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_get(struct io_kiocb
*req
)
1160 WARN_ON_ONCE(!(req
->flags
& REQ_F_REFCOUNT
));
1161 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req
));
1162 atomic_inc(&req
->refs
);
1165 static inline void __io_req_set_refcount(struct io_kiocb
*req
, int nr
)
1167 if (!(req
->flags
& REQ_F_REFCOUNT
)) {
1168 req
->flags
|= REQ_F_REFCOUNT
;
1169 atomic_set(&req
->refs
, nr
);
1173 static inline void io_req_set_refcount(struct io_kiocb
*req
)
1175 __io_req_set_refcount(req
, 1);
1178 static inline void io_req_set_rsrc_node(struct io_kiocb
*req
)
1180 struct io_ring_ctx
*ctx
= req
->ctx
;
1182 if (!req
->fixed_rsrc_refs
) {
1183 req
->fixed_rsrc_refs
= &ctx
->rsrc_node
->refs
;
1184 percpu_ref_get(req
->fixed_rsrc_refs
);
1188 static void io_refs_resurrect(struct percpu_ref
*ref
, struct completion
*compl)
1190 bool got
= percpu_ref_tryget(ref
);
1192 /* already at zero, wait for ->release() */
1194 wait_for_completion(compl);
1195 percpu_ref_resurrect(ref
);
1197 percpu_ref_put(ref
);
1200 static bool io_match_task(struct io_kiocb
*head
, struct task_struct
*task
,
1202 __must_hold(&req
->ctx
->timeout_lock
)
1204 struct io_kiocb
*req
;
1206 if (task
&& head
->task
!= task
)
1211 io_for_each_link(req
, head
) {
1212 if (req
->flags
& REQ_F_INFLIGHT
)
1218 static bool io_match_linked(struct io_kiocb
*head
)
1220 struct io_kiocb
*req
;
1222 io_for_each_link(req
, head
) {
1223 if (req
->flags
& REQ_F_INFLIGHT
)
1230 * As io_match_task() but protected against racing with linked timeouts.
1231 * User must not hold timeout_lock.
1233 static bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
1238 if (task
&& head
->task
!= task
)
1243 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
1244 struct io_ring_ctx
*ctx
= head
->ctx
;
1246 /* protect against races with linked timeouts */
1247 spin_lock_irq(&ctx
->timeout_lock
);
1248 matched
= io_match_linked(head
);
1249 spin_unlock_irq(&ctx
->timeout_lock
);
1251 matched
= io_match_linked(head
);
1256 static inline void req_set_fail(struct io_kiocb
*req
)
1258 req
->flags
|= REQ_F_FAIL
;
1261 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
1267 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
1269 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
1271 complete(&ctx
->ref_comp
);
1274 static inline bool io_is_timeout_noseq(struct io_kiocb
*req
)
1276 return !req
->timeout
.off
;
1279 static void io_fallback_req_func(struct work_struct
*work
)
1281 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
1282 fallback_work
.work
);
1283 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
1284 struct io_kiocb
*req
, *tmp
;
1285 bool locked
= false;
1287 percpu_ref_get(&ctx
->refs
);
1288 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.fallback_node
)
1289 req
->io_task_work
.func(req
, &locked
);
1292 if (ctx
->submit_state
.compl_nr
)
1293 io_submit_flush_completions(ctx
);
1294 mutex_unlock(&ctx
->uring_lock
);
1296 percpu_ref_put(&ctx
->refs
);
1300 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
1302 struct io_ring_ctx
*ctx
;
1305 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
1310 * Use 5 bits less than the max cq entries, that should give us around
1311 * 32 entries per hash list if totally full and uniformly spread.
1313 hash_bits
= ilog2(p
->cq_entries
);
1317 ctx
->cancel_hash_bits
= hash_bits
;
1318 ctx
->cancel_hash
= kmalloc((1U << hash_bits
) * sizeof(struct hlist_head
),
1320 if (!ctx
->cancel_hash
)
1322 __hash_init(ctx
->cancel_hash
, 1U << hash_bits
);
1324 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
1325 if (!ctx
->dummy_ubuf
)
1327 /* set invalid range, so io_import_fixed() fails meeting it */
1328 ctx
->dummy_ubuf
->ubuf
= -1UL;
1330 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
1331 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
))
1334 ctx
->flags
= p
->flags
;
1335 init_waitqueue_head(&ctx
->sqo_sq_wait
);
1336 INIT_LIST_HEAD(&ctx
->sqd_list
);
1337 init_waitqueue_head(&ctx
->poll_wait
);
1338 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
1339 init_completion(&ctx
->ref_comp
);
1340 xa_init_flags(&ctx
->io_buffers
, XA_FLAGS_ALLOC1
);
1341 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
1342 mutex_init(&ctx
->uring_lock
);
1343 init_waitqueue_head(&ctx
->cq_wait
);
1344 spin_lock_init(&ctx
->completion_lock
);
1345 spin_lock_init(&ctx
->timeout_lock
);
1346 INIT_LIST_HEAD(&ctx
->iopoll_list
);
1347 INIT_LIST_HEAD(&ctx
->defer_list
);
1348 INIT_LIST_HEAD(&ctx
->timeout_list
);
1349 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
1350 spin_lock_init(&ctx
->rsrc_ref_lock
);
1351 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
1352 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
1353 init_llist_head(&ctx
->rsrc_put_llist
);
1354 INIT_LIST_HEAD(&ctx
->tctx_list
);
1355 INIT_LIST_HEAD(&ctx
->submit_state
.free_list
);
1356 INIT_LIST_HEAD(&ctx
->locked_free_list
);
1357 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
1360 kfree(ctx
->dummy_ubuf
);
1361 kfree(ctx
->cancel_hash
);
1366 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
1368 struct io_rings
*r
= ctx
->rings
;
1370 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
1374 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
1376 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
1377 struct io_ring_ctx
*ctx
= req
->ctx
;
1379 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
1385 #define FFS_ASYNC_READ 0x1UL
1386 #define FFS_ASYNC_WRITE 0x2UL
1388 #define FFS_ISREG 0x4UL
1390 #define FFS_ISREG 0x0UL
1392 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1394 static inline bool io_req_ffs_set(struct io_kiocb
*req
)
1396 return IS_ENABLED(CONFIG_64BIT
) && (req
->flags
& REQ_F_FIXED_FILE
);
1399 static void io_req_track_inflight(struct io_kiocb
*req
)
1401 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
1402 req
->flags
|= REQ_F_INFLIGHT
;
1403 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
1407 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
1409 if (WARN_ON_ONCE(!req
->link
))
1412 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
1413 req
->flags
|= REQ_F_LINK_TIMEOUT
;
1415 /* linked timeouts should have two refs once prep'ed */
1416 io_req_set_refcount(req
);
1417 __io_req_set_refcount(req
->link
, 2);
1421 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
1423 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
1425 return __io_prep_linked_timeout(req
);
1428 static void io_prep_async_work(struct io_kiocb
*req
)
1430 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1431 struct io_ring_ctx
*ctx
= req
->ctx
;
1433 if (!(req
->flags
& REQ_F_CREDS
)) {
1434 req
->flags
|= REQ_F_CREDS
;
1435 req
->creds
= get_current_cred();
1438 req
->work
.list
.next
= NULL
;
1439 req
->work
.flags
= 0;
1440 if (req
->flags
& REQ_F_FORCE_ASYNC
)
1441 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
1443 if (req
->flags
& REQ_F_ISREG
) {
1444 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
1445 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
1446 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
1447 if (def
->unbound_nonreg_file
)
1448 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
1452 static void io_prep_async_link(struct io_kiocb
*req
)
1454 struct io_kiocb
*cur
;
1456 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
1457 struct io_ring_ctx
*ctx
= req
->ctx
;
1459 spin_lock_irq(&ctx
->timeout_lock
);
1460 io_for_each_link(cur
, req
)
1461 io_prep_async_work(cur
);
1462 spin_unlock_irq(&ctx
->timeout_lock
);
1464 io_for_each_link(cur
, req
)
1465 io_prep_async_work(cur
);
1469 static void io_queue_async_work(struct io_kiocb
*req
, bool *locked
)
1471 struct io_ring_ctx
*ctx
= req
->ctx
;
1472 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
1473 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1475 /* must not take the lock, NULL it as a precaution */
1479 BUG_ON(!tctx
->io_wq
);
1481 /* init ->work of the whole link before punting */
1482 io_prep_async_link(req
);
1485 * Not expected to happen, but if we do have a bug where this _can_
1486 * happen, catch it here and ensure the request is marked as
1487 * canceled. That will make io-wq go through the usual work cancel
1488 * procedure rather than attempt to run this request (or create a new
1491 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
1492 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
1494 trace_io_uring_queue_async_work(ctx
, io_wq_is_hashed(&req
->work
), req
,
1495 &req
->work
, req
->flags
);
1496 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
1498 io_queue_linked_timeout(link
);
1501 static void io_kill_timeout(struct io_kiocb
*req
, int status
)
1502 __must_hold(&req
->ctx
->completion_lock
)
1503 __must_hold(&req
->ctx
->timeout_lock
)
1505 struct io_timeout_data
*io
= req
->async_data
;
1507 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
1510 atomic_set(&req
->ctx
->cq_timeouts
,
1511 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
1512 list_del_init(&req
->timeout
.list
);
1513 io_fill_cqe_req(req
, status
, 0);
1514 io_put_req_deferred(req
);
1518 static void io_queue_deferred(struct io_ring_ctx
*ctx
)
1520 while (!list_empty(&ctx
->defer_list
)) {
1521 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
1522 struct io_defer_entry
, list
);
1524 if (req_need_defer(de
->req
, de
->seq
))
1526 list_del_init(&de
->list
);
1527 io_req_task_queue(de
->req
);
1532 static void io_flush_timeouts(struct io_ring_ctx
*ctx
)
1533 __must_hold(&ctx
->completion_lock
)
1535 u32 seq
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
1536 struct io_kiocb
*req
, *tmp
;
1538 spin_lock_irq(&ctx
->timeout_lock
);
1539 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
1540 u32 events_needed
, events_got
;
1542 if (io_is_timeout_noseq(req
))
1546 * Since seq can easily wrap around over time, subtract
1547 * the last seq at which timeouts were flushed before comparing.
1548 * Assuming not more than 2^31-1 events have happened since,
1549 * these subtractions won't have wrapped, so we can check if
1550 * target is in [last_seq, current_seq] by comparing the two.
1552 events_needed
= req
->timeout
.target_seq
- ctx
->cq_last_tm_flush
;
1553 events_got
= seq
- ctx
->cq_last_tm_flush
;
1554 if (events_got
< events_needed
)
1557 io_kill_timeout(req
, 0);
1559 ctx
->cq_last_tm_flush
= seq
;
1560 spin_unlock_irq(&ctx
->timeout_lock
);
1563 static void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
1565 if (ctx
->off_timeout_used
)
1566 io_flush_timeouts(ctx
);
1567 if (ctx
->drain_active
)
1568 io_queue_deferred(ctx
);
1571 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
1573 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
))
1574 __io_commit_cqring_flush(ctx
);
1575 /* order cqe stores with ring update */
1576 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
1579 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
1581 struct io_rings
*r
= ctx
->rings
;
1583 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
1586 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
1588 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
1591 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
1593 struct io_rings
*rings
= ctx
->rings
;
1594 unsigned tail
, mask
= ctx
->cq_entries
- 1;
1597 * writes to the cq entry need to come after reading head; the
1598 * control dependency is enough as we're using WRITE_ONCE to
1601 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
1604 tail
= ctx
->cached_cq_tail
++;
1605 return &rings
->cqes
[tail
& mask
];
1608 static inline bool io_should_trigger_evfd(struct io_ring_ctx
*ctx
)
1610 if (likely(!ctx
->cq_ev_fd
))
1612 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
1614 return !ctx
->eventfd_async
|| io_wq_current_is_worker();
1618 * This should only get called when at least one event has been posted.
1619 * Some applications rely on the eventfd notification count only changing
1620 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1621 * 1:1 relationship between how many times this function is called (and
1622 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1624 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
1627 * wake_up_all() may seem excessive, but io_wake_function() and
1628 * io_should_wake() handle the termination of the loop and only
1629 * wake as many waiters as we need to.
1631 if (wq_has_sleeper(&ctx
->cq_wait
))
1632 wake_up_all(&ctx
->cq_wait
);
1633 if (ctx
->sq_data
&& waitqueue_active(&ctx
->sq_data
->wait
))
1634 wake_up(&ctx
->sq_data
->wait
);
1635 if (io_should_trigger_evfd(ctx
))
1636 eventfd_signal(ctx
->cq_ev_fd
, 1);
1637 if (waitqueue_active(&ctx
->poll_wait
))
1638 wake_up_interruptible(&ctx
->poll_wait
);
1641 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx
*ctx
)
1643 /* see waitqueue_active() comment */
1646 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
1647 if (waitqueue_active(&ctx
->cq_wait
))
1648 wake_up_all(&ctx
->cq_wait
);
1650 if (io_should_trigger_evfd(ctx
))
1651 eventfd_signal(ctx
->cq_ev_fd
, 1);
1652 if (waitqueue_active(&ctx
->poll_wait
))
1653 wake_up_interruptible(&ctx
->poll_wait
);
1656 /* Returns true if there are no backlogged entries after the flush */
1657 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
1659 bool all_flushed
, posted
;
1661 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
1665 spin_lock(&ctx
->completion_lock
);
1666 while (!list_empty(&ctx
->cq_overflow_list
)) {
1667 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
1668 struct io_overflow_cqe
*ocqe
;
1672 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
1673 struct io_overflow_cqe
, list
);
1675 memcpy(cqe
, &ocqe
->cqe
, sizeof(*cqe
));
1677 io_account_cq_overflow(ctx
);
1680 list_del(&ocqe
->list
);
1684 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
1686 clear_bit(0, &ctx
->check_cq_overflow
);
1687 WRITE_ONCE(ctx
->rings
->sq_flags
,
1688 ctx
->rings
->sq_flags
& ~IORING_SQ_CQ_OVERFLOW
);
1692 io_commit_cqring(ctx
);
1693 spin_unlock(&ctx
->completion_lock
);
1695 io_cqring_ev_posted(ctx
);
1699 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
1703 if (test_bit(0, &ctx
->check_cq_overflow
)) {
1704 /* iopoll syncs against uring_lock, not completion_lock */
1705 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1706 mutex_lock(&ctx
->uring_lock
);
1707 ret
= __io_cqring_overflow_flush(ctx
, false);
1708 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
1709 mutex_unlock(&ctx
->uring_lock
);
1715 /* must to be called somewhat shortly after putting a request */
1716 static inline void io_put_task(struct task_struct
*task
, int nr
)
1718 struct io_uring_task
*tctx
= task
->io_uring
;
1720 if (likely(task
== current
)) {
1721 tctx
->cached_refs
+= nr
;
1723 percpu_counter_sub(&tctx
->inflight
, nr
);
1724 if (unlikely(atomic_read(&tctx
->in_idle
)))
1725 wake_up(&tctx
->wait
);
1726 put_task_struct_many(task
, nr
);
1730 static void io_task_refs_refill(struct io_uring_task
*tctx
)
1732 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
1734 percpu_counter_add(&tctx
->inflight
, refill
);
1735 refcount_add(refill
, ¤t
->usage
);
1736 tctx
->cached_refs
+= refill
;
1739 static inline void io_get_task_refs(int nr
)
1741 struct io_uring_task
*tctx
= current
->io_uring
;
1743 tctx
->cached_refs
-= nr
;
1744 if (unlikely(tctx
->cached_refs
< 0))
1745 io_task_refs_refill(tctx
);
1748 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
1750 struct io_uring_task
*tctx
= task
->io_uring
;
1751 unsigned int refs
= tctx
->cached_refs
;
1754 tctx
->cached_refs
= 0;
1755 percpu_counter_sub(&tctx
->inflight
, refs
);
1756 put_task_struct_many(task
, refs
);
1760 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
1761 s32 res
, u32 cflags
)
1763 struct io_overflow_cqe
*ocqe
;
1765 ocqe
= kmalloc(sizeof(*ocqe
), GFP_ATOMIC
| __GFP_ACCOUNT
);
1768 * If we're in ring overflow flush mode, or in task cancel mode,
1769 * or cannot allocate an overflow entry, then we need to drop it
1772 io_account_cq_overflow(ctx
);
1775 if (list_empty(&ctx
->cq_overflow_list
)) {
1776 set_bit(0, &ctx
->check_cq_overflow
);
1777 WRITE_ONCE(ctx
->rings
->sq_flags
,
1778 ctx
->rings
->sq_flags
| IORING_SQ_CQ_OVERFLOW
);
1781 ocqe
->cqe
.user_data
= user_data
;
1782 ocqe
->cqe
.res
= res
;
1783 ocqe
->cqe
.flags
= cflags
;
1784 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
1788 static inline bool __io_fill_cqe(struct io_ring_ctx
*ctx
, u64 user_data
,
1789 s32 res
, u32 cflags
)
1791 struct io_uring_cqe
*cqe
;
1793 trace_io_uring_complete(ctx
, user_data
, res
, cflags
);
1796 * If we can't get a cq entry, userspace overflowed the
1797 * submission (by quite a lot). Increment the overflow count in
1800 cqe
= io_get_cqe(ctx
);
1802 WRITE_ONCE(cqe
->user_data
, user_data
);
1803 WRITE_ONCE(cqe
->res
, res
);
1804 WRITE_ONCE(cqe
->flags
, cflags
);
1807 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
);
1810 static noinline
void io_fill_cqe_req(struct io_kiocb
*req
, s32 res
, u32 cflags
)
1812 __io_fill_cqe(req
->ctx
, req
->user_data
, res
, cflags
);
1815 static noinline
bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
,
1816 s32 res
, u32 cflags
)
1819 return __io_fill_cqe(ctx
, user_data
, res
, cflags
);
1822 static void io_req_complete_post(struct io_kiocb
*req
, s32 res
,
1825 struct io_ring_ctx
*ctx
= req
->ctx
;
1827 spin_lock(&ctx
->completion_lock
);
1828 __io_fill_cqe(ctx
, req
->user_data
, res
, cflags
);
1830 * If we're the last reference to this request, add to our locked
1833 if (req_ref_put_and_test(req
)) {
1834 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
1835 if (req
->flags
& IO_DISARM_MASK
)
1836 io_disarm_next(req
);
1838 io_req_task_queue(req
->link
);
1842 io_dismantle_req(req
);
1843 io_put_task(req
->task
, 1);
1844 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
1845 ctx
->locked_free_nr
++;
1847 if (!percpu_ref_tryget(&ctx
->refs
))
1850 io_commit_cqring(ctx
);
1851 spin_unlock(&ctx
->completion_lock
);
1854 io_cqring_ev_posted(ctx
);
1855 percpu_ref_put(&ctx
->refs
);
1859 static inline bool io_req_needs_clean(struct io_kiocb
*req
)
1861 return req
->flags
& IO_REQ_CLEAN_FLAGS
;
1864 static inline void io_req_complete_state(struct io_kiocb
*req
, s32 res
,
1867 if (io_req_needs_clean(req
))
1870 req
->compl.cflags
= cflags
;
1871 req
->flags
|= REQ_F_COMPLETE_INLINE
;
1874 static inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
,
1875 s32 res
, u32 cflags
)
1877 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1878 io_req_complete_state(req
, res
, cflags
);
1880 io_req_complete_post(req
, res
, cflags
);
1883 static inline void io_req_complete(struct io_kiocb
*req
, s32 res
)
1885 __io_req_complete(req
, 0, res
, 0);
1888 static void io_req_complete_failed(struct io_kiocb
*req
, s32 res
)
1891 io_req_complete_post(req
, res
, 0);
1894 static void io_req_complete_fail_submit(struct io_kiocb
*req
)
1897 * We don't submit, fail them all, for that replace hardlinks with
1898 * normal links. Extra REQ_F_LINK is tolerated.
1900 req
->flags
&= ~REQ_F_HARDLINK
;
1901 req
->flags
|= REQ_F_LINK
;
1902 io_req_complete_failed(req
, req
->result
);
1906 * Don't initialise the fields below on every allocation, but do that in
1907 * advance and keep them valid across allocations.
1909 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1913 req
->async_data
= NULL
;
1914 /* not necessary, but safer to zero */
1918 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1919 struct io_submit_state
*state
)
1921 spin_lock(&ctx
->completion_lock
);
1922 list_splice_init(&ctx
->locked_free_list
, &state
->free_list
);
1923 ctx
->locked_free_nr
= 0;
1924 spin_unlock(&ctx
->completion_lock
);
1927 /* Returns true IFF there are requests in the cache */
1928 static bool io_flush_cached_reqs(struct io_ring_ctx
*ctx
)
1930 struct io_submit_state
*state
= &ctx
->submit_state
;
1934 * If we have more than a batch's worth of requests in our IRQ side
1935 * locked cache, grab the lock and move them over to our submission
1938 if (READ_ONCE(ctx
->locked_free_nr
) > IO_COMPL_BATCH
)
1939 io_flush_cached_locked_reqs(ctx
, state
);
1941 nr
= state
->free_reqs
;
1942 while (!list_empty(&state
->free_list
)) {
1943 struct io_kiocb
*req
= list_first_entry(&state
->free_list
,
1944 struct io_kiocb
, inflight_entry
);
1946 list_del(&req
->inflight_entry
);
1947 state
->reqs
[nr
++] = req
;
1948 if (nr
== ARRAY_SIZE(state
->reqs
))
1952 state
->free_reqs
= nr
;
1957 * A request might get retired back into the request caches even before opcode
1958 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1959 * Because of that, io_alloc_req() should be called only under ->uring_lock
1960 * and with extra caution to not get a request that is still worked on.
1962 static struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
1963 __must_hold(&ctx
->uring_lock
)
1965 struct io_submit_state
*state
= &ctx
->submit_state
;
1966 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1969 BUILD_BUG_ON(ARRAY_SIZE(state
->reqs
) < IO_REQ_ALLOC_BATCH
);
1971 if (likely(state
->free_reqs
|| io_flush_cached_reqs(ctx
)))
1974 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, IO_REQ_ALLOC_BATCH
,
1978 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1979 * retry single alloc to be on the safe side.
1981 if (unlikely(ret
<= 0)) {
1982 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1983 if (!state
->reqs
[0])
1988 for (i
= 0; i
< ret
; i
++)
1989 io_preinit_req(state
->reqs
[i
], ctx
);
1990 state
->free_reqs
= ret
;
1993 return state
->reqs
[state
->free_reqs
];
1996 static inline void io_put_file(struct file
*file
)
2002 static void io_dismantle_req(struct io_kiocb
*req
)
2004 unsigned int flags
= req
->flags
;
2006 if (io_req_needs_clean(req
))
2008 if (!(flags
& REQ_F_FIXED_FILE
))
2009 io_put_file(req
->file
);
2010 if (req
->fixed_rsrc_refs
)
2011 percpu_ref_put(req
->fixed_rsrc_refs
);
2012 if (req
->async_data
) {
2013 kfree(req
->async_data
);
2014 req
->async_data
= NULL
;
2018 static void __io_free_req(struct io_kiocb
*req
)
2020 struct io_ring_ctx
*ctx
= req
->ctx
;
2022 io_dismantle_req(req
);
2023 io_put_task(req
->task
, 1);
2025 spin_lock(&ctx
->completion_lock
);
2026 list_add(&req
->inflight_entry
, &ctx
->locked_free_list
);
2027 ctx
->locked_free_nr
++;
2028 spin_unlock(&ctx
->completion_lock
);
2030 percpu_ref_put(&ctx
->refs
);
2033 static inline void io_remove_next_linked(struct io_kiocb
*req
)
2035 struct io_kiocb
*nxt
= req
->link
;
2037 req
->link
= nxt
->link
;
2041 static bool io_kill_linked_timeout(struct io_kiocb
*req
)
2042 __must_hold(&req
->ctx
->completion_lock
)
2043 __must_hold(&req
->ctx
->timeout_lock
)
2045 struct io_kiocb
*link
= req
->link
;
2047 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2048 struct io_timeout_data
*io
= link
->async_data
;
2050 io_remove_next_linked(req
);
2051 link
->timeout
.head
= NULL
;
2052 if (hrtimer_try_to_cancel(&io
->timer
) != -1) {
2053 list_del(&link
->timeout
.list
);
2054 io_fill_cqe_req(link
, -ECANCELED
, 0);
2055 io_put_req_deferred(link
);
2062 static void io_fail_links(struct io_kiocb
*req
)
2063 __must_hold(&req
->ctx
->completion_lock
)
2065 struct io_kiocb
*nxt
, *link
= req
->link
;
2069 long res
= -ECANCELED
;
2071 if (link
->flags
& REQ_F_FAIL
)
2077 trace_io_uring_fail_link(req
, link
);
2078 io_fill_cqe_req(link
, res
, 0);
2079 io_put_req_deferred(link
);
2084 static bool io_disarm_next(struct io_kiocb
*req
)
2085 __must_hold(&req
->ctx
->completion_lock
)
2087 bool posted
= false;
2089 if (req
->flags
& REQ_F_ARM_LTIMEOUT
) {
2090 struct io_kiocb
*link
= req
->link
;
2092 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
2093 if (link
&& link
->opcode
== IORING_OP_LINK_TIMEOUT
) {
2094 io_remove_next_linked(req
);
2095 io_fill_cqe_req(link
, -ECANCELED
, 0);
2096 io_put_req_deferred(link
);
2099 } else if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
2100 struct io_ring_ctx
*ctx
= req
->ctx
;
2102 spin_lock_irq(&ctx
->timeout_lock
);
2103 posted
= io_kill_linked_timeout(req
);
2104 spin_unlock_irq(&ctx
->timeout_lock
);
2106 if (unlikely((req
->flags
& REQ_F_FAIL
) &&
2107 !(req
->flags
& REQ_F_HARDLINK
))) {
2108 posted
|= (req
->link
!= NULL
);
2114 static struct io_kiocb
*__io_req_find_next(struct io_kiocb
*req
)
2116 struct io_kiocb
*nxt
;
2119 * If LINK is set, we have dependent requests in this chain. If we
2120 * didn't fail this request, queue the first one up, moving any other
2121 * dependencies to the next request. In case of failure, fail the rest
2124 if (req
->flags
& IO_DISARM_MASK
) {
2125 struct io_ring_ctx
*ctx
= req
->ctx
;
2128 spin_lock(&ctx
->completion_lock
);
2129 posted
= io_disarm_next(req
);
2131 io_commit_cqring(req
->ctx
);
2132 spin_unlock(&ctx
->completion_lock
);
2134 io_cqring_ev_posted(ctx
);
2141 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
2143 if (likely(!(req
->flags
& (REQ_F_LINK
|REQ_F_HARDLINK
))))
2145 return __io_req_find_next(req
);
2148 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, bool *locked
)
2153 if (ctx
->submit_state
.compl_nr
)
2154 io_submit_flush_completions(ctx
);
2155 mutex_unlock(&ctx
->uring_lock
);
2158 percpu_ref_put(&ctx
->refs
);
2161 static void tctx_task_work(struct callback_head
*cb
)
2163 bool locked
= false;
2164 struct io_ring_ctx
*ctx
= NULL
;
2165 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
2169 struct io_wq_work_node
*node
;
2171 if (!tctx
->task_list
.first
&& locked
&& ctx
->submit_state
.compl_nr
)
2172 io_submit_flush_completions(ctx
);
2174 spin_lock_irq(&tctx
->task_lock
);
2175 node
= tctx
->task_list
.first
;
2176 INIT_WQ_LIST(&tctx
->task_list
);
2178 tctx
->task_running
= false;
2179 spin_unlock_irq(&tctx
->task_lock
);
2184 struct io_wq_work_node
*next
= node
->next
;
2185 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
2188 if (req
->ctx
!= ctx
) {
2189 ctx_flush_and_put(ctx
, &locked
);
2191 /* if not contended, grab and improve batching */
2192 locked
= mutex_trylock(&ctx
->uring_lock
);
2193 percpu_ref_get(&ctx
->refs
);
2195 req
->io_task_work
.func(req
, &locked
);
2202 ctx_flush_and_put(ctx
, &locked
);
2204 /* relaxed read is enough as only the task itself sets ->in_idle */
2205 if (unlikely(atomic_read(&tctx
->in_idle
)))
2206 io_uring_drop_tctx_refs(current
);
2209 static void io_req_task_work_add(struct io_kiocb
*req
)
2211 struct task_struct
*tsk
= req
->task
;
2212 struct io_uring_task
*tctx
= tsk
->io_uring
;
2213 enum task_work_notify_mode notify
;
2214 struct io_wq_work_node
*node
;
2215 unsigned long flags
;
2218 WARN_ON_ONCE(!tctx
);
2220 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2221 wq_list_add_tail(&req
->io_task_work
.node
, &tctx
->task_list
);
2222 running
= tctx
->task_running
;
2224 tctx
->task_running
= true;
2225 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2227 /* task_work already pending, we're done */
2232 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2233 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2234 * processing task_work. There's no reliable way to tell if TWA_RESUME
2237 notify
= (req
->ctx
->flags
& IORING_SETUP_SQPOLL
) ? TWA_NONE
: TWA_SIGNAL
;
2238 if (!task_work_add(tsk
, &tctx
->task_work
, notify
)) {
2239 wake_up_process(tsk
);
2243 spin_lock_irqsave(&tctx
->task_lock
, flags
);
2244 tctx
->task_running
= false;
2245 node
= tctx
->task_list
.first
;
2246 INIT_WQ_LIST(&tctx
->task_list
);
2247 spin_unlock_irqrestore(&tctx
->task_lock
, flags
);
2250 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
2252 if (llist_add(&req
->io_task_work
.fallback_node
,
2253 &req
->ctx
->fallback_llist
))
2254 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
2258 static void io_req_task_cancel(struct io_kiocb
*req
, bool *locked
)
2260 struct io_ring_ctx
*ctx
= req
->ctx
;
2262 /* not needed for normal modes, but SQPOLL depends on it */
2263 io_tw_lock(ctx
, locked
);
2264 io_req_complete_failed(req
, req
->result
);
2267 static void io_req_task_submit(struct io_kiocb
*req
, bool *locked
)
2269 struct io_ring_ctx
*ctx
= req
->ctx
;
2271 io_tw_lock(ctx
, locked
);
2272 /* req->task == current here, checking PF_EXITING is safe */
2273 if (likely(!(req
->task
->flags
& PF_EXITING
)))
2274 __io_queue_sqe(req
);
2276 io_req_complete_failed(req
, -EFAULT
);
2279 static void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
2282 req
->io_task_work
.func
= io_req_task_cancel
;
2283 io_req_task_work_add(req
);
2286 static void io_req_task_queue(struct io_kiocb
*req
)
2288 req
->io_task_work
.func
= io_req_task_submit
;
2289 io_req_task_work_add(req
);
2292 static void io_req_task_queue_reissue(struct io_kiocb
*req
)
2294 req
->io_task_work
.func
= io_queue_async_work
;
2295 io_req_task_work_add(req
);
2298 static inline void io_queue_next(struct io_kiocb
*req
)
2300 struct io_kiocb
*nxt
= io_req_find_next(req
);
2303 io_req_task_queue(nxt
);
2306 static void io_free_req(struct io_kiocb
*req
)
2312 static void io_free_req_work(struct io_kiocb
*req
, bool *locked
)
2318 struct task_struct
*task
;
2323 static inline void io_init_req_batch(struct req_batch
*rb
)
2330 static void io_req_free_batch_finish(struct io_ring_ctx
*ctx
,
2331 struct req_batch
*rb
)
2334 percpu_ref_put_many(&ctx
->refs
, rb
->ctx_refs
);
2336 io_put_task(rb
->task
, rb
->task_refs
);
2339 static void io_req_free_batch(struct req_batch
*rb
, struct io_kiocb
*req
,
2340 struct io_submit_state
*state
)
2343 io_dismantle_req(req
);
2345 if (req
->task
!= rb
->task
) {
2347 io_put_task(rb
->task
, rb
->task_refs
);
2348 rb
->task
= req
->task
;
2354 if (state
->free_reqs
!= ARRAY_SIZE(state
->reqs
))
2355 state
->reqs
[state
->free_reqs
++] = req
;
2357 list_add(&req
->inflight_entry
, &state
->free_list
);
2360 static void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
2361 __must_hold(&ctx
->uring_lock
)
2363 struct io_submit_state
*state
= &ctx
->submit_state
;
2364 int i
, nr
= state
->compl_nr
;
2365 struct req_batch rb
;
2367 spin_lock(&ctx
->completion_lock
);
2368 for (i
= 0; i
< nr
; i
++) {
2369 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2371 __io_fill_cqe(ctx
, req
->user_data
, req
->result
,
2374 io_commit_cqring(ctx
);
2375 spin_unlock(&ctx
->completion_lock
);
2376 io_cqring_ev_posted(ctx
);
2378 io_init_req_batch(&rb
);
2379 for (i
= 0; i
< nr
; i
++) {
2380 struct io_kiocb
*req
= state
->compl_reqs
[i
];
2382 if (req_ref_put_and_test(req
))
2383 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2386 io_req_free_batch_finish(ctx
, &rb
);
2387 state
->compl_nr
= 0;
2391 * Drop reference to request, return next in chain (if there is one) if this
2392 * was the last reference to this request.
