
= io_req_prep_async(req
);
7137 if (unlikely(ret
)) {
7138 req_fail_link_node(req
, ret
);
7139 if (!(head
->flags
& REQ_F_FAIL
))
7140 req_fail_link_node(head
, -ECANCELED
);
7143 trace_io_uring_link(ctx
, req
, head
);
7144 link
->last
->link
= req
;
7147 /* last request of a link, enqueue the link */
7148 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7153 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
7165 * Batched submission is done, ensure local IO is flushed out.
7167 static void io_submit_state_end(struct io_submit_state
*state
,
7168 struct io_ring_ctx
*ctx
)
7170 if (state
->link
.head
)
7171 io_queue_sqe(state
->link
.head
);
7172 if (state
->compl_nr
)
7173 io_submit_flush_completions(ctx
);
7174 if (state
->plug_started
)
7175 blk_finish_plug(&state
->plug
);
7179 * Start submission side cache.
7181 static void io_submit_state_start(struct io_submit_state
*state
,
7182 unsigned int max_ios
)
7184 state
->plug_started
= false;
7185 state
->ios_left
= max_ios
;
7186 /* set only head, no need to init link_last in advance */
7187 state
->link
.head
= NULL
;
7190 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
7192 struct io_rings
*rings
= ctx
->rings
;
7195 * Ensure any loads from the SQEs are done at this point,
7196 * since once we write the new head, the application could
7197 * write new data to them.
7199 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
7203 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7204 * that is mapped by userspace. This means that care needs to be taken to
7205 * ensure that reads are stable, as we cannot rely on userspace always
7206 * being a good citizen. If members of the sqe are validated and then later
7207 * used, it's important that those reads are done through READ_ONCE() to
7208 * prevent a re-load down the line.
7210 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
7212 unsigned head
, mask
= ctx
->sq_entries
- 1;
7213 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
7216 * The cached sq head (or cq tail) serves two purposes:
7218 * 1) allows us to batch the cost of updating the user visible
7220 * 2) allows the kernel side to track the head on its own, even
7221 * though the application is the one updating it.
7223 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
7224 if (likely(head
< ctx
->sq_entries
))
7225 return &ctx
->sq_sqes
[head
];
7227 /* drop invalid entries */
7229 WRITE_ONCE(ctx
->rings
->sq_dropped
,
7230 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
7234 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
7235 __must_hold(&ctx
->uring_lock
)
7239 /* make sure SQ entry isn't read before tail */
7240 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
7241 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
7243 io_get_task_refs(nr
);
7245 io_submit_state_start(&ctx
->submit_state
, nr
);
7246 while (submitted
< nr
) {
7247 const struct io_uring_sqe
*sqe
;
7248 struct io_kiocb
*req
;
7250 req
= io_alloc_req(ctx
);
7251 if (unlikely(!req
)) {
7253 submitted
= -EAGAIN
;
7256 sqe
= io_get_sqe(ctx
);
7257 if (unlikely(!sqe
)) {
7258 list_add(&req
->inflight_entry
, &ctx
->submit_state
.free_list
);
7261 /* will complete beyond this point, count as submitted */
7263 if (io_submit_sqe(ctx
, req
, sqe
))
7267 if (unlikely(submitted
!= nr
)) {
7268 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
7269 int unused
= nr
- ref_used
;
7271 current
->io_uring
->cached_refs
+= unused
;
7272 percpu_ref_put_many(&ctx
->refs
, unused
);
7275 io_submit_state_end(&ctx
->submit_state
, ctx
);
7276 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7277 io_commit_sqring(ctx
);
7282 static inline bool io_sqd_events_pending(struct io_sq_data
*sqd
)
7284 return READ_ONCE(sqd
->state
);
7287 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
7289 /* Tell userspace we may need a wakeup call */
7290 spin_lock(&ctx
->completion_lock
);
7291 WRITE_ONCE(ctx
->rings
->sq_flags
,
7292 ctx
->rings
->sq_flags
| IORING_SQ_NEED_WAKEUP
);
7293 spin_unlock(&ctx
->completion_lock
);
7296 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
7298 spin_lock(&ctx
->completion_lock
);
7299 WRITE_ONCE(ctx
->rings
->sq_flags
,
7300 ctx
->rings
->sq_flags
& ~IORING_SQ_NEED_WAKEUP
);
7301 spin_unlock(&ctx
->completion_lock
);
7304 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
7306 unsigned int to_submit
;
7309 to_submit
= io_sqring_entries(ctx
);
7310 /* if we're handling multiple rings, cap submit size for fairness */
7311 if (cap_entries
&& to_submit
> IORING_SQPOLL_CAP_ENTRIES_VALUE
)
7312 to_submit
= IORING_SQPOLL_CAP_ENTRIES_VALUE
;
7314 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
7315 unsigned nr_events
= 0;
7316 const struct cred
*creds
= NULL
;
7318 if (ctx
->sq_creds
!= current_cred())
7319 creds
= override_creds(ctx
->sq_creds
);
7321 mutex_lock(&ctx
->uring_lock
);
7322 if (!list_empty(&ctx
->iopoll_list
))
7323 io_do_iopoll(ctx
, &nr_events
, 0);
7326 * Don't submit if refs are dying, good for io_uring_register(),
7327 * but also it is relied upon by io_ring_exit_work()
7329 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
7330 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
7331 ret
= io_submit_sqes(ctx
, to_submit
);
7332 mutex_unlock(&ctx
->uring_lock
);
7334 if (to_submit
&& wq_has_sleeper(&ctx
->sqo_sq_wait
))
7335 wake_up(&ctx
->sqo_sq_wait
);
7337 revert_creds(creds
);
7343 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7345 struct io_ring_ctx
*ctx
;
7346 unsigned sq_thread_idle
= 0;
7348 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7349 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
7350 sqd
->sq_thread_idle
= sq_thread_idle
;
7353 static bool io_sqd_handle_event(struct io_sq_data
*sqd
)
7355 bool did_sig
= false;
7356 struct ksignal ksig
;
7358 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
7359 signal_pending(current
)) {
7360 mutex_unlock(&sqd
->lock
);
7361 if (signal_pending(current
))
7362 did_sig
= get_signal(&ksig
);
7364 mutex_lock(&sqd
->lock
);
7366 return did_sig
|| test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7369 static int io_sq_thread(void *data
)
7371 struct io_sq_data
*sqd
= data
;
7372 struct io_ring_ctx
*ctx
;
7373 unsigned long timeout
= 0;
7374 char buf
[TASK_COMM_LEN
];
7377 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
7378 set_task_comm(current
, buf
);
7380 if (sqd
->sq_cpu
!= -1)
7381 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
7383 set_cpus_allowed_ptr(current
, cpu_online_mask
);
7384 current
->flags
|= PF_NO_SETAFFINITY
;
7386 mutex_lock(&sqd
->lock
);
7388 bool cap_entries
, sqt_spin
= false;
7390 if (io_sqd_events_pending(sqd
) || signal_pending(current
)) {
7391 if (io_sqd_handle_event(sqd
))
7393 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7396 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7397 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7398 int ret
= __io_sq_thread(ctx
, cap_entries
);
7400 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7403 if (io_run_task_work())
7406 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7409 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7413 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7414 if (!io_sqd_events_pending(sqd
) && !current
->task_works
) {
7415 bool needs_sched
= true;
7417 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7418 io_ring_set_wakeup_flag(ctx
);
7420 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7421 !list_empty_careful(&ctx
->iopoll_list
)) {
7422 needs_sched
= false;
7425 if (io_sqring_entries(ctx
)) {
7426 needs_sched
= false;
7432 mutex_unlock(&sqd
->lock
);
7434 mutex_lock(&sqd
->lock
);
7436 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7437 io_ring_clear_wakeup_flag(ctx
);
7440 finish_wait(&sqd
->wait
, &wait
);
7441 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7444 io_uring_cancel_generic(true, sqd
);
7446 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7447 io_ring_set_wakeup_flag(ctx
);
7449 mutex_unlock(&sqd
->lock
);
7451 complete(&sqd
->exited
);
7455 struct io_wait_queue
{
7456 struct wait_queue_entry wq
;
7457 struct io_ring_ctx
*ctx
;
7459 unsigned nr_timeouts
;
7462 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7464 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7465 int dist
= ctx
->cached_cq_tail
- (int) iowq
->cq_tail
;
7468 * Wake up if we have enough events, or if a timeout occurred since we
7469 * started waiting. For timeouts, we always want to return to userspace,
7470 * regardless of event count.
7472 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7475 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7476 int wake_flags
, void *key
)
7478 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7482 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7483 * the task, and the next invocation will do it.
7485 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->check_cq_overflow
))
7486 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7490 static int io_run_task_work_sig(void)
7492 if (io_run_task_work())
7494 if (!signal_pending(current
))
7496 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
7497 return -ERESTARTSYS
;
7501 /* when returns >0, the caller should retry */
7502 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7503 struct io_wait_queue
*iowq
,
7508 /* make sure we run task_work before checking for signals */
7509 ret
= io_run_task_work_sig();
7510 if (ret
|| io_should_wake(iowq
))
7512 /* let the caller flush overflows, retry */
7513 if (test_bit(0, &ctx
->check_cq_overflow
))
7516 if (!schedule_hrtimeout(&timeout
, HRTIMER_MODE_ABS
))
7522 * Wait until events become available, if we don't already have some. The
7523 * application must reap them itself, as they reside on the shared cq ring.