2394 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
2396 struct io_kiocb
*nxt
= NULL
;
2398 if (req_ref_put_and_test(req
)) {
2399 nxt
= io_req_find_next(req
);
2405 static inline void io_put_req(struct io_kiocb
*req
)
2407 if (req_ref_put_and_test(req
))
2411 static inline void io_put_req_deferred(struct io_kiocb
*req
)
2413 if (req_ref_put_and_test(req
)) {
2414 req
->io_task_work
.func
= io_free_req_work
;
2415 io_req_task_work_add(req
);
2419 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
2421 /* See comment at the top of this file */
2423 return __io_cqring_events(ctx
);
2426 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
2428 struct io_rings
*rings
= ctx
->rings
;
2430 /* make sure SQ entry isn't read before tail */
2431 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
2434 static unsigned int io_put_kbuf(struct io_kiocb
*req
, struct io_buffer
*kbuf
)
2436 unsigned int cflags
;
2438 cflags
= kbuf
->bid
<< IORING_CQE_BUFFER_SHIFT
;
2439 cflags
|= IORING_CQE_F_BUFFER
;
2440 req
->flags
&= ~REQ_F_BUFFER_SELECTED
;
2445 static inline unsigned int io_put_rw_kbuf(struct io_kiocb
*req
)
2447 struct io_buffer
*kbuf
;
2449 if (likely(!(req
->flags
& REQ_F_BUFFER_SELECTED
)))
2451 kbuf
= (struct io_buffer
*) (unsigned long) req
->rw
.addr
;
2452 return io_put_kbuf(req
, kbuf
);
2455 static inline bool io_run_task_work(void)
2457 if (test_thread_flag(TIF_NOTIFY_SIGNAL
) || current
->task_works
) {
2458 __set_current_state(TASK_RUNNING
);
2459 tracehook_notify_signal();
2467 * Find and free completed poll iocbs
2469 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2470 struct list_head
*done
)
2472 struct req_batch rb
;
2473 struct io_kiocb
*req
;
2475 /* order with ->result store in io_complete_rw_iopoll() */
2478 io_init_req_batch(&rb
);
2479 while (!list_empty(done
)) {
2480 req
= list_first_entry(done
, struct io_kiocb
, inflight_entry
);
2481 list_del(&req
->inflight_entry
);
2483 io_fill_cqe_req(req
, req
->result
, io_put_rw_kbuf(req
));
2486 if (req_ref_put_and_test(req
))
2487 io_req_free_batch(&rb
, req
, &ctx
->submit_state
);
2490 io_commit_cqring(ctx
);
2491 io_cqring_ev_posted_iopoll(ctx
);
2492 io_req_free_batch_finish(ctx
, &rb
);
2495 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
2498 struct io_kiocb
*req
, *tmp
;
2503 * Only spin for completions if we don't have multiple devices hanging
2504 * off our complete list, and we're under the requested amount.
2506 spin
= !ctx
->poll_multi_queue
&& *nr_events
< min
;
2508 list_for_each_entry_safe(req
, tmp
, &ctx
->iopoll_list
, inflight_entry
) {
2509 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2513 * Move completed and retryable entries to our local lists.
2514 * If we find a request that requires polling, break out
2515 * and complete those lists first, if we have entries there.
2517 if (READ_ONCE(req
->iopoll_completed
)) {
2518 list_move_tail(&req
->inflight_entry
, &done
);
2521 if (!list_empty(&done
))
2524 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
2525 if (unlikely(ret
< 0))
2530 /* iopoll may have completed current req */
2531 if (READ_ONCE(req
->iopoll_completed
))
2532 list_move_tail(&req
->inflight_entry
, &done
);
2535 if (!list_empty(&done
))
2536 io_iopoll_complete(ctx
, nr_events
, &done
);
2542 * We can't just wait for polled events to come to us, we have to actively
2543 * find and complete them.
2545 static void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
2547 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
2550 mutex_lock(&ctx
->uring_lock
);
2551 while (!list_empty(&ctx
->iopoll_list
)) {
2552 unsigned int nr_events
= 0;
2554 io_do_iopoll(ctx
, &nr_events
, 0);
2556 /* let it sleep and repeat later if can't complete a request */
2560 * Ensure we allow local-to-the-cpu processing to take place,
2561 * in this case we need to ensure that we reap all events.
2562 * Also let task_work, etc. to progress by releasing the mutex
2564 if (need_resched()) {
2565 mutex_unlock(&ctx
->uring_lock
);
2567 mutex_lock(&ctx
->uring_lock
);
2570 mutex_unlock(&ctx
->uring_lock
);
2573 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
2575 unsigned int nr_events
= 0;
2579 * We disallow the app entering submit/complete with polling, but we
2580 * still need to lock the ring to prevent racing with polled issue
2581 * that got punted to a workqueue.
2583 mutex_lock(&ctx
->uring_lock
);
2585 * Don't enter poll loop if we already have events pending.
2586 * If we do, we can potentially be spinning for commands that
2587 * already triggered a CQE (eg in error).
2589 if (test_bit(0, &ctx
->check_cq_overflow
))
2590 __io_cqring_overflow_flush(ctx
, false);
2591 if (io_cqring_events(ctx
))
2595 * If a submit got punted to a workqueue, we can have the
2596 * application entering polling for a command before it gets
2597 * issued. That app will hold the uring_lock for the duration
2598 * of the poll right here, so we need to take a breather every
2599 * now and then to ensure that the issue has a chance to add
2600 * the poll to the issued list. Otherwise we can spin here
2601 * forever, while the workqueue is stuck trying to acquire the
2604 if (list_empty(&ctx
->iopoll_list
)) {
2605 u32 tail
= ctx
->cached_cq_tail
;
2607 mutex_unlock(&ctx
->uring_lock
);
2609 mutex_lock(&ctx
->uring_lock
);
2611 /* some requests don't go through iopoll_list */
2612 if (tail
!= ctx
->cached_cq_tail
||
2613 list_empty(&ctx
->iopoll_list
))
2616 ret
= io_do_iopoll(ctx
, &nr_events
, min
);
2617 } while (!ret
&& nr_events
< min
&& !need_resched());
2619 mutex_unlock(&ctx
->uring_lock
);
2623 static void kiocb_end_write(struct io_kiocb
*req
)
2626 * Tell lockdep we inherited freeze protection from submission
2629 if (req
->flags
& REQ_F_ISREG
) {
2630 struct super_block
*sb
= file_inode(req
->file
)->i_sb
;
2632 __sb_writers_acquired(sb
, SB_FREEZE_WRITE
);
2638 static bool io_resubmit_prep(struct io_kiocb
*req
)
2640 struct io_async_rw
*rw
= req
->async_data
;
2643 return !io_req_prep_async(req
);
2644 iov_iter_restore(&rw
->iter
, &rw
->iter_state
);
2648 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2650 umode_t mode
= file_inode(req
->file
)->i_mode
;
2651 struct io_ring_ctx
*ctx
= req
->ctx
;
2653 if (!S_ISBLK(mode
) && !S_ISREG(mode
))
2655 if ((req
->flags
& REQ_F_NOWAIT
) || (io_wq_current_is_worker() &&
2656 !(ctx
->flags
& IORING_SETUP_IOPOLL
)))
2659 * If ref is dying, we might be running poll reap from the exit work.
2660 * Don't attempt to reissue from that path, just let it fail with
2663 if (percpu_ref_is_dying(&ctx
->refs
))
2666 * Play it safe and assume not safe to re-import and reissue if we're
2667 * not in the original thread group (or in task context).
2669 if (!same_thread_group(req
->task
, current
) || !in_task())
2674 static bool io_resubmit_prep(struct io_kiocb
*req
)
2678 static bool io_rw_should_reissue(struct io_kiocb
*req
)
2684 static bool __io_complete_rw_common(struct io_kiocb
*req
, long res
)
2686 if (req
->rw
.kiocb
.ki_flags
& IOCB_WRITE
) {
2687 kiocb_end_write(req
);
2688 fsnotify_modify(req
->file
);
2690 fsnotify_access(req
->file
);
2692 if (res
!= req
->result
) {
2693 if ((res
== -EAGAIN
|| res
== -EOPNOTSUPP
) &&
2694 io_rw_should_reissue(req
)) {
2695 req
->flags
|= REQ_F_REISSUE
;
2704 static inline int io_fixup_rw_res(struct io_kiocb
*req
, unsigned res
)
2706 struct io_async_rw
*io
= req
->async_data
;
2708 /* add previously done IO, if any */
2709 if (io
&& io
->bytes_done
> 0) {
2711 res
= io
->bytes_done
;
2713 res
+= io
->bytes_done
;
2718 static void io_req_task_complete(struct io_kiocb
*req
, bool *locked
)
2720 unsigned int cflags
= io_put_rw_kbuf(req
);
2721 int res
= req
->result
;
2724 struct io_ring_ctx
*ctx
= req
->ctx
;
2725 struct io_submit_state
*state
= &ctx
->submit_state
;
2727 io_req_complete_state(req
, res
, cflags
);
2728 state
->compl_reqs
[state
->compl_nr
++] = req
;
2729 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
2730 io_submit_flush_completions(ctx
);
2732 io_req_complete_post(req
, res
, cflags
);
2736 static void __io_complete_rw(struct io_kiocb
*req
, long res
, long res2
,
2737 unsigned int issue_flags
)
2739 if (__io_complete_rw_common(req
, res
))
2741 __io_req_complete(req
, issue_flags
, io_fixup_rw_res(req
, res
), io_put_rw_kbuf(req
));
2744 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
2746 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2748 if (__io_complete_rw_common(req
, res
))
2750 req
->result
= io_fixup_rw_res(req
, res
);
2751 req
->io_task_work
.func
= io_req_task_complete
;
2752 io_req_task_work_add(req
);
2755 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
2757 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2759 if (kiocb
->ki_flags
& IOCB_WRITE
)
2760 kiocb_end_write(req
);
2761 if (unlikely(res
!= req
->result
)) {
2762 if (res
== -EAGAIN
&& io_rw_should_reissue(req
)) {
2763 req
->flags
|= REQ_F_REISSUE
;
2768 WRITE_ONCE(req
->result
, res
);
2769 /* order with io_iopoll_complete() checking ->result */
2771 WRITE_ONCE(req
->iopoll_completed
, 1);
2775 * After the iocb has been issued, it's safe to be found on the poll list.
2776 * Adding the kiocb to the list AFTER submission ensures that we don't
2777 * find it from a io_do_iopoll() thread before the issuer is done
2778 * accessing the kiocb cookie.
2780 static void io_iopoll_req_issued(struct io_kiocb
*req
)
2782 struct io_ring_ctx
*ctx
= req
->ctx
;
2783 const bool in_async
= io_wq_current_is_worker();
2785 /* workqueue context doesn't hold uring_lock, grab it now */
2786 if (unlikely(in_async
))
2787 mutex_lock(&ctx
->uring_lock
);
2790 * Track whether we have multiple files in our lists. This will impact
2791 * how we do polling eventually, not spinning if we're on potentially
2792 * different devices.
2794 if (list_empty(&ctx
->iopoll_list
)) {
2795 ctx
->poll_multi_queue
= false;
2796 } else if (!ctx
->poll_multi_queue
) {
2797 struct io_kiocb
*list_req
;
2798 unsigned int queue_num0
, queue_num1
;
2800 list_req
= list_first_entry(&ctx
->iopoll_list
, struct io_kiocb
,
2803 if (list_req
->file
!= req
->file
) {
2804 ctx
->poll_multi_queue
= true;
2806 queue_num0
= blk_qc_t_to_queue_num(list_req
->rw
.kiocb
.ki_cookie
);
2807 queue_num1
= blk_qc_t_to_queue_num(req
->rw
.kiocb
.ki_cookie
);
2808 if (queue_num0
!= queue_num1
)
2809 ctx
->poll_multi_queue
= true;
2814 * For fast devices, IO may have already completed. If it has, add
2815 * it to the front so we find it first.
2817 if (READ_ONCE(req
->iopoll_completed
))
2818 list_add(&req
->inflight_entry
, &ctx
->iopoll_list
);
2820 list_add_tail(&req
->inflight_entry
, &ctx
->iopoll_list
);
2822 if (unlikely(in_async
)) {
2824 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2825 * in sq thread task context or in io worker task context. If
2826 * current task context is sq thread, we don't need to check
2827 * whether should wake up sq thread.
2829 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
2830 wq_has_sleeper(&ctx
->sq_data
->wait
))
2831 wake_up(&ctx
->sq_data
->wait
);
2833 mutex_unlock(&ctx
->uring_lock
);
2837 static bool io_bdev_nowait(struct block_device
*bdev
)
2839 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
2843 * If we tracked the file through the SCM inflight mechanism, we could support
2844 * any file. For now, just ensure that anything potentially problematic is done
2847 static bool __io_file_supports_nowait(struct file
*file
, int rw
)
2849 umode_t mode
= file_inode(file
)->i_mode
;
2851 if (S_ISBLK(mode
)) {
2852 if (IS_ENABLED(CONFIG_BLOCK
) &&
2853 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
2859 if (S_ISREG(mode
)) {
2860 if (IS_ENABLED(CONFIG_BLOCK
) &&
2861 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
2862 file
->f_op
!= &io_uring_fops
)
2867 /* any ->read/write should understand O_NONBLOCK */
2868 if (file
->f_flags
& O_NONBLOCK
)
2871 if (!(file
->f_mode
& FMODE_NOWAIT
))
2875 return file
->f_op
->read_iter
!= NULL
;
2877 return file
->f_op
->write_iter
!= NULL
;
2880 static bool io_file_supports_nowait(struct io_kiocb
*req
, int rw
)
2882 if (rw
== READ
&& (req
->flags
& REQ_F_NOWAIT_READ
))
2884 else if (rw
== WRITE
&& (req
->flags
& REQ_F_NOWAIT_WRITE
))
2887 return __io_file_supports_nowait(req
->file
, rw
);
2890 static int io_prep_rw(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
2893 struct io_ring_ctx
*ctx
= req
->ctx
;
2894 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
2895 struct file
*file
= req
->file
;
2899 if (!io_req_ffs_set(req
) && S_ISREG(file_inode(file
)->i_mode
))
2900 req
->flags
|= REQ_F_ISREG
;
2902 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
2903 if (kiocb
->ki_pos
== -1) {
2904 if (!(file
->f_mode
& FMODE_STREAM
)) {
2905 req
->flags
|= REQ_F_CUR_POS
;
2906 kiocb
->ki_pos
= file
->f_pos
;
2911 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
2912 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
2913 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
2918 * If the file is marked O_NONBLOCK, still allow retry for it if it
2919 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2920 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2922 if ((kiocb
->ki_flags
& IOCB_NOWAIT
) ||
2923 ((file
->f_flags
& O_NONBLOCK
) && !io_file_supports_nowait(req
, rw
)))
2924 req
->flags
|= REQ_F_NOWAIT
;
2926 ioprio
= READ_ONCE(sqe
->ioprio
);
2928 ret
= ioprio_check_cap(ioprio
);
2932 kiocb
->ki_ioprio
= ioprio
;
2934 kiocb
->ki_ioprio
= get_current_ioprio();
2936 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2937 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
2938 !kiocb
->ki_filp
->f_op
->iopoll
)
2941 kiocb
->ki_flags
|= IOCB_HIPRI
| IOCB_ALLOC_CACHE
;
2942 kiocb
->ki_complete
= io_complete_rw_iopoll
;
2943 req
->iopoll_completed
= 0;
2945 if (kiocb
->ki_flags
& IOCB_HIPRI
)
2947 kiocb
->ki_complete
= io_complete_rw
;
2950 /* used for fixed read/write too - just read unconditionally */
2951 req
->buf_index
= READ_ONCE(sqe
->buf_index
);
2954 if (req
->opcode
== IORING_OP_READ_FIXED
||
2955 req
->opcode
== IORING_OP_WRITE_FIXED
) {
2956 struct io_ring_ctx
*ctx
= req
->ctx
;
2959 if (unlikely(req
->buf_index
>= ctx
->nr_user_bufs
))
2961 index
= array_index_nospec(req
->buf_index
, ctx
->nr_user_bufs
);
2962 req
->imu
= ctx
->user_bufs
[index
];
2963 io_req_set_rsrc_node(req
);
2966 req
->rw
.addr
= READ_ONCE(sqe
->addr
);
2967 req
->rw
.len
= READ_ONCE(sqe
->len
);
2971 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
2977 case -ERESTARTNOINTR
:
2978 case -ERESTARTNOHAND
:
2979 case -ERESTART_RESTARTBLOCK
:
2981 * We can't just restart the syscall, since previously
2982 * submitted sqes may already be in progress. Just fail this
2988 kiocb
->ki_complete(kiocb
, ret
, 0);
2992 static void kiocb_done(struct kiocb
*kiocb
, ssize_t ret
,
2993 unsigned int issue_flags
)
2995 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
.kiocb
);
2997 if (req
->flags
& REQ_F_CUR_POS
)
2998 req
->file
->f_pos
= kiocb
->ki_pos
;
2999 if (ret
>= 0 && (kiocb
->ki_complete
== io_complete_rw
))
3000 __io_complete_rw(req
, ret
, 0, issue_flags
);
3002 io_rw_done(kiocb
, ret
);
3004 if (req
->flags
& REQ_F_REISSUE
) {
3005 req
->flags
&= ~REQ_F_REISSUE
;
3006 if (io_resubmit_prep(req
)) {
3007 io_req_task_queue_reissue(req
);
3009 unsigned int cflags
= io_put_rw_kbuf(req
);
3010 struct io_ring_ctx
*ctx
= req
->ctx
;
3012 ret
= io_fixup_rw_res(req
, ret
);
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
;
3609 req
->result
= iov_iter_count(iter
);
3611 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3612 * we get -EIOCBQUEUED, then we'll get a notification when the
3613 * desired page gets unlocked. We can also get a partial read
3614 * here, and if we do, then just retry at the new offset.
3616 ret
= io_iter_do_read(req
, iter
);
3617 if (ret
== -EIOCBQUEUED
)
3619 /* we got some bytes, but not all. retry. */
3620 kiocb
->ki_flags
&= ~IOCB_WAITQ
;
3621 iov_iter_restore(iter
, state
);
3624 kiocb_done(kiocb
, ret
, issue_flags
);
3626 /* it's faster to check here then delegate to kfree */
3632 static int io_write_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
3634 if (unlikely(!(req
->file
->f_mode
& FMODE_WRITE
)))
3636 return io_prep_rw(req
, sqe
, WRITE
);
3639 static int io_write(struct io_kiocb
*req
, unsigned int issue_flags
)
3641 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
3642 struct kiocb
*kiocb
= &req
->rw
.kiocb
;
3643 struct iov_iter __iter
, *iter
= &__iter
;
3644 struct io_async_rw
*rw
= req
->async_data
;
3645 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
3646 struct iov_iter_state __state
, *state
;
3651 state
= &rw
->iter_state
;
3652 iov_iter_restore(iter
, state
);
3655 ret
= io_import_iovec(WRITE
, req
, &iovec
, iter
, !force_nonblock
);
3659 iov_iter_save_state(iter
, state
);
3661 req
->result
= iov_iter_count(iter
);
3663 /* Ensure we clear previously set non-block flag */
3664 if (!force_nonblock
)
3665 kiocb
->ki_flags
&= ~IOCB_NOWAIT
;
3667 kiocb
->ki_flags
|= IOCB_NOWAIT
;
3669 /* If the file doesn't support async, just async punt */
3670 if (force_nonblock
&& !io_file_supports_nowait(req
, WRITE
))
3673 /* file path doesn't support NOWAIT for non-direct_IO */
3674 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
) &&
3675 (req
->flags
& REQ_F_ISREG
))
3678 ret
= rw_verify_area(WRITE
, req
->file
, io_kiocb_ppos(kiocb
), req
->result
);
3683 * Open-code file_start_write here to grab freeze protection,
3684 * which will be released by another thread in
3685 * io_complete_rw(). Fool lockdep by telling it the lock got
3686 * released so that it doesn't complain about the held lock when
3687 * we return to userspace.
3689 if (req
->flags
& REQ_F_ISREG
) {
3690 sb_start_write(file_inode(req
->file
)->i_sb
);
3691 __sb_writers_release(file_inode(req
->file
)->i_sb
,
3694 kiocb
->ki_flags
|= IOCB_WRITE
;
3696 if (req
->file
->f_op
->write_iter
)
3697 ret2
= call_write_iter(req
->file
, kiocb
, iter
);
3698 else if (req
->file
->f_op
->write
)
3699 ret2
= loop_rw_iter(WRITE
, req
, iter
);
3703 if (req
->flags
& REQ_F_REISSUE
) {
3704 req
->flags
&= ~REQ_F_REISSUE
;
3709 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3710 * retry them without IOCB_NOWAIT.
3712 if (ret2
== -EOPNOTSUPP
&& (kiocb
->ki_flags
& IOCB_NOWAIT
))
3714 /* no retry on NONBLOCK nor RWF_NOWAIT */
3715 if (ret2
== -EAGAIN
&& (req
->flags
& REQ_F_NOWAIT
))
3717 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
3718 /* IOPOLL retry should happen for io-wq threads */
3719 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && ret2
== -EAGAIN
)
3722 kiocb_done(kiocb
, ret2
, issue_flags
);
3725 iov_iter_restore(iter
, state
);
3726 ret
= io_setup_async_rw(req
, iovec
, inline_vecs
, iter
, false);
3728 if (kiocb
->ki_flags
& IOCB_WRITE
)
3729 kiocb_end_write(req
);
3735 /* it's reportedly faster than delegating the null check to kfree() */
3741 static int io_renameat_prep(struct io_kiocb
*req
,
3742 const struct io_uring_sqe
*sqe
)
3744 struct io_rename
*ren
= &req
->rename
;
3745 const char __user
*oldf
, *newf
;
3747 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3749 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3751 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3754 ren
->old_dfd
= READ_ONCE(sqe
->fd
);
3755 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3756 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3757 ren
->new_dfd
= READ_ONCE(sqe
->len
);
3758 ren
->flags
= READ_ONCE(sqe
->rename_flags
);
3760 ren
->oldpath
= getname(oldf
);
3761 if (IS_ERR(ren
->oldpath
))
3762 return PTR_ERR(ren
->oldpath
);
3764 ren
->newpath
= getname(newf
);
3765 if (IS_ERR(ren
->newpath
)) {
3766 putname(ren
->oldpath
);
3767 return PTR_ERR(ren
->newpath
);
3770 req
->flags
|= REQ_F_NEED_CLEANUP
;
3774 static int io_renameat(struct io_kiocb
*req
, unsigned int issue_flags
)
3776 struct io_rename
*ren
= &req
->rename
;
3779 if (issue_flags
& IO_URING_F_NONBLOCK
)
3782 ret
= do_renameat2(ren
->old_dfd
, ren
->oldpath
, ren
->new_dfd
,
3783 ren
->newpath
, ren
->flags
);
3785 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3788 io_req_complete(req
, ret
);
3792 static int io_unlinkat_prep(struct io_kiocb
*req
,
3793 const struct io_uring_sqe
*sqe
)
3795 struct io_unlink
*un
= &req
->unlink
;
3796 const char __user
*fname
;
3798 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3800 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
3803 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3806 un
->dfd
= READ_ONCE(sqe
->fd
);
3808 un
->flags
= READ_ONCE(sqe
->unlink_flags
);
3809 if (un
->flags
& ~AT_REMOVEDIR
)
3812 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3813 un
->filename
= getname(fname
);
3814 if (IS_ERR(un
->filename
))
3815 return PTR_ERR(un
->filename
);
3817 req
->flags
|= REQ_F_NEED_CLEANUP
;
3821 static int io_unlinkat(struct io_kiocb
*req
, unsigned int issue_flags
)
3823 struct io_unlink
*un
= &req
->unlink
;
3826 if (issue_flags
& IO_URING_F_NONBLOCK
)
3829 if (un
->flags
& AT_REMOVEDIR
)
3830 ret
= do_rmdir(un
->dfd
, un
->filename
);
3832 ret
= do_unlinkat(un
->dfd
, un
->filename
);
3834 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3837 io_req_complete(req
, ret
);
3841 static int io_mkdirat_prep(struct io_kiocb
*req
,
3842 const struct io_uring_sqe
*sqe
)
3844 struct io_mkdir
*mkd
= &req
->mkdir
;
3845 const char __user
*fname
;
3847 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3849 if (sqe
->ioprio
|| sqe
->off
|| sqe
->rw_flags
|| sqe
->buf_index
||
3852 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3855 mkd
->dfd
= READ_ONCE(sqe
->fd
);
3856 mkd
->mode
= READ_ONCE(sqe
->len
);
3858 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3859 mkd
->filename
= getname(fname
);
3860 if (IS_ERR(mkd
->filename
))
3861 return PTR_ERR(mkd
->filename
);
3863 req
->flags
|= REQ_F_NEED_CLEANUP
;
3867 static int io_mkdirat(struct io_kiocb
*req
, int issue_flags
)
3869 struct io_mkdir
*mkd
= &req
->mkdir
;
3872 if (issue_flags
& IO_URING_F_NONBLOCK
)
3875 ret
= do_mkdirat(mkd
->dfd
, mkd
->filename
, mkd
->mode
);
3877 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3880 io_req_complete(req
, ret
);
3884 static int io_symlinkat_prep(struct io_kiocb
*req
,
3885 const struct io_uring_sqe
*sqe
)
3887 struct io_symlink
*sl
= &req
->symlink
;
3888 const char __user
*oldpath
, *newpath
;
3890 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3892 if (sqe
->ioprio
|| sqe
->len
|| sqe
->rw_flags
|| sqe
->buf_index
||
3895 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3898 sl
->new_dfd
= READ_ONCE(sqe
->fd
);
3899 oldpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3900 newpath
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3902 sl
->oldpath
= getname(oldpath
);
3903 if (IS_ERR(sl
->oldpath
))
3904 return PTR_ERR(sl
->oldpath
);
3906 sl
->newpath
= getname(newpath
);
3907 if (IS_ERR(sl
->newpath
)) {
3908 putname(sl
->oldpath
);
3909 return PTR_ERR(sl
->newpath
);
3912 req
->flags
|= REQ_F_NEED_CLEANUP
;
3916 static int io_symlinkat(struct io_kiocb
*req
, int issue_flags
)
3918 struct io_symlink
*sl
= &req
->symlink
;
3921 if (issue_flags
& IO_URING_F_NONBLOCK
)
3924 ret
= do_symlinkat(sl
->oldpath
, sl
->new_dfd
, sl
->newpath
);
3926 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3929 io_req_complete(req
, ret
);
3933 static int io_linkat_prep(struct io_kiocb
*req
,
3934 const struct io_uring_sqe
*sqe
)
3936 struct io_hardlink
*lnk
= &req
->hardlink
;
3937 const char __user
*oldf
, *newf
;
3939 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3941 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
3943 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
3946 lnk
->old_dfd
= READ_ONCE(sqe
->fd
);
3947 lnk
->new_dfd
= READ_ONCE(sqe
->len
);
3948 oldf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
3949 newf
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
3950 lnk
->flags
= READ_ONCE(sqe
->hardlink_flags
);
3952 lnk
->oldpath
= getname(oldf
);
3953 if (IS_ERR(lnk
->oldpath
))
3954 return PTR_ERR(lnk
->oldpath
);
3956 lnk
->newpath
= getname(newf
);
3957 if (IS_ERR(lnk
->newpath
)) {
3958 putname(lnk
->oldpath
);
3959 return PTR_ERR(lnk
->newpath
);
3962 req
->flags
|= REQ_F_NEED_CLEANUP
;
3966 static int io_linkat(struct io_kiocb
*req
, int issue_flags
)
3968 struct io_hardlink
*lnk
= &req
->hardlink
;
3971 if (issue_flags
& IO_URING_F_NONBLOCK
)
3974 ret
= do_linkat(lnk
->old_dfd
, lnk
->oldpath
, lnk
->new_dfd
,
3975 lnk
->newpath
, lnk
->flags
);
3977 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
3980 io_req_complete(req
, ret
);
3984 static int io_shutdown_prep(struct io_kiocb
*req
,
3985 const struct io_uring_sqe
*sqe
)
3987 #if defined(CONFIG_NET)
3988 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
3990 if (unlikely(sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->rw_flags
||
3991 sqe
->buf_index
|| sqe
->splice_fd_in
))
3994 req
->shutdown
.how
= READ_ONCE(sqe
->len
);
4001 static int io_shutdown(struct io_kiocb
*req
, unsigned int issue_flags
)
4003 #if defined(CONFIG_NET)
4004 struct socket
*sock
;
4007 if (issue_flags
& IO_URING_F_NONBLOCK
)
4010 sock
= sock_from_file(req
->file
);
4011 if (unlikely(!sock
))
4014 ret
= __sys_shutdown_sock(sock
, req
->shutdown
.how
);
4017 io_req_complete(req
, ret
);
4024 static int __io_splice_prep(struct io_kiocb
*req
,
4025 const struct io_uring_sqe
*sqe
)
4027 struct io_splice
*sp
= &req
->splice
;
4028 unsigned int valid_flags
= SPLICE_F_FD_IN_FIXED
| SPLICE_F_ALL
;
4030 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4033 sp
->len
= READ_ONCE(sqe
->len
);
4034 sp
->flags
= READ_ONCE(sqe
->splice_flags
);
4035 if (unlikely(sp
->flags
& ~valid_flags
))
4037 sp
->splice_fd_in
= READ_ONCE(sqe
->splice_fd_in
);
4041 static int io_tee_prep(struct io_kiocb
*req
,
4042 const struct io_uring_sqe
*sqe
)
4044 if (READ_ONCE(sqe
->splice_off_in
) || READ_ONCE(sqe
->off
))
4046 return __io_splice_prep(req
, sqe
);
4049 static int io_tee(struct io_kiocb
*req
, unsigned int issue_flags
)
4051 struct io_splice
*sp
= &req
->splice
;
4052 struct file
*out
= sp
->file_out
;
4053 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4057 if (issue_flags
& IO_URING_F_NONBLOCK
)
4060 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4061 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4068 ret
= do_tee(in
, out
, sp
->len
, flags
);
4070 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4075 io_req_complete(req
, ret
);
4079 static int io_splice_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4081 struct io_splice
*sp
= &req
->splice
;
4083 sp
->off_in
= READ_ONCE(sqe
->splice_off_in
);
4084 sp
->off_out
= READ_ONCE(sqe
->off
);
4085 return __io_splice_prep(req
, sqe
);
4088 static int io_splice(struct io_kiocb
*req
, unsigned int issue_flags
)
4090 struct io_splice
*sp
= &req
->splice
;
4091 struct file
*out
= sp
->file_out
;
4092 unsigned int flags
= sp
->flags
& ~SPLICE_F_FD_IN_FIXED
;
4093 loff_t
*poff_in
, *poff_out
;
4097 if (issue_flags
& IO_URING_F_NONBLOCK
)
4100 in
= io_file_get(req
->ctx
, req
, sp
->splice_fd_in
,
4101 (sp
->flags
& SPLICE_F_FD_IN_FIXED
));
4107 poff_in
= (sp
->off_in
== -1) ? NULL
: &sp
->off_in
;
4108 poff_out
= (sp
->off_out
== -1) ? NULL
: &sp
->off_out
;
4111 ret
= do_splice(in
, poff_in
, out
, poff_out
, sp
->len
, flags
);
4113 if (!(sp
->flags
& SPLICE_F_FD_IN_FIXED
))
4118 io_req_complete(req
, ret
);
4123 * IORING_OP_NOP just posts a completion event, nothing else.
4125 static int io_nop(struct io_kiocb
*req
, unsigned int issue_flags
)
4127 struct io_ring_ctx
*ctx
= req
->ctx
;
4129 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4132 __io_req_complete(req
, issue_flags
, 0, 0);
4136 static int io_fsync_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4138 struct io_ring_ctx
*ctx
= req
->ctx
;
4140 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4142 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4146 req
->sync
.flags
= READ_ONCE(sqe
->fsync_flags
);
4147 if (unlikely(req
->sync
.flags
& ~IORING_FSYNC_DATASYNC
))
4150 req
->sync
.off
= READ_ONCE(sqe
->off
);
4151 req
->sync
.len
= READ_ONCE(sqe
->len
);
4155 static int io_fsync(struct io_kiocb
*req
, unsigned int issue_flags
)
4157 loff_t end
= req
->sync
.off
+ req
->sync
.len
;
4160 /* fsync always requires a blocking context */
4161 if (issue_flags
& IO_URING_F_NONBLOCK
)
4164 ret
= vfs_fsync_range(req
->file
, req
->sync
.off
,
4165 end
> 0 ? end
: LLONG_MAX
,
4166 req
->sync
.flags
& IORING_FSYNC_DATASYNC
);
4169 io_req_complete(req
, ret
);
4173 static int io_fallocate_prep(struct io_kiocb
*req
,
4174 const struct io_uring_sqe
*sqe
)
4176 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->rw_flags
||
4179 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4182 req
->sync
.off
= READ_ONCE(sqe
->off
);
4183 req
->sync
.len
= READ_ONCE(sqe
->addr
);
4184 req
->sync
.mode
= READ_ONCE(sqe
->len
);
4188 static int io_fallocate(struct io_kiocb
*req
, unsigned int issue_flags
)
4192 /* fallocate always requiring blocking context */
4193 if (issue_flags
& IO_URING_F_NONBLOCK
)
4195 ret
= vfs_fallocate(req
->file
, req
->sync
.mode
, req
->sync
.off
,
4200 fsnotify_modify(req
->file
);
4201 io_req_complete(req
, ret
);
4205 static int __io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4207 const char __user
*fname
;
4210 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4212 if (unlikely(sqe
->ioprio
|| sqe
->buf_index
))
4214 if (unlikely(req
->flags
& REQ_F_FIXED_FILE
))
4217 /* open.how should be already initialised */
4218 if (!(req
->open
.how
.flags
& O_PATH
) && force_o_largefile())
4219 req
->open
.how
.flags
|= O_LARGEFILE
;
4221 req
->open
.dfd
= READ_ONCE(sqe
->fd
);
4222 fname
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4223 req
->open
.filename
= getname(fname
);
4224 if (IS_ERR(req
->open
.filename
)) {
4225 ret
= PTR_ERR(req
->open
.filename
);
4226 req
->open
.filename
= NULL
;
4230 req
->open
.file_slot
= READ_ONCE(sqe
->file_index
);
4231 if (req
->open
.file_slot
&& (req
->open
.how
.flags
& O_CLOEXEC
))
4234 req
->open
.nofile
= rlimit(RLIMIT_NOFILE
);
4235 req
->flags
|= REQ_F_NEED_CLEANUP
;
4239 static int io_openat_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4241 u64 mode
= READ_ONCE(sqe
->len
);
4242 u64 flags
= READ_ONCE(sqe
->open_flags
);
4244 req
->open
.how
= build_open_how(flags
, mode
);
4245 return __io_openat_prep(req
, sqe
);
4248 static int io_openat2_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4250 struct open_how __user
*how
;
4254 how
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4255 len
= READ_ONCE(sqe
->len
);
4256 if (len
< OPEN_HOW_SIZE_VER0
)
4259 ret
= copy_struct_from_user(&req
->open
.how
, sizeof(req
->open
.how
), how
,
4264 return __io_openat_prep(req
, sqe
);
4267 static int io_openat2(struct io_kiocb
*req
, unsigned int issue_flags
)
4269 struct open_flags op
;
4271 bool resolve_nonblock
, nonblock_set
;
4272 bool fixed
= !!req
->open
.file_slot
;
4275 ret
= build_open_flags(&req
->open
.how
, &op
);
4278 nonblock_set
= op
.open_flag
& O_NONBLOCK
;
4279 resolve_nonblock
= req
->open
.how
.resolve
& RESOLVE_CACHED
;
4280 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4282 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4283 * it'll always -EAGAIN
4285 if (req
->open
.how
.flags
& (O_TRUNC
| O_CREAT
| O_TMPFILE
))
4287 op
.lookup_flags
|= LOOKUP_CACHED
;
4288 op
.open_flag
|= O_NONBLOCK
;
4292 ret
= __get_unused_fd_flags(req
->open
.how
.flags
, req
->open
.nofile
);
4297 file
= do_filp_open(req
->open
.dfd
, req
->open
.filename
, &op
);
4300 * We could hang on to this 'fd' on retrying, but seems like
4301 * marginal gain for something that is now known to be a slower
4302 * path. So just put it, and we'll get a new one when we retry.
4307 ret
= PTR_ERR(file
);
4308 /* only retry if RESOLVE_CACHED wasn't already set by application */
4309 if (ret
== -EAGAIN
&&
4310 (!resolve_nonblock
&& (issue_flags
& IO_URING_F_NONBLOCK
)))
4315 if ((issue_flags
& IO_URING_F_NONBLOCK
) && !nonblock_set
)
4316 file
->f_flags
&= ~O_NONBLOCK
;
4317 fsnotify_open(file
);
4320 fd_install(ret
, file
);
4322 ret
= io_install_fixed_file(req
, file
, issue_flags
,
4323 req
->open
.file_slot
- 1);
4325 putname(req
->open
.filename
);
4326 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4329 __io_req_complete(req
, issue_flags
, ret
, 0);
4333 static int io_openat(struct io_kiocb
*req
, unsigned int issue_flags
)
4335 return io_openat2(req
, issue_flags
);
4338 static int io_remove_buffers_prep(struct io_kiocb
*req
,
4339 const struct io_uring_sqe
*sqe
)
4341 struct io_provide_buf
*p
= &req
->pbuf
;
4344 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->addr
|| sqe
->len
|| sqe
->off
||
4348 tmp
= READ_ONCE(sqe
->fd
);
4349 if (!tmp
|| tmp
> USHRT_MAX
)
4352 memset(p
, 0, sizeof(*p
));
4354 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4358 static int __io_remove_buffers(struct io_ring_ctx
*ctx
, struct io_buffer
*buf
,
4359 int bgid
, unsigned nbufs
)
4363 /* shouldn't happen */
4367 /* the head kbuf is the list itself */
4368 while (!list_empty(&buf
->list
)) {
4369 struct io_buffer
*nxt
;
4371 nxt
= list_first_entry(&buf
->list
, struct io_buffer
, list
);
4372 list_del(&nxt
->list
);
4380 xa_erase(&ctx
->io_buffers
, bgid
);
4385 static int io_remove_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4387 struct io_provide_buf
*p
= &req
->pbuf
;
4388 struct io_ring_ctx
*ctx
= req
->ctx
;
4389 struct io_buffer
*head
;
4391 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4393 io_ring_submit_lock(ctx
, !force_nonblock
);
4395 lockdep_assert_held(&ctx
->uring_lock
);
4398 head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4400 ret
= __io_remove_buffers(ctx
, head
, p
->bgid
, p
->nbufs
);
4404 /* complete before unlock, IOPOLL may need the lock */
4405 __io_req_complete(req
, issue_flags
, ret
, 0);
4406 io_ring_submit_unlock(ctx
, !force_nonblock
);
4410 static int io_provide_buffers_prep(struct io_kiocb
*req
,
4411 const struct io_uring_sqe
*sqe
)
4413 unsigned long size
, tmp_check
;
4414 struct io_provide_buf
*p
= &req
->pbuf
;
4417 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
4420 tmp
= READ_ONCE(sqe
->fd
);
4421 if (!tmp
|| tmp
> USHRT_MAX
)
4424 p
->addr
= READ_ONCE(sqe
->addr
);
4425 p
->len
= READ_ONCE(sqe
->len
);
4427 if (check_mul_overflow((unsigned long)p
->len
, (unsigned long)p
->nbufs
,
4430 if (check_add_overflow((unsigned long)p
->addr
, size
, &tmp_check
))
4433 size
= (unsigned long)p
->len
* p
->nbufs
;
4434 if (!access_ok(u64_to_user_ptr(p
->addr
), size
))
4437 p
->bgid
= READ_ONCE(sqe
->buf_group
);
4438 tmp
= READ_ONCE(sqe
->off
);
4439 if (tmp
> USHRT_MAX
)
4445 static int io_add_buffers(struct io_provide_buf
*pbuf
, struct io_buffer
**head
)
4447 struct io_buffer
*buf
;
4448 u64 addr
= pbuf
->addr
;
4449 int i
, bid
= pbuf
->bid
;
4451 for (i
= 0; i
< pbuf
->nbufs
; i
++) {
4452 buf
= kmalloc(sizeof(*buf
), GFP_KERNEL_ACCOUNT
);
4457 buf
->len
= min_t(__u32
, pbuf
->len
, MAX_RW_COUNT
);
4462 INIT_LIST_HEAD(&buf
->list
);
4465 list_add_tail(&buf
->list
, &(*head
)->list
);
4470 return i
? i
: -ENOMEM
;
4473 static int io_provide_buffers(struct io_kiocb
*req
, unsigned int issue_flags
)
4475 struct io_provide_buf
*p
= &req
->pbuf
;
4476 struct io_ring_ctx
*ctx
= req
->ctx
;
4477 struct io_buffer
*head
, *list
;
4479 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4481 io_ring_submit_lock(ctx
, !force_nonblock
);
4483 lockdep_assert_held(&ctx
->uring_lock
);
4485 list
= head
= xa_load(&ctx
->io_buffers
, p
->bgid
);
4487 ret
= io_add_buffers(p
, &head
);
4488 if (ret
>= 0 && !list
) {
4489 ret
= xa_insert(&ctx
->io_buffers
, p
->bgid
, head
,
4490 GFP_KERNEL_ACCOUNT
);
4492 __io_remove_buffers(ctx
, head
, p
->bgid
, -1U);
4496 /* complete before unlock, IOPOLL may need the lock */
4497 __io_req_complete(req
, issue_flags
, ret
, 0);
4498 io_ring_submit_unlock(ctx
, !force_nonblock
);
4502 static int io_epoll_ctl_prep(struct io_kiocb
*req
,
4503 const struct io_uring_sqe
*sqe
)
4505 #if defined(CONFIG_EPOLL)
4506 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4508 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4511 req
->epoll
.epfd
= READ_ONCE(sqe
->fd
);
4512 req
->epoll
.op
= READ_ONCE(sqe
->len
);
4513 req
->epoll
.fd
= READ_ONCE(sqe
->off
);
4515 if (ep_op_has_event(req
->epoll
.op
)) {
4516 struct epoll_event __user
*ev
;
4518 ev
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4519 if (copy_from_user(&req
->epoll
.event
, ev
, sizeof(*ev
)))
4529 static int io_epoll_ctl(struct io_kiocb
*req
, unsigned int issue_flags
)
4531 #if defined(CONFIG_EPOLL)
4532 struct io_epoll
*ie
= &req
->epoll
;
4534 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
4536 ret
= do_epoll_ctl(ie
->epfd
, ie
->op
, ie
->fd
, &ie
->event
, force_nonblock
);
4537 if (force_nonblock
&& ret
== -EAGAIN
)
4542 __io_req_complete(req
, issue_flags
, ret
, 0);
4549 static int io_madvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4551 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4552 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->splice_fd_in
)
4554 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4557 req
->madvise
.addr
= READ_ONCE(sqe
->addr
);
4558 req
->madvise
.len
= READ_ONCE(sqe
->len
);
4559 req
->madvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4566 static int io_madvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4568 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4569 struct io_madvise
*ma
= &req
->madvise
;
4572 if (issue_flags
& IO_URING_F_NONBLOCK
)
4575 ret
= do_madvise(current
->mm
, ma
->addr
, ma
->len
, ma
->advice
);
4578 io_req_complete(req
, ret
);
4585 static int io_fadvise_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4587 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->addr
|| sqe
->splice_fd_in
)
4589 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4592 req
->fadvise
.offset
= READ_ONCE(sqe
->off
);
4593 req
->fadvise
.len
= READ_ONCE(sqe
->len
);
4594 req
->fadvise
.advice
= READ_ONCE(sqe
->fadvise_advice
);
4598 static int io_fadvise(struct io_kiocb
*req
, unsigned int issue_flags
)
4600 struct io_fadvise
*fa
= &req
->fadvise
;
4603 if (issue_flags
& IO_URING_F_NONBLOCK
) {
4604 switch (fa
->advice
) {
4605 case POSIX_FADV_NORMAL
:
4606 case POSIX_FADV_RANDOM
:
4607 case POSIX_FADV_SEQUENTIAL
:
4614 ret
= vfs_fadvise(req
->file
, fa
->offset
, fa
->len
, fa
->advice
);
4617 __io_req_complete(req
, issue_flags
, ret
, 0);
4621 static int io_statx_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4623 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4625 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
4627 if (req
->flags
& REQ_F_FIXED_FILE
)
4630 req
->statx
.dfd
= READ_ONCE(sqe
->fd
);
4631 req
->statx
.mask
= READ_ONCE(sqe
->len
);
4632 req
->statx
.filename
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4633 req
->statx
.buffer
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
4634 req
->statx
.flags
= READ_ONCE(sqe
->statx_flags
);
4639 static int io_statx(struct io_kiocb
*req
, unsigned int issue_flags
)
4641 struct io_statx
*ctx
= &req
->statx
;
4644 if (issue_flags
& IO_URING_F_NONBLOCK
)
4647 ret
= do_statx(ctx
->dfd
, ctx
->filename
, ctx
->flags
, ctx
->mask
,
4652 io_req_complete(req
, ret
);
4656 static int io_close_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4658 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4660 if (sqe
->ioprio
|| sqe
->off
|| sqe
->addr
|| sqe
->len
||
4661 sqe
->rw_flags
|| sqe
->buf_index
)
4663 if (req
->flags
& REQ_F_FIXED_FILE
)
4666 req
->close
.fd
= READ_ONCE(sqe
->fd
);
4667 req
->close
.file_slot
= READ_ONCE(sqe
->file_index
);
4668 if (req
->close
.file_slot
&& req
->close
.fd
)
4674 static int io_close(struct io_kiocb
*req
, unsigned int issue_flags
)
4676 struct files_struct
*files
= current
->files
;
4677 struct io_close
*close
= &req
->close
;
4678 struct fdtable
*fdt
;
4679 struct file
*file
= NULL
;
4682 if (req
->close
.file_slot
) {
4683 ret
= io_close_fixed(req
, issue_flags
);
4687 spin_lock(&files
->file_lock
);
4688 fdt
= files_fdtable(files
);
4689 if (close
->fd
>= fdt
->max_fds
) {
4690 spin_unlock(&files
->file_lock
);
4693 file
= fdt
->fd
[close
->fd
];
4694 if (!file
|| file
->f_op
== &io_uring_fops
) {
4695 spin_unlock(&files
->file_lock
);
4700 /* if the file has a flush method, be safe and punt to async */
4701 if (file
->f_op
->flush
&& (issue_flags
& IO_URING_F_NONBLOCK
)) {
4702 spin_unlock(&files
->file_lock
);
4706 ret
= __close_fd_get_file(close
->fd
, &file
);
4707 spin_unlock(&files
->file_lock
);
4714 /* No ->flush() or already async, safely close from here */
4715 ret
= filp_close(file
, current
->files
);
4721 __io_req_complete(req
, issue_flags
, ret
, 0);
4725 static int io_sfr_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4727 struct io_ring_ctx
*ctx
= req
->ctx
;
4729 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
4731 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
||
4735 req
->sync
.off
= READ_ONCE(sqe
->off
);
4736 req
->sync
.len
= READ_ONCE(sqe
->len
);
4737 req
->sync
.flags
= READ_ONCE(sqe
->sync_range_flags
);
4741 static int io_sync_file_range(struct io_kiocb
*req
, unsigned int issue_flags
)
4745 /* sync_file_range always requires a blocking context */
4746 if (issue_flags
& IO_URING_F_NONBLOCK
)
4749 ret
= sync_file_range(req
->file
, req
->sync
.off
, req
->sync
.len
,
4753 io_req_complete(req
, ret
);
4757 #if defined(CONFIG_NET)
4758 static int io_setup_async_msg(struct io_kiocb
*req
,
4759 struct io_async_msghdr
*kmsg
)
4761 struct io_async_msghdr
*async_msg
= req
->async_data
;
4765 if (io_alloc_async_data(req
)) {
4766 kfree(kmsg
->free_iov
);
4769 async_msg
= req
->async_data
;
4770 req
->flags
|= REQ_F_NEED_CLEANUP
;
4771 memcpy(async_msg
, kmsg
, sizeof(*kmsg
));
4772 if (async_msg
->msg
.msg_name
)
4773 async_msg
->msg
.msg_name
= &async_msg
->addr
;
4774 /* if were using fast_iov, set it to the new one */
4775 if (!async_msg
->free_iov
)
4776 async_msg
->msg
.msg_iter
.iov
= async_msg
->fast_iov
;
4781 static int io_sendmsg_copy_hdr(struct io_kiocb
*req
,
4782 struct io_async_msghdr
*iomsg
)
4784 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4785 iomsg
->free_iov
= iomsg
->fast_iov
;
4786 return sendmsg_copy_msghdr(&iomsg
->msg
, req
->sr_msg
.umsg
,
4787 req
->sr_msg
.msg_flags
, &iomsg
->free_iov
);
4790 static int io_sendmsg_prep_async(struct io_kiocb
*req
)
4794 ret
= io_sendmsg_copy_hdr(req
, req
->async_data
);
4796 req
->flags
|= REQ_F_NEED_CLEANUP
;
4800 static int io_sendmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
4802 struct io_sr_msg
*sr
= &req
->sr_msg
;
4804 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
4806 if (unlikely(sqe
->addr2
|| sqe
->file_index
))
4808 if (unlikely(sqe
->addr2
|| sqe
->file_index
|| sqe
->ioprio
))
4811 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
4812 sr
->len
= READ_ONCE(sqe
->len
);
4813 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
4814 if (sr
->msg_flags
& MSG_DONTWAIT
)
4815 req
->flags
|= REQ_F_NOWAIT
;
4817 #ifdef CONFIG_COMPAT
4818 if (req
->ctx
->compat
)
4819 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
4824 static int io_sendmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
4826 struct io_async_msghdr iomsg
, *kmsg
;
4827 struct socket
*sock
;
4832 sock
= sock_from_file(req
->file
);
4833 if (unlikely(!sock
))
4836 kmsg
= req
->async_data
;
4838 ret
= io_sendmsg_copy_hdr(req
, &iomsg
);
4844 flags
= req
->sr_msg
.msg_flags
;
4845 if (issue_flags
& IO_URING_F_NONBLOCK
)
4846 flags
|= MSG_DONTWAIT
;
4847 if (flags
& MSG_WAITALL
)
4848 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
4850 ret
= __sys_sendmsg_sock(sock
, &kmsg
->msg
, flags
);
4851 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4852 return io_setup_async_msg(req
, kmsg
);
4853 if (ret
== -ERESTARTSYS
)
4856 /* fast path, check for non-NULL to avoid function call */
4858 kfree(kmsg
->free_iov
);
4859 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
4862 __io_req_complete(req
, issue_flags
, ret
, 0);
4866 static int io_send(struct io_kiocb
*req
, unsigned int issue_flags
)
4868 struct io_sr_msg
*sr
= &req
->sr_msg
;
4871 struct socket
*sock
;
4876 sock
= sock_from_file(req
->file
);
4877 if (unlikely(!sock
))
4880 ret
= import_single_range(WRITE
, sr
->buf
, sr
->len
, &iov
, &msg
.msg_iter
);
4884 msg
.msg_name
= NULL
;
4885 msg
.msg_control
= NULL
;
4886 msg
.msg_controllen
= 0;
4887 msg
.msg_namelen
= 0;
4889 flags
= req
->sr_msg
.msg_flags
;
4890 if (issue_flags
& IO_URING_F_NONBLOCK
)
4891 flags
|= MSG_DONTWAIT
;
4892 if (flags
& MSG_WAITALL
)
4893 min_ret
= iov_iter_count(&msg
.msg_iter
);
4895 msg
.msg_flags
= flags
;
4896 ret
= sock_sendmsg(sock
, &msg
);
4897 if ((issue_flags
& IO_URING_F_NONBLOCK
) && ret
== -EAGAIN
)
4899 if (ret
== -ERESTARTSYS
)
4904 __io_req_complete(req
, issue_flags
, ret
, 0);
4908 static int __io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4909 struct io_async_msghdr
*iomsg
)
4911 struct io_sr_msg
*sr
= &req
->sr_msg
;
4912 struct iovec __user
*uiov
;
4916 ret
= __copy_msghdr_from_user(&iomsg
->msg
, sr
->umsg
,
4917 &iomsg
->uaddr
, &uiov
, &iov_len
);
4921 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4924 if (copy_from_user(iomsg
->fast_iov
, uiov
, sizeof(*uiov
)))
4926 sr
->len
= iomsg
->fast_iov
[0].iov_len
;
4927 iomsg
->free_iov
= NULL
;
4929 iomsg
->free_iov
= iomsg
->fast_iov
;
4930 ret
= __import_iovec(READ
, uiov
, iov_len
, UIO_FASTIOV
,
4931 &iomsg
->free_iov
, &iomsg
->msg
.msg_iter
,
4940 #ifdef CONFIG_COMPAT
4941 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb
*req
,
4942 struct io_async_msghdr
*iomsg
)
4944 struct io_sr_msg
*sr
= &req
->sr_msg
;
4945 struct compat_iovec __user
*uiov
;
4950 ret
= __get_compat_msghdr(&iomsg
->msg
, sr
->umsg_compat
, &iomsg
->uaddr
,
4955 uiov
= compat_ptr(ptr
);
4956 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
4957 compat_ssize_t clen
;
4961 if (!access_ok(uiov
, sizeof(*uiov
)))
4963 if (__get_user(clen
, &uiov
->iov_len
))
4968 iomsg
->free_iov
= NULL
;
4970 iomsg
->free_iov
= iomsg
->fast_iov
;
4971 ret
= __import_iovec(READ
, (struct iovec __user
*)uiov
, len
,
4972 UIO_FASTIOV
, &iomsg
->free_iov
,
4973 &iomsg
->msg
.msg_iter
, true);
4982 static int io_recvmsg_copy_hdr(struct io_kiocb
*req
,
4983 struct io_async_msghdr
*iomsg
)
4985 iomsg
->msg
.msg_name
= &iomsg
->addr
;
4987 #ifdef CONFIG_COMPAT
4988 if (req
->ctx
->compat
)
4989 return __io_compat_recvmsg_copy_hdr(req
, iomsg
);
4992 return __io_recvmsg_copy_hdr(req
, iomsg
);
4995 static struct io_buffer
*io_recv_buffer_select(struct io_kiocb
*req
,
4998 struct io_sr_msg
*sr
= &req
->sr_msg
;
4999 struct io_buffer
*kbuf
;
5001 kbuf
= io_buffer_select(req
, &sr
->len
, sr
->bgid
, sr
->kbuf
, needs_lock
);
5006 req
->flags
|= REQ_F_BUFFER_SELECTED
;
5010 static inline unsigned int io_put_recv_kbuf(struct io_kiocb
*req
)
5012 return io_put_kbuf(req
, req
->sr_msg
.kbuf
);
5015 static int io_recvmsg_prep_async(struct io_kiocb
*req
)
5019 ret
= io_recvmsg_copy_hdr(req
, req
->async_data
);
5021 req
->flags
|= REQ_F_NEED_CLEANUP
;
5025 static int io_recvmsg_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5027 struct io_sr_msg
*sr
= &req
->sr_msg
;
5029 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5031 if (unlikely(sqe
->addr2
|| sqe
->file_index
))
5033 if (unlikely(sqe
->addr2
|| sqe
->file_index
|| sqe
->ioprio
))
5036 sr
->umsg
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5037 sr
->len
= READ_ONCE(sqe
->len
);
5038 sr
->bgid
= READ_ONCE(sqe
->buf_group
);
5039 sr
->msg_flags
= READ_ONCE(sqe
->msg_flags
) | MSG_NOSIGNAL
;
5040 if (sr
->msg_flags
& MSG_DONTWAIT
)
5041 req
->flags
|= REQ_F_NOWAIT
;
5043 #ifdef CONFIG_COMPAT
5044 if (req
->ctx
->compat
)
5045 sr
->msg_flags
|= MSG_CMSG_COMPAT
;
5050 static int io_recvmsg(struct io_kiocb
*req
, unsigned int issue_flags
)
5052 struct io_async_msghdr iomsg
, *kmsg
;
5053 struct socket
*sock
;
5054 struct io_buffer
*kbuf
;
5057 int ret
, cflags
= 0;
5058 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5060 sock
= sock_from_file(req
->file
);
5061 if (unlikely(!sock
))
5064 kmsg
= req
->async_data
;
5066 ret
= io_recvmsg_copy_hdr(req
, &iomsg
);
5072 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5073 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5075 return PTR_ERR(kbuf
);
5076 kmsg
->fast_iov
[0].iov_base
= u64_to_user_ptr(kbuf
->addr
);
5077 kmsg
->fast_iov
[0].iov_len
= req
->sr_msg
.len
;
5078 iov_iter_init(&kmsg
->msg
.msg_iter
, READ
, kmsg
->fast_iov
,
5079 1, req
->sr_msg
.len
);
5082 flags
= req
->sr_msg
.msg_flags
;
5084 flags
|= MSG_DONTWAIT
;
5085 if (flags
& MSG_WAITALL
)
5086 min_ret
= iov_iter_count(&kmsg
->msg
.msg_iter
);
5088 ret
= __sys_recvmsg_sock(sock
, &kmsg
->msg
, req
->sr_msg
.umsg
,
5089 kmsg
->uaddr
, flags
);
5090 if (force_nonblock
&& ret
== -EAGAIN
)
5091 return io_setup_async_msg(req
, kmsg
);
5092 if (ret
== -ERESTARTSYS
)
5095 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5096 cflags
= io_put_recv_kbuf(req
);
5097 /* fast path, check for non-NULL to avoid function call */
5099 kfree(kmsg
->free_iov
);
5100 req
->flags
&= ~REQ_F_NEED_CLEANUP
;
5101 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (kmsg
->msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5103 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5107 static int io_recv(struct io_kiocb
*req
, unsigned int issue_flags
)
5109 struct io_buffer
*kbuf
;
5110 struct io_sr_msg
*sr
= &req
->sr_msg
;
5112 void __user
*buf
= sr
->buf
;
5113 struct socket
*sock
;
5117 int ret
, cflags
= 0;
5118 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5120 sock
= sock_from_file(req
->file
);
5121 if (unlikely(!sock
))
5124 if (req
->flags
& REQ_F_BUFFER_SELECT
) {
5125 kbuf
= io_recv_buffer_select(req
, !force_nonblock
);
5127 return PTR_ERR(kbuf
);
5128 buf
= u64_to_user_ptr(kbuf
->addr
);
5131 ret
= import_single_range(READ
, buf
, sr
->len
, &iov
, &msg
.msg_iter
);
5135 msg
.msg_name
= NULL
;
5136 msg
.msg_control
= NULL
;
5137 msg
.msg_controllen
= 0;
5138 msg
.msg_namelen
= 0;
5139 msg
.msg_iocb
= NULL
;
5142 flags
= req
->sr_msg
.msg_flags
;
5144 flags
|= MSG_DONTWAIT
;
5145 if (flags
& MSG_WAITALL
)
5146 min_ret
= iov_iter_count(&msg
.msg_iter
);
5148 ret
= sock_recvmsg(sock
, &msg
, flags
);
5149 if (force_nonblock
&& ret
== -EAGAIN
)
5151 if (ret
== -ERESTARTSYS
)
5154 if (req
->flags
& REQ_F_BUFFER_SELECTED
)
5155 cflags
= io_put_recv_kbuf(req
);
5156 if (ret
< min_ret
|| ((flags
& MSG_WAITALL
) && (msg
.msg_flags
& (MSG_TRUNC
| MSG_CTRUNC
))))
5158 __io_req_complete(req
, issue_flags
, ret
, cflags
);
5162 static int io_accept_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5164 struct io_accept
*accept
= &req
->accept
;
5166 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5168 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
)
5171 accept
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5172 accept
->addr_len
= u64_to_user_ptr(READ_ONCE(sqe
->addr2
));
5173 accept
->flags
= READ_ONCE(sqe
->accept_flags
);
5174 accept
->nofile
= rlimit(RLIMIT_NOFILE
);
5176 accept
->file_slot
= READ_ONCE(sqe
->file_index
);
5177 if (accept
->file_slot
&& (accept
->flags
& SOCK_CLOEXEC
))
5179 if (accept
->flags
& ~(SOCK_CLOEXEC
| SOCK_NONBLOCK
))
5181 if (SOCK_NONBLOCK
!= O_NONBLOCK
&& (accept
->flags
& SOCK_NONBLOCK
))
5182 accept
->flags
= (accept
->flags
& ~SOCK_NONBLOCK
) | O_NONBLOCK
;
5186 static int io_accept(struct io_kiocb
*req
, unsigned int issue_flags
)
5188 struct io_accept
*accept
= &req
->accept
;
5189 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5190 unsigned int file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5191 bool fixed
= !!accept
->file_slot
;
5195 if (req
->file
->f_flags
& O_NONBLOCK
)
5196 req
->flags
|= REQ_F_NOWAIT
;
5199 fd
= __get_unused_fd_flags(accept
->flags
, accept
->nofile
);
5200 if (unlikely(fd
< 0))
5203 file
= do_accept(req
->file
, file_flags
, accept
->addr
, accept
->addr_len
,
5208 ret
= PTR_ERR(file
);
5209 if (ret
== -EAGAIN
&& force_nonblock
)
5211 if (ret
== -ERESTARTSYS
)
5214 } else if (!fixed
) {
5215 fd_install(fd
, file
);
5218 ret
= io_install_fixed_file(req
, file
, issue_flags
,
5219 accept
->file_slot
- 1);
5221 __io_req_complete(req
, issue_flags
, ret
, 0);
5225 static int io_connect_prep_async(struct io_kiocb
*req
)
5227 struct io_async_connect
*io
= req
->async_data
;
5228 struct io_connect
*conn
= &req
->connect
;
5230 return move_addr_to_kernel(conn
->addr
, conn
->addr_len
, &io
->address
);
5233 static int io_connect_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5235 struct io_connect
*conn
= &req
->connect
;
5237 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5239 if (sqe
->ioprio
|| sqe
->len
|| sqe
->buf_index
|| sqe
->rw_flags
||
5243 conn
->addr
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
5244 conn
->addr_len
= READ_ONCE(sqe
->addr2
);
5248 static int io_connect(struct io_kiocb
*req
, unsigned int issue_flags
)
5250 struct io_async_connect __io
, *io
;
5251 unsigned file_flags
;
5253 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
5255 if (req
->async_data
) {
5256 io
= req
->async_data
;
5258 ret
= move_addr_to_kernel(req
->connect
.addr
,
5259 req
->connect
.addr_len
,
5266 file_flags
= force_nonblock
? O_NONBLOCK
: 0;
5268 ret
= __sys_connect_file(req
->file
, &io
->address
,
5269 req
->connect
.addr_len
, file_flags
);
5270 if ((ret
== -EAGAIN
|| ret
== -EINPROGRESS
) && force_nonblock
) {
5271 if (req
->async_data
)
5273 if (io_alloc_async_data(req
)) {
5277 memcpy(req
->async_data
, &__io
, sizeof(__io
));
5280 if (ret
== -ERESTARTSYS
)
5285 __io_req_complete(req
, issue_flags
, ret
, 0);
5288 #else /* !CONFIG_NET */
5289 #define IO_NETOP_FN(op) \
5290 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5292 return -EOPNOTSUPP; \
5295 #define IO_NETOP_PREP(op) \
5297 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5299 return -EOPNOTSUPP; \
5302 #define IO_NETOP_PREP_ASYNC(op) \
5304 static int io_##op##_prep_async(struct io_kiocb *req) \
5306 return -EOPNOTSUPP; \
5309 IO_NETOP_PREP_ASYNC(sendmsg
);
5310 IO_NETOP_PREP_ASYNC(recvmsg
);
5311 IO_NETOP_PREP_ASYNC(connect
);
5312 IO_NETOP_PREP(accept
);
5315 #endif /* CONFIG_NET */
5317 struct io_poll_table
{
5318 struct poll_table_struct pt
;
5319 struct io_kiocb
*req
;
5324 #define IO_POLL_CANCEL_FLAG BIT(31)
5325 #define IO_POLL_RETRY_FLAG BIT(30)
5326 #define IO_POLL_REF_MASK GENMASK(29, 0)
5329 * We usually have 1-2 refs taken, 128 is more than enough and we want to
5330 * maximise the margin between this amount and the moment when it overflows.