7525 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7526 const sigset_t __user
*sig
, size_t sigsz
,
7527 struct __kernel_timespec __user
*uts
)
7529 struct io_wait_queue iowq
;
7530 struct io_rings
*rings
= ctx
->rings
;
7531 ktime_t timeout
= KTIME_MAX
;
7535 io_cqring_overflow_flush(ctx
);
7536 if (io_cqring_events(ctx
) >= min_events
)
7538 if (!io_run_task_work())
7543 struct timespec64 ts
;
7545 if (get_timespec64(&ts
, uts
))
7547 timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
7551 #ifdef CONFIG_COMPAT
7552 if (in_compat_syscall())
7553 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7557 ret
= set_user_sigmask(sig
, sigsz
);
7563 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
7564 iowq
.wq
.private = current
;
7565 INIT_LIST_HEAD(&iowq
.wq
.entry
);
7567 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7568 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
7570 trace_io_uring_cqring_wait(ctx
, min_events
);
7572 /* if we can't even flush overflow, don't wait for more */
7573 if (!io_cqring_overflow_flush(ctx
)) {
7577 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
7578 TASK_INTERRUPTIBLE
);
7579 ret
= io_cqring_wait_schedule(ctx
, &iowq
, timeout
);
7580 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
7584 restore_saved_sigmask_unless(ret
== -EINTR
);
7586 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7589 static void io_free_page_table(void **table
, size_t size
)
7591 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7593 for (i
= 0; i
< nr_tables
; i
++)
7598 static void **io_alloc_page_table(size_t size
)
7600 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7601 size_t init_size
= size
;
7604 table
= kcalloc(nr_tables
, sizeof(*table
), GFP_KERNEL_ACCOUNT
);
7608 for (i
= 0; i
< nr_tables
; i
++) {
7609 unsigned int this_size
= min_t(size_t, size
, PAGE_SIZE
);
7611 table
[i
] = kzalloc(this_size
, GFP_KERNEL_ACCOUNT
);
7613 io_free_page_table(table
, init_size
);
7621 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
7623 percpu_ref_exit(&ref_node
->refs
);
7627 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7629 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
7630 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
7631 unsigned long flags
;
7632 bool first_add
= false;
7633 unsigned long delay
= HZ
;
7635 spin_lock_irqsave(&ctx
->rsrc_ref_lock
, flags
);
7638 /* if we are mid-quiesce then do not delay */
7639 if (node
->rsrc_data
->quiesce
)
7642 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7643 node
= list_first_entry(&ctx
->rsrc_ref_list
,
7644 struct io_rsrc_node
, node
);
7645 /* recycle ref nodes in order */
7648 list_del(&node
->node
);
7649 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
7651 spin_unlock_irqrestore(&ctx
->rsrc_ref_lock
, flags
);
7654 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
7657 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
)
7659 struct io_rsrc_node
*ref_node
;
7661 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7665 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7670 INIT_LIST_HEAD(&ref_node
->node
);
7671 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7672 ref_node
->done
= false;
7676 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
7677 struct io_rsrc_data
*data_to_kill
)
7679 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
7680 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
7683 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
7685 rsrc_node
->rsrc_data
= data_to_kill
;
7686 spin_lock_irq(&ctx
->rsrc_ref_lock
);
7687 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
7688 spin_unlock_irq(&ctx
->rsrc_ref_lock
);
7690 atomic_inc(&data_to_kill
->refs
);
7691 percpu_ref_kill(&rsrc_node
->refs
);
7692 ctx
->rsrc_node
= NULL
;
7695 if (!ctx
->rsrc_node
) {
7696 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
7697 ctx
->rsrc_backup_node
= NULL
;
7701 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
7703 if (ctx
->rsrc_backup_node
)
7705 ctx
->rsrc_backup_node
= io_rsrc_node_alloc(ctx
);
7706 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
7709 static int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
, struct io_ring_ctx
*ctx
)
7713 /* As we may drop ->uring_lock, other task may have started quiesce */
7717 data
->quiesce
= true;
7719 ret
= io_rsrc_node_switch_start(ctx
);
7722 io_rsrc_node_switch(ctx
, data
);
7724 /* kill initial ref, already quiesced if zero */
7725 if (atomic_dec_and_test(&data
->refs
))
7727 mutex_unlock(&ctx
->uring_lock
);
7728 flush_delayed_work(&ctx
->rsrc_put_work
);
7729 ret
= wait_for_completion_interruptible(&data
->done
);
7731 mutex_lock(&ctx
->uring_lock
);
7732 if (atomic_read(&data
->refs
) > 0) {
7734 * it has been revived by another thread while
7737 mutex_unlock(&ctx
->uring_lock
);
7743 atomic_inc(&data
->refs
);
7744 /* wait for all works potentially completing data->done */
7745 flush_delayed_work(&ctx
->rsrc_put_work
);
7746 reinit_completion(&data
->done
);
7748 ret
= io_run_task_work_sig();
7749 mutex_lock(&ctx
->uring_lock
);
7751 data
->quiesce
= false;
7756 static u64
*io_get_tag_slot(struct io_rsrc_data
*data
, unsigned int idx
)
7758 unsigned int off
= idx
& IO_RSRC_TAG_TABLE_MASK
;
7759 unsigned int table_idx
= idx
>> IO_RSRC_TAG_TABLE_SHIFT
;
7761 return &data
->tags
[table_idx
][off
];
7764 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
7766 size_t size
= data
->nr
* sizeof(data
->tags
[0][0]);
7769 io_free_page_table((void **)data
->tags
, size
);
7773 static int io_rsrc_data_alloc(struct io_ring_ctx
*ctx
, rsrc_put_fn
*do_put
,
7774 u64 __user
*utags
, unsigned nr
,
7775 struct io_rsrc_data
**pdata
)
7777 struct io_rsrc_data
*data
;
7781 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7784 data
->tags
= (u64
**)io_alloc_page_table(nr
* sizeof(data
->tags
[0][0]));
7792 data
->do_put
= do_put
;
7795 for (i
= 0; i
< nr
; i
++) {
7796 u64
*tag_slot
= io_get_tag_slot(data
, i
);
7798 if (copy_from_user(tag_slot
, &utags
[i
],
7804 atomic_set(&data
->refs
, 1);
7805 init_completion(&data
->done
);
7809 io_rsrc_data_free(data
);
7813 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7815 table
->files
= kvcalloc(nr_files
, sizeof(table
->files
[0]),
7816 GFP_KERNEL_ACCOUNT
);
7817 return !!table
->files
;
7820 static void io_free_file_tables(struct io_file_table
*table
)
7822 kvfree(table
->files
);
7823 table
->files
= NULL
;
7826 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7828 #if defined(CONFIG_UNIX)
7829 if (ctx
->ring_sock
) {
7830 struct sock
*sock
= ctx
->ring_sock
->sk
;
7831 struct sk_buff
*skb
;
7833 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7839 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7842 file
= io_file_from_index(ctx
, i
);
7847 io_free_file_tables(&ctx
->file_table
);
7848 io_rsrc_data_free(ctx
->file_data
);
7849 ctx
->file_data
= NULL
;
7850 ctx
->nr_user_files
= 0;
7853 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7855 unsigned nr
= ctx
->nr_user_files
;
7858 if (!ctx
->file_data
)
7862 * Quiesce may unlock ->uring_lock, and while it's not held
7863 * prevent new requests using the table.
7865 ctx
->nr_user_files
= 0;
7866 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
7867 ctx
->nr_user_files
= nr
;
7869 __io_sqe_files_unregister(ctx
);
7873 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7874 __releases(&sqd
->lock
)
7876 WARN_ON_ONCE(sqd
->thread
== current
);
7879 * Do the dance but not conditional clear_bit() because it'd race with
7880 * other threads incrementing park_pending and setting the bit.
7882 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7883 if (atomic_dec_return(&sqd
->park_pending
))
7884 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7885 mutex_unlock(&sqd
->lock
);
7888 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7889 __acquires(&sqd
->lock
)
7891 WARN_ON_ONCE(sqd
->thread
== current
);
7893 atomic_inc(&sqd
->park_pending
);
7894 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7895 mutex_lock(&sqd
->lock
);
7897 wake_up_process(sqd
->thread
);
7900 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7902 WARN_ON_ONCE(sqd
->thread
== current
);
7903 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
7905 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7906 mutex_lock(&sqd
->lock
);
7908 wake_up_process(sqd
->thread
);
7909 mutex_unlock(&sqd
->lock
);
7910 wait_for_completion(&sqd
->exited
);
7913 static void io_put_sq_data(struct io_sq_data
*sqd
)
7915 if (refcount_dec_and_test(&sqd
->refs
)) {
7916 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
7918 io_sq_thread_stop(sqd
);
7923 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
7925 struct io_sq_data
*sqd
= ctx
->sq_data
;
7928 io_sq_thread_park(sqd
);
7929 list_del_init(&ctx
->sqd_list
);
7930 io_sqd_update_thread_idle(sqd
);
7931 io_sq_thread_unpark(sqd
);
7933 io_put_sq_data(sqd
);
7934 ctx
->sq_data
= NULL
;
7938 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
7940 struct io_ring_ctx
*ctx_attach
;
7941 struct io_sq_data
*sqd
;
7944 f
= fdget(p
->wq_fd
);
7946 return ERR_PTR(-ENXIO
);
7947 if (f
.file
->f_op
!= &io_uring_fops
) {
7949 return ERR_PTR(-EINVAL
);
7952 ctx_attach
= f
.file
->private_data
;
7953 sqd
= ctx_attach
->sq_data
;
7956 return ERR_PTR(-EINVAL
);
7958 if (sqd
->task_tgid
!= current
->tgid
) {
7960 return ERR_PTR(-EPERM
);
7963 refcount_inc(&sqd
->refs
);
7968 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
7971 struct io_sq_data
*sqd
;
7974 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
7975 sqd
= io_attach_sq_data(p
);
7980 /* fall through for EPERM case, setup new sqd/task */
7981 if (PTR_ERR(sqd
) != -EPERM
)
7985 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
7987 return ERR_PTR(-ENOMEM
);
7989 atomic_set(&sqd
->park_pending
, 0);
7990 refcount_set(&sqd
->refs
, 1);
7991 INIT_LIST_HEAD(&sqd
->ctx_list
);
7992 mutex_init(&sqd
->lock
);
7993 init_waitqueue_head(&sqd
->wait
);
7994 init_completion(&sqd
->exited
);
7998 #if defined(CONFIG_UNIX)
8000 * Ensure the UNIX gc is aware of our file set, so we are certain that
8001 * the io_uring can be safely unregistered on process exit, even if we have
8002 * loops in the file referencing.