5332 #define IO_POLL_REF_BIAS 128
5334 static bool io_poll_get_ownership_slowpath(struct io_kiocb
*req
)
5339 * poll_refs are already elevated and we don't have much hope for
5340 * grabbing the ownership. Instead of incrementing set a retry flag
5341 * to notify the loop that there might have been some change.
5343 v
= atomic_fetch_or(IO_POLL_RETRY_FLAG
, &req
->poll_refs
);
5344 if (v
& IO_POLL_REF_MASK
)
5346 return !(atomic_fetch_inc(&req
->poll_refs
) & IO_POLL_REF_MASK
);
5350 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
5351 * bump it and acquire ownership. It's disallowed to modify requests while not
5352 * owning it, that prevents from races for enqueueing task_work's and b/w
5353 * arming poll and wakeups.
5355 static inline bool io_poll_get_ownership(struct io_kiocb
*req
)
5357 if (unlikely(atomic_read(&req
->poll_refs
) >= IO_POLL_REF_BIAS
))
5358 return io_poll_get_ownership_slowpath(req
);
5359 return !(atomic_fetch_inc(&req
->poll_refs
) & IO_POLL_REF_MASK
);
5362 static void io_poll_mark_cancelled(struct io_kiocb
*req
)
5364 atomic_or(IO_POLL_CANCEL_FLAG
, &req
->poll_refs
);
5367 static struct io_poll_iocb
*io_poll_get_double(struct io_kiocb
*req
)
5369 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5370 if (req
->opcode
== IORING_OP_POLL_ADD
)
5371 return req
->async_data
;
5372 return req
->apoll
->double_poll
;
5375 static struct io_poll_iocb
*io_poll_get_single(struct io_kiocb
*req
)
5377 if (req
->opcode
== IORING_OP_POLL_ADD
)
5379 return &req
->apoll
->poll
;
5382 static void io_poll_req_insert(struct io_kiocb
*req
)
5384 struct io_ring_ctx
*ctx
= req
->ctx
;
5385 struct hlist_head
*list
;
5387 list
= &ctx
->cancel_hash
[hash_long(req
->user_data
, ctx
->cancel_hash_bits
)];
5388 hlist_add_head(&req
->hash_node
, list
);
5391 static void io_init_poll_iocb(struct io_poll_iocb
*poll
, __poll_t events
,
5392 wait_queue_func_t wake_func
)
5395 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5396 /* mask in events that we always want/need */
5397 poll
->events
= events
| IO_POLL_UNMASK
;
5398 INIT_LIST_HEAD(&poll
->wait
.entry
);
5399 init_waitqueue_func_entry(&poll
->wait
, wake_func
);
5402 static inline void io_poll_remove_entry(struct io_poll_iocb
*poll
)
5404 struct wait_queue_head
*head
= smp_load_acquire(&poll
->head
);
5407 spin_lock_irq(&head
->lock
);
5408 list_del_init(&poll
->wait
.entry
);
5410 spin_unlock_irq(&head
->lock
);
5414 static void io_poll_remove_entries(struct io_kiocb
*req
)
5416 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5417 struct io_poll_iocb
*poll_double
= io_poll_get_double(req
);
5420 * While we hold the waitqueue lock and the waitqueue is nonempty,
5421 * wake_up_pollfree() will wait for us. However, taking the waitqueue
5422 * lock in the first place can race with the waitqueue being freed.
5424 * We solve this as eventpoll does: by taking advantage of the fact that
5425 * all users of wake_up_pollfree() will RCU-delay the actual free. If
5426 * we enter rcu_read_lock() and see that the pointer to the queue is
5427 * non-NULL, we can then lock it without the memory being freed out from
5430 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
5431 * case the caller deletes the entry from the queue, leaving it empty.
5432 * In that case, only RCU prevents the queue memory from being freed.
5435 io_poll_remove_entry(poll
);
5437 io_poll_remove_entry(poll_double
);
5442 * All poll tw should go through this. Checks for poll events, manages
5443 * references, does rewait, etc.
5445 * Returns a negative error on failure. >0 when no action require, which is
5446 * either spurious wakeup or multishot CQE is served. 0 when it's done with
5447 * the request, then the mask is stored in req->result.
5449 static int io_poll_check_events(struct io_kiocb
*req
)
5451 struct io_ring_ctx
*ctx
= req
->ctx
;
5452 struct io_poll_iocb
*poll
= io_poll_get_single(req
);
5455 /* req->task == current here, checking PF_EXITING is safe */
5456 if (unlikely(req
->task
->flags
& PF_EXITING
))
5457 io_poll_mark_cancelled(req
);
5460 v
= atomic_read(&req
->poll_refs
);
5462 /* tw handler should be the owner, and so have some references */
5463 if (WARN_ON_ONCE(!(v
& IO_POLL_REF_MASK
)))
5465 if (v
& IO_POLL_CANCEL_FLAG
)
5468 * cqe.res contains only events of the first wake up
5469 * and all others are be lost. Redo vfs_poll() to get
5472 if ((v
& IO_POLL_REF_MASK
) != 1)
5474 if (v
& IO_POLL_RETRY_FLAG
) {
5477 * We won't find new events that came in between
5478 * vfs_poll and the ref put unless we clear the
5481 atomic_andnot(IO_POLL_RETRY_FLAG
, &req
->poll_refs
);
5482 v
&= ~IO_POLL_RETRY_FLAG
;
5486 struct poll_table_struct pt
= { ._key
= poll
->events
};
5488 req
->result
= vfs_poll(req
->file
, &pt
) & poll
->events
;
5491 /* multishot, just fill an CQE and proceed */
5492 if (req
->result
&& !(poll
->events
& EPOLLONESHOT
)) {
5493 __poll_t mask
= mangle_poll(req
->result
& poll
->events
);
5496 spin_lock(&ctx
->completion_lock
);
5497 filled
= io_fill_cqe_aux(ctx
, req
->user_data
, mask
,
5499 io_commit_cqring(ctx
);
5500 spin_unlock(&ctx
->completion_lock
);
5501 if (unlikely(!filled
))
5503 io_cqring_ev_posted(ctx
);
5504 } else if (req
->result
) {
5508 /* force the next iteration to vfs_poll() */
5512 * Release all references, retry if someone tried to restart
5513 * task_work while we were executing it.
5515 } while (atomic_sub_return(v
& IO_POLL_REF_MASK
, &req
->poll_refs
) &
5521 static void io_poll_task_func(struct io_kiocb
*req
, bool *locked
)
5523 struct io_ring_ctx
*ctx
= req
->ctx
;
5526 ret
= io_poll_check_events(req
);
5531 req
->result
= mangle_poll(req
->result
& req
->poll
.events
);
5537 io_poll_remove_entries(req
);
5538 spin_lock(&ctx
->completion_lock
);
5539 hash_del(&req
->hash_node
);
5540 spin_unlock(&ctx
->completion_lock
);
5541 io_req_complete_post(req
, req
->result
, 0);
5544 static void io_apoll_task_func(struct io_kiocb
*req
, bool *locked
)
5546 struct io_ring_ctx
*ctx
= req
->ctx
;
5549 ret
= io_poll_check_events(req
);
5553 io_poll_remove_entries(req
);
5554 spin_lock(&ctx
->completion_lock
);
5555 hash_del(&req
->hash_node
);
5556 spin_unlock(&ctx
->completion_lock
);
5559 io_req_task_submit(req
, locked
);
5561 io_req_complete_failed(req
, ret
);
5564 static void __io_poll_execute(struct io_kiocb
*req
, int mask
)
5567 if (req
->opcode
== IORING_OP_POLL_ADD
)
5568 req
->io_task_work
.func
= io_poll_task_func
;
5570 req
->io_task_work
.func
= io_apoll_task_func
;
5572 trace_io_uring_task_add(req
->ctx
, req
->opcode
, req
->user_data
, mask
);
5573 io_req_task_work_add(req
);
5576 static inline void io_poll_execute(struct io_kiocb
*req
, int res
)
5578 if (io_poll_get_ownership(req
))
5579 __io_poll_execute(req
, res
);
5582 static void io_poll_cancel_req(struct io_kiocb
*req
)
5584 io_poll_mark_cancelled(req
);
5585 /* kick tw, which should complete the request */
5586 io_poll_execute(req
, 0);
5589 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
5592 struct io_kiocb
*req
= wait
->private;
5593 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
5595 __poll_t mask
= key_to_poll(key
);
5597 if (unlikely(mask
& POLLFREE
)) {
5598 io_poll_mark_cancelled(req
);
5599 /* we have to kick tw in case it's not already */
5600 io_poll_execute(req
, 0);
5603 * If the waitqueue is being freed early but someone is already
5604 * holds ownership over it, we have to tear down the request as
5605 * best we can. That means immediately removing the request from
5606 * its waitqueue and preventing all further accesses to the
5607 * waitqueue via the request.
5609 list_del_init(&poll
->wait
.entry
);
5612 * Careful: this *must* be the last step, since as soon
5613 * as req->head is NULL'ed out, the request can be
5614 * completed and freed, since aio_poll_complete_work()
5615 * will no longer need to take the waitqueue lock.
5617 smp_store_release(&poll
->head
, NULL
);
5621 /* for instances that support it check for an event match first */
5622 if (mask
&& !(mask
& poll
->events
))
5625 if (io_poll_get_ownership(req
))
5626 __io_poll_execute(req
, mask
);
5630 static void __io_queue_proc(struct io_poll_iocb
*poll
, struct io_poll_table
*pt
,
5631 struct wait_queue_head
*head
,
5632 struct io_poll_iocb
**poll_ptr
)
5634 struct io_kiocb
*req
= pt
->req
;
5637 * The file being polled uses multiple waitqueues for poll handling
5638 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5641 if (unlikely(pt
->nr_entries
)) {
5642 struct io_poll_iocb
*first
= poll
;
5644 /* double add on the same waitqueue head, ignore */
5645 if (first
->head
== head
)
5647 /* already have a 2nd entry, fail a third attempt */
5649 if ((*poll_ptr
)->head
== head
)
5651 pt
->error
= -EINVAL
;
5655 poll
= kmalloc(sizeof(*poll
), GFP_ATOMIC
);
5657 pt
->error
= -ENOMEM
;
5660 io_init_poll_iocb(poll
, first
->events
, first
->wait
.func
);
5666 poll
->wait
.private = req
;
5668 if (poll
->events
& EPOLLEXCLUSIVE
)
5669 add_wait_queue_exclusive(head
, &poll
->wait
);
5671 add_wait_queue(head
, &poll
->wait
);
5674 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5675 struct poll_table_struct
*p
)
5677 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5679 __io_queue_proc(&pt
->req
->poll
, pt
, head
,
5680 (struct io_poll_iocb
**) &pt
->req
->async_data
);
5683 static int __io_arm_poll_handler(struct io_kiocb
*req
,
5684 struct io_poll_iocb
*poll
,
5685 struct io_poll_table
*ipt
, __poll_t mask
)
5687 struct io_ring_ctx
*ctx
= req
->ctx
;
5689 INIT_HLIST_NODE(&req
->hash_node
);
5690 io_init_poll_iocb(poll
, mask
, io_poll_wake
);
5691 poll
->file
= req
->file
;
5692 poll
->wait
.private = req
;
5694 ipt
->pt
._key
= mask
;
5697 ipt
->nr_entries
= 0;
5700 * Take the ownership to delay any tw execution up until we're done
5701 * with poll arming. see io_poll_get_ownership().
5703 atomic_set(&req
->poll_refs
, 1);
5704 mask
= vfs_poll(req
->file
, &ipt
->pt
) & poll
->events
;
5706 if (mask
&& (poll
->events
& EPOLLONESHOT
)) {
5707 io_poll_remove_entries(req
);
5708 /* no one else has access to the req, forget about the ref */
5711 if (!mask
&& unlikely(ipt
->error
|| !ipt
->nr_entries
)) {
5712 io_poll_remove_entries(req
);
5714 ipt
->error
= -EINVAL
;
5718 spin_lock(&ctx
->completion_lock
);
5719 io_poll_req_insert(req
);
5720 spin_unlock(&ctx
->completion_lock
);
5723 /* can't multishot if failed, just queue the event we've got */
5724 if (unlikely(ipt
->error
|| !ipt
->nr_entries
)) {
5725 poll
->events
|= EPOLLONESHOT
;
5728 __io_poll_execute(req
, mask
);
5733 * Try to release ownership. If we see a change of state, e.g.
5734 * poll was waken up, queue up a tw, it'll deal with it.
5736 if (atomic_cmpxchg(&req
->poll_refs
, 1, 0) != 1)
5737 __io_poll_execute(req
, 0);
5741 static void io_async_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
5742 struct poll_table_struct
*p
)
5744 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
5745 struct async_poll
*apoll
= pt
->req
->apoll
;
5747 __io_queue_proc(&apoll
->poll
, pt
, head
, &apoll
->double_poll
);
5756 static int io_arm_poll_handler(struct io_kiocb
*req
)
5758 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
5759 struct io_ring_ctx
*ctx
= req
->ctx
;
5760 struct async_poll
*apoll
;
5761 struct io_poll_table ipt
;
5762 __poll_t mask
= EPOLLONESHOT
| POLLERR
| POLLPRI
;
5765 if (!req
->file
|| !file_can_poll(req
->file
))
5766 return IO_APOLL_ABORTED
;
5767 if (req
->flags
& REQ_F_POLLED
)
5768 return IO_APOLL_ABORTED
;
5769 if (!def
->pollin
&& !def
->pollout
)
5770 return IO_APOLL_ABORTED
;
5773 mask
|= POLLIN
| POLLRDNORM
;
5775 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5776 if ((req
->opcode
== IORING_OP_RECVMSG
) &&
5777 (req
->sr_msg
.msg_flags
& MSG_ERRQUEUE
))
5780 mask
|= POLLOUT
| POLLWRNORM
;
5783 apoll
= kmalloc(sizeof(*apoll
), GFP_ATOMIC
);
5784 if (unlikely(!apoll
))
5785 return IO_APOLL_ABORTED
;
5786 apoll
->double_poll
= NULL
;
5788 req
->flags
|= REQ_F_POLLED
;
5789 ipt
.pt
._qproc
= io_async_queue_proc
;
5791 ret
= __io_arm_poll_handler(req
, &apoll
->poll
, &ipt
, mask
);
5792 if (ret
|| ipt
.error
)
5793 return ret
? IO_APOLL_READY
: IO_APOLL_ABORTED
;
5795 trace_io_uring_poll_arm(ctx
, req
, req
->opcode
, req
->user_data
,
5796 mask
, apoll
->poll
.events
);
5801 * Returns true if we found and killed one or more poll requests
5803 static bool io_poll_remove_all(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
5806 struct hlist_node
*tmp
;
5807 struct io_kiocb
*req
;
5811 spin_lock(&ctx
->completion_lock
);
5812 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
5813 struct hlist_head
*list
;
5815 list
= &ctx
->cancel_hash
[i
];
5816 hlist_for_each_entry_safe(req
, tmp
, list
, hash_node
) {
5817 if (io_match_task_safe(req
, tsk
, cancel_all
)) {
5818 hlist_del_init(&req
->hash_node
);
5819 io_poll_cancel_req(req
);
5824 spin_unlock(&ctx
->completion_lock
);
5828 static struct io_kiocb
*io_poll_find(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5830 __must_hold(&ctx
->completion_lock
)
5832 struct hlist_head
*list
;
5833 struct io_kiocb
*req
;
5835 list
= &ctx
->cancel_hash
[hash_long(sqe_addr
, ctx
->cancel_hash_bits
)];
5836 hlist_for_each_entry(req
, list
, hash_node
) {
5837 if (sqe_addr
!= req
->user_data
)
5839 if (poll_only
&& req
->opcode
!= IORING_OP_POLL_ADD
)
5846 static bool io_poll_disarm(struct io_kiocb
*req
)
5847 __must_hold(&ctx
->completion_lock
)
5849 if (!io_poll_get_ownership(req
))
5851 io_poll_remove_entries(req
);
5852 hash_del(&req
->hash_node
);
5856 static int io_poll_cancel(struct io_ring_ctx
*ctx
, __u64 sqe_addr
,
5858 __must_hold(&ctx
->completion_lock
)
5860 struct io_kiocb
*req
= io_poll_find(ctx
, sqe_addr
, poll_only
);
5864 io_poll_cancel_req(req
);
5868 static __poll_t
io_poll_parse_events(const struct io_uring_sqe
*sqe
,
5873 events
= READ_ONCE(sqe
->poll32_events
);
5875 events
= swahw32(events
);
5877 if (!(flags
& IORING_POLL_ADD_MULTI
))
5878 events
|= EPOLLONESHOT
;
5879 return demangle_poll(events
) | (events
& (EPOLLEXCLUSIVE
|EPOLLONESHOT
));
5882 static int io_poll_update_prep(struct io_kiocb
*req
,
5883 const struct io_uring_sqe
*sqe
)
5885 struct io_poll_update
*upd
= &req
->poll_update
;
5888 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5890 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->splice_fd_in
)
5892 flags
= READ_ONCE(sqe
->len
);
5893 if (flags
& ~(IORING_POLL_UPDATE_EVENTS
| IORING_POLL_UPDATE_USER_DATA
|
5894 IORING_POLL_ADD_MULTI
))
5896 /* meaningless without update */
5897 if (flags
== IORING_POLL_ADD_MULTI
)
5900 upd
->old_user_data
= READ_ONCE(sqe
->addr
);
5901 upd
->update_events
= flags
& IORING_POLL_UPDATE_EVENTS
;
5902 upd
->update_user_data
= flags
& IORING_POLL_UPDATE_USER_DATA
;
5904 upd
->new_user_data
= READ_ONCE(sqe
->off
);
5905 if (!upd
->update_user_data
&& upd
->new_user_data
)
5907 if (upd
->update_events
)
5908 upd
->events
= io_poll_parse_events(sqe
, flags
);
5909 else if (sqe
->poll32_events
)
5915 static int io_poll_add_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
5917 struct io_poll_iocb
*poll
= &req
->poll
;
5920 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
5922 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->off
|| sqe
->addr
)
5924 flags
= READ_ONCE(sqe
->len
);
5925 if (flags
& ~IORING_POLL_ADD_MULTI
)
5928 io_req_set_refcount(req
);
5929 poll
->events
= io_poll_parse_events(sqe
, flags
);
5933 static int io_poll_add(struct io_kiocb
*req
, unsigned int issue_flags
)
5935 struct io_poll_iocb
*poll
= &req
->poll
;
5936 struct io_poll_table ipt
;
5939 ipt
.pt
._qproc
= io_poll_queue_proc
;
5941 ret
= __io_arm_poll_handler(req
, &req
->poll
, &ipt
, poll
->events
);
5942 if (!ret
&& ipt
.error
)
5944 ret
= ret
?: ipt
.error
;
5946 __io_req_complete(req
, issue_flags
, ret
, 0);
5950 static int io_poll_update(struct io_kiocb
*req
, unsigned int issue_flags
)
5952 struct io_ring_ctx
*ctx
= req
->ctx
;
5953 struct io_kiocb
*preq
;
5956 spin_lock(&ctx
->completion_lock
);
5957 preq
= io_poll_find(ctx
, req
->poll_update
.old_user_data
, true);
5958 if (!preq
|| !io_poll_disarm(preq
)) {
5959 spin_unlock(&ctx
->completion_lock
);
5960 ret
= preq
? -EALREADY
: -ENOENT
;
5963 spin_unlock(&ctx
->completion_lock
);
5965 if (req
->poll_update
.update_events
|| req
->poll_update
.update_user_data
) {
5966 /* only mask one event flags, keep behavior flags */
5967 if (req
->poll_update
.update_events
) {
5968 preq
->poll
.events
&= ~0xffff;
5969 preq
->poll
.events
|= req
->poll_update
.events
& 0xffff;
5970 preq
->poll
.events
|= IO_POLL_UNMASK
;
5972 if (req
->poll_update
.update_user_data
)
5973 preq
->user_data
= req
->poll_update
.new_user_data
;
5975 ret2
= io_poll_add(preq
, issue_flags
);
5976 /* successfully updated, don't complete poll request */
5981 io_req_complete(preq
, -ECANCELED
);
5985 /* complete update request, we're done with it */
5986 io_req_complete(req
, ret
);
5990 static void io_req_task_timeout(struct io_kiocb
*req
, bool *locked
)
5993 io_req_complete_post(req
, -ETIME
, 0);
5996 static enum hrtimer_restart
io_timeout_fn(struct hrtimer
*timer
)
5998 struct io_timeout_data
*data
= container_of(timer
,
5999 struct io_timeout_data
, timer
);
6000 struct io_kiocb
*req
= data
->req
;
6001 struct io_ring_ctx
*ctx
= req
->ctx
;
6002 unsigned long flags
;
6004 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
6005 list_del_init(&req
->timeout
.list
);
6006 atomic_set(&req
->ctx
->cq_timeouts
,
6007 atomic_read(&req
->ctx
->cq_timeouts
) + 1);
6008 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
6010 req
->io_task_work
.func
= io_req_task_timeout
;
6011 io_req_task_work_add(req
);
6012 return HRTIMER_NORESTART
;
6015 static struct io_kiocb
*io_timeout_extract(struct io_ring_ctx
*ctx
,
6017 __must_hold(&ctx
->timeout_lock
)
6019 struct io_timeout_data
*io
;
6020 struct io_kiocb
*req
;
6023 list_for_each_entry(req
, &ctx
->timeout_list
, timeout
.list
) {
6024 found
= user_data
== req
->user_data
;
6029 return ERR_PTR(-ENOENT
);
6031 io
= req
->async_data
;
6032 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6033 return ERR_PTR(-EALREADY
);
6034 list_del_init(&req
->timeout
.list
);
6038 static int io_timeout_cancel(struct io_ring_ctx
*ctx
, __u64 user_data
)
6039 __must_hold(&ctx
->completion_lock
)
6040 __must_hold(&ctx
->timeout_lock
)
6042 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6045 return PTR_ERR(req
);
6048 io_fill_cqe_req(req
, -ECANCELED
, 0);
6049 io_put_req_deferred(req
);
6053 static clockid_t
io_timeout_get_clock(struct io_timeout_data
*data
)
6055 switch (data
->flags
& IORING_TIMEOUT_CLOCK_MASK
) {
6056 case IORING_TIMEOUT_BOOTTIME
:
6057 return CLOCK_BOOTTIME
;
6058 case IORING_TIMEOUT_REALTIME
:
6059 return CLOCK_REALTIME
;
6061 /* can't happen, vetted at prep time */
6065 return CLOCK_MONOTONIC
;
6069 static int io_linked_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6070 struct timespec64
*ts
, enum hrtimer_mode mode
)
6071 __must_hold(&ctx
->timeout_lock
)
6073 struct io_timeout_data
*io
;
6074 struct io_kiocb
*req
;
6077 list_for_each_entry(req
, &ctx
->ltimeout_list
, timeout
.list
) {
6078 found
= user_data
== req
->user_data
;
6085 io
= req
->async_data
;
6086 if (hrtimer_try_to_cancel(&io
->timer
) == -1)
6088 hrtimer_init(&io
->timer
, io_timeout_get_clock(io
), mode
);
6089 io
->timer
.function
= io_link_timeout_fn
;
6090 hrtimer_start(&io
->timer
, timespec64_to_ktime(*ts
), mode
);
6094 static int io_timeout_update(struct io_ring_ctx
*ctx
, __u64 user_data
,
6095 struct timespec64
*ts
, enum hrtimer_mode mode
)
6096 __must_hold(&ctx
->timeout_lock
)
6098 struct io_kiocb
*req
= io_timeout_extract(ctx
, user_data
);
6099 struct io_timeout_data
*data
;
6102 return PTR_ERR(req
);
6104 req
->timeout
.off
= 0; /* noseq */
6105 data
= req
->async_data
;
6106 list_add_tail(&req
->timeout
.list
, &ctx
->timeout_list
);
6107 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), mode
);
6108 data
->timer
.function
= io_timeout_fn
;
6109 hrtimer_start(&data
->timer
, timespec64_to_ktime(*ts
), mode
);
6113 static int io_timeout_remove_prep(struct io_kiocb
*req
,
6114 const struct io_uring_sqe
*sqe
)
6116 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6118 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6120 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6122 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
|| sqe
->splice_fd_in
)
6125 tr
->ltimeout
= false;
6126 tr
->addr
= READ_ONCE(sqe
->addr
);
6127 tr
->flags
= READ_ONCE(sqe
->timeout_flags
);
6128 if (tr
->flags
& IORING_TIMEOUT_UPDATE_MASK
) {
6129 if (hweight32(tr
->flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6131 if (tr
->flags
& IORING_LINK_TIMEOUT_UPDATE
)
6132 tr
->ltimeout
= true;
6133 if (tr
->flags
& ~(IORING_TIMEOUT_UPDATE_MASK
|IORING_TIMEOUT_ABS
))
6135 if (get_timespec64(&tr
->ts
, u64_to_user_ptr(sqe
->addr2
)))
6137 } else if (tr
->flags
) {
6138 /* timeout removal doesn't support flags */
6145 static inline enum hrtimer_mode
io_translate_timeout_mode(unsigned int flags
)
6147 return (flags
& IORING_TIMEOUT_ABS
) ? HRTIMER_MODE_ABS
6152 * Remove or update an existing timeout command
6154 static int io_timeout_remove(struct io_kiocb
*req
, unsigned int issue_flags
)
6156 struct io_timeout_rem
*tr
= &req
->timeout_rem
;
6157 struct io_ring_ctx
*ctx
= req
->ctx
;
6160 if (!(req
->timeout_rem
.flags
& IORING_TIMEOUT_UPDATE
)) {
6161 spin_lock(&ctx
->completion_lock
);
6162 spin_lock_irq(&ctx
->timeout_lock
);
6163 ret
= io_timeout_cancel(ctx
, tr
->addr
);
6164 spin_unlock_irq(&ctx
->timeout_lock
);
6165 spin_unlock(&ctx
->completion_lock
);
6167 enum hrtimer_mode mode
= io_translate_timeout_mode(tr
->flags
);
6169 spin_lock_irq(&ctx
->timeout_lock
);
6171 ret
= io_linked_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6173 ret
= io_timeout_update(ctx
, tr
->addr
, &tr
->ts
, mode
);
6174 spin_unlock_irq(&ctx
->timeout_lock
);
6179 io_req_complete_post(req
, ret
, 0);
6183 static int io_timeout_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
6184 bool is_timeout_link
)
6186 struct io_timeout_data
*data
;
6188 u32 off
= READ_ONCE(sqe
->off
);
6190 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6192 if (sqe
->ioprio
|| sqe
->buf_index
|| sqe
->len
!= 1 ||
6195 if (off
&& is_timeout_link
)
6197 flags
= READ_ONCE(sqe
->timeout_flags
);
6198 if (flags
& ~(IORING_TIMEOUT_ABS
| IORING_TIMEOUT_CLOCK_MASK
))
6200 /* more than one clock specified is invalid, obviously */
6201 if (hweight32(flags
& IORING_TIMEOUT_CLOCK_MASK
) > 1)
6204 INIT_LIST_HEAD(&req
->timeout
.list
);
6205 req
->timeout
.off
= off
;
6206 if (unlikely(off
&& !req
->ctx
->off_timeout_used
))
6207 req
->ctx
->off_timeout_used
= true;
6209 if (!req
->async_data
&& io_alloc_async_data(req
))
6212 data
= req
->async_data
;
6214 data
->flags
= flags
;
6216 if (get_timespec64(&data
->ts
, u64_to_user_ptr(sqe
->addr
)))
6219 INIT_LIST_HEAD(&req
->timeout
.list
);
6220 data
->mode
= io_translate_timeout_mode(flags
);
6221 hrtimer_init(&data
->timer
, io_timeout_get_clock(data
), data
->mode
);
6223 if (is_timeout_link
) {
6224 struct io_submit_link
*link
= &req
->ctx
->submit_state
.link
;
6228 if (link
->last
->opcode
== IORING_OP_LINK_TIMEOUT
)
6230 req
->timeout
.head
= link
->last
;
6231 link
->last
->flags
|= REQ_F_ARM_LTIMEOUT
;
6236 static int io_timeout(struct io_kiocb
*req
, unsigned int issue_flags
)
6238 struct io_ring_ctx
*ctx
= req
->ctx
;
6239 struct io_timeout_data
*data
= req
->async_data
;
6240 struct list_head
*entry
;
6241 u32 tail
, off
= req
->timeout
.off
;
6243 spin_lock_irq(&ctx
->timeout_lock
);
6246 * sqe->off holds how many events that need to occur for this
6247 * timeout event to be satisfied. If it isn't set, then this is
6248 * a pure timeout request, sequence isn't used.