8004 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
8006 struct sock
*sk
= ctx
->ring_sock
->sk
;
8007 struct scm_fp_list
*fpl
;
8008 struct sk_buff
*skb
;
8011 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
8015 skb
= alloc_skb(0, GFP_KERNEL
);
8024 fpl
->user
= get_uid(current_user());
8025 for (i
= 0; i
< nr
; i
++) {
8026 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8030 fpl
->fp
[nr_files
] = get_file(file
);
8031 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
8036 fpl
->max
= SCM_MAX_FD
;
8037 fpl
->count
= nr_files
;
8038 UNIXCB(skb
).fp
= fpl
;
8039 skb
->destructor
= unix_destruct_scm
;
8040 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
8041 skb_queue_head(&sk
->sk_receive_queue
, skb
);
8043 for (i
= 0; i
< nr
; i
++) {
8044 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8051 free_uid(fpl
->user
);
8059 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8060 * causes regular reference counting to break down. We rely on the UNIX
8061 * garbage collection to take care of this problem for us.
8063 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8065 unsigned left
, total
;
8069 left
= ctx
->nr_user_files
;
8071 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
8073 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
8077 total
+= this_files
;
8083 while (total
< ctx
->nr_user_files
) {
8084 struct file
*file
= io_file_from_index(ctx
, total
);
8094 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8100 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8102 struct file
*file
= prsrc
->file
;
8103 #if defined(CONFIG_UNIX)
8104 struct sock
*sock
= ctx
->ring_sock
->sk
;
8105 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
8106 struct sk_buff
*skb
;
8109 __skb_queue_head_init(&list
);
8112 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8113 * remove this entry and rearrange the file array.
8115 skb
= skb_dequeue(head
);
8117 struct scm_fp_list
*fp
;
8119 fp
= UNIXCB(skb
).fp
;
8120 for (i
= 0; i
< fp
->count
; i
++) {
8123 if (fp
->fp
[i
] != file
)
8126 unix_notinflight(fp
->user
, fp
->fp
[i
]);
8127 left
= fp
->count
- 1 - i
;
8129 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
8130 left
* sizeof(struct file
*));
8137 __skb_queue_tail(&list
, skb
);
8147 __skb_queue_tail(&list
, skb
);
8149 skb
= skb_dequeue(head
);
8152 if (skb_peek(&list
)) {
8153 spin_lock_irq(&head
->lock
);
8154 while ((skb
= __skb_dequeue(&list
)) != NULL
)
8155 __skb_queue_tail(head
, skb
);
8156 spin_unlock_irq(&head
->lock
);
8163 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
8165 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
8166 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
8167 struct io_rsrc_put
*prsrc
, *tmp
;
8169 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
8170 list_del(&prsrc
->list
);
8173 bool lock_ring
= ctx
->flags
& IORING_SETUP_IOPOLL
;
8175 io_ring_submit_lock(ctx
, lock_ring
);
8176 spin_lock(&ctx
->completion_lock
);
8177 io_fill_cqe_aux(ctx
, prsrc
->tag
, 0, 0);
8178 io_commit_cqring(ctx
);
8179 spin_unlock(&ctx
->completion_lock
);
8180 io_cqring_ev_posted(ctx
);
8181 io_ring_submit_unlock(ctx
, lock_ring
);
8184 rsrc_data
->do_put(ctx
, prsrc
);
8188 io_rsrc_node_destroy(ref_node
);
8189 if (atomic_dec_and_test(&rsrc_data
->refs
))
8190 complete(&rsrc_data
->done
);
8193 static void io_rsrc_put_work(struct work_struct
*work
)
8195 struct io_ring_ctx
*ctx
;
8196 struct llist_node
*node
;
8198 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
8199 node
= llist_del_all(&ctx
->rsrc_put_llist
);
8202 struct io_rsrc_node
*ref_node
;
8203 struct llist_node
*next
= node
->next
;
8205 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
8206 __io_rsrc_put_work(ref_node
);
8211 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8212 unsigned nr_args
, u64 __user
*tags
)
8214 __s32 __user
*fds
= (__s32 __user
*) arg
;
8223 if (nr_args
> IORING_MAX_FIXED_FILES
)
8225 if (nr_args
> rlimit(RLIMIT_NOFILE
))
8227 ret
= io_rsrc_node_switch_start(ctx
);
8230 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, tags
, nr_args
,
8236 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
))
8239 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
8240 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
8244 /* allow sparse sets */
8247 if (unlikely(*io_get_tag_slot(ctx
->file_data
, i
)))
8254 if (unlikely(!file
))
8258 * Don't allow io_uring instances to be registered. If UNIX
8259 * isn't enabled, then this causes a reference cycle and this
8260 * instance can never get freed. If UNIX is enabled we'll
8261 * handle it just fine, but there's still no point in allowing
8262 * a ring fd as it doesn't support regular read/write anyway.
8264 if (file
->f_op
== &io_uring_fops
) {
8268 io_fixed_file_set(io_fixed_file_slot(&ctx
->file_table
, i
), file
);
8271 ret
= io_sqe_files_scm(ctx
);
8273 __io_sqe_files_unregister(ctx
);
8277 io_rsrc_node_switch(ctx
, NULL
);
8280 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
8281 file
= io_file_from_index(ctx
, i
);
8285 io_free_file_tables(&ctx
->file_table
);
8286 ctx
->nr_user_files
= 0;
8288 io_rsrc_data_free(ctx
->file_data
);
8289 ctx
->file_data
= NULL
;
8293 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
8296 #if defined(CONFIG_UNIX)
8297 struct sock
*sock
= ctx
->ring_sock
->sk
;
8298 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
8299 struct sk_buff
*skb
;
8302 * See if we can merge this file into an existing skb SCM_RIGHTS
8303 * file set. If there's no room, fall back to allocating a new skb
8304 * and filling it in.
8306 spin_lock_irq(&head
->lock
);
8307 skb
= skb_peek(head
);
8309 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
8311 if (fpl
->count
< SCM_MAX_FD
) {
8312 __skb_unlink(skb
, head
);
8313 spin_unlock_irq(&head
->lock
);
8314 fpl
->fp
[fpl
->count
] = get_file(file
);
8315 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
8317 spin_lock_irq(&head
->lock
);
8318 __skb_queue_head(head
, skb
);
8323 spin_unlock_irq(&head
->lock
);
8330 return __io_sqe_files_scm(ctx
, 1, index
);
8336 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
8337 struct io_rsrc_node
*node
, void *rsrc
)
8339 u64
*tag_slot
= io_get_tag_slot(data
, idx
);
8340 struct io_rsrc_put
*prsrc
;
8342 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
8346 prsrc
->tag
= *tag_slot
;
8349 list_add(&prsrc
->list
, &node
->rsrc_list
);
8353 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
8354 unsigned int issue_flags
, u32 slot_index
)
8356 struct io_ring_ctx
*ctx
= req
->ctx
;
8357 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
8358 bool needs_switch
= false;
8359 struct io_fixed_file
*file_slot
;
8362 io_ring_submit_lock(ctx
, !force_nonblock
);
8363 if (file
->f_op
== &io_uring_fops
)
8366 if (!ctx
->file_data
)
8369 if (slot_index
>= ctx
->nr_user_files
)
8372 slot_index
= array_index_nospec(slot_index
, ctx
->nr_user_files
);
8373 file_slot
= io_fixed_file_slot(&ctx
->file_table
, slot_index
);
8375 if (file_slot
->file_ptr
) {
8376 struct file
*old_file
;
8378 ret
= io_rsrc_node_switch_start(ctx
);
8382 old_file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8383 ret
= io_queue_rsrc_removal(ctx
->file_data
, slot_index
,
8384 ctx
->rsrc_node
, old_file
);
8387 file_slot
->file_ptr
= 0;
8388 needs_switch
= true;
8391 *io_get_tag_slot(ctx
->file_data
, slot_index
) = 0;
8392 io_fixed_file_set(file_slot
, file
);
8393 ret
= io_sqe_file_register(ctx
, file
, slot_index
);
8395 file_slot
->file_ptr
= 0;
8402 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8403 io_ring_submit_unlock(ctx
, !force_nonblock
);
8409 static int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
)
8411 unsigned int offset
= req
->close
.file_slot
- 1;
8412 struct io_ring_ctx
*ctx
= req
->ctx
;
8413 struct io_fixed_file
*file_slot
;
8417 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8419 if (unlikely(!ctx
->file_data
))
8422 if (offset
>= ctx
->nr_user_files
)
8424 ret
= io_rsrc_node_switch_start(ctx
);
8428 offset
= array_index_nospec(offset
, ctx
->nr_user_files
);
8429 file_slot
= io_fixed_file_slot(&ctx
->file_table
, offset
);
8431 if (!file_slot
->file_ptr
)
8434 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8435 ret
= io_queue_rsrc_removal(ctx
->file_data
, offset
, ctx
->rsrc_node
, file
);
8439 file_slot
->file_ptr
= 0;
8440 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8443 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8447 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
8448 struct io_uring_rsrc_update2
*up
,
8451 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
8452 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
8453 struct io_rsrc_data
*data
= ctx
->file_data
;
8454 struct io_fixed_file
*file_slot
;
8458 bool needs_switch
= false;
8460 if (!ctx
->file_data
)
8462 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
8465 for (done
= 0; done
< nr_args
; done
++) {
8468 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
8469 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8473 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
8477 if (fd
== IORING_REGISTER_FILES_SKIP
)
8480 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
8481 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
8483 if (file_slot
->file_ptr
) {
8484 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8485 err
= io_queue_rsrc_removal(data
, i
, ctx
->rsrc_node
, file
);
8488 file_slot
->file_ptr
= 0;
8489 needs_switch
= true;
8498 * Don't allow io_uring instances to be registered. If
8499 * UNIX isn't enabled, then this causes a reference
8500 * cycle and this instance can never get freed. If UNIX
8501 * is enabled we'll handle it just fine, but there's
8502 * still no point in allowing a ring fd as it doesn't
8503 * support regular read/write anyway.