6250 if (io_is_timeout_noseq(req
)) {
6251 entry
= ctx
->timeout_list
.prev
;
6255 tail
= ctx
->cached_cq_tail
- atomic_read(&ctx
->cq_timeouts
);
6256 req
->timeout
.target_seq
= tail
+ off
;
6258 /* Update the last seq here in case io_flush_timeouts() hasn't.
6259 * This is safe because ->completion_lock is held, and submissions
6260 * and completions are never mixed in the same ->completion_lock section.
6262 ctx
->cq_last_tm_flush
= tail
;
6265 * Insertion sort, ensuring the first entry in the list is always
6266 * the one we need first.
6268 list_for_each_prev(entry
, &ctx
->timeout_list
) {
6269 struct io_kiocb
*nxt
= list_entry(entry
, struct io_kiocb
,
6272 if (io_is_timeout_noseq(nxt
))
6274 /* nxt.seq is behind @tail, otherwise would've been completed */
6275 if (off
>= nxt
->timeout
.target_seq
- tail
)
6279 list_add(&req
->timeout
.list
, entry
);
6280 data
->timer
.function
= io_timeout_fn
;
6281 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
), data
->mode
);
6282 spin_unlock_irq(&ctx
->timeout_lock
);
6286 struct io_cancel_data
{
6287 struct io_ring_ctx
*ctx
;
6291 static bool io_cancel_cb(struct io_wq_work
*work
, void *data
)
6293 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6294 struct io_cancel_data
*cd
= data
;
6296 return req
->ctx
== cd
->ctx
&& req
->user_data
== cd
->user_data
;
6299 static int io_async_cancel_one(struct io_uring_task
*tctx
, u64 user_data
,
6300 struct io_ring_ctx
*ctx
)
6302 struct io_cancel_data data
= { .ctx
= ctx
, .user_data
= user_data
, };
6303 enum io_wq_cancel cancel_ret
;
6306 if (!tctx
|| !tctx
->io_wq
)
6309 cancel_ret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_cb
, &data
, false);
6310 switch (cancel_ret
) {
6311 case IO_WQ_CANCEL_OK
:
6314 case IO_WQ_CANCEL_RUNNING
:
6317 case IO_WQ_CANCEL_NOTFOUND
:
6325 static int io_try_cancel_userdata(struct io_kiocb
*req
, u64 sqe_addr
)
6327 struct io_ring_ctx
*ctx
= req
->ctx
;
6330 WARN_ON_ONCE(!io_wq_current_is_worker() && req
->task
!= current
);
6332 ret
= io_async_cancel_one(req
->task
->io_uring
, sqe_addr
, ctx
);
6336 spin_lock(&ctx
->completion_lock
);
6337 spin_lock_irq(&ctx
->timeout_lock
);
6338 ret
= io_timeout_cancel(ctx
, sqe_addr
);
6339 spin_unlock_irq(&ctx
->timeout_lock
);
6342 ret
= io_poll_cancel(ctx
, sqe_addr
, false);
6344 spin_unlock(&ctx
->completion_lock
);
6348 static int io_async_cancel_prep(struct io_kiocb
*req
,
6349 const struct io_uring_sqe
*sqe
)
6351 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6353 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6355 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->cancel_flags
||
6359 req
->cancel
.addr
= READ_ONCE(sqe
->addr
);
6363 static int io_async_cancel(struct io_kiocb
*req
, unsigned int issue_flags
)
6365 struct io_ring_ctx
*ctx
= req
->ctx
;
6366 u64 sqe_addr
= req
->cancel
.addr
;
6367 struct io_tctx_node
*node
;
6370 ret
= io_try_cancel_userdata(req
, sqe_addr
);
6374 /* slow path, try all io-wq's */
6375 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6377 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
6378 struct io_uring_task
*tctx
= node
->task
->io_uring
;
6380 ret
= io_async_cancel_one(tctx
, req
->cancel
.addr
, ctx
);
6384 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6388 io_req_complete_post(req
, ret
, 0);
6392 static int io_rsrc_update_prep(struct io_kiocb
*req
,
6393 const struct io_uring_sqe
*sqe
)
6395 if (unlikely(req
->flags
& (REQ_F_FIXED_FILE
| REQ_F_BUFFER_SELECT
)))
6397 if (sqe
->ioprio
|| sqe
->rw_flags
|| sqe
->splice_fd_in
)
6400 req
->rsrc_update
.offset
= READ_ONCE(sqe
->off
);
6401 req
->rsrc_update
.nr_args
= READ_ONCE(sqe
->len
);
6402 if (!req
->rsrc_update
.nr_args
)
6404 req
->rsrc_update
.arg
= READ_ONCE(sqe
->addr
);
6408 static int io_files_update(struct io_kiocb
*req
, unsigned int issue_flags
)
6410 struct io_ring_ctx
*ctx
= req
->ctx
;
6411 struct io_uring_rsrc_update2 up
;
6414 up
.offset
= req
->rsrc_update
.offset
;
6415 up
.data
= req
->rsrc_update
.arg
;
6421 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6422 ret
= __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
,
6423 &up
, req
->rsrc_update
.nr_args
);
6424 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
6428 __io_req_complete(req
, issue_flags
, ret
, 0);
6432 static int io_req_prep(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
6434 switch (req
->opcode
) {
6437 case IORING_OP_READV
:
6438 case IORING_OP_READ_FIXED
:
6439 case IORING_OP_READ
:
6440 return io_read_prep(req
, sqe
);
6441 case IORING_OP_WRITEV
:
6442 case IORING_OP_WRITE_FIXED
:
6443 case IORING_OP_WRITE
:
6444 return io_write_prep(req
, sqe
);
6445 case IORING_OP_POLL_ADD
:
6446 return io_poll_add_prep(req
, sqe
);
6447 case IORING_OP_POLL_REMOVE
:
6448 return io_poll_update_prep(req
, sqe
);
6449 case IORING_OP_FSYNC
:
6450 return io_fsync_prep(req
, sqe
);
6451 case IORING_OP_SYNC_FILE_RANGE
:
6452 return io_sfr_prep(req
, sqe
);
6453 case IORING_OP_SENDMSG
:
6454 case IORING_OP_SEND
:
6455 return io_sendmsg_prep(req
, sqe
);
6456 case IORING_OP_RECVMSG
:
6457 case IORING_OP_RECV
:
6458 return io_recvmsg_prep(req
, sqe
);
6459 case IORING_OP_CONNECT
:
6460 return io_connect_prep(req
, sqe
);
6461 case IORING_OP_TIMEOUT
:
6462 return io_timeout_prep(req
, sqe
, false);
6463 case IORING_OP_TIMEOUT_REMOVE
:
6464 return io_timeout_remove_prep(req
, sqe
);
6465 case IORING_OP_ASYNC_CANCEL
:
6466 return io_async_cancel_prep(req
, sqe
);
6467 case IORING_OP_LINK_TIMEOUT
:
6468 return io_timeout_prep(req
, sqe
, true);
6469 case IORING_OP_ACCEPT
:
6470 return io_accept_prep(req
, sqe
);
6471 case IORING_OP_FALLOCATE
:
6472 return io_fallocate_prep(req
, sqe
);
6473 case IORING_OP_OPENAT
:
6474 return io_openat_prep(req
, sqe
);
6475 case IORING_OP_CLOSE
:
6476 return io_close_prep(req
, sqe
);
6477 case IORING_OP_FILES_UPDATE
:
6478 return io_rsrc_update_prep(req
, sqe
);
6479 case IORING_OP_STATX
:
6480 return io_statx_prep(req
, sqe
);
6481 case IORING_OP_FADVISE
:
6482 return io_fadvise_prep(req
, sqe
);
6483 case IORING_OP_MADVISE
:
6484 return io_madvise_prep(req
, sqe
);
6485 case IORING_OP_OPENAT2
:
6486 return io_openat2_prep(req
, sqe
);
6487 case IORING_OP_EPOLL_CTL
:
6488 return io_epoll_ctl_prep(req
, sqe
);
6489 case IORING_OP_SPLICE
:
6490 return io_splice_prep(req
, sqe
);
6491 case IORING_OP_PROVIDE_BUFFERS
:
6492 return io_provide_buffers_prep(req
, sqe
);
6493 case IORING_OP_REMOVE_BUFFERS
:
6494 return io_remove_buffers_prep(req
, sqe
);
6496 return io_tee_prep(req
, sqe
);
6497 case IORING_OP_SHUTDOWN
:
6498 return io_shutdown_prep(req
, sqe
);
6499 case IORING_OP_RENAMEAT
:
6500 return io_renameat_prep(req
, sqe
);
6501 case IORING_OP_UNLINKAT
:
6502 return io_unlinkat_prep(req
, sqe
);
6503 case IORING_OP_MKDIRAT
:
6504 return io_mkdirat_prep(req
, sqe
);
6505 case IORING_OP_SYMLINKAT
:
6506 return io_symlinkat_prep(req
, sqe
);
6507 case IORING_OP_LINKAT
:
6508 return io_linkat_prep(req
, sqe
);
6511 printk_once(KERN_WARNING
"io_uring: unhandled opcode %d\n",
6516 static int io_req_prep_async(struct io_kiocb
*req
)
6518 if (!io_op_defs
[req
->opcode
].needs_async_setup
)
6520 if (WARN_ON_ONCE(req
->async_data
))
6522 if (io_alloc_async_data(req
))
6525 switch (req
->opcode
) {
6526 case IORING_OP_READV
:
6527 return io_rw_prep_async(req
, READ
);
6528 case IORING_OP_WRITEV
:
6529 return io_rw_prep_async(req
, WRITE
);
6530 case IORING_OP_SENDMSG
:
6531 return io_sendmsg_prep_async(req
);
6532 case IORING_OP_RECVMSG
:
6533 return io_recvmsg_prep_async(req
);
6534 case IORING_OP_CONNECT
:
6535 return io_connect_prep_async(req
);
6537 printk_once(KERN_WARNING
"io_uring: prep_async() bad opcode %d\n",
6542 static u32
io_get_sequence(struct io_kiocb
*req
)
6544 u32 seq
= req
->ctx
->cached_sq_head
;
6546 /* need original cached_sq_head, but it was increased for each req */
6547 io_for_each_link(req
, req
)
6552 static bool io_drain_req(struct io_kiocb
*req
)
6554 struct io_kiocb
*pos
;
6555 struct io_ring_ctx
*ctx
= req
->ctx
;
6556 struct io_defer_entry
*de
;
6560 if (req
->flags
& REQ_F_FAIL
) {
6561 io_req_complete_fail_submit(req
);
6566 * If we need to drain a request in the middle of a link, drain the
6567 * head request and the next request/link after the current link.
6568 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6569 * maintained for every request of our link.
6571 if (ctx
->drain_next
) {
6572 req
->flags
|= REQ_F_IO_DRAIN
;
6573 ctx
->drain_next
= false;
6575 /* not interested in head, start from the first linked */
6576 io_for_each_link(pos
, req
->link
) {
6577 if (pos
->flags
& REQ_F_IO_DRAIN
) {
6578 ctx
->drain_next
= true;
6579 req
->flags
|= REQ_F_IO_DRAIN
;
6584 /* Still need defer if there is pending req in defer list. */
6585 spin_lock(&ctx
->completion_lock
);
6586 if (likely(list_empty_careful(&ctx
->defer_list
) &&
6587 !(req
->flags
& REQ_F_IO_DRAIN
))) {
6588 spin_unlock(&ctx
->completion_lock
);
6589 ctx
->drain_active
= false;
6592 spin_unlock(&ctx
->completion_lock
);
6594 seq
= io_get_sequence(req
);
6595 /* Still a chance to pass the sequence check */
6596 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
))
6599 ret
= io_req_prep_async(req
);
6602 io_prep_async_link(req
);
6603 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
6607 io_req_complete_failed(req
, ret
);
6611 spin_lock(&ctx
->completion_lock
);
6612 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
6613 spin_unlock(&ctx
->completion_lock
);
6615 io_queue_async_work(req
, NULL
);
6619 trace_io_uring_defer(ctx
, req
, req
->user_data
);
6622 list_add_tail(&de
->list
, &ctx
->defer_list
);
6623 spin_unlock(&ctx
->completion_lock
);
6627 static void io_clean_op(struct io_kiocb
*req
)
6629 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
6630 switch (req
->opcode
) {
6631 case IORING_OP_READV
:
6632 case IORING_OP_READ_FIXED
:
6633 case IORING_OP_READ
:
6634 kfree((void *)(unsigned long)req
->rw
.addr
);
6636 case IORING_OP_RECVMSG
:
6637 case IORING_OP_RECV
:
6638 kfree(req
->sr_msg
.kbuf
);
6643 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
6644 switch (req
->opcode
) {
6645 case IORING_OP_READV
:
6646 case IORING_OP_READ_FIXED
:
6647 case IORING_OP_READ
:
6648 case IORING_OP_WRITEV
:
6649 case IORING_OP_WRITE_FIXED
:
6650 case IORING_OP_WRITE
: {
6651 struct io_async_rw
*io
= req
->async_data
;
6653 kfree(io
->free_iovec
);
6656 case IORING_OP_RECVMSG
:
6657 case IORING_OP_SENDMSG
: {
6658 struct io_async_msghdr
*io
= req
->async_data
;
6660 kfree(io
->free_iov
);
6663 case IORING_OP_OPENAT
:
6664 case IORING_OP_OPENAT2
:
6665 if (req
->open
.filename
)
6666 putname(req
->open
.filename
);
6668 case IORING_OP_RENAMEAT
:
6669 putname(req
->rename
.oldpath
);
6670 putname(req
->rename
.newpath
);
6672 case IORING_OP_UNLINKAT
:
6673 putname(req
->unlink
.filename
);
6675 case IORING_OP_MKDIRAT
:
6676 putname(req
->mkdir
.filename
);
6678 case IORING_OP_SYMLINKAT
:
6679 putname(req
->symlink
.oldpath
);
6680 putname(req
->symlink
.newpath
);
6682 case IORING_OP_LINKAT
:
6683 putname(req
->hardlink
.oldpath
);
6684 putname(req
->hardlink
.newpath
);
6688 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
6689 kfree(req
->apoll
->double_poll
);
6693 if (req
->flags
& REQ_F_INFLIGHT
) {
6694 struct io_uring_task
*tctx
= req
->task
->io_uring
;
6696 atomic_dec(&tctx
->inflight_tracked
);
6698 if (req
->flags
& REQ_F_CREDS
)
6699 put_cred(req
->creds
);
6701 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
6704 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
6706 struct io_ring_ctx
*ctx
= req
->ctx
;
6707 const struct cred
*creds
= NULL
;
6710 if ((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred())
6711 creds
= override_creds(req
->creds
);
6713 switch (req
->opcode
) {
6715 ret
= io_nop(req
, issue_flags
);
6717 case IORING_OP_READV
:
6718 case IORING_OP_READ_FIXED
:
6719 case IORING_OP_READ
:
6720 ret
= io_read(req
, issue_flags
);
6722 case IORING_OP_WRITEV
:
6723 case IORING_OP_WRITE_FIXED
:
6724 case IORING_OP_WRITE
:
6725 ret
= io_write(req
, issue_flags
);
6727 case IORING_OP_FSYNC
:
6728 ret
= io_fsync(req
, issue_flags
);
6730 case IORING_OP_POLL_ADD
:
6731 ret
= io_poll_add(req
, issue_flags
);
6733 case IORING_OP_POLL_REMOVE
:
6734 ret
= io_poll_update(req
, issue_flags
);
6736 case IORING_OP_SYNC_FILE_RANGE
:
6737 ret
= io_sync_file_range(req
, issue_flags
);
6739 case IORING_OP_SENDMSG
:
6740 ret
= io_sendmsg(req
, issue_flags
);
6742 case IORING_OP_SEND
:
6743 ret
= io_send(req
, issue_flags
);
6745 case IORING_OP_RECVMSG
:
6746 ret
= io_recvmsg(req
, issue_flags
);
6748 case IORING_OP_RECV
:
6749 ret
= io_recv(req
, issue_flags
);
6751 case IORING_OP_TIMEOUT
:
6752 ret
= io_timeout(req
, issue_flags
);
6754 case IORING_OP_TIMEOUT_REMOVE
:
6755 ret
= io_timeout_remove(req
, issue_flags
);
6757 case IORING_OP_ACCEPT
:
6758 ret
= io_accept(req
, issue_flags
);
6760 case IORING_OP_CONNECT
:
6761 ret
= io_connect(req
, issue_flags
);
6763 case IORING_OP_ASYNC_CANCEL
:
6764 ret
= io_async_cancel(req
, issue_flags
);
6766 case IORING_OP_FALLOCATE
:
6767 ret
= io_fallocate(req
, issue_flags
);
6769 case IORING_OP_OPENAT
:
6770 ret
= io_openat(req
, issue_flags
);
6772 case IORING_OP_CLOSE
:
6773 ret
= io_close(req
, issue_flags
);
6775 case IORING_OP_FILES_UPDATE
:
6776 ret
= io_files_update(req
, issue_flags
);
6778 case IORING_OP_STATX
:
6779 ret
= io_statx(req
, issue_flags
);
6781 case IORING_OP_FADVISE
:
6782 ret
= io_fadvise(req
, issue_flags
);
6784 case IORING_OP_MADVISE
:
6785 ret
= io_madvise(req
, issue_flags
);
6787 case IORING_OP_OPENAT2
:
6788 ret
= io_openat2(req
, issue_flags
);
6790 case IORING_OP_EPOLL_CTL
:
6791 ret
= io_epoll_ctl(req
, issue_flags
);
6793 case IORING_OP_SPLICE
:
6794 ret
= io_splice(req
, issue_flags
);
6796 case IORING_OP_PROVIDE_BUFFERS
:
6797 ret
= io_provide_buffers(req
, issue_flags
);
6799 case IORING_OP_REMOVE_BUFFERS
:
6800 ret
= io_remove_buffers(req
, issue_flags
);
6803 ret
= io_tee(req
, issue_flags
);
6805 case IORING_OP_SHUTDOWN
:
6806 ret
= io_shutdown(req
, issue_flags
);
6808 case IORING_OP_RENAMEAT
:
6809 ret
= io_renameat(req
, issue_flags
);
6811 case IORING_OP_UNLINKAT
:
6812 ret
= io_unlinkat(req
, issue_flags
);
6814 case IORING_OP_MKDIRAT
:
6815 ret
= io_mkdirat(req
, issue_flags
);
6817 case IORING_OP_SYMLINKAT
:
6818 ret
= io_symlinkat(req
, issue_flags
);
6820 case IORING_OP_LINKAT
:
6821 ret
= io_linkat(req
, issue_flags
);
6829 revert_creds(creds
);
6832 /* If the op doesn't have a file, we're not polling for it */
6833 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
6834 io_iopoll_req_issued(req
);
6839 static struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
6841 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6843 req
= io_put_req_find_next(req
);
6844 return req
? &req
->work
: NULL
;
6847 static void io_wq_submit_work(struct io_wq_work
*work
)
6849 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
6850 struct io_kiocb
*timeout
;
6853 /* one will be dropped by ->io_free_work() after returning to io-wq */
6854 if (!(req
->flags
& REQ_F_REFCOUNT
))
6855 __io_req_set_refcount(req
, 2);
6859 timeout
= io_prep_linked_timeout(req
);
6861 io_queue_linked_timeout(timeout
);
6863 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6864 if (work
->flags
& IO_WQ_WORK_CANCEL
)
6869 ret
= io_issue_sqe(req
, 0);
6871 * We can get EAGAIN for polled IO even though we're
6872 * forcing a sync submission from here, since we can't
6873 * wait for request slots on the block side.
6875 if (ret
!= -EAGAIN
|| !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
6881 /* avoid locking problems by failing it from a clean context */
6883 io_req_task_queue_fail(req
, ret
);
6886 static inline struct io_fixed_file
*io_fixed_file_slot(struct io_file_table
*table
,
6889 return &table
->files
[i
];
6892 static inline struct file
*io_file_from_index(struct io_ring_ctx
*ctx
,
6895 struct io_fixed_file
*slot
= io_fixed_file_slot(&ctx
->file_table
, index
);
6897 return (struct file
*) (slot
->file_ptr
& FFS_MASK
);
6900 static void io_fixed_file_set(struct io_fixed_file
*file_slot
, struct file
*file
)
6902 unsigned long file_ptr
= (unsigned long) file
;
6904 if (__io_file_supports_nowait(file
, READ
))
6905 file_ptr
|= FFS_ASYNC_READ
;
6906 if (__io_file_supports_nowait(file
, WRITE
))
6907 file_ptr
|= FFS_ASYNC_WRITE
;
6908 if (S_ISREG(file_inode(file
)->i_mode
))
6909 file_ptr
|= FFS_ISREG
;
6910 file_slot
->file_ptr
= file_ptr
;
6913 static inline struct file
*io_file_get_fixed(struct io_ring_ctx
*ctx
,
6914 struct io_kiocb
*req
, int fd
)
6917 unsigned long file_ptr
;
6919 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
6921 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
6922 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
6923 file
= (struct file
*) (file_ptr
& FFS_MASK
);
6924 file_ptr
&= ~FFS_MASK
;
6925 /* mask in overlapping REQ_F and FFS bits */
6926 req
->flags
|= (file_ptr
<< REQ_F_NOWAIT_READ_BIT
);
6927 io_req_set_rsrc_node(req
);
6931 static struct file
*io_file_get_normal(struct io_ring_ctx
*ctx
,
6932 struct io_kiocb
*req
, int fd
)
6934 struct file
*file
= fget(fd
);
6936 trace_io_uring_file_get(ctx
, fd
);
6938 /* we don't allow fixed io_uring files */
6939 if (file
&& unlikely(file
->f_op
== &io_uring_fops
))
6940 io_req_track_inflight(req
);
6944 static inline struct file
*io_file_get(struct io_ring_ctx
*ctx
,
6945 struct io_kiocb
*req
, int fd
, bool fixed
)
6948 return io_file_get_fixed(ctx
, req
, fd
);
6950 return io_file_get_normal(ctx
, req
, fd
);
6953 static void io_req_task_link_timeout(struct io_kiocb
*req
, bool *locked
)
6955 struct io_kiocb
*prev
= req
->timeout
.prev
;
6959 if (!(req
->task
->flags
& PF_EXITING
))
6960 ret
= io_try_cancel_userdata(req
, prev
->user_data
);
6961 io_req_complete_post(req
, ret
?: -ETIME
, 0);
6964 io_req_complete_post(req
, -ETIME
, 0);
6968 static enum hrtimer_restart
io_link_timeout_fn(struct hrtimer
*timer
)
6970 struct io_timeout_data
*data
= container_of(timer
,
6971 struct io_timeout_data
, timer
);
6972 struct io_kiocb
*prev
, *req
= data
->req
;
6973 struct io_ring_ctx
*ctx
= req
->ctx
;
6974 unsigned long flags
;
6976 spin_lock_irqsave(&ctx
->timeout_lock
, flags
);
6977 prev
= req
->timeout
.head
;
6978 req
->timeout
.head
= NULL
;
6981 * We don't expect the list to be empty, that will only happen if we
6982 * race with the completion of the linked work.