8505 if (file
->f_op
== &io_uring_fops
) {
8510 *io_get_tag_slot(data
, i
) = tag
;
8511 io_fixed_file_set(file_slot
, file
);
8512 err
= io_sqe_file_register(ctx
, file
, i
);
8514 file_slot
->file_ptr
= 0;
8522 io_rsrc_node_switch(ctx
, data
);
8523 return done
? done
: err
;
8526 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
8527 struct task_struct
*task
)
8529 struct io_wq_hash
*hash
;
8530 struct io_wq_data data
;
8531 unsigned int concurrency
;
8533 mutex_lock(&ctx
->uring_lock
);
8534 hash
= ctx
->hash_map
;
8536 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
8538 mutex_unlock(&ctx
->uring_lock
);
8539 return ERR_PTR(-ENOMEM
);
8541 refcount_set(&hash
->refs
, 1);
8542 init_waitqueue_head(&hash
->wait
);
8543 ctx
->hash_map
= hash
;
8545 mutex_unlock(&ctx
->uring_lock
);
8549 data
.free_work
= io_wq_free_work
;
8550 data
.do_work
= io_wq_submit_work
;
8552 /* Do QD, or 4 * CPUS, whatever is smallest */
8553 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8555 return io_wq_create(concurrency
, &data
);
8558 static int io_uring_alloc_task_context(struct task_struct
*task
,
8559 struct io_ring_ctx
*ctx
)
8561 struct io_uring_task
*tctx
;
8564 tctx
= kzalloc(sizeof(*tctx
), GFP_KERNEL
);
8565 if (unlikely(!tctx
))
8568 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8569 if (unlikely(ret
)) {
8574 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
8575 if (IS_ERR(tctx
->io_wq
)) {
8576 ret
= PTR_ERR(tctx
->io_wq
);
8577 percpu_counter_destroy(&tctx
->inflight
);
8583 init_waitqueue_head(&tctx
->wait
);
8584 atomic_set(&tctx
->in_idle
, 0);
8585 atomic_set(&tctx
->inflight_tracked
, 0);
8586 task
->io_uring
= tctx
;
8587 spin_lock_init(&tctx
->task_lock
);
8588 INIT_WQ_LIST(&tctx
->task_list
);
8589 init_task_work(&tctx
->task_work
, tctx_task_work
);
8593 void __io_uring_free(struct task_struct
*tsk
)
8595 struct io_uring_task
*tctx
= tsk
->io_uring
;
8597 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8598 WARN_ON_ONCE(tctx
->io_wq
);
8599 WARN_ON_ONCE(tctx
->cached_refs
);
8601 percpu_counter_destroy(&tctx
->inflight
);
8603 tsk
->io_uring
= NULL
;
8606 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8607 struct io_uring_params
*p
)
8611 /* Retain compatibility with failing for an invalid attach attempt */
8612 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
8613 IORING_SETUP_ATTACH_WQ
) {
8616 f
= fdget(p
->wq_fd
);
8619 if (f
.file
->f_op
!= &io_uring_fops
) {
8625 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8626 struct task_struct
*tsk
;
8627 struct io_sq_data
*sqd
;
8630 sqd
= io_get_sq_data(p
, &attached
);
8636 ctx
->sq_creds
= get_current_cred();
8638 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8639 if (!ctx
->sq_thread_idle
)
8640 ctx
->sq_thread_idle
= HZ
;
8642 io_sq_thread_park(sqd
);
8643 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
8644 io_sqd_update_thread_idle(sqd
);
8645 /* don't attach to a dying SQPOLL thread, would be racy */
8646 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
8647 io_sq_thread_unpark(sqd
);
8654 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8655 int cpu
= p
->sq_thread_cpu
;
8658 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
8665 sqd
->task_pid
= current
->pid
;
8666 sqd
->task_tgid
= current
->tgid
;
8667 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
8674 ret
= io_uring_alloc_task_context(tsk
, ctx
);
8675 wake_up_new_task(tsk
);
8678 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8679 /* Can't have SQ_AFF without SQPOLL */
8686 complete(&ctx
->sq_data
->exited
);
8688 io_sq_thread_finish(ctx
);
8692 static inline void __io_unaccount_mem(struct user_struct
*user
,
8693 unsigned long nr_pages
)
8695 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8698 static inline int __io_account_mem(struct user_struct
*user
,
8699 unsigned long nr_pages
)
8701 unsigned long page_limit
, cur_pages
, new_pages
;
8703 /* Don't allow more pages than we can safely lock */
8704 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8707 cur_pages
= atomic_long_read(&user
->locked_vm
);
8708 new_pages
= cur_pages
+ nr_pages
;
8709 if (new_pages
> page_limit
)
8711 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8712 new_pages
) != cur_pages
);
8717 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8720 __io_unaccount_mem(ctx
->user
, nr_pages
);
8722 if (ctx
->mm_account
)
8723 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8726 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8731 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8736 if (ctx
->mm_account
)
8737 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8742 static void io_mem_free(void *ptr
)
8749 page
= virt_to_head_page(ptr
);
8750 if (put_page_testzero(page
))
8751 free_compound_page(page
);
8754 static void *io_mem_alloc(size_t size
)
8756 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
8758 return (void *) __get_free_pages(gfp
, get_order(size
));
8761 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8764 struct io_rings
*rings
;
8765 size_t off
, sq_array_size
;
8767 off
= struct_size(rings
, cqes
, cq_entries
);
8768 if (off
== SIZE_MAX
)
8772 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8780 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8781 if (sq_array_size
== SIZE_MAX
)
8784 if (check_add_overflow(off
, sq_array_size
, &off
))
8790 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
8792 struct io_mapped_ubuf
*imu
= *slot
;
8795 if (imu
!= ctx
->dummy_ubuf
) {
8796 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
8797 unpin_user_page(imu
->bvec
[i
].bv_page
);
8798 if (imu
->acct_pages
)
8799 io_unaccount_mem(ctx
, imu
->acct_pages
);
8805 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8807 io_buffer_unmap(ctx
, &prsrc
->buf
);
8811 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8815 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
8816 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
8817 kfree(ctx
->user_bufs
);
8818 io_rsrc_data_free(ctx
->buf_data
);
8819 ctx
->user_bufs
= NULL
;
8820 ctx
->buf_data
= NULL
;
8821 ctx
->nr_user_bufs
= 0;
8824 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8826 unsigned nr
= ctx
->nr_user_bufs
;
8833 * Quiesce may unlock ->uring_lock, and while it's not held
8834 * prevent new requests using the table.
8836 ctx
->nr_user_bufs
= 0;
8837 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
8838 ctx
->nr_user_bufs
= nr
;
8840 __io_sqe_buffers_unregister(ctx
);
8844 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8845 void __user
*arg
, unsigned index
)
8847 struct iovec __user
*src
;
8849 #ifdef CONFIG_COMPAT
8851 struct compat_iovec __user
*ciovs
;
8852 struct compat_iovec ciov
;
8854 ciovs
= (struct compat_iovec __user
*) arg
;
8855 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8858 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8859 dst
->iov_len
= ciov
.iov_len
;
8863 src
= (struct iovec __user
*) arg
;
8864 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8870 * Not super efficient, but this is just a registration time. And we do cache
8871 * the last compound head, so generally we'll only do a full search if we don't
8874 * We check if the given compound head page has already been accounted, to
8875 * avoid double accounting it. This allows us to account the full size of the
8876 * page, not just the constituent pages of a huge page.
8878 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8879 int nr_pages
, struct page
*hpage
)
8883 /* check current page array */
8884 for (i
= 0; i
< nr_pages
; i
++) {
8885 if (!PageCompound(pages
[i
]))
8887 if (compound_head(pages
[i
]) == hpage
)
8891 /* check previously registered pages */
8892 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8893 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
8895 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8896 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8898 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
8906 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
8907 int nr_pages
, struct io_mapped_ubuf
*imu
,
8908 struct page
**last_hpage
)
8912 imu
->acct_pages
= 0;
8913 for (i
= 0; i
< nr_pages
; i
++) {
8914 if (!PageCompound(pages
[i
])) {
8919 hpage
= compound_head(pages
[i
]);
8920 if (hpage
== *last_hpage
)
8922 *last_hpage
= hpage
;
8923 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
8925 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
8929 if (!imu
->acct_pages
)
8932 ret
= io_account_mem(ctx
, imu
->acct_pages
);
8934 imu
->acct_pages
= 0;
8938 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
8939 struct io_mapped_ubuf
**pimu
,
8940 struct page
**last_hpage
)
8942 struct io_mapped_ubuf
*imu
= NULL
;
8943 struct vm_area_struct
**vmas
= NULL
;
8944 struct page
**pages
= NULL
;
8945 unsigned long off
, start
, end
, ubuf
;
8947 int ret
, pret
, nr_pages
, i
;
8949 if (!iov
->iov_base
) {
8950 *pimu
= ctx
->dummy_ubuf
;
8954 ubuf
= (unsigned long) iov
->iov_base
;
8955 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
8956 start
= ubuf
>> PAGE_SHIFT
;
8957 nr_pages
= end
- start
;
8962 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
8966 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
8971 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
8976 mmap_read_lock(current
->mm
);
8977 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
8979 if (pret
== nr_pages
) {
8980 /* don't support file backed memory */
8981 for (i
= 0; i
< nr_pages
; i
++) {
8982 struct vm_area_struct
*vma
= vmas
[i
];
8984 if (vma_is_shmem(vma
))
8987 !is_file_hugepages(vma
->vm_file
)) {
8993 ret
= pret
< 0 ? pret
: -EFAULT
;
8995 mmap_read_unlock(current
->mm
);
8998 * if we did partial map, or found file backed vmas,
8999 * release any pages we did get
9002 unpin_user_pages(pages
, pret
);
9006 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
9008 unpin_user_pages(pages
, pret
);
9012 off
= ubuf
& ~PAGE_MASK
;
9013 size
= iov
->iov_len
;
9014 for (i
= 0; i
< nr_pages
; i
++) {
9017 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
9018 imu
->bvec
[i
].bv_page
= pages
[i
];
9019 imu
->bvec
[i
].bv_len
= vec_len
;
9020 imu
->bvec
[i
].bv_offset
= off
;
9024 /* store original address for later verification */
9026 imu
->ubuf_end
= ubuf
+ iov
->iov_len
;
9027 imu
->nr_bvecs
= nr_pages
;
9038 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
9040 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
9041 return ctx
->user_bufs
? 0 : -ENOMEM
;
9044 static int io_buffer_validate(struct iovec
*iov
)
9046 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
9049 * Don't impose further limits on the size and buffer
9050 * constraints here, we'll -EINVAL later when IO is
9051 * submitted if they are wrong.
9054 return iov
->iov_len
? -EFAULT
: 0;
9058 /* arbitrary limit, but we need something */
9059 if (iov
->iov_len
> SZ_1G
)
9062 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
9068 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
9069 unsigned int nr_args
, u64 __user
*tags
)
9071 struct page
*last_hpage
= NULL
;
9072 struct io_rsrc_data
*data
;
9078 if (!nr_args
|| nr_args
> IORING_MAX_REG_BUFFERS
)
9080 ret
= io_rsrc_node_switch_start(ctx
);
9083 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, tags
, nr_args
, &data
);
9086 ret
= io_buffers_map_alloc(ctx
, nr_args
);
9088 io_rsrc_data_free(data
);
9092 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
9093 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
9096 ret
= io_buffer_validate(&iov
);
9099 if (!iov
.iov_base
&& *io_get_tag_slot(data
, i
)) {
9104 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
9110 WARN_ON_ONCE(ctx
->buf_data
);
9112 ctx
->buf_data
= data
;
9114 __io_sqe_buffers_unregister(ctx
);
9116 io_rsrc_node_switch(ctx
, NULL
);
9120 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
9121 struct io_uring_rsrc_update2
*up
,
9122 unsigned int nr_args
)
9124 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
9125 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
9126 struct page
*last_hpage
= NULL
;
9127 bool needs_switch
= false;
9133 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
9136 for (done
= 0; done
< nr_args
; done
++) {
9137 struct io_mapped_ubuf
*imu
;
9138 int offset
= up
->offset
+ done
;
9141 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
9144 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
9148 err
= io_buffer_validate(&iov
);
9151 if (!iov
.iov_base
&& tag
) {
9155 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
9159 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
9160 if (ctx
->user_bufs
[i
] != ctx
->dummy_ubuf
) {
9161 err
= io_queue_rsrc_removal(ctx
->buf_data
, i
,
9162 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
9163 if (unlikely(err
)) {
9164 io_buffer_unmap(ctx
, &imu
);
9167 ctx
->user_bufs
[i
] = NULL
;
9168 needs_switch
= true;
9171 ctx
->user_bufs
[i
] = imu
;
9172 *io_get_tag_slot(ctx
->buf_data
, offset
) = tag
;
9176 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
9177 return done
? done
: err
;
9180 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
9182 __s32 __user
*fds
= arg
;
9188 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
9191 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
9192 if (IS_ERR(ctx
->cq_ev_fd
)) {
9193 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
9195 ctx
->cq_ev_fd
= NULL
;
9202 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
9204 if (ctx
->cq_ev_fd
) {
9205 eventfd_ctx_put(ctx
->cq_ev_fd
);
9206 ctx
->cq_ev_fd
= NULL
;
9213 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
9215 struct io_buffer
*buf
;
9216 unsigned long index
;
9218 xa_for_each(&ctx
->io_buffers
, index
, buf
)
9219 __io_remove_buffers(ctx
, buf
, index
, -1U);
9222 static void io_req_cache_free(struct list_head
*list
)
9224 struct io_kiocb
*req
, *nxt
;
9226 list_for_each_entry_safe(req
, nxt
, list
, inflight_entry
) {
9227 list_del(&req
->inflight_entry
);
9228 kmem_cache_free(req_cachep
, req
);
9232 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
9234 struct io_submit_state
*state
= &ctx
->submit_state
;
9236 mutex_lock(&ctx
->uring_lock
);
9238 if (state
->free_reqs
) {
9239 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
9240 state
->free_reqs
= 0;
9243 io_flush_cached_locked_reqs(ctx
, state
);
9244 io_req_cache_free(&state
->free_list
);
9245 mutex_unlock(&ctx
->uring_lock
);
9248 static void io_wait_rsrc_data(struct io_rsrc_data
*data
)
9250 if (data
&& !atomic_dec_and_test(&data
->refs
))
9251 wait_for_completion(&data
->done
);
9254 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
9256 io_sq_thread_finish(ctx
);
9258 if (ctx
->mm_account
) {
9259 mmdrop(ctx
->mm_account
);
9260 ctx
->mm_account
= NULL
;
9263 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9264 io_wait_rsrc_data(ctx
->buf_data
);
9265 io_wait_rsrc_data(ctx
->file_data
);
9267 mutex_lock(&ctx
->uring_lock
);
9269 __io_sqe_buffers_unregister(ctx
);
9271 __io_sqe_files_unregister(ctx
);
9273 __io_cqring_overflow_flush(ctx
, true);
9274 mutex_unlock(&ctx
->uring_lock
);
9275 io_eventfd_unregister(ctx
);
9276 io_destroy_buffers(ctx
);
9278 put_cred(ctx
->sq_creds
);
9280 /* there are no registered resources left, nobody uses it */
9282 io_rsrc_node_destroy(ctx
->rsrc_node
);
9283 if (ctx
->rsrc_backup_node
)
9284 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
9285 flush_delayed_work(&ctx
->rsrc_put_work
);
9287 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
9288 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
9290 #if defined(CONFIG_UNIX)
9291 if (ctx
->ring_sock
) {
9292 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
9293 sock_release(ctx
->ring_sock
);
9296 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
9298 io_mem_free(ctx
->rings
);
9299 io_mem_free(ctx
->sq_sqes
);
9301 percpu_ref_exit(&ctx
->refs
);
9302 free_uid(ctx
->user
);
9303 io_req_caches_free(ctx
);
9305 io_wq_put_hash(ctx
->hash_map
);
9306 kfree(ctx
->cancel_hash
);
9307 kfree(ctx
->dummy_ubuf
);
9311 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
9313 struct io_ring_ctx
*ctx
= file
->private_data
;
9316 poll_wait(file
, &ctx
->poll_wait
, wait
);
9318 * synchronizes with barrier from wq_has_sleeper call in
9322 if (!io_sqring_full(ctx
))
9323 mask
|= EPOLLOUT
| EPOLLWRNORM
;
9326 * Don't flush cqring overflow list here, just do a simple check.
9327 * Otherwise there could possible be ABBA deadlock:
9330 * lock(&ctx->uring_lock);
9332 * lock(&ctx->uring_lock);
9335 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9336 * pushs them to do the flush.
9338 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->check_cq_overflow
))
9339 mask
|= EPOLLIN
| EPOLLRDNORM
;
9344 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9346 const struct cred
*creds
;
9348 creds
= xa_erase(&ctx
->personalities
, id
);
9357 struct io_tctx_exit
{
9358 struct callback_head task_work
;
9359 struct completion completion
;
9360 struct io_ring_ctx
*ctx
;
9363 static void io_tctx_exit_cb(struct callback_head
*cb
)
9365 struct io_uring_task
*tctx
= current
->io_uring
;
9366 struct io_tctx_exit
*work
;
9368 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
9370 * When @in_idle, we're in cancellation and it's racy to remove the
9371 * node. It'll be removed by the end of cancellation, just ignore it.
9373 if (!atomic_read(&tctx
->in_idle
))
9374 io_uring_del_tctx_node((unsigned long)work
->ctx
);
9375 complete(&work
->completion
);
9378 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
9380 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9382 return req
->ctx
== data
;
9385 static void io_ring_exit_work(struct work_struct
*work
)
9387 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
9388 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
9389 unsigned long interval
= HZ
/ 20;
9390 struct io_tctx_exit exit
;
9391 struct io_tctx_node
*node
;
9395 * If we're doing polled IO and end up having requests being
9396 * submitted async (out-of-line), then completions can come in while
9397 * we're waiting for refs to drop. We need to reap these manually,
9398 * as nobody else will be looking for them.
9401 io_uring_try_cancel_requests(ctx
, NULL
, true);
9403 struct io_sq_data
*sqd
= ctx
->sq_data
;
9404 struct task_struct
*tsk
;
9406 io_sq_thread_park(sqd
);
9408 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
9409 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
9410 io_cancel_ctx_cb
, ctx
, true);
9411 io_sq_thread_unpark(sqd
);
9414 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
9415 /* there is little hope left, don't run it too often */
9418 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
9420 init_completion(&exit
.completion
);
9421 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
9424 * Some may use context even when all refs and requests have been put,
9425 * and they are free to do so while still holding uring_lock or
9426 * completion_lock, see io_req_task_submit(). Apart from other work,
9427 * this lock/unlock section also waits them to finish.
9429 mutex_lock(&ctx
->uring_lock
);
9430 while (!list_empty(&ctx
->tctx_list
)) {
9431 WARN_ON_ONCE(time_after(jiffies
, timeout
));
9433 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
9435 /* don't spin on a single task if cancellation failed */
9436 list_rotate_left(&ctx
->tctx_list
);
9437 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
9438 if (WARN_ON_ONCE(ret
))
9440 wake_up_process(node
->task
);
9442 mutex_unlock(&ctx
->uring_lock
);
9443 wait_for_completion(&exit
.completion
);
9444 mutex_lock(&ctx
->uring_lock
);
9446 mutex_unlock(&ctx
->uring_lock
);
9447 spin_lock(&ctx
->completion_lock
);
9448 spin_unlock(&ctx
->completion_lock
);
9450 io_ring_ctx_free(ctx
);
9453 /* Returns true if we found and killed one or more timeouts */
9454 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
9457 struct io_kiocb
*req
, *tmp
;
9460 spin_lock(&ctx
->completion_lock
);
9461 spin_lock_irq(&ctx
->timeout_lock
);
9462 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
9463 if (io_match_task(req
, tsk
, cancel_all
)) {
9464 io_kill_timeout(req
, -ECANCELED
);
9468 spin_unlock_irq(&ctx
->timeout_lock
);
9470 io_commit_cqring(ctx
);
9471 spin_unlock(&ctx
->completion_lock
);
9473 io_cqring_ev_posted(ctx
);
9474 return canceled
!= 0;
9477 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
9479 unsigned long index
;
9480 struct creds
*creds
;
9482 mutex_lock(&ctx
->uring_lock
);
9483 percpu_ref_kill(&ctx
->refs
);
9485 __io_cqring_overflow_flush(ctx
, true);
9486 xa_for_each(&ctx
->personalities
, index
, creds
)
9487 io_unregister_personality(ctx
, index
);
9488 mutex_unlock(&ctx
->uring_lock
);
9490 io_kill_timeouts(ctx
, NULL
, true);
9491 io_poll_remove_all(ctx
, NULL
, true);
9493 /* if we failed setting up the ctx, we might not have any rings */
9494 io_iopoll_try_reap_events(ctx
);
9496 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
9498 * Use system_unbound_wq to avoid spawning tons of event kworkers
9499 * if we're exiting a ton of rings at the same time. It just adds
9500 * noise and overhead, there's no discernable change in runtime
9501 * over using system_wq.