6985 io_remove_next_linked(prev
);
6986 if (!req_ref_inc_not_zero(prev
))
6989 list_del(&req
->timeout
.list
);
6990 req
->timeout
.prev
= prev
;
6991 spin_unlock_irqrestore(&ctx
->timeout_lock
, flags
);
6993 req
->io_task_work
.func
= io_req_task_link_timeout
;
6994 io_req_task_work_add(req
);
6995 return HRTIMER_NORESTART
;
6998 static void io_queue_linked_timeout(struct io_kiocb
*req
)
7000 struct io_ring_ctx
*ctx
= req
->ctx
;
7002 spin_lock_irq(&ctx
->timeout_lock
);
7004 * If the back reference is NULL, then our linked request finished
7005 * before we got a chance to setup the timer
7007 if (req
->timeout
.head
) {
7008 struct io_timeout_data
*data
= req
->async_data
;
7010 data
->timer
.function
= io_link_timeout_fn
;
7011 hrtimer_start(&data
->timer
, timespec64_to_ktime(data
->ts
),
7013 list_add_tail(&req
->timeout
.list
, &ctx
->ltimeout_list
);
7015 spin_unlock_irq(&ctx
->timeout_lock
);
7016 /* drop submission reference */
7020 static void __io_queue_sqe(struct io_kiocb
*req
)
7021 __must_hold(&req
->ctx
->uring_lock
)
7023 struct io_kiocb
*linked_timeout
;
7027 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
7030 * We async punt it if the file wasn't marked NOWAIT, or if the file
7031 * doesn't support non-blocking read/write attempts
7034 if (req
->flags
& REQ_F_COMPLETE_INLINE
) {
7035 struct io_ring_ctx
*ctx
= req
->ctx
;
7036 struct io_submit_state
*state
= &ctx
->submit_state
;
7038 state
->compl_reqs
[state
->compl_nr
++] = req
;
7039 if (state
->compl_nr
== ARRAY_SIZE(state
->compl_reqs
))
7040 io_submit_flush_completions(ctx
);
7044 linked_timeout
= io_prep_linked_timeout(req
);
7046 io_queue_linked_timeout(linked_timeout
);
7047 } else if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
7048 linked_timeout
= io_prep_linked_timeout(req
);
7050 switch (io_arm_poll_handler(req
)) {
7051 case IO_APOLL_READY
:
7053 io_queue_linked_timeout(linked_timeout
);
7055 case IO_APOLL_ABORTED
:
7057 * Queued up for async execution, worker will release
7058 * submit reference when the iocb is actually submitted.
7060 io_queue_async_work(req
, NULL
);
7065 io_queue_linked_timeout(linked_timeout
);
7067 io_req_complete_failed(req
, ret
);
7071 static inline void io_queue_sqe(struct io_kiocb
*req
)
7072 __must_hold(&req
->ctx
->uring_lock
)
7074 if (unlikely(req
->ctx
->drain_active
) && io_drain_req(req
))
7077 if (likely(!(req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
)))) {
7078 __io_queue_sqe(req
);
7079 } else if (req
->flags
& REQ_F_FAIL
) {
7080 io_req_complete_fail_submit(req
);
7082 int ret
= io_req_prep_async(req
);
7085 io_req_complete_failed(req
, ret
);
7087 io_queue_async_work(req
, NULL
);
7092 * Check SQE restrictions (opcode and flags).
7094 * Returns 'true' if SQE is allowed, 'false' otherwise.
7096 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
7097 struct io_kiocb
*req
,
7098 unsigned int sqe_flags
)
7100 if (likely(!ctx
->restricted
))
7103 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
7106 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
7107 ctx
->restrictions
.sqe_flags_required
)
7110 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
7111 ctx
->restrictions
.sqe_flags_required
))
7117 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7118 const struct io_uring_sqe
*sqe
)
7119 __must_hold(&ctx
->uring_lock
)
7121 struct io_submit_state
*state
;
7122 unsigned int sqe_flags
;
7123 int personality
, ret
= 0;
7125 /* req is partially pre-initialised, see io_preinit_req() */
7126 req
->opcode
= READ_ONCE(sqe
->opcode
);
7127 /* same numerical values with corresponding REQ_F_*, safe to copy */
7128 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
7129 req
->user_data
= READ_ONCE(sqe
->user_data
);
7131 req
->fixed_rsrc_refs
= NULL
;
7132 req
->task
= current
;
7134 /* enforce forwards compatibility on users */
7135 if (unlikely(sqe_flags
& ~SQE_VALID_FLAGS
))
7137 if (unlikely(req
->opcode
>= IORING_OP_LAST
))
7139 if (!io_check_restriction(ctx
, req
, sqe_flags
))
7142 if ((sqe_flags
& IOSQE_BUFFER_SELECT
) &&
7143 !io_op_defs
[req
->opcode
].buffer_select
)
7145 if (unlikely(sqe_flags
& IOSQE_IO_DRAIN
))
7146 ctx
->drain_active
= true;
7148 personality
= READ_ONCE(sqe
->personality
);
7150 req
->creds
= xa_load(&ctx
->personalities
, personality
);
7153 get_cred(req
->creds
);
7154 req
->flags
|= REQ_F_CREDS
;
7156 state
= &ctx
->submit_state
;
7159 * Plug now if we have more than 1 IO left after this, and the target
7160 * is potentially a read/write to block based storage.
7162 if (!state
->plug_started
&& state
->ios_left
> 1 &&
7163 io_op_defs
[req
->opcode
].plug
) {
7164 blk_start_plug(&state
->plug
);
7165 state
->plug_started
= true;
7168 if (io_op_defs
[req
->opcode
].needs_file
) {
7169 req
->file
= io_file_get(ctx
, req
, READ_ONCE(sqe
->fd
),
7170 (sqe_flags
& IOSQE_FIXED_FILE
));
7171 if (unlikely(!req
->file
))
7179 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
7180 const struct io_uring_sqe
*sqe
)
7181 __must_hold(&ctx
->uring_lock
)
7183 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
7186 ret
= io_init_req(ctx
, req
, sqe
);
7187 if (unlikely(ret
)) {
7189 /* fail even hard links since we don't submit */
7192 * we can judge a link req is failed or cancelled by if
7193 * REQ_F_FAIL is set, but the head is an exception since
7194 * it may be set REQ_F_FAIL because of other req's failure
7195 * so let's leverage req->result to distinguish if a head
7196 * is set REQ_F_FAIL because of its failure or other req's
7197 * failure so that we can set the correct ret code for it.
7198 * init result here to avoid affecting the normal path.
7200 if (!(link
->head
->flags
& REQ_F_FAIL
))
7201 req_fail_link_node(link
->head
, -ECANCELED
);
7202 } else if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7204 * the current req is a normal req, we should return
7205 * error and thus break the submittion loop.
7207 io_req_complete_failed(req
, ret
);
7210 req_fail_link_node(req
, ret
);
7212 ret
= io_req_prep(req
, sqe
);
7217 /* don't need @sqe from now on */
7218 trace_io_uring_submit_sqe(ctx
, req
, req
->opcode
, req
->user_data
,
7220 ctx
->flags
& IORING_SETUP_SQPOLL
);
7223 * If we already have a head request, queue this one for async
7224 * submittal once the head completes. If we don't have a head but
7225 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7226 * submitted sync once the chain is complete. If none of those
7227 * conditions are true (normal request), then just queue it.
7230 struct io_kiocb
*head
= link
->head
;
7232 if (!(req
->flags
& REQ_F_FAIL
)) {
7233 ret
= io_req_prep_async(req
);
7234 if (unlikely(ret
)) {
7235 req_fail_link_node(req
, ret
);
7236 if (!(head
->flags
& REQ_F_FAIL
))
7237 req_fail_link_node(head
, -ECANCELED
);
7240 trace_io_uring_link(ctx
, req
, head
);
7241 link
->last
->link
= req
;
7244 /* last request of a link, enqueue the link */
7245 if (!(req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))) {
7250 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
)) {
7262 * Batched submission is done, ensure local IO is flushed out.
7264 static void io_submit_state_end(struct io_submit_state
*state
,
7265 struct io_ring_ctx
*ctx
)
7267 if (state
->link
.head
)
7268 io_queue_sqe(state
->link
.head
);
7269 if (state
->compl_nr
)
7270 io_submit_flush_completions(ctx
);
7271 if (state
->plug_started
)
7272 blk_finish_plug(&state
->plug
);
7276 * Start submission side cache.
7278 static void io_submit_state_start(struct io_submit_state
*state
,
7279 unsigned int max_ios
)
7281 state
->plug_started
= false;
7282 state
->ios_left
= max_ios
;
7283 /* set only head, no need to init link_last in advance */
7284 state
->link
.head
= NULL
;
7287 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
7289 struct io_rings
*rings
= ctx
->rings
;
7292 * Ensure any loads from the SQEs are done at this point,
7293 * since once we write the new head, the application could
7294 * write new data to them.
7296 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
7300 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7301 * that is mapped by userspace. This means that care needs to be taken to
7302 * ensure that reads are stable, as we cannot rely on userspace always
7303 * being a good citizen. If members of the sqe are validated and then later
7304 * used, it's important that those reads are done through READ_ONCE() to
7305 * prevent a re-load down the line.
7307 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
7309 unsigned head
, mask
= ctx
->sq_entries
- 1;
7310 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
7313 * The cached sq head (or cq tail) serves two purposes:
7315 * 1) allows us to batch the cost of updating the user visible
7317 * 2) allows the kernel side to track the head on its own, even
7318 * though the application is the one updating it.
7320 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
7321 if (likely(head
< ctx
->sq_entries
))
7322 return &ctx
->sq_sqes
[head
];
7324 /* drop invalid entries */
7326 WRITE_ONCE(ctx
->rings
->sq_dropped
,
7327 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
7331 static int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
7332 __must_hold(&ctx
->uring_lock
)
7336 /* make sure SQ entry isn't read before tail */
7337 nr
= min3(nr
, ctx
->sq_entries
, io_sqring_entries(ctx
));
7338 if (!percpu_ref_tryget_many(&ctx
->refs
, nr
))
7340 io_get_task_refs(nr
);
7342 io_submit_state_start(&ctx
->submit_state
, nr
);
7343 while (submitted
< nr
) {
7344 const struct io_uring_sqe
*sqe
;
7345 struct io_kiocb
*req
;
7347 req
= io_alloc_req(ctx
);
7348 if (unlikely(!req
)) {
7350 submitted
= -EAGAIN
;
7353 sqe
= io_get_sqe(ctx
);
7354 if (unlikely(!sqe
)) {
7355 list_add(&req
->inflight_entry
, &ctx
->submit_state
.free_list
);
7358 /* will complete beyond this point, count as submitted */
7360 if (io_submit_sqe(ctx
, req
, sqe
))
7364 if (unlikely(submitted
!= nr
)) {
7365 int ref_used
= (submitted
== -EAGAIN
) ? 0 : submitted
;
7366 int unused
= nr
- ref_used
;
7368 current
->io_uring
->cached_refs
+= unused
;
7369 percpu_ref_put_many(&ctx
->refs
, unused
);
7372 io_submit_state_end(&ctx
->submit_state
, ctx
);
7373 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7374 io_commit_sqring(ctx
);
7379 static inline bool io_sqd_events_pending(struct io_sq_data
*sqd
)
7381 return READ_ONCE(sqd
->state
);
7384 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx
*ctx
)
7386 /* Tell userspace we may need a wakeup call */
7387 spin_lock(&ctx
->completion_lock
);
7388 WRITE_ONCE(ctx
->rings
->sq_flags
,
7389 ctx
->rings
->sq_flags
| IORING_SQ_NEED_WAKEUP
);
7390 spin_unlock(&ctx
->completion_lock
);
7393 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx
*ctx
)
7395 spin_lock(&ctx
->completion_lock
);
7396 WRITE_ONCE(ctx
->rings
->sq_flags
,
7397 ctx
->rings
->sq_flags
& ~IORING_SQ_NEED_WAKEUP
);
7398 spin_unlock(&ctx
->completion_lock
);
7401 static int __io_sq_thread(struct io_ring_ctx
*ctx
, bool cap_entries
)
7403 unsigned int to_submit
;
7406 to_submit
= io_sqring_entries(ctx
);
7407 /* if we're handling multiple rings, cap submit size for fairness */
7408 if (cap_entries
&& to_submit
> IORING_SQPOLL_CAP_ENTRIES_VALUE
)
7409 to_submit
= IORING_SQPOLL_CAP_ENTRIES_VALUE
;
7411 if (!list_empty(&ctx
->iopoll_list
) || to_submit
) {
7412 unsigned nr_events
= 0;
7413 const struct cred
*creds
= NULL
;
7415 if (ctx
->sq_creds
!= current_cred())
7416 creds
= override_creds(ctx
->sq_creds
);
7418 mutex_lock(&ctx
->uring_lock
);
7419 if (!list_empty(&ctx
->iopoll_list
))
7420 io_do_iopoll(ctx
, &nr_events
, 0);
7423 * Don't submit if refs are dying, good for io_uring_register(),
7424 * but also it is relied upon by io_ring_exit_work()
7426 if (to_submit
&& likely(!percpu_ref_is_dying(&ctx
->refs
)) &&
7427 !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
7428 ret
= io_submit_sqes(ctx
, to_submit
);
7429 mutex_unlock(&ctx
->uring_lock
);
7431 if (to_submit
&& wq_has_sleeper(&ctx
->sqo_sq_wait
))
7432 wake_up(&ctx
->sqo_sq_wait
);
7434 revert_creds(creds
);
7440 static void io_sqd_update_thread_idle(struct io_sq_data
*sqd
)
7442 struct io_ring_ctx
*ctx
;
7443 unsigned sq_thread_idle
= 0;
7445 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7446 sq_thread_idle
= max(sq_thread_idle
, ctx
->sq_thread_idle
);
7447 sqd
->sq_thread_idle
= sq_thread_idle
;
7450 static bool io_sqd_handle_event(struct io_sq_data
*sqd
)
7452 bool did_sig
= false;
7453 struct ksignal ksig
;
7455 if (test_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
) ||
7456 signal_pending(current
)) {
7457 mutex_unlock(&sqd
->lock
);
7458 if (signal_pending(current
))
7459 did_sig
= get_signal(&ksig
);
7461 mutex_lock(&sqd
->lock
);
7463 return did_sig
|| test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
7466 static int io_sq_thread(void *data
)
7468 struct io_sq_data
*sqd
= data
;
7469 struct io_ring_ctx
*ctx
;
7470 unsigned long timeout
= 0;
7471 char buf
[TASK_COMM_LEN
];
7474 snprintf(buf
, sizeof(buf
), "iou-sqp-%d", sqd
->task_pid
);
7475 set_task_comm(current
, buf
);
7477 if (sqd
->sq_cpu
!= -1)
7478 set_cpus_allowed_ptr(current
, cpumask_of(sqd
->sq_cpu
));
7480 set_cpus_allowed_ptr(current
, cpu_online_mask
);
7481 current
->flags
|= PF_NO_SETAFFINITY
;
7483 mutex_lock(&sqd
->lock
);
7485 bool cap_entries
, sqt_spin
= false;
7487 if (io_sqd_events_pending(sqd
) || signal_pending(current
)) {
7488 if (io_sqd_handle_event(sqd
))
7490 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7493 cap_entries
= !list_is_singular(&sqd
->ctx_list
);
7494 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7495 int ret
= __io_sq_thread(ctx
, cap_entries
);
7497 if (!sqt_spin
&& (ret
> 0 || !list_empty(&ctx
->iopoll_list
)))
7500 if (io_run_task_work())
7503 if (sqt_spin
|| !time_after(jiffies
, timeout
)) {
7506 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7510 prepare_to_wait(&sqd
->wait
, &wait
, TASK_INTERRUPTIBLE
);
7511 if (!io_sqd_events_pending(sqd
) && !current
->task_works
) {
7512 bool needs_sched
= true;
7514 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
) {
7515 io_ring_set_wakeup_flag(ctx
);
7517 if ((ctx
->flags
& IORING_SETUP_IOPOLL
) &&
7518 !list_empty_careful(&ctx
->iopoll_list
)) {
7519 needs_sched
= false;
7522 if (io_sqring_entries(ctx
)) {
7523 needs_sched
= false;
7529 mutex_unlock(&sqd
->lock
);
7531 mutex_lock(&sqd
->lock
);
7533 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7534 io_ring_clear_wakeup_flag(ctx
);
7537 finish_wait(&sqd
->wait
, &wait
);
7538 timeout
= jiffies
+ sqd
->sq_thread_idle
;
7541 io_uring_cancel_generic(true, sqd
);
7543 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
7544 io_ring_set_wakeup_flag(ctx
);
7546 mutex_unlock(&sqd
->lock
);
7548 complete(&sqd
->exited
);
7552 struct io_wait_queue
{
7553 struct wait_queue_entry wq
;
7554 struct io_ring_ctx
*ctx
;
7556 unsigned nr_timeouts
;
7559 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
7561 struct io_ring_ctx
*ctx
= iowq
->ctx
;
7562 int dist
= ctx
->cached_cq_tail
- (int) iowq
->cq_tail
;
7565 * Wake up if we have enough events, or if a timeout occurred since we
7566 * started waiting. For timeouts, we always want to return to userspace,
7567 * regardless of event count.
7569 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
7572 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
7573 int wake_flags
, void *key
)
7575 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
7579 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7580 * the task, and the next invocation will do it.
7582 if (io_should_wake(iowq
) || test_bit(0, &iowq
->ctx
->check_cq_overflow
))
7583 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
7587 static int io_run_task_work_sig(void)
7589 if (io_run_task_work())
7591 if (!signal_pending(current
))
7593 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
7594 return -ERESTARTSYS
;
7598 /* when returns >0, the caller should retry */
7599 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
7600 struct io_wait_queue
*iowq
,
7605 /* make sure we run task_work before checking for signals */
7606 ret
= io_run_task_work_sig();
7607 if (ret
|| io_should_wake(iowq
))
7609 /* let the caller flush overflows, retry */
7610 if (test_bit(0, &ctx
->check_cq_overflow
))
7613 if (!schedule_hrtimeout(&timeout
, HRTIMER_MODE_ABS
))
7619 * Wait until events become available, if we don't already have some. The
7620 * application must reap them itself, as they reside on the shared cq ring.
7622 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
7623 const sigset_t __user
*sig
, size_t sigsz
,
7624 struct __kernel_timespec __user
*uts
)
7626 struct io_wait_queue iowq
;
7627 struct io_rings
*rings
= ctx
->rings
;
7628 ktime_t timeout
= KTIME_MAX
;
7632 io_cqring_overflow_flush(ctx
);
7633 if (io_cqring_events(ctx
) >= min_events
)
7635 if (!io_run_task_work())
7640 struct timespec64 ts
;
7642 if (get_timespec64(&ts
, uts
))
7644 timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
7648 #ifdef CONFIG_COMPAT
7649 if (in_compat_syscall())
7650 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
7654 ret
= set_user_sigmask(sig
, sigsz
);
7660 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
7661 iowq
.wq
.private = current
;
7662 INIT_LIST_HEAD(&iowq
.wq
.entry
);
7664 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
7665 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
7667 trace_io_uring_cqring_wait(ctx
, min_events
);
7669 /* if we can't even flush overflow, don't wait for more */
7670 if (!io_cqring_overflow_flush(ctx
)) {
7674 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
7675 TASK_INTERRUPTIBLE
);
7676 ret
= io_cqring_wait_schedule(ctx
, &iowq
, timeout
);
7677 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
7681 restore_saved_sigmask_unless(ret
== -EINTR
);
7683 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
7686 static void io_free_page_table(void **table
, size_t size
)
7688 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7690 for (i
= 0; i
< nr_tables
; i
++)
7695 static void **io_alloc_page_table(size_t size
)
7697 unsigned i
, nr_tables
= DIV_ROUND_UP(size
, PAGE_SIZE
);
7698 size_t init_size
= size
;
7701 table
= kcalloc(nr_tables
, sizeof(*table
), GFP_KERNEL_ACCOUNT
);
7705 for (i
= 0; i
< nr_tables
; i
++) {
7706 unsigned int this_size
= min_t(size_t, size
, PAGE_SIZE
);
7708 table
[i
] = kzalloc(this_size
, GFP_KERNEL_ACCOUNT
);
7710 io_free_page_table(table
, init_size
);
7718 static void io_rsrc_node_destroy(struct io_rsrc_node
*ref_node
)
7720 percpu_ref_exit(&ref_node
->refs
);
7724 static void io_rsrc_node_ref_zero(struct percpu_ref
*ref
)
7726 struct io_rsrc_node
*node
= container_of(ref
, struct io_rsrc_node
, refs
);
7727 struct io_ring_ctx
*ctx
= node
->rsrc_data
->ctx
;
7728 unsigned long flags
;
7729 bool first_add
= false;
7730 unsigned long delay
= HZ
;
7732 spin_lock_irqsave(&ctx
->rsrc_ref_lock
, flags
);
7735 /* if we are mid-quiesce then do not delay */
7736 if (node
->rsrc_data
->quiesce
)
7739 while (!list_empty(&ctx
->rsrc_ref_list
)) {
7740 node
= list_first_entry(&ctx
->rsrc_ref_list
,
7741 struct io_rsrc_node
, node
);
7742 /* recycle ref nodes in order */
7745 list_del(&node
->node
);
7746 first_add
|= llist_add(&node
->llist
, &ctx
->rsrc_put_llist
);
7748 spin_unlock_irqrestore(&ctx
->rsrc_ref_lock
, flags
);
7751 mod_delayed_work(system_wq
, &ctx
->rsrc_put_work
, delay
);
7754 static struct io_rsrc_node
*io_rsrc_node_alloc(struct io_ring_ctx
*ctx
)
7756 struct io_rsrc_node
*ref_node
;
7758 ref_node
= kzalloc(sizeof(*ref_node
), GFP_KERNEL
);
7762 if (percpu_ref_init(&ref_node
->refs
, io_rsrc_node_ref_zero
,
7767 INIT_LIST_HEAD(&ref_node
->node
);
7768 INIT_LIST_HEAD(&ref_node
->rsrc_list
);
7769 ref_node
->done
= false;
7773 static void io_rsrc_node_switch(struct io_ring_ctx
*ctx
,
7774 struct io_rsrc_data
*data_to_kill
)
7776 WARN_ON_ONCE(!ctx
->rsrc_backup_node
);
7777 WARN_ON_ONCE(data_to_kill
&& !ctx
->rsrc_node
);
7780 struct io_rsrc_node
*rsrc_node
= ctx
->rsrc_node
;
7782 rsrc_node
->rsrc_data
= data_to_kill
;
7783 spin_lock_irq(&ctx
->rsrc_ref_lock
);
7784 list_add_tail(&rsrc_node
->node
, &ctx
->rsrc_ref_list
);
7785 spin_unlock_irq(&ctx
->rsrc_ref_lock
);
7787 atomic_inc(&data_to_kill
->refs
);
7788 percpu_ref_kill(&rsrc_node
->refs
);
7789 ctx
->rsrc_node
= NULL
;
7792 if (!ctx
->rsrc_node
) {
7793 ctx
->rsrc_node
= ctx
->rsrc_backup_node
;
7794 ctx
->rsrc_backup_node
= NULL
;
7798 static int io_rsrc_node_switch_start(struct io_ring_ctx
*ctx
)
7800 if (ctx
->rsrc_backup_node
)
7802 ctx
->rsrc_backup_node
= io_rsrc_node_alloc(ctx
);
7803 return ctx
->rsrc_backup_node
? 0 : -ENOMEM
;
7806 static int io_rsrc_ref_quiesce(struct io_rsrc_data
*data
, struct io_ring_ctx
*ctx
)
7810 /* As we may drop ->uring_lock, other task may have started quiesce */
7814 data
->quiesce
= true;
7816 ret
= io_rsrc_node_switch_start(ctx
);
7819 io_rsrc_node_switch(ctx
, data
);
7821 /* kill initial ref, already quiesced if zero */
7822 if (atomic_dec_and_test(&data
->refs
))
7824 mutex_unlock(&ctx
->uring_lock
);
7825 flush_delayed_work(&ctx
->rsrc_put_work
);
7826 ret
= wait_for_completion_interruptible(&data
->done
);
7828 mutex_lock(&ctx
->uring_lock
);
7829 if (atomic_read(&data
->refs
) > 0) {
7831 * it has been revived by another thread while
7834 mutex_unlock(&ctx
->uring_lock
);
7840 atomic_inc(&data
->refs
);
7841 /* wait for all works potentially completing data->done */
7842 flush_delayed_work(&ctx
->rsrc_put_work
);
7843 reinit_completion(&data
->done
);
7845 ret
= io_run_task_work_sig();
7846 mutex_lock(&ctx
->uring_lock
);
7848 data
->quiesce
= false;
7853 static u64
*io_get_tag_slot(struct io_rsrc_data
*data
, unsigned int idx
)
7855 unsigned int off
= idx
& IO_RSRC_TAG_TABLE_MASK
;
7856 unsigned int table_idx
= idx
>> IO_RSRC_TAG_TABLE_SHIFT
;
7858 return &data
->tags
[table_idx
][off
];
7861 static void io_rsrc_data_free(struct io_rsrc_data
*data
)
7863 size_t size
= data
->nr
* sizeof(data
->tags
[0][0]);
7866 io_free_page_table((void **)data
->tags
, size
);
7870 static int io_rsrc_data_alloc(struct io_ring_ctx
*ctx
, rsrc_put_fn
*do_put
,
7871 u64 __user
*utags
, unsigned nr
,
7872 struct io_rsrc_data
**pdata
)
7874 struct io_rsrc_data
*data
;
7878 data
= kzalloc(sizeof(*data
), GFP_KERNEL
);
7881 data
->tags
= (u64
**)io_alloc_page_table(nr
* sizeof(data
->tags
[0][0]));
7889 data
->do_put
= do_put
;
7892 for (i
= 0; i
< nr
; i
++) {
7893 u64
*tag_slot
= io_get_tag_slot(data
, i
);
7895 if (copy_from_user(tag_slot
, &utags
[i
],
7901 atomic_set(&data
->refs
, 1);
7902 init_completion(&data
->done
);
7906 io_rsrc_data_free(data
);
7910 static bool io_alloc_file_tables(struct io_file_table
*table
, unsigned nr_files
)
7912 table
->files
= kvcalloc(nr_files
, sizeof(table
->files
[0]),
7913 GFP_KERNEL_ACCOUNT
);
7914 return !!table
->files
;
7917 static void io_free_file_tables(struct io_file_table
*table
)
7919 kvfree(table
->files
);
7920 table
->files
= NULL
;
7923 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7925 #if defined(CONFIG_UNIX)
7926 if (ctx
->ring_sock
) {
7927 struct sock
*sock
= ctx
->ring_sock
->sk
;
7928 struct sk_buff
*skb
;
7930 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
7936 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
7939 file
= io_file_from_index(ctx
, i
);
7944 io_free_file_tables(&ctx
->file_table
);
7945 io_rsrc_data_free(ctx
->file_data
);
7946 ctx
->file_data
= NULL
;
7947 ctx
->nr_user_files
= 0;
7950 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
7952 unsigned nr
= ctx
->nr_user_files
;
7955 if (!ctx
->file_data
)
7959 * Quiesce may unlock ->uring_lock, and while it's not held
7960 * prevent new requests using the table.
7962 ctx
->nr_user_files
= 0;
7963 ret
= io_rsrc_ref_quiesce(ctx
->file_data
, ctx
);
7964 ctx
->nr_user_files
= nr
;
7966 __io_sqe_files_unregister(ctx
);
7970 static void io_sq_thread_unpark(struct io_sq_data
*sqd
)
7971 __releases(&sqd
->lock
)
7973 WARN_ON_ONCE(sqd
->thread
== current
);
7976 * Do the dance but not conditional clear_bit() because it'd race with
7977 * other threads incrementing park_pending and setting the bit.