9503 queue_work(system_unbound_wq
, &ctx
->exit_work
);
9506 static int io_uring_release(struct inode
*inode
, struct file
*file
)
9508 struct io_ring_ctx
*ctx
= file
->private_data
;
9510 file
->private_data
= NULL
;
9511 io_ring_ctx_wait_and_kill(ctx
);
9515 struct io_task_cancel
{
9516 struct task_struct
*task
;
9520 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
9522 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9523 struct io_task_cancel
*cancel
= data
;
9525 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
9528 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
9529 struct task_struct
*task
, bool cancel_all
)
9531 struct io_defer_entry
*de
;
9534 spin_lock(&ctx
->completion_lock
);
9535 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
9536 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
9537 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
9541 spin_unlock(&ctx
->completion_lock
);
9542 if (list_empty(&list
))
9545 while (!list_empty(&list
)) {
9546 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
9547 list_del_init(&de
->list
);
9548 io_req_complete_failed(de
->req
, -ECANCELED
);
9554 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
9556 struct io_tctx_node
*node
;
9557 enum io_wq_cancel cret
;
9560 mutex_lock(&ctx
->uring_lock
);
9561 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
9562 struct io_uring_task
*tctx
= node
->task
->io_uring
;
9565 * io_wq will stay alive while we hold uring_lock, because it's
9566 * killed after ctx nodes, which requires to take the lock.
9568 if (!tctx
|| !tctx
->io_wq
)
9570 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
9571 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9573 mutex_unlock(&ctx
->uring_lock
);
9578 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
9579 struct task_struct
*task
,
9582 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
9583 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
9586 enum io_wq_cancel cret
;
9590 ret
|= io_uring_try_cancel_iowq(ctx
);
9591 } else if (tctx
&& tctx
->io_wq
) {
9593 * Cancels requests of all rings, not only @ctx, but
9594 * it's fine as the task is in exit/exec.
9596 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
9598 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9601 /* SQPOLL thread does its own polling */
9602 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
9603 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
9604 while (!list_empty_careful(&ctx
->iopoll_list
)) {
9605 io_iopoll_try_reap_events(ctx
);
9610 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
9611 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
9612 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
9614 ret
|= io_run_task_work();
9621 static int __io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9623 struct io_uring_task
*tctx
= current
->io_uring
;
9624 struct io_tctx_node
*node
;
9627 if (unlikely(!tctx
)) {
9628 ret
= io_uring_alloc_task_context(current
, ctx
);
9632 tctx
= current
->io_uring
;
9633 if (ctx
->iowq_limits_set
) {
9634 unsigned int limits
[2] = { ctx
->iowq_limits
[0],
9635 ctx
->iowq_limits
[1], };
9637 ret
= io_wq_max_workers(tctx
->io_wq
, limits
);
9642 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
9643 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
9647 node
->task
= current
;
9649 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
9656 mutex_lock(&ctx
->uring_lock
);
9657 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
9658 mutex_unlock(&ctx
->uring_lock
);
9665 * Note that this task has used io_uring. We use it for cancelation purposes.
9667 static inline int io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9669 struct io_uring_task
*tctx
= current
->io_uring
;
9671 if (likely(tctx
&& tctx
->last
== ctx
))
9673 return __io_uring_add_tctx_node(ctx
);
9677 * Remove this io_uring_file -> task mapping.
9679 static void io_uring_del_tctx_node(unsigned long index
)
9681 struct io_uring_task
*tctx
= current
->io_uring
;
9682 struct io_tctx_node
*node
;
9686 node
= xa_erase(&tctx
->xa
, index
);
9690 WARN_ON_ONCE(current
!= node
->task
);
9691 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
9693 mutex_lock(&node
->ctx
->uring_lock
);
9694 list_del(&node
->ctx_node
);
9695 mutex_unlock(&node
->ctx
->uring_lock
);
9697 if (tctx
->last
== node
->ctx
)
9702 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
9704 struct io_wq
*wq
= tctx
->io_wq
;
9705 struct io_tctx_node
*node
;
9706 unsigned long index
;
9708 xa_for_each(&tctx
->xa
, index
, node
) {
9709 io_uring_del_tctx_node(index
);
9714 * Must be after io_uring_del_task_file() (removes nodes under
9715 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9717 io_wq_put_and_exit(wq
);
9722 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
9725 return atomic_read(&tctx
->inflight_tracked
);
9726 return percpu_counter_sum(&tctx
->inflight
);
9730 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9731 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9733 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
9735 struct io_uring_task
*tctx
= current
->io_uring
;
9736 struct io_ring_ctx
*ctx
;
9740 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
9742 if (!current
->io_uring
)
9745 io_wq_exit_start(tctx
->io_wq
);
9747 atomic_inc(&tctx
->in_idle
);
9749 io_uring_drop_tctx_refs(current
);
9750 /* read completions before cancelations */
9751 inflight
= tctx_inflight(tctx
, !cancel_all
);
9756 struct io_tctx_node
*node
;
9757 unsigned long index
;
9759 xa_for_each(&tctx
->xa
, index
, node
) {
9760 /* sqpoll task will cancel all its requests */
9761 if (node
->ctx
->sq_data
)
9763 io_uring_try_cancel_requests(node
->ctx
, current
,
9767 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9768 io_uring_try_cancel_requests(ctx
, current
,
9772 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
9774 io_uring_drop_tctx_refs(current
);
9777 * If we've seen completions, retry without waiting. This
9778 * avoids a race where a completion comes in before we did
9779 * prepare_to_wait().
9781 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
9783 finish_wait(&tctx
->wait
, &wait
);
9786 io_uring_clean_tctx(tctx
);
9789 * We shouldn't run task_works after cancel, so just leave
9790 * ->in_idle set for normal exit.
9792 atomic_dec(&tctx
->in_idle
);
9793 /* for exec all current's requests should be gone, kill tctx */
9794 __io_uring_free(current
);
9798 void __io_uring_cancel(bool cancel_all
)
9800 io_uring_cancel_generic(cancel_all
, NULL
);
9803 static void *io_uring_validate_mmap_request(struct file
*file
,
9804 loff_t pgoff
, size_t sz
)
9806 struct io_ring_ctx
*ctx
= file
->private_data
;
9807 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9812 case IORING_OFF_SQ_RING
:
9813 case IORING_OFF_CQ_RING
:
9816 case IORING_OFF_SQES
:
9820 return ERR_PTR(-EINVAL
);
9823 page
= virt_to_head_page(ptr
);
9824 if (sz
> page_size(page
))
9825 return ERR_PTR(-EINVAL
);
9832 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9834 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9838 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9840 return PTR_ERR(ptr
);
9842 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9843 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9846 #else /* !CONFIG_MMU */
9848 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9850 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9853 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9855 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9858 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9859 unsigned long addr
, unsigned long len
,
9860 unsigned long pgoff
, unsigned long flags
)
9864 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9866 return PTR_ERR(ptr
);
9868 return (unsigned long) ptr
;
9871 #endif /* !CONFIG_MMU */
9873 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9878 if (!io_sqring_full(ctx
))
9880 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9882 if (!io_sqring_full(ctx
))
9885 } while (!signal_pending(current
));
9887 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9891 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9892 struct __kernel_timespec __user
**ts
,
9893 const sigset_t __user
**sig
)
9895 struct io_uring_getevents_arg arg
;
9898 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9899 * is just a pointer to the sigset_t.
9901 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
9902 *sig
= (const sigset_t __user
*) argp
;
9908 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9909 * timespec and sigset_t pointers if good.
9911 if (*argsz
!= sizeof(arg
))
9913 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
9917 *sig
= u64_to_user_ptr(arg
.sigmask
);
9918 *argsz
= arg
.sigmask_sz
;
9919 *ts
= u64_to_user_ptr(arg
.ts
);
9923 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
9924 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
9927 struct io_ring_ctx
*ctx
;
9934 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
9935 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
)))
9939 if (unlikely(!f
.file
))
9943 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
9947 ctx
= f
.file
->private_data
;
9948 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
9952 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
9956 * For SQ polling, the thread will do all submissions and completions.
9957 * Just return the requested submit count, and wake the thread if
9961 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
9962 io_cqring_overflow_flush(ctx
);
9964 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
9968 if (flags
& IORING_ENTER_SQ_WAKEUP
)
9969 wake_up(&ctx
->sq_data
->wait
);
9970 if (flags
& IORING_ENTER_SQ_WAIT
) {
9971 ret
= io_sqpoll_wait_sq(ctx
);
9975 submitted
= to_submit
;
9976 } else if (to_submit
) {
9977 ret
= io_uring_add_tctx_node(ctx
);
9980 mutex_lock(&ctx
->uring_lock
);
9981 submitted
= io_submit_sqes(ctx
, to_submit
);
9982 mutex_unlock(&ctx
->uring_lock
);
9984 if (submitted
!= to_submit
)
9987 if (flags
& IORING_ENTER_GETEVENTS
) {
9988 const sigset_t __user
*sig
;
9989 struct __kernel_timespec __user
*ts
;
9991 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
9995 min_complete
= min(min_complete
, ctx
->cq_entries
);
9998 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9999 * space applications don't need to do io completion events
10000 * polling again, they can rely on io_sq_thread to do polling
10001 * work, which can reduce cpu usage and uring_lock contention.
10003 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
10004 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10005 ret
= io_iopoll_check(ctx
, min_complete
);
10007 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
10012 percpu_ref_put(&ctx
->refs
);
10015 return submitted
? submitted
: ret
;
10018 #ifdef CONFIG_PROC_FS
10019 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
10020 const struct cred
*cred
)
10022 struct user_namespace
*uns
= seq_user_ns(m
);
10023 struct group_info
*gi
;
10028 seq_printf(m
, "%5d\n", id
);
10029 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
10030 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
10031 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
10032 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
10033 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
10034 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
10035 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
10036 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
10037 seq_puts(m
, "\n\tGroups:\t");
10038 gi
= cred
->group_info
;
10039 for (g
= 0; g
< gi
->ngroups
; g
++) {
10040 seq_put_decimal_ull(m
, g
? " " : "",
10041 from_kgid_munged(uns
, gi
->gid
[g
]));
10043 seq_puts(m
, "\n\tCapEff:\t");
10044 cap
= cred
->cap_effective
;
10045 CAP_FOR_EACH_U32(__capi
)
10046 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
10051 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
10053 struct io_sq_data
*sq
= NULL
;
10058 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10059 * since fdinfo case grabs it in the opposite direction of normal use
10060 * cases. If we fail to get the lock, we just don't iterate any
10061 * structures that could be going away outside the io_uring mutex.
10063 has_lock
= mutex_trylock(&ctx
->uring_lock
);
10065 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10071 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
10072 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
10073 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
10074 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
10075 struct file
*f
= io_file_from_index(ctx
, i
);
10078 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
10080 seq_printf(m
, "%5u: <none>\n", i
);
10082 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
10083 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
10084 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
10085 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
10087 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
10089 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
10090 unsigned long index
;
10091 const struct cred
*cred
;
10093 seq_printf(m
, "Personalities:\n");
10094 xa_for_each(&ctx
->personalities
, index
, cred
)
10095 io_uring_show_cred(m
, index
, cred
);
10097 seq_printf(m
, "PollList:\n");
10098 spin_lock(&ctx
->completion_lock
);
10099 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
10100 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
10101 struct io_kiocb
*req
;
10103 hlist_for_each_entry(req
, list
, hash_node
)
10104 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
10105 req
->task
->task_works
!= NULL
);
10107 spin_unlock(&ctx
->completion_lock
);
10109 mutex_unlock(&ctx
->uring_lock
);
10112 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
10114 struct io_ring_ctx
*ctx
= f
->private_data
;
10116 if (percpu_ref_tryget(&ctx
->refs
)) {
10117 __io_uring_show_fdinfo(ctx
, m
);
10118 percpu_ref_put(&ctx
->refs
);
10123 static const struct file_operations io_uring_fops
= {
10124 .release
= io_uring_release
,
10125 .mmap
= io_uring_mmap
,
10127 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
10128 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
10130 .poll
= io_uring_poll
,
10131 #ifdef CONFIG_PROC_FS
10132 .show_fdinfo
= io_uring_show_fdinfo
,
10136 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
10137 struct io_uring_params
*p
)
10139 struct io_rings
*rings
;
10140 size_t size
, sq_array_offset
;
10142 /* make sure these are sane, as we already accounted them */
10143 ctx
->sq_entries
= p
->sq_entries
;
10144 ctx
->cq_entries
= p
->cq_entries
;
10146 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
10147 if (size
== SIZE_MAX
)
10150 rings
= io_mem_alloc(size
);
10154 ctx
->rings
= rings
;
10155 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
10156 rings
->sq_ring_mask
= p
->sq_entries
- 1;
10157 rings
->cq_ring_mask
= p
->cq_entries
- 1;
10158 rings
->sq_ring_entries
= p
->sq_entries
;
10159 rings
->cq_ring_entries
= p
->cq_entries
;
10161 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
10162 if (size
== SIZE_MAX
) {
10163 io_mem_free(ctx
->rings
);
10168 ctx
->sq_sqes
= io_mem_alloc(size
);
10169 if (!ctx
->sq_sqes
) {
10170 io_mem_free(ctx
->rings
);
10178 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
10182 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
10186 ret
= io_uring_add_tctx_node(ctx
);
10191 fd_install(fd
, file
);
10196 * Allocate an anonymous fd, this is what constitutes the application
10197 * visible backing of an io_uring instance. The application mmaps this
10198 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10199 * we have to tie this fd to a socket for file garbage collection purposes.
10201 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
10204 #if defined(CONFIG_UNIX)
10207 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
10210 return ERR_PTR(ret
);
10213 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
10214 O_RDWR
| O_CLOEXEC
);
10215 #if defined(CONFIG_UNIX)
10216 if (IS_ERR(file
)) {
10217 sock_release(ctx
->ring_sock
);
10218 ctx
->ring_sock
= NULL
;
10220 ctx
->ring_sock
->file
= file
;
10226 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
10227 struct io_uring_params __user
*params
)
10229 struct io_ring_ctx
*ctx
;
10235 if (entries
> IORING_MAX_ENTRIES
) {
10236 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10238 entries
= IORING_MAX_ENTRIES
;
10242 * Use twice as many entries for the CQ ring. It's possible for the
10243 * application to drive a higher depth than the size of the SQ ring,
10244 * since the sqes are only used at submission time. This allows for
10245 * some flexibility in overcommitting a bit. If the application has
10246 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10247 * of CQ ring entries manually.
10249 p
->sq_entries
= roundup_pow_of_two(entries
);
10250 if (p
->flags
& IORING_SETUP_CQSIZE
) {
10252 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10253 * to a power-of-two, if it isn't already. We do NOT impose
10254 * any cq vs sq ring sizing.
10256 if (!p
->cq_entries
)
10258 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
10259 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10261 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
10263 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
10264 if (p
->cq_entries
< p
->sq_entries
)
10267 p
->cq_entries
= 2 * p
->sq_entries
;
10270 ctx
= io_ring_ctx_alloc(p
);
10273 ctx
->compat
= in_compat_syscall();
10274 if (!capable(CAP_IPC_LOCK
))
10275 ctx
->user
= get_uid(current_user());
10278 * This is just grabbed for accounting purposes. When a process exits,
10279 * the mm is exited and dropped before the files, hence we need to hang
10280 * on to this mm purely for the purposes of being able to unaccount
10281 * memory (locked/pinned vm). It's not used for anything else.
10283 mmgrab(current
->mm
);
10284 ctx
->mm_account
= current
->mm
;
10286 ret
= io_allocate_scq_urings(ctx
, p
);
10290 ret
= io_sq_offload_create(ctx
, p
);
10293 /* always set a rsrc node */
10294 ret
= io_rsrc_node_switch_start(ctx
);
10297 io_rsrc_node_switch(ctx
, NULL
);
10299 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
10300 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
10301 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
10302 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
10303 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
10304 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
10305 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
10306 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
10308 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
10309 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
10310 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
10311 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
10312 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
10313 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
10314 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
10315 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
10317 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
10318 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
10319 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
10320 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
10321 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
10322 IORING_FEAT_RSRC_TAGS
;
10324 if (copy_to_user(params
, p
, sizeof(*p
))) {
10329 file
= io_uring_get_file(ctx
);
10330 if (IS_ERR(file
)) {
10331 ret
= PTR_ERR(file
);
10336 * Install ring fd as the very last thing, so we don't risk someone
10337 * having closed it before we finish setup
10339 ret
= io_uring_install_fd(ctx
, file
);
10341 /* fput will clean it up */
10346 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
10349 io_ring_ctx_wait_and_kill(ctx
);
10354 * Sets up an aio uring context, and returns the fd. Applications asks for a
10355 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10356 * params structure passed in.
10358 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
10360 struct io_uring_params p
;
10363 if (copy_from_user(&p
, params
, sizeof(p
)))
10365 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
10370 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
10371 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
10372 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
10373 IORING_SETUP_R_DISABLED
))
10376 return io_uring_create(entries
, &p
, params
);
10379 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
10380 struct io_uring_params __user
*, params
)
10382 return io_uring_setup(entries
, params
);
10385 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
10387 struct io_uring_probe
*p
;
10391 size
= struct_size(p
, ops
, nr_args
);
10392 if (size
== SIZE_MAX
)
10394 p
= kzalloc(size
, GFP_KERNEL
);
10399 if (copy_from_user(p
, arg
, size
))
10402 if (memchr_inv(p
, 0, size
))
10405 p
->last_op
= IORING_OP_LAST
- 1;
10406 if (nr_args
> IORING_OP_LAST
)
10407 nr_args
= IORING_OP_LAST
;
10409 for (i
= 0; i
< nr_args
; i
++) {
10411 if (!io_op_defs
[i
].not_supported
)
10412 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
10417 if (copy_to_user(arg
, p
, size
))
10424 static int io_register_personality(struct io_ring_ctx
*ctx
)
10426 const struct cred
*creds
;
10430 creds
= get_current_cred();
10432 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
10433 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
10441 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
10442 unsigned int nr_args
)
10444 struct io_uring_restriction
*res
;
10448 /* Restrictions allowed only if rings started disabled */
10449 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10452 /* We allow only a single restrictions registration */
10453 if (ctx
->restrictions
.registered
)
10456 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
10459 size
= array_size(nr_args
, sizeof(*res
));
10460 if (size
== SIZE_MAX
)
10463 res
= memdup_user(arg
, size
);
10465 return PTR_ERR(res
);
10469 for (i
= 0; i
< nr_args
; i
++) {
10470 switch (res
[i
].opcode
) {
10471 case IORING_RESTRICTION_REGISTER_OP
:
10472 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
10477 __set_bit(res
[i
].register_op
,
10478 ctx
->restrictions
.register_op
);
10480 case IORING_RESTRICTION_SQE_OP
:
10481 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
10486 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
10488 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
10489 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
10491 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
10492 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
10501 /* Reset all restrictions if an error happened */
10503 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
10505 ctx
->restrictions
.registered
= true;
10511 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10513 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10516 if (ctx
->restrictions
.registered
)
10517 ctx
->restricted
= 1;
10519 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
10520 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
10521 wake_up(&ctx
->sq_data
->wait
);
10525 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
10526 struct io_uring_rsrc_update2
*up
,
10532 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
10534 err
= io_rsrc_node_switch_start(ctx
);
10539 case IORING_RSRC_FILE
:
10540 return __io_sqe_files_update(ctx
, up
, nr_args
);
10541 case IORING_RSRC_BUFFER
:
10542 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
10547 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10550 struct io_uring_rsrc_update2 up
;
10554 memset(&up
, 0, sizeof(up
));
10555 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
10557 if (up
.resv
|| up
.resv2
)
10559 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
10562 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10563 unsigned size
, unsigned type
)
10565 struct io_uring_rsrc_update2 up
;
10567 if (size
!= sizeof(up
))
10569 if (copy_from_user(&up
, arg
, sizeof(up
)))
10571 if (!up
.nr
|| up
.resv
|| up
.resv2
)
10573 return __io_register_rsrc_update(ctx
, type
, &up
, up
.nr
);
10576 static int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
10577 unsigned int size
, unsigned int type
)
10579 struct io_uring_rsrc_register rr
;
10581 /* keep it extendible */
10582 if (size
!= sizeof(rr
))
10585 memset(&rr
, 0, sizeof(rr
));
10586 if (copy_from_user(&rr
, arg
, size
))
10588 if (!rr
.nr
|| rr
.resv
|| rr
.resv2
)
10592 case IORING_RSRC_FILE
:
10593 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
10594 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10595 case IORING_RSRC_BUFFER
:
10596 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
10597 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10602 static int io_register_iowq_aff(struct io_ring_ctx
*ctx
, void __user
*arg
,
10605 struct io_uring_task
*tctx
= current
->io_uring
;
10606 cpumask_var_t new_mask
;
10609 if (!tctx
|| !tctx
->io_wq
)
10612 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
10615 cpumask_clear(new_mask
);
10616 if (len
> cpumask_size())
10617 len
= cpumask_size();
10619 if (in_compat_syscall()) {
10620 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
10621 (const compat_ulong_t __user
*)arg
,
10622 len
* 8 /* CHAR_BIT */);
10624 ret
= copy_from_user(new_mask
, arg
, len
);
10628 free_cpumask_var(new_mask
);
10632 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
10633 free_cpumask_var(new_mask
);
10637 static int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
10639 struct io_uring_task
*tctx
= current
->io_uring
;
10641 if (!tctx
|| !tctx
->io_wq
)
10644 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
10647 static int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
10649 __must_hold(&ctx
->uring_lock
)
10651 struct io_tctx_node
*node
;
10652 struct io_uring_task
*tctx
= NULL
;
10653 struct io_sq_data
*sqd
= NULL
;
10654 __u32 new_count
[2];
10657 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
10659 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10660 if (new_count
[i
] > INT_MAX
)
10663 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10664 sqd
= ctx
->sq_data
;
10667 * Observe the correct sqd->lock -> ctx->uring_lock
10668 * ordering. Fine to drop uring_lock here, we hold
10669 * a ref to the ctx.