7979 clear_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7980 if (atomic_dec_return(&sqd
->park_pending
))
7981 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7982 mutex_unlock(&sqd
->lock
);
7985 static void io_sq_thread_park(struct io_sq_data
*sqd
)
7986 __acquires(&sqd
->lock
)
7988 WARN_ON_ONCE(sqd
->thread
== current
);
7990 atomic_inc(&sqd
->park_pending
);
7991 set_bit(IO_SQ_THREAD_SHOULD_PARK
, &sqd
->state
);
7992 mutex_lock(&sqd
->lock
);
7994 wake_up_process(sqd
->thread
);
7997 static void io_sq_thread_stop(struct io_sq_data
*sqd
)
7999 WARN_ON_ONCE(sqd
->thread
== current
);
8000 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
));
8002 set_bit(IO_SQ_THREAD_SHOULD_STOP
, &sqd
->state
);
8003 mutex_lock(&sqd
->lock
);
8005 wake_up_process(sqd
->thread
);
8006 mutex_unlock(&sqd
->lock
);
8007 wait_for_completion(&sqd
->exited
);
8010 static void io_put_sq_data(struct io_sq_data
*sqd
)
8012 if (refcount_dec_and_test(&sqd
->refs
)) {
8013 WARN_ON_ONCE(atomic_read(&sqd
->park_pending
));
8015 io_sq_thread_stop(sqd
);
8020 static void io_sq_thread_finish(struct io_ring_ctx
*ctx
)
8022 struct io_sq_data
*sqd
= ctx
->sq_data
;
8025 io_sq_thread_park(sqd
);
8026 list_del_init(&ctx
->sqd_list
);
8027 io_sqd_update_thread_idle(sqd
);
8028 io_sq_thread_unpark(sqd
);
8030 io_put_sq_data(sqd
);
8031 ctx
->sq_data
= NULL
;
8035 static struct io_sq_data
*io_attach_sq_data(struct io_uring_params
*p
)
8037 struct io_ring_ctx
*ctx_attach
;
8038 struct io_sq_data
*sqd
;
8041 f
= fdget(p
->wq_fd
);
8043 return ERR_PTR(-ENXIO
);
8044 if (f
.file
->f_op
!= &io_uring_fops
) {
8046 return ERR_PTR(-EINVAL
);
8049 ctx_attach
= f
.file
->private_data
;
8050 sqd
= ctx_attach
->sq_data
;
8053 return ERR_PTR(-EINVAL
);
8055 if (sqd
->task_tgid
!= current
->tgid
) {
8057 return ERR_PTR(-EPERM
);
8060 refcount_inc(&sqd
->refs
);
8065 static struct io_sq_data
*io_get_sq_data(struct io_uring_params
*p
,
8068 struct io_sq_data
*sqd
;
8071 if (p
->flags
& IORING_SETUP_ATTACH_WQ
) {
8072 sqd
= io_attach_sq_data(p
);
8077 /* fall through for EPERM case, setup new sqd/task */
8078 if (PTR_ERR(sqd
) != -EPERM
)
8082 sqd
= kzalloc(sizeof(*sqd
), GFP_KERNEL
);
8084 return ERR_PTR(-ENOMEM
);
8086 atomic_set(&sqd
->park_pending
, 0);
8087 refcount_set(&sqd
->refs
, 1);
8088 INIT_LIST_HEAD(&sqd
->ctx_list
);
8089 mutex_init(&sqd
->lock
);
8090 init_waitqueue_head(&sqd
->wait
);
8091 init_completion(&sqd
->exited
);
8095 #if defined(CONFIG_UNIX)
8097 * Ensure the UNIX gc is aware of our file set, so we are certain that
8098 * the io_uring can be safely unregistered on process exit, even if we have
8099 * loops in the file referencing.
8101 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
8103 struct sock
*sk
= ctx
->ring_sock
->sk
;
8104 struct scm_fp_list
*fpl
;
8105 struct sk_buff
*skb
;
8108 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
8112 skb
= alloc_skb(0, GFP_KERNEL
);
8121 fpl
->user
= get_uid(current_user());
8122 for (i
= 0; i
< nr
; i
++) {
8123 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8127 fpl
->fp
[nr_files
] = get_file(file
);
8128 unix_inflight(fpl
->user
, fpl
->fp
[nr_files
]);
8133 fpl
->max
= SCM_MAX_FD
;
8134 fpl
->count
= nr_files
;
8135 UNIXCB(skb
).fp
= fpl
;
8136 skb
->scm_io_uring
= 1;
8137 skb
->destructor
= unix_destruct_scm
;
8138 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
8139 skb_queue_head(&sk
->sk_receive_queue
, skb
);
8141 for (i
= 0; i
< nr
; i
++) {
8142 struct file
*file
= io_file_from_index(ctx
, i
+ offset
);
8149 free_uid(fpl
->user
);
8157 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8158 * causes regular reference counting to break down. We rely on the UNIX
8159 * garbage collection to take care of this problem for us.
8161 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8163 unsigned left
, total
;
8167 left
= ctx
->nr_user_files
;
8169 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
8171 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
8175 total
+= this_files
;
8181 while (total
< ctx
->nr_user_files
) {
8182 struct file
*file
= io_file_from_index(ctx
, total
);
8192 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
8198 static void io_rsrc_file_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8200 struct file
*file
= prsrc
->file
;
8201 #if defined(CONFIG_UNIX)
8202 struct sock
*sock
= ctx
->ring_sock
->sk
;
8203 struct sk_buff_head list
, *head
= &sock
->sk_receive_queue
;
8204 struct sk_buff
*skb
;
8207 __skb_queue_head_init(&list
);
8210 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8211 * remove this entry and rearrange the file array.
8213 skb
= skb_dequeue(head
);
8215 struct scm_fp_list
*fp
;
8217 fp
= UNIXCB(skb
).fp
;
8218 for (i
= 0; i
< fp
->count
; i
++) {
8221 if (fp
->fp
[i
] != file
)
8224 unix_notinflight(fp
->user
, fp
->fp
[i
]);
8225 left
= fp
->count
- 1 - i
;
8227 memmove(&fp
->fp
[i
], &fp
->fp
[i
+ 1],
8228 left
* sizeof(struct file
*));
8235 __skb_queue_tail(&list
, skb
);
8245 __skb_queue_tail(&list
, skb
);
8247 skb
= skb_dequeue(head
);
8250 if (skb_peek(&list
)) {
8251 spin_lock_irq(&head
->lock
);
8252 while ((skb
= __skb_dequeue(&list
)) != NULL
)
8253 __skb_queue_tail(head
, skb
);
8254 spin_unlock_irq(&head
->lock
);
8261 static void __io_rsrc_put_work(struct io_rsrc_node
*ref_node
)
8263 struct io_rsrc_data
*rsrc_data
= ref_node
->rsrc_data
;
8264 struct io_ring_ctx
*ctx
= rsrc_data
->ctx
;
8265 struct io_rsrc_put
*prsrc
, *tmp
;
8267 list_for_each_entry_safe(prsrc
, tmp
, &ref_node
->rsrc_list
, list
) {
8268 list_del(&prsrc
->list
);
8271 bool lock_ring
= ctx
->flags
& IORING_SETUP_IOPOLL
;
8273 io_ring_submit_lock(ctx
, lock_ring
);
8274 spin_lock(&ctx
->completion_lock
);
8275 io_fill_cqe_aux(ctx
, prsrc
->tag
, 0, 0);
8276 io_commit_cqring(ctx
);
8277 spin_unlock(&ctx
->completion_lock
);
8278 io_cqring_ev_posted(ctx
);
8279 io_ring_submit_unlock(ctx
, lock_ring
);
8282 rsrc_data
->do_put(ctx
, prsrc
);
8286 io_rsrc_node_destroy(ref_node
);
8287 if (atomic_dec_and_test(&rsrc_data
->refs
))
8288 complete(&rsrc_data
->done
);
8291 static void io_rsrc_put_work(struct work_struct
*work
)
8293 struct io_ring_ctx
*ctx
;
8294 struct llist_node
*node
;
8296 ctx
= container_of(work
, struct io_ring_ctx
, rsrc_put_work
.work
);
8297 node
= llist_del_all(&ctx
->rsrc_put_llist
);
8300 struct io_rsrc_node
*ref_node
;
8301 struct llist_node
*next
= node
->next
;
8303 ref_node
= llist_entry(node
, struct io_rsrc_node
, llist
);
8304 __io_rsrc_put_work(ref_node
);
8309 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
8310 unsigned nr_args
, u64 __user
*tags
)
8312 __s32 __user
*fds
= (__s32 __user
*) arg
;
8321 if (nr_args
> IORING_MAX_FIXED_FILES
)
8323 if (nr_args
> rlimit(RLIMIT_NOFILE
))
8325 ret
= io_rsrc_node_switch_start(ctx
);
8328 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_file_put
, tags
, nr_args
,
8334 if (!io_alloc_file_tables(&ctx
->file_table
, nr_args
))
8337 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_files
++) {
8338 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
))) {
8342 /* allow sparse sets */
8345 if (unlikely(*io_get_tag_slot(ctx
->file_data
, i
)))
8352 if (unlikely(!file
))
8356 * Don't allow io_uring instances to be registered. If UNIX
8357 * isn't enabled, then this causes a reference cycle and this
8358 * instance can never get freed. If UNIX is enabled we'll
8359 * handle it just fine, but there's still no point in allowing
8360 * a ring fd as it doesn't support regular read/write anyway.
8362 if (file
->f_op
== &io_uring_fops
) {
8366 io_fixed_file_set(io_fixed_file_slot(&ctx
->file_table
, i
), file
);
8369 ret
= io_sqe_files_scm(ctx
);
8371 __io_sqe_files_unregister(ctx
);
8375 io_rsrc_node_switch(ctx
, NULL
);
8378 for (i
= 0; i
< ctx
->nr_user_files
; i
++) {
8379 file
= io_file_from_index(ctx
, i
);
8383 io_free_file_tables(&ctx
->file_table
);
8384 ctx
->nr_user_files
= 0;
8386 io_rsrc_data_free(ctx
->file_data
);
8387 ctx
->file_data
= NULL
;
8391 static int io_sqe_file_register(struct io_ring_ctx
*ctx
, struct file
*file
,
8394 #if defined(CONFIG_UNIX)
8395 struct sock
*sock
= ctx
->ring_sock
->sk
;
8396 struct sk_buff_head
*head
= &sock
->sk_receive_queue
;
8397 struct sk_buff
*skb
;
8400 * See if we can merge this file into an existing skb SCM_RIGHTS
8401 * file set. If there's no room, fall back to allocating a new skb
8402 * and filling it in.
8404 spin_lock_irq(&head
->lock
);
8405 skb
= skb_peek(head
);
8407 struct scm_fp_list
*fpl
= UNIXCB(skb
).fp
;
8409 if (fpl
->count
< SCM_MAX_FD
) {
8410 __skb_unlink(skb
, head
);
8411 spin_unlock_irq(&head
->lock
);
8412 fpl
->fp
[fpl
->count
] = get_file(file
);
8413 unix_inflight(fpl
->user
, fpl
->fp
[fpl
->count
]);
8415 spin_lock_irq(&head
->lock
);
8416 __skb_queue_head(head
, skb
);
8421 spin_unlock_irq(&head
->lock
);
8428 return __io_sqe_files_scm(ctx
, 1, index
);
8434 static int io_queue_rsrc_removal(struct io_rsrc_data
*data
, unsigned idx
,
8435 struct io_rsrc_node
*node
, void *rsrc
)
8437 u64
*tag_slot
= io_get_tag_slot(data
, idx
);
8438 struct io_rsrc_put
*prsrc
;
8440 prsrc
= kzalloc(sizeof(*prsrc
), GFP_KERNEL
);
8444 prsrc
->tag
= *tag_slot
;
8447 list_add(&prsrc
->list
, &node
->rsrc_list
);
8451 static int io_install_fixed_file(struct io_kiocb
*req
, struct file
*file
,
8452 unsigned int issue_flags
, u32 slot_index
)
8454 struct io_ring_ctx
*ctx
= req
->ctx
;
8455 bool force_nonblock
= issue_flags
& IO_URING_F_NONBLOCK
;
8456 bool needs_switch
= false;
8457 struct io_fixed_file
*file_slot
;
8460 io_ring_submit_lock(ctx
, !force_nonblock
);
8461 if (file
->f_op
== &io_uring_fops
)
8464 if (!ctx
->file_data
)
8467 if (slot_index
>= ctx
->nr_user_files
)
8470 slot_index
= array_index_nospec(slot_index
, ctx
->nr_user_files
);
8471 file_slot
= io_fixed_file_slot(&ctx
->file_table
, slot_index
);
8473 if (file_slot
->file_ptr
) {
8474 struct file
*old_file
;
8476 ret
= io_rsrc_node_switch_start(ctx
);
8480 old_file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8481 ret
= io_queue_rsrc_removal(ctx
->file_data
, slot_index
,
8482 ctx
->rsrc_node
, old_file
);
8485 file_slot
->file_ptr
= 0;
8486 needs_switch
= true;
8489 *io_get_tag_slot(ctx
->file_data
, slot_index
) = 0;
8490 io_fixed_file_set(file_slot
, file
);
8491 ret
= io_sqe_file_register(ctx
, file
, slot_index
);
8493 file_slot
->file_ptr
= 0;
8500 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8501 io_ring_submit_unlock(ctx
, !force_nonblock
);
8507 static int io_close_fixed(struct io_kiocb
*req
, unsigned int issue_flags
)
8509 unsigned int offset
= req
->close
.file_slot
- 1;
8510 struct io_ring_ctx
*ctx
= req
->ctx
;
8511 struct io_fixed_file
*file_slot
;
8515 io_ring_submit_lock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8517 if (unlikely(!ctx
->file_data
))
8520 if (offset
>= ctx
->nr_user_files
)
8522 ret
= io_rsrc_node_switch_start(ctx
);
8526 offset
= array_index_nospec(offset
, ctx
->nr_user_files
);
8527 file_slot
= io_fixed_file_slot(&ctx
->file_table
, offset
);
8529 if (!file_slot
->file_ptr
)
8532 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8533 ret
= io_queue_rsrc_removal(ctx
->file_data
, offset
, ctx
->rsrc_node
, file
);
8537 file_slot
->file_ptr
= 0;
8538 io_rsrc_node_switch(ctx
, ctx
->file_data
);
8541 io_ring_submit_unlock(ctx
, !(issue_flags
& IO_URING_F_NONBLOCK
));
8545 static int __io_sqe_files_update(struct io_ring_ctx
*ctx
,
8546 struct io_uring_rsrc_update2
*up
,
8549 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
8550 __s32 __user
*fds
= u64_to_user_ptr(up
->data
);
8551 struct io_rsrc_data
*data
= ctx
->file_data
;
8552 struct io_fixed_file
*file_slot
;
8556 bool needs_switch
= false;
8558 if (!ctx
->file_data
)
8560 if (up
->offset
+ nr_args
> ctx
->nr_user_files
)
8563 for (done
= 0; done
< nr_args
; done
++) {
8566 if ((tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) ||
8567 copy_from_user(&fd
, &fds
[done
], sizeof(fd
))) {
8571 if ((fd
== IORING_REGISTER_FILES_SKIP
|| fd
== -1) && tag
) {
8575 if (fd
== IORING_REGISTER_FILES_SKIP
)
8578 i
= array_index_nospec(up
->offset
+ done
, ctx
->nr_user_files
);
8579 file_slot
= io_fixed_file_slot(&ctx
->file_table
, i
);
8581 if (file_slot
->file_ptr
) {
8582 file
= (struct file
*)(file_slot
->file_ptr
& FFS_MASK
);
8583 err
= io_queue_rsrc_removal(data
, i
, ctx
->rsrc_node
, file
);
8586 file_slot
->file_ptr
= 0;
8587 needs_switch
= true;
8596 * Don't allow io_uring instances to be registered. If
8597 * UNIX isn't enabled, then this causes a reference
8598 * cycle and this instance can never get freed. If UNIX
8599 * is enabled we'll handle it just fine, but there's
8600 * still no point in allowing a ring fd as it doesn't
8601 * support regular read/write anyway.
8603 if (file
->f_op
== &io_uring_fops
) {
8608 *io_get_tag_slot(data
, i
) = tag
;
8609 io_fixed_file_set(file_slot
, file
);
8610 err
= io_sqe_file_register(ctx
, file
, i
);
8612 file_slot
->file_ptr
= 0;
8620 io_rsrc_node_switch(ctx
, data
);
8621 return done
? done
: err
;
8624 static struct io_wq
*io_init_wq_offload(struct io_ring_ctx
*ctx
,
8625 struct task_struct
*task
)
8627 struct io_wq_hash
*hash
;
8628 struct io_wq_data data
;
8629 unsigned int concurrency
;
8631 mutex_lock(&ctx
->uring_lock
);
8632 hash
= ctx
->hash_map
;
8634 hash
= kzalloc(sizeof(*hash
), GFP_KERNEL
);
8636 mutex_unlock(&ctx
->uring_lock
);
8637 return ERR_PTR(-ENOMEM
);
8639 refcount_set(&hash
->refs
, 1);
8640 init_waitqueue_head(&hash
->wait
);
8641 ctx
->hash_map
= hash
;
8643 mutex_unlock(&ctx
->uring_lock
);
8647 data
.free_work
= io_wq_free_work
;
8648 data
.do_work
= io_wq_submit_work
;
8650 /* Do QD, or 4 * CPUS, whatever is smallest */
8651 concurrency
= min(ctx
->sq_entries
, 4 * num_online_cpus());
8653 return io_wq_create(concurrency
, &data
);
8656 static int io_uring_alloc_task_context(struct task_struct
*task
,
8657 struct io_ring_ctx
*ctx
)
8659 struct io_uring_task
*tctx
;
8662 tctx
= kzalloc(sizeof(*tctx
), GFP_KERNEL
);
8663 if (unlikely(!tctx
))
8666 ret
= percpu_counter_init(&tctx
->inflight
, 0, GFP_KERNEL
);
8667 if (unlikely(ret
)) {
8672 tctx
->io_wq
= io_init_wq_offload(ctx
, task
);
8673 if (IS_ERR(tctx
->io_wq
)) {
8674 ret
= PTR_ERR(tctx
->io_wq
);
8675 percpu_counter_destroy(&tctx
->inflight
);
8681 init_waitqueue_head(&tctx
->wait
);
8682 atomic_set(&tctx
->in_idle
, 0);
8683 atomic_set(&tctx
->inflight_tracked
, 0);
8684 task
->io_uring
= tctx
;
8685 spin_lock_init(&tctx
->task_lock
);
8686 INIT_WQ_LIST(&tctx
->task_list
);
8687 init_task_work(&tctx
->task_work
, tctx_task_work
);
8691 void __io_uring_free(struct task_struct
*tsk
)
8693 struct io_uring_task
*tctx
= tsk
->io_uring
;
8695 WARN_ON_ONCE(!xa_empty(&tctx
->xa
));
8696 WARN_ON_ONCE(tctx
->io_wq
);
8697 WARN_ON_ONCE(tctx
->cached_refs
);
8699 percpu_counter_destroy(&tctx
->inflight
);
8701 tsk
->io_uring
= NULL
;
8704 static int io_sq_offload_create(struct io_ring_ctx
*ctx
,
8705 struct io_uring_params
*p
)
8709 /* Retain compatibility with failing for an invalid attach attempt */
8710 if ((ctx
->flags
& (IORING_SETUP_ATTACH_WQ
| IORING_SETUP_SQPOLL
)) ==
8711 IORING_SETUP_ATTACH_WQ
) {
8714 f
= fdget(p
->wq_fd
);
8717 if (f
.file
->f_op
!= &io_uring_fops
) {
8723 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
8724 struct task_struct
*tsk
;
8725 struct io_sq_data
*sqd
;
8728 sqd
= io_get_sq_data(p
, &attached
);
8734 ctx
->sq_creds
= get_current_cred();
8736 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
8737 if (!ctx
->sq_thread_idle
)
8738 ctx
->sq_thread_idle
= HZ
;
8740 io_sq_thread_park(sqd
);
8741 list_add(&ctx
->sqd_list
, &sqd
->ctx_list
);
8742 io_sqd_update_thread_idle(sqd
);
8743 /* don't attach to a dying SQPOLL thread, would be racy */
8744 ret
= (attached
&& !sqd
->thread
) ? -ENXIO
: 0;
8745 io_sq_thread_unpark(sqd
);
8752 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8753 int cpu
= p
->sq_thread_cpu
;
8756 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
8763 sqd
->task_pid
= current
->pid
;
8764 sqd
->task_tgid
= current
->tgid
;
8765 tsk
= create_io_thread(io_sq_thread
, sqd
, NUMA_NO_NODE
);
8772 ret
= io_uring_alloc_task_context(tsk
, ctx
);
8773 wake_up_new_task(tsk
);
8776 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
8777 /* Can't have SQ_AFF without SQPOLL */
8784 complete(&ctx
->sq_data
->exited
);
8786 io_sq_thread_finish(ctx
);
8790 static inline void __io_unaccount_mem(struct user_struct
*user
,
8791 unsigned long nr_pages
)
8793 atomic_long_sub(nr_pages
, &user
->locked_vm
);
8796 static inline int __io_account_mem(struct user_struct
*user
,
8797 unsigned long nr_pages
)
8799 unsigned long page_limit
, cur_pages
, new_pages
;
8801 /* Don't allow more pages than we can safely lock */
8802 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
8805 cur_pages
= atomic_long_read(&user
->locked_vm
);
8806 new_pages
= cur_pages
+ nr_pages
;
8807 if (new_pages
> page_limit
)
8809 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
8810 new_pages
) != cur_pages
);
8815 static void io_unaccount_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8818 __io_unaccount_mem(ctx
->user
, nr_pages
);
8820 if (ctx
->mm_account
)
8821 atomic64_sub(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8824 static int io_account_mem(struct io_ring_ctx
*ctx
, unsigned long nr_pages
)
8829 ret
= __io_account_mem(ctx
->user
, nr_pages
);
8834 if (ctx
->mm_account
)
8835 atomic64_add(nr_pages
, &ctx
->mm_account
->pinned_vm
);
8840 static void io_mem_free(void *ptr
)
8847 page
= virt_to_head_page(ptr
);
8848 if (put_page_testzero(page
))
8849 free_compound_page(page
);
8852 static void *io_mem_alloc(size_t size
)
8854 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
8856 return (void *) __get_free_pages(gfp
, get_order(size
));
8859 static unsigned long rings_size(unsigned sq_entries
, unsigned cq_entries
,
8862 struct io_rings
*rings
;
8863 size_t off
, sq_array_size
;
8865 off
= struct_size(rings
, cqes
, cq_entries
);
8866 if (off
== SIZE_MAX
)
8870 off
= ALIGN(off
, SMP_CACHE_BYTES
);
8878 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
8879 if (sq_array_size
== SIZE_MAX
)
8882 if (check_add_overflow(off
, sq_array_size
, &off
))
8888 static void io_buffer_unmap(struct io_ring_ctx
*ctx
, struct io_mapped_ubuf
**slot
)
8890 struct io_mapped_ubuf
*imu
= *slot
;
8893 if (imu
!= ctx
->dummy_ubuf
) {
8894 for (i
= 0; i
< imu
->nr_bvecs
; i
++)
8895 unpin_user_page(imu
->bvec
[i
].bv_page
);
8896 if (imu
->acct_pages
)
8897 io_unaccount_mem(ctx
, imu
->acct_pages
);
8903 static void io_rsrc_buf_put(struct io_ring_ctx
*ctx
, struct io_rsrc_put
*prsrc
)
8905 io_buffer_unmap(ctx
, &prsrc
->buf
);
8909 static void __io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8913 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++)
8914 io_buffer_unmap(ctx
, &ctx
->user_bufs
[i
]);
8915 kfree(ctx
->user_bufs
);
8916 io_rsrc_data_free(ctx
->buf_data
);
8917 ctx
->user_bufs
= NULL
;
8918 ctx
->buf_data
= NULL
;
8919 ctx
->nr_user_bufs
= 0;
8922 static int io_sqe_buffers_unregister(struct io_ring_ctx
*ctx
)
8924 unsigned nr
= ctx
->nr_user_bufs
;
8931 * Quiesce may unlock ->uring_lock, and while it's not held
8932 * prevent new requests using the table.
8934 ctx
->nr_user_bufs
= 0;
8935 ret
= io_rsrc_ref_quiesce(ctx
->buf_data
, ctx
);
8936 ctx
->nr_user_bufs
= nr
;
8938 __io_sqe_buffers_unregister(ctx
);
8942 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
8943 void __user
*arg
, unsigned index
)
8945 struct iovec __user
*src
;
8947 #ifdef CONFIG_COMPAT
8949 struct compat_iovec __user
*ciovs
;
8950 struct compat_iovec ciov
;
8952 ciovs
= (struct compat_iovec __user
*) arg
;
8953 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
8956 dst
->iov_base
= u64_to_user_ptr((u64
)ciov
.iov_base
);
8957 dst
->iov_len
= ciov
.iov_len
;
8961 src
= (struct iovec __user
*) arg
;
8962 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
8968 * Not super efficient, but this is just a registration time. And we do cache
8969 * the last compound head, so generally we'll only do a full search if we don't
8972 * We check if the given compound head page has already been accounted, to
8973 * avoid double accounting it. This allows us to account the full size of the
8974 * page, not just the constituent pages of a huge page.
8976 static bool headpage_already_acct(struct io_ring_ctx
*ctx
, struct page
**pages
,
8977 int nr_pages
, struct page
*hpage
)
8981 /* check current page array */
8982 for (i
= 0; i
< nr_pages
; i
++) {
8983 if (!PageCompound(pages
[i
]))
8985 if (compound_head(pages
[i
]) == hpage
)
8989 /* check previously registered pages */
8990 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
8991 struct io_mapped_ubuf
*imu
= ctx
->user_bufs
[i
];
8993 for (j
= 0; j
< imu
->nr_bvecs
; j
++) {
8994 if (!PageCompound(imu
->bvec
[j
].bv_page
))
8996 if (compound_head(imu
->bvec
[j
].bv_page
) == hpage
)
9004 static int io_buffer_account_pin(struct io_ring_ctx
*ctx
, struct page
**pages
,
9005 int nr_pages
, struct io_mapped_ubuf
*imu
,
9006 struct page
**last_hpage
)
9010 imu
->acct_pages
= 0;
9011 for (i
= 0; i
< nr_pages
; i
++) {
9012 if (!PageCompound(pages
[i
])) {
9017 hpage
= compound_head(pages
[i
]);
9018 if (hpage
== *last_hpage
)
9020 *last_hpage
= hpage
;
9021 if (headpage_already_acct(ctx
, pages
, i
, hpage
))
9023 imu
->acct_pages
+= page_size(hpage
) >> PAGE_SHIFT
;
9027 if (!imu
->acct_pages
)
9030 ret
= io_account_mem(ctx
, imu
->acct_pages
);
9032 imu
->acct_pages
= 0;
9036 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, struct iovec
*iov
,
9037 struct io_mapped_ubuf
**pimu
,
9038 struct page
**last_hpage
)
9040 struct io_mapped_ubuf
*imu
= NULL
;
9041 struct vm_area_struct
**vmas
= NULL
;
9042 struct page
**pages
= NULL
;
9043 unsigned long off
, start
, end
, ubuf
;
9045 int ret
, pret
, nr_pages
, i
;
9047 if (!iov
->iov_base
) {
9048 *pimu
= ctx
->dummy_ubuf
;
9052 ubuf
= (unsigned long) iov
->iov_base
;
9053 end
= (ubuf
+ iov
->iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
9054 start
= ubuf
>> PAGE_SHIFT
;
9055 nr_pages
= end
- start
;
9060 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
9064 vmas
= kvmalloc_array(nr_pages
, sizeof(struct vm_area_struct
*),
9069 imu
= kvmalloc(struct_size(imu
, bvec
, nr_pages
), GFP_KERNEL
);
9074 mmap_read_lock(current
->mm
);
9075 pret
= pin_user_pages(ubuf
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
9077 if (pret
== nr_pages
) {
9078 /* don't support file backed memory */
9079 for (i
= 0; i
< nr_pages
; i
++) {
9080 struct vm_area_struct
*vma
= vmas
[i
];
9082 if (vma_is_shmem(vma
))
9085 !is_file_hugepages(vma
->vm_file
)) {
9091 ret
= pret
< 0 ? pret
: -EFAULT
;
9093 mmap_read_unlock(current
->mm
);
9096 * if we did partial map, or found file backed vmas,
9097 * release any pages we did get
9100 unpin_user_pages(pages
, pret
);
9104 ret
= io_buffer_account_pin(ctx
, pages
, pret
, imu
, last_hpage
);
9106 unpin_user_pages(pages
, pret
);
9110 off
= ubuf
& ~PAGE_MASK
;
9111 size
= iov
->iov_len
;
9112 for (i
= 0; i
< nr_pages
; i
++) {
9115 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
9116 imu
->bvec
[i
].bv_page
= pages
[i
];
9117 imu
->bvec
[i
].bv_len
= vec_len
;
9118 imu
->bvec
[i
].bv_offset
= off
;
9122 /* store original address for later verification */
9124 imu
->ubuf_end
= ubuf
+ iov
->iov_len
;
9125 imu
->nr_bvecs
= nr_pages
;
9136 static int io_buffers_map_alloc(struct io_ring_ctx
*ctx
, unsigned int nr_args
)
9138 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(*ctx
->user_bufs
), GFP_KERNEL
);
9139 return ctx
->user_bufs
? 0 : -ENOMEM
;
9142 static int io_buffer_validate(struct iovec
*iov
)
9144 unsigned long tmp
, acct_len
= iov
->iov_len
+ (PAGE_SIZE
- 1);
9147 * Don't impose further limits on the size and buffer
9148 * constraints here, we'll -EINVAL later when IO is
9149 * submitted if they are wrong.