10671 refcount_inc(&sqd
->refs
);
10672 mutex_unlock(&ctx
->uring_lock
);
10673 mutex_lock(&sqd
->lock
);
10674 mutex_lock(&ctx
->uring_lock
);
10676 tctx
= sqd
->thread
->io_uring
;
10679 tctx
= current
->io_uring
;
10682 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
10684 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10686 ctx
->iowq_limits
[i
] = new_count
[i
];
10687 ctx
->iowq_limits_set
= true;
10690 if (tctx
&& tctx
->io_wq
) {
10691 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
10695 memset(new_count
, 0, sizeof(new_count
));
10699 mutex_unlock(&sqd
->lock
);
10700 io_put_sq_data(sqd
);
10703 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
10706 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10710 /* now propagate the restriction to all registered users */
10711 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
10712 struct io_uring_task
*tctx
= node
->task
->io_uring
;
10714 if (WARN_ON_ONCE(!tctx
->io_wq
))
10717 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10718 new_count
[i
] = ctx
->iowq_limits
[i
];
10719 /* ignore errors, it always returns zero anyway */
10720 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
10725 mutex_unlock(&sqd
->lock
);
10726 io_put_sq_data(sqd
);
10731 static bool io_register_op_must_quiesce(int op
)
10734 case IORING_REGISTER_BUFFERS
:
10735 case IORING_UNREGISTER_BUFFERS
:
10736 case IORING_REGISTER_FILES
:
10737 case IORING_UNREGISTER_FILES
:
10738 case IORING_REGISTER_FILES_UPDATE
:
10739 case IORING_REGISTER_PROBE
:
10740 case IORING_REGISTER_PERSONALITY
:
10741 case IORING_UNREGISTER_PERSONALITY
:
10742 case IORING_REGISTER_FILES2
:
10743 case IORING_REGISTER_FILES_UPDATE2
:
10744 case IORING_REGISTER_BUFFERS2
:
10745 case IORING_REGISTER_BUFFERS_UPDATE
:
10746 case IORING_REGISTER_IOWQ_AFF
:
10747 case IORING_UNREGISTER_IOWQ_AFF
:
10748 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10755 static int io_ctx_quiesce(struct io_ring_ctx
*ctx
)
10759 percpu_ref_kill(&ctx
->refs
);
10762 * Drop uring mutex before waiting for references to exit. If another
10763 * thread is currently inside io_uring_enter() it might need to grab the
10764 * uring_lock to make progress. If we hold it here across the drain
10765 * wait, then we can deadlock. It's safe to drop the mutex here, since
10766 * no new references will come in after we've killed the percpu ref.
10768 mutex_unlock(&ctx
->uring_lock
);
10770 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10773 ret
= io_run_task_work_sig();
10774 } while (ret
>= 0);
10775 mutex_lock(&ctx
->uring_lock
);
10778 io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
);
10782 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10783 void __user
*arg
, unsigned nr_args
)
10784 __releases(ctx
->uring_lock
)
10785 __acquires(ctx
->uring_lock
)
10790 * We're inside the ring mutex, if the ref is already dying, then
10791 * someone else killed the ctx or is already going through
10792 * io_uring_register().
10794 if (percpu_ref_is_dying(&ctx
->refs
))
10797 if (ctx
->restricted
) {
10798 if (opcode
>= IORING_REGISTER_LAST
)
10800 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
10801 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
10805 if (io_register_op_must_quiesce(opcode
)) {
10806 ret
= io_ctx_quiesce(ctx
);
10812 case IORING_REGISTER_BUFFERS
:
10813 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
10815 case IORING_UNREGISTER_BUFFERS
:
10817 if (arg
|| nr_args
)
10819 ret
= io_sqe_buffers_unregister(ctx
);
10821 case IORING_REGISTER_FILES
:
10822 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
10824 case IORING_UNREGISTER_FILES
:
10826 if (arg
|| nr_args
)
10828 ret
= io_sqe_files_unregister(ctx
);
10830 case IORING_REGISTER_FILES_UPDATE
:
10831 ret
= io_register_files_update(ctx
, arg
, nr_args
);
10833 case IORING_REGISTER_EVENTFD
:
10834 case IORING_REGISTER_EVENTFD_ASYNC
:
10838 ret
= io_eventfd_register(ctx
, arg
);
10841 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10842 ctx
->eventfd_async
= 1;
10844 ctx
->eventfd_async
= 0;
10846 case IORING_UNREGISTER_EVENTFD
:
10848 if (arg
|| nr_args
)
10850 ret
= io_eventfd_unregister(ctx
);
10852 case IORING_REGISTER_PROBE
:
10854 if (!arg
|| nr_args
> 256)
10856 ret
= io_probe(ctx
, arg
, nr_args
);
10858 case IORING_REGISTER_PERSONALITY
:
10860 if (arg
|| nr_args
)
10862 ret
= io_register_personality(ctx
);
10864 case IORING_UNREGISTER_PERSONALITY
:
10868 ret
= io_unregister_personality(ctx
, nr_args
);
10870 case IORING_REGISTER_ENABLE_RINGS
:
10872 if (arg
|| nr_args
)
10874 ret
= io_register_enable_rings(ctx
);
10876 case IORING_REGISTER_RESTRICTIONS
:
10877 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10879 case IORING_REGISTER_FILES2
:
10880 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
10882 case IORING_REGISTER_FILES_UPDATE2
:
10883 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10886 case IORING_REGISTER_BUFFERS2
:
10887 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
10889 case IORING_REGISTER_BUFFERS_UPDATE
:
10890 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10891 IORING_RSRC_BUFFER
);
10893 case IORING_REGISTER_IOWQ_AFF
:
10895 if (!arg
|| !nr_args
)
10897 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
10899 case IORING_UNREGISTER_IOWQ_AFF
:
10901 if (arg
|| nr_args
)
10903 ret
= io_unregister_iowq_aff(ctx
);
10905 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10907 if (!arg
|| nr_args
!= 2)
10909 ret
= io_register_iowq_max_workers(ctx
, arg
);
10916 if (io_register_op_must_quiesce(opcode
)) {
10917 /* bring the ctx back to life */
10918 percpu_ref_reinit(&ctx
->refs
);
10919 reinit_completion(&ctx
->ref_comp
);
10924 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
10925 void __user
*, arg
, unsigned int, nr_args
)
10927 struct io_ring_ctx
*ctx
;
10936 if (f
.file
->f_op
!= &io_uring_fops
)
10939 ctx
= f
.file
->private_data
;
10941 io_run_task_work();
10943 mutex_lock(&ctx
->uring_lock
);
10944 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
10945 mutex_unlock(&ctx
->uring_lock
);
10946 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
10947 ctx
->cq_ev_fd
!= NULL
, ret
);
10953 static int __init
io_uring_init(void)
10955 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10956 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10957 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10960 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10961 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10962 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
10963 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
10964 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
10965 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
10966 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
10967 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
10968 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
10969 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
10970 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
10971 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
10972 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
10973 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
10974 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
10975 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
10976 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
10977 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
10978 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
10979 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
10980 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
10981 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
10982 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
10983 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
10984 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
10985 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
10986 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
10987 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
10988 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
10989 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
10990 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
10991 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
10992 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
10994 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
10995 sizeof(struct io_uring_rsrc_update
));
10996 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
10997 sizeof(struct io_uring_rsrc_update2
));
10999 /* ->buf_index is u16 */
11000 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS
>= (1u << 16));
11002 /* should fit into one byte */
11003 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
11005 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
11006 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
11008 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
11012 __initcall(io_uring_init
);