9152 return iov
->iov_len
? -EFAULT
: 0;
9156 /* arbitrary limit, but we need something */
9157 if (iov
->iov_len
> SZ_1G
)
9160 if (check_add_overflow((unsigned long)iov
->iov_base
, acct_len
, &tmp
))
9166 static int io_sqe_buffers_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
9167 unsigned int nr_args
, u64 __user
*tags
)
9169 struct page
*last_hpage
= NULL
;
9170 struct io_rsrc_data
*data
;
9176 if (!nr_args
|| nr_args
> IORING_MAX_REG_BUFFERS
)
9178 ret
= io_rsrc_node_switch_start(ctx
);
9181 ret
= io_rsrc_data_alloc(ctx
, io_rsrc_buf_put
, tags
, nr_args
, &data
);
9184 ret
= io_buffers_map_alloc(ctx
, nr_args
);
9186 io_rsrc_data_free(data
);
9190 for (i
= 0; i
< nr_args
; i
++, ctx
->nr_user_bufs
++) {
9191 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
9194 ret
= io_buffer_validate(&iov
);
9197 if (!iov
.iov_base
&& *io_get_tag_slot(data
, i
)) {
9202 ret
= io_sqe_buffer_register(ctx
, &iov
, &ctx
->user_bufs
[i
],
9208 WARN_ON_ONCE(ctx
->buf_data
);
9210 ctx
->buf_data
= data
;
9212 __io_sqe_buffers_unregister(ctx
);
9214 io_rsrc_node_switch(ctx
, NULL
);
9218 static int __io_sqe_buffers_update(struct io_ring_ctx
*ctx
,
9219 struct io_uring_rsrc_update2
*up
,
9220 unsigned int nr_args
)
9222 u64 __user
*tags
= u64_to_user_ptr(up
->tags
);
9223 struct iovec iov
, __user
*iovs
= u64_to_user_ptr(up
->data
);
9224 struct page
*last_hpage
= NULL
;
9225 bool needs_switch
= false;
9231 if (up
->offset
+ nr_args
> ctx
->nr_user_bufs
)
9234 for (done
= 0; done
< nr_args
; done
++) {
9235 struct io_mapped_ubuf
*imu
;
9236 int offset
= up
->offset
+ done
;
9239 err
= io_copy_iov(ctx
, &iov
, iovs
, done
);
9242 if (tags
&& copy_from_user(&tag
, &tags
[done
], sizeof(tag
))) {
9246 err
= io_buffer_validate(&iov
);
9249 if (!iov
.iov_base
&& tag
) {
9253 err
= io_sqe_buffer_register(ctx
, &iov
, &imu
, &last_hpage
);
9257 i
= array_index_nospec(offset
, ctx
->nr_user_bufs
);
9258 if (ctx
->user_bufs
[i
] != ctx
->dummy_ubuf
) {
9259 err
= io_queue_rsrc_removal(ctx
->buf_data
, i
,
9260 ctx
->rsrc_node
, ctx
->user_bufs
[i
]);
9261 if (unlikely(err
)) {
9262 io_buffer_unmap(ctx
, &imu
);
9265 ctx
->user_bufs
[i
] = NULL
;
9266 needs_switch
= true;
9269 ctx
->user_bufs
[i
] = imu
;
9270 *io_get_tag_slot(ctx
->buf_data
, offset
) = tag
;
9274 io_rsrc_node_switch(ctx
, ctx
->buf_data
);
9275 return done
? done
: err
;
9278 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
9280 __s32 __user
*fds
= arg
;
9286 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
9289 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
9290 if (IS_ERR(ctx
->cq_ev_fd
)) {
9291 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
9293 ctx
->cq_ev_fd
= NULL
;
9300 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
9302 if (ctx
->cq_ev_fd
) {
9303 eventfd_ctx_put(ctx
->cq_ev_fd
);
9304 ctx
->cq_ev_fd
= NULL
;
9311 static void io_destroy_buffers(struct io_ring_ctx
*ctx
)
9313 struct io_buffer
*buf
;
9314 unsigned long index
;
9316 xa_for_each(&ctx
->io_buffers
, index
, buf
)
9317 __io_remove_buffers(ctx
, buf
, index
, -1U);
9320 static void io_req_cache_free(struct list_head
*list
)
9322 struct io_kiocb
*req
, *nxt
;
9324 list_for_each_entry_safe(req
, nxt
, list
, inflight_entry
) {
9325 list_del(&req
->inflight_entry
);
9326 kmem_cache_free(req_cachep
, req
);
9330 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
9332 struct io_submit_state
*state
= &ctx
->submit_state
;
9334 mutex_lock(&ctx
->uring_lock
);
9336 if (state
->free_reqs
) {
9337 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
, state
->reqs
);
9338 state
->free_reqs
= 0;
9341 io_flush_cached_locked_reqs(ctx
, state
);
9342 io_req_cache_free(&state
->free_list
);
9343 mutex_unlock(&ctx
->uring_lock
);
9346 static void io_wait_rsrc_data(struct io_rsrc_data
*data
)
9348 if (data
&& !atomic_dec_and_test(&data
->refs
))
9349 wait_for_completion(&data
->done
);
9352 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
9354 io_sq_thread_finish(ctx
);
9356 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9357 io_wait_rsrc_data(ctx
->buf_data
);
9358 io_wait_rsrc_data(ctx
->file_data
);
9360 mutex_lock(&ctx
->uring_lock
);
9362 __io_sqe_buffers_unregister(ctx
);
9364 __io_sqe_files_unregister(ctx
);
9366 __io_cqring_overflow_flush(ctx
, true);
9367 mutex_unlock(&ctx
->uring_lock
);
9368 io_eventfd_unregister(ctx
);
9369 io_destroy_buffers(ctx
);
9371 put_cred(ctx
->sq_creds
);
9373 /* there are no registered resources left, nobody uses it */
9375 io_rsrc_node_destroy(ctx
->rsrc_node
);
9376 if (ctx
->rsrc_backup_node
)
9377 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
9378 flush_delayed_work(&ctx
->rsrc_put_work
);
9380 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
9381 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
9383 #if defined(CONFIG_UNIX)
9384 if (ctx
->ring_sock
) {
9385 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
9386 sock_release(ctx
->ring_sock
);
9389 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
9391 if (ctx
->mm_account
) {
9392 mmdrop(ctx
->mm_account
);
9393 ctx
->mm_account
= NULL
;
9396 io_mem_free(ctx
->rings
);
9397 io_mem_free(ctx
->sq_sqes
);
9399 percpu_ref_exit(&ctx
->refs
);
9400 free_uid(ctx
->user
);
9401 io_req_caches_free(ctx
);
9403 io_wq_put_hash(ctx
->hash_map
);
9404 kfree(ctx
->cancel_hash
);
9405 kfree(ctx
->dummy_ubuf
);
9409 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
9411 struct io_ring_ctx
*ctx
= file
->private_data
;
9414 poll_wait(file
, &ctx
->poll_wait
, wait
);
9416 * synchronizes with barrier from wq_has_sleeper call in
9420 if (!io_sqring_full(ctx
))
9421 mask
|= EPOLLOUT
| EPOLLWRNORM
;
9424 * Don't flush cqring overflow list here, just do a simple check.
9425 * Otherwise there could possible be ABBA deadlock:
9428 * lock(&ctx->uring_lock);
9430 * lock(&ctx->uring_lock);
9433 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9434 * pushs them to do the flush.
9436 if (io_cqring_events(ctx
) || test_bit(0, &ctx
->check_cq_overflow
))
9437 mask
|= EPOLLIN
| EPOLLRDNORM
;
9442 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
9444 const struct cred
*creds
;
9446 creds
= xa_erase(&ctx
->personalities
, id
);
9455 struct io_tctx_exit
{
9456 struct callback_head task_work
;
9457 struct completion completion
;
9458 struct io_ring_ctx
*ctx
;
9461 static void io_tctx_exit_cb(struct callback_head
*cb
)
9463 struct io_uring_task
*tctx
= current
->io_uring
;
9464 struct io_tctx_exit
*work
;
9466 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
9468 * When @in_idle, we're in cancellation and it's racy to remove the
9469 * node. It'll be removed by the end of cancellation, just ignore it.
9471 if (!atomic_read(&tctx
->in_idle
))
9472 io_uring_del_tctx_node((unsigned long)work
->ctx
);
9473 complete(&work
->completion
);
9476 static bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
9478 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9480 return req
->ctx
== data
;
9483 static void io_ring_exit_work(struct work_struct
*work
)
9485 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
9486 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
9487 unsigned long interval
= HZ
/ 20;
9488 struct io_tctx_exit exit
;
9489 struct io_tctx_node
*node
;
9493 * If we're doing polled IO and end up having requests being
9494 * submitted async (out-of-line), then completions can come in while
9495 * we're waiting for refs to drop. We need to reap these manually,
9496 * as nobody else will be looking for them.
9499 io_uring_try_cancel_requests(ctx
, NULL
, true);
9501 struct io_sq_data
*sqd
= ctx
->sq_data
;
9502 struct task_struct
*tsk
;
9504 io_sq_thread_park(sqd
);
9506 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
9507 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
9508 io_cancel_ctx_cb
, ctx
, true);
9509 io_sq_thread_unpark(sqd
);
9512 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
9513 /* there is little hope left, don't run it too often */
9516 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
9518 init_completion(&exit
.completion
);
9519 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
9522 * Some may use context even when all refs and requests have been put,
9523 * and they are free to do so while still holding uring_lock or
9524 * completion_lock, see io_req_task_submit(). Apart from other work,
9525 * this lock/unlock section also waits them to finish.
9527 mutex_lock(&ctx
->uring_lock
);
9528 while (!list_empty(&ctx
->tctx_list
)) {
9529 WARN_ON_ONCE(time_after(jiffies
, timeout
));
9531 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
9533 /* don't spin on a single task if cancellation failed */
9534 list_rotate_left(&ctx
->tctx_list
);
9535 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
9536 if (WARN_ON_ONCE(ret
))
9538 wake_up_process(node
->task
);
9540 mutex_unlock(&ctx
->uring_lock
);
9541 wait_for_completion(&exit
.completion
);
9542 mutex_lock(&ctx
->uring_lock
);
9544 mutex_unlock(&ctx
->uring_lock
);
9545 spin_lock(&ctx
->completion_lock
);
9546 spin_unlock(&ctx
->completion_lock
);
9548 io_ring_ctx_free(ctx
);
9551 /* Returns true if we found and killed one or more timeouts */
9552 static bool io_kill_timeouts(struct io_ring_ctx
*ctx
, struct task_struct
*tsk
,
9555 struct io_kiocb
*req
, *tmp
;
9558 spin_lock(&ctx
->completion_lock
);
9559 spin_lock_irq(&ctx
->timeout_lock
);
9560 list_for_each_entry_safe(req
, tmp
, &ctx
->timeout_list
, timeout
.list
) {
9561 if (io_match_task(req
, tsk
, cancel_all
)) {
9562 io_kill_timeout(req
, -ECANCELED
);
9566 spin_unlock_irq(&ctx
->timeout_lock
);
9568 io_commit_cqring(ctx
);
9569 spin_unlock(&ctx
->completion_lock
);
9571 io_cqring_ev_posted(ctx
);
9572 return canceled
!= 0;
9575 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
9577 unsigned long index
;
9578 struct creds
*creds
;
9580 mutex_lock(&ctx
->uring_lock
);
9581 percpu_ref_kill(&ctx
->refs
);
9583 __io_cqring_overflow_flush(ctx
, true);
9584 xa_for_each(&ctx
->personalities
, index
, creds
)
9585 io_unregister_personality(ctx
, index
);
9586 mutex_unlock(&ctx
->uring_lock
);
9588 io_kill_timeouts(ctx
, NULL
, true);
9589 io_poll_remove_all(ctx
, NULL
, true);
9591 /* if we failed setting up the ctx, we might not have any rings */
9592 io_iopoll_try_reap_events(ctx
);
9594 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
9596 * Use system_unbound_wq to avoid spawning tons of event kworkers
9597 * if we're exiting a ton of rings at the same time. It just adds
9598 * noise and overhead, there's no discernable change in runtime
9599 * over using system_wq.
9601 queue_work(system_unbound_wq
, &ctx
->exit_work
);
9604 static int io_uring_release(struct inode
*inode
, struct file
*file
)
9606 struct io_ring_ctx
*ctx
= file
->private_data
;
9608 file
->private_data
= NULL
;
9609 io_ring_ctx_wait_and_kill(ctx
);
9613 struct io_task_cancel
{
9614 struct task_struct
*task
;
9618 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
9620 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
9621 struct io_task_cancel
*cancel
= data
;
9623 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
9626 static bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
9627 struct task_struct
*task
, bool cancel_all
)
9629 struct io_defer_entry
*de
;
9632 spin_lock(&ctx
->completion_lock
);
9633 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
9634 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
9635 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
9639 spin_unlock(&ctx
->completion_lock
);
9640 if (list_empty(&list
))
9643 while (!list_empty(&list
)) {
9644 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
9645 list_del_init(&de
->list
);
9646 io_req_complete_failed(de
->req
, -ECANCELED
);
9652 static bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
9654 struct io_tctx_node
*node
;
9655 enum io_wq_cancel cret
;
9658 mutex_lock(&ctx
->uring_lock
);
9659 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
9660 struct io_uring_task
*tctx
= node
->task
->io_uring
;
9663 * io_wq will stay alive while we hold uring_lock, because it's
9664 * killed after ctx nodes, which requires to take the lock.
9666 if (!tctx
|| !tctx
->io_wq
)
9668 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
9669 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9671 mutex_unlock(&ctx
->uring_lock
);
9676 static void io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
9677 struct task_struct
*task
,
9680 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
9681 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
9684 enum io_wq_cancel cret
;
9688 ret
|= io_uring_try_cancel_iowq(ctx
);
9689 } else if (tctx
&& tctx
->io_wq
) {
9691 * Cancels requests of all rings, not only @ctx, but
9692 * it's fine as the task is in exit/exec.
9694 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
9696 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
9699 /* SQPOLL thread does its own polling */
9700 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
9701 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
9702 while (!list_empty_careful(&ctx
->iopoll_list
)) {
9703 io_iopoll_try_reap_events(ctx
);
9708 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
9709 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
9710 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
9712 ret
|= io_run_task_work();
9719 static int __io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9721 struct io_uring_task
*tctx
= current
->io_uring
;
9722 struct io_tctx_node
*node
;
9725 if (unlikely(!tctx
)) {
9726 ret
= io_uring_alloc_task_context(current
, ctx
);
9730 tctx
= current
->io_uring
;
9731 if (ctx
->iowq_limits_set
) {
9732 unsigned int limits
[2] = { ctx
->iowq_limits
[0],
9733 ctx
->iowq_limits
[1], };
9735 ret
= io_wq_max_workers(tctx
->io_wq
, limits
);
9740 if (!xa_load(&tctx
->xa
, (unsigned long)ctx
)) {
9741 node
= kmalloc(sizeof(*node
), GFP_KERNEL
);
9745 node
->task
= current
;
9747 ret
= xa_err(xa_store(&tctx
->xa
, (unsigned long)ctx
,
9754 mutex_lock(&ctx
->uring_lock
);
9755 list_add(&node
->ctx_node
, &ctx
->tctx_list
);
9756 mutex_unlock(&ctx
->uring_lock
);
9763 * Note that this task has used io_uring. We use it for cancelation purposes.
9765 static inline int io_uring_add_tctx_node(struct io_ring_ctx
*ctx
)
9767 struct io_uring_task
*tctx
= current
->io_uring
;
9769 if (likely(tctx
&& tctx
->last
== ctx
))
9771 return __io_uring_add_tctx_node(ctx
);
9775 * Remove this io_uring_file -> task mapping.
9777 static void io_uring_del_tctx_node(unsigned long index
)
9779 struct io_uring_task
*tctx
= current
->io_uring
;
9780 struct io_tctx_node
*node
;
9784 node
= xa_erase(&tctx
->xa
, index
);
9788 WARN_ON_ONCE(current
!= node
->task
);
9789 WARN_ON_ONCE(list_empty(&node
->ctx_node
));
9791 mutex_lock(&node
->ctx
->uring_lock
);
9792 list_del(&node
->ctx_node
);
9793 mutex_unlock(&node
->ctx
->uring_lock
);
9795 if (tctx
->last
== node
->ctx
)
9800 static void io_uring_clean_tctx(struct io_uring_task
*tctx
)
9802 struct io_wq
*wq
= tctx
->io_wq
;
9803 struct io_tctx_node
*node
;
9804 unsigned long index
;
9806 xa_for_each(&tctx
->xa
, index
, node
) {
9807 io_uring_del_tctx_node(index
);
9812 * Must be after io_uring_del_task_file() (removes nodes under
9813 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9815 io_wq_put_and_exit(wq
);
9820 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
9823 return atomic_read(&tctx
->inflight_tracked
);
9824 return percpu_counter_sum(&tctx
->inflight
);
9828 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9829 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9831 static void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
9833 struct io_uring_task
*tctx
= current
->io_uring
;
9834 struct io_ring_ctx
*ctx
;
9838 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
9840 if (!current
->io_uring
)
9843 io_wq_exit_start(tctx
->io_wq
);
9845 atomic_inc(&tctx
->in_idle
);
9847 io_uring_drop_tctx_refs(current
);
9848 /* read completions before cancelations */
9849 inflight
= tctx_inflight(tctx
, !cancel_all
);
9854 struct io_tctx_node
*node
;
9855 unsigned long index
;
9857 xa_for_each(&tctx
->xa
, index
, node
) {
9858 /* sqpoll task will cancel all its requests */
9859 if (node
->ctx
->sq_data
)
9861 io_uring_try_cancel_requests(node
->ctx
, current
,
9865 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
9866 io_uring_try_cancel_requests(ctx
, current
,
9870 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
9872 io_uring_drop_tctx_refs(current
);
9875 * If we've seen completions, retry without waiting. This
9876 * avoids a race where a completion comes in before we did
9877 * prepare_to_wait().
9879 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
9881 finish_wait(&tctx
->wait
, &wait
);
9884 io_uring_clean_tctx(tctx
);
9887 * We shouldn't run task_works after cancel, so just leave
9888 * ->in_idle set for normal exit.
9890 atomic_dec(&tctx
->in_idle
);
9891 /* for exec all current's requests should be gone, kill tctx */
9892 __io_uring_free(current
);
9896 void __io_uring_cancel(bool cancel_all
)
9898 io_uring_cancel_generic(cancel_all
, NULL
);
9901 static void *io_uring_validate_mmap_request(struct file
*file
,
9902 loff_t pgoff
, size_t sz
)
9904 struct io_ring_ctx
*ctx
= file
->private_data
;
9905 loff_t offset
= pgoff
<< PAGE_SHIFT
;
9910 case IORING_OFF_SQ_RING
:
9911 case IORING_OFF_CQ_RING
:
9914 case IORING_OFF_SQES
:
9918 return ERR_PTR(-EINVAL
);
9921 page
= virt_to_head_page(ptr
);
9922 if (sz
> page_size(page
))
9923 return ERR_PTR(-EINVAL
);
9930 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9932 size_t sz
= vma
->vm_end
- vma
->vm_start
;
9936 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
9938 return PTR_ERR(ptr
);
9940 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
9941 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
9944 #else /* !CONFIG_MMU */
9946 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
9948 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
9951 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
9953 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
9956 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
9957 unsigned long addr
, unsigned long len
,
9958 unsigned long pgoff
, unsigned long flags
)
9962 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
9964 return PTR_ERR(ptr
);
9966 return (unsigned long) ptr
;
9969 #endif /* !CONFIG_MMU */
9971 static int io_sqpoll_wait_sq(struct io_ring_ctx
*ctx
)
9976 if (!io_sqring_full(ctx
))
9978 prepare_to_wait(&ctx
->sqo_sq_wait
, &wait
, TASK_INTERRUPTIBLE
);
9980 if (!io_sqring_full(ctx
))
9983 } while (!signal_pending(current
));
9985 finish_wait(&ctx
->sqo_sq_wait
, &wait
);
9989 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
9990 struct __kernel_timespec __user
**ts
,
9991 const sigset_t __user
**sig
)
9993 struct io_uring_getevents_arg arg
;
9996 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9997 * is just a pointer to the sigset_t.
9999 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
10000 *sig
= (const sigset_t __user
*) argp
;
10006 * EXT_ARG is set - ensure we agree on the size of it and copy in our
10007 * timespec and sigset_t pointers if good.
10009 if (*argsz
!= sizeof(arg
))
10011 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
10015 *sig
= u64_to_user_ptr(arg
.sigmask
);
10016 *argsz
= arg
.sigmask_sz
;
10017 *ts
= u64_to_user_ptr(arg
.ts
);
10021 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
10022 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
10025 struct io_ring_ctx
*ctx
;
10030 io_run_task_work();
10032 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
10033 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
)))
10037 if (unlikely(!f
.file
))
10041 if (unlikely(f
.file
->f_op
!= &io_uring_fops
))
10045 ctx
= f
.file
->private_data
;
10046 if (unlikely(!percpu_ref_tryget(&ctx
->refs
)))
10050 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10054 * For SQ polling, the thread will do all submissions and completions.
10055 * Just return the requested submit count, and wake the thread if
10056 * we were asked to.
10059 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10060 io_cqring_overflow_flush(ctx
);
10062 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
10066 if (flags
& IORING_ENTER_SQ_WAKEUP
)
10067 wake_up(&ctx
->sq_data
->wait
);
10068 if (flags
& IORING_ENTER_SQ_WAIT
) {
10069 ret
= io_sqpoll_wait_sq(ctx
);
10073 submitted
= to_submit
;
10074 } else if (to_submit
) {
10075 ret
= io_uring_add_tctx_node(ctx
);
10078 mutex_lock(&ctx
->uring_lock
);
10079 submitted
= io_submit_sqes(ctx
, to_submit
);
10080 mutex_unlock(&ctx
->uring_lock
);
10082 if (submitted
!= to_submit
)
10085 if (flags
& IORING_ENTER_GETEVENTS
) {
10086 const sigset_t __user
*sig
;
10087 struct __kernel_timespec __user
*ts
;
10089 ret
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
10093 min_complete
= min(min_complete
, ctx
->cq_entries
);
10096 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10097 * space applications don't need to do io completion events
10098 * polling again, they can rely on io_sq_thread to do polling
10099 * work, which can reduce cpu usage and uring_lock contention.
10101 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
10102 !(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10103 ret
= io_iopoll_check(ctx
, min_complete
);
10105 ret
= io_cqring_wait(ctx
, min_complete
, sig
, argsz
, ts
);
10110 percpu_ref_put(&ctx
->refs
);
10113 return submitted
? submitted
: ret
;
10116 #ifdef CONFIG_PROC_FS
10117 static int io_uring_show_cred(struct seq_file
*m
, unsigned int id
,
10118 const struct cred
*cred
)
10120 struct user_namespace
*uns
= seq_user_ns(m
);
10121 struct group_info
*gi
;
10126 seq_printf(m
, "%5d\n", id
);
10127 seq_put_decimal_ull(m
, "\tUid:\t", from_kuid_munged(uns
, cred
->uid
));
10128 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->euid
));
10129 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->suid
));
10130 seq_put_decimal_ull(m
, "\t\t", from_kuid_munged(uns
, cred
->fsuid
));
10131 seq_put_decimal_ull(m
, "\n\tGid:\t", from_kgid_munged(uns
, cred
->gid
));
10132 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->egid
));
10133 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->sgid
));
10134 seq_put_decimal_ull(m
, "\t\t", from_kgid_munged(uns
, cred
->fsgid
));
10135 seq_puts(m
, "\n\tGroups:\t");
10136 gi
= cred
->group_info
;
10137 for (g
= 0; g
< gi
->ngroups
; g
++) {
10138 seq_put_decimal_ull(m
, g
? " " : "",
10139 from_kgid_munged(uns
, gi
->gid
[g
]));
10141 seq_puts(m
, "\n\tCapEff:\t");
10142 cap
= cred
->cap_effective
;
10143 CAP_FOR_EACH_U32(__capi
)
10144 seq_put_hex_ll(m
, NULL
, cap
.cap
[CAP_LAST_U32
- __capi
], 8);
10149 static void __io_uring_show_fdinfo(struct io_ring_ctx
*ctx
, struct seq_file
*m
)
10151 struct io_sq_data
*sq
= NULL
;
10156 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10157 * since fdinfo case grabs it in the opposite direction of normal use
10158 * cases. If we fail to get the lock, we just don't iterate any
10159 * structures that could be going away outside the io_uring mutex.
10161 has_lock
= mutex_trylock(&ctx
->uring_lock
);
10163 if (has_lock
&& (ctx
->flags
& IORING_SETUP_SQPOLL
)) {
10169 seq_printf(m
, "SqThread:\t%d\n", sq
? task_pid_nr(sq
->thread
) : -1);
10170 seq_printf(m
, "SqThreadCpu:\t%d\n", sq
? task_cpu(sq
->thread
) : -1);
10171 seq_printf(m
, "UserFiles:\t%u\n", ctx
->nr_user_files
);
10172 for (i
= 0; has_lock
&& i
< ctx
->nr_user_files
; i
++) {
10173 struct file
*f
= io_file_from_index(ctx
, i
);
10176 seq_printf(m
, "%5u: %s\n", i
, file_dentry(f
)->d_iname
);
10178 seq_printf(m
, "%5u: <none>\n", i
);
10180 seq_printf(m
, "UserBufs:\t%u\n", ctx
->nr_user_bufs
);
10181 for (i
= 0; has_lock
&& i
< ctx
->nr_user_bufs
; i
++) {
10182 struct io_mapped_ubuf
*buf
= ctx
->user_bufs
[i
];
10183 unsigned int len
= buf
->ubuf_end
- buf
->ubuf
;
10185 seq_printf(m
, "%5u: 0x%llx/%u\n", i
, buf
->ubuf
, len
);
10187 if (has_lock
&& !xa_empty(&ctx
->personalities
)) {
10188 unsigned long index
;
10189 const struct cred
*cred
;
10191 seq_printf(m
, "Personalities:\n");
10192 xa_for_each(&ctx
->personalities
, index
, cred
)
10193 io_uring_show_cred(m
, index
, cred
);
10195 seq_printf(m
, "PollList:\n");
10196 spin_lock(&ctx
->completion_lock
);
10197 for (i
= 0; i
< (1U << ctx
->cancel_hash_bits
); i
++) {
10198 struct hlist_head
*list
= &ctx
->cancel_hash
[i
];
10199 struct io_kiocb
*req
;
10201 hlist_for_each_entry(req
, list
, hash_node
)
10202 seq_printf(m
, " op=%d, task_works=%d\n", req
->opcode
,
10203 req
->task
->task_works
!= NULL
);
10205 spin_unlock(&ctx
->completion_lock
);
10207 mutex_unlock(&ctx
->uring_lock
);
10210 static void io_uring_show_fdinfo(struct seq_file
*m
, struct file
*f
)
10212 struct io_ring_ctx
*ctx
= f
->private_data
;
10214 if (percpu_ref_tryget(&ctx
->refs
)) {
10215 __io_uring_show_fdinfo(ctx
, m
);
10216 percpu_ref_put(&ctx
->refs
);
10221 static const struct file_operations io_uring_fops
= {
10222 .release
= io_uring_release
,
10223 .mmap
= io_uring_mmap
,
10225 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
10226 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
10228 .poll
= io_uring_poll
,
10229 #ifdef CONFIG_PROC_FS
10230 .show_fdinfo
= io_uring_show_fdinfo
,
10234 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
10235 struct io_uring_params
*p
)
10237 struct io_rings
*rings
;
10238 size_t size
, sq_array_offset
;
10240 /* make sure these are sane, as we already accounted them */
10241 ctx
->sq_entries
= p
->sq_entries
;
10242 ctx
->cq_entries
= p
->cq_entries
;
10244 size
= rings_size(p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
10245 if (size
== SIZE_MAX
)
10248 rings
= io_mem_alloc(size
);
10252 ctx
->rings
= rings
;
10253 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
10254 rings
->sq_ring_mask
= p
->sq_entries
- 1;
10255 rings
->cq_ring_mask
= p
->cq_entries
- 1;
10256 rings
->sq_ring_entries
= p
->sq_entries
;
10257 rings
->cq_ring_entries
= p
->cq_entries
;
10259 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
10260 if (size
== SIZE_MAX
) {
10261 io_mem_free(ctx
->rings
);
10266 ctx
->sq_sqes
= io_mem_alloc(size
);
10267 if (!ctx
->sq_sqes
) {
10268 io_mem_free(ctx
->rings
);
10276 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
10280 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
10284 ret
= io_uring_add_tctx_node(ctx
);
10289 fd_install(fd
, file
);
10294 * Allocate an anonymous fd, this is what constitutes the application
10295 * visible backing of an io_uring instance. The application mmaps this
10296 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10297 * we have to tie this fd to a socket for file garbage collection purposes.
10299 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
10302 #if defined(CONFIG_UNIX)
10305 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
10308 return ERR_PTR(ret
);
10311 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
10312 O_RDWR
| O_CLOEXEC
);
10313 #if defined(CONFIG_UNIX)
10314 if (IS_ERR(file
)) {
10315 sock_release(ctx
->ring_sock
);
10316 ctx
->ring_sock
= NULL
;
10318 ctx
->ring_sock
->file
= file
;
10324 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
10325 struct io_uring_params __user
*params
)
10327 struct io_ring_ctx
*ctx
;
10333 if (entries
> IORING_MAX_ENTRIES
) {
10334 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10336 entries
= IORING_MAX_ENTRIES
;
10340 * Use twice as many entries for the CQ ring. It's possible for the
10341 * application to drive a higher depth than the size of the SQ ring,
10342 * since the sqes are only used at submission time. This allows for
10343 * some flexibility in overcommitting a bit. If the application has
10344 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10345 * of CQ ring entries manually.
10347 p
->sq_entries
= roundup_pow_of_two(entries
);
10348 if (p
->flags
& IORING_SETUP_CQSIZE
) {
10350 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10351 * to a power-of-two, if it isn't already. We do NOT impose
10352 * any cq vs sq ring sizing.
10354 if (!p
->cq_entries
)
10356 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
10357 if (!(p
->flags
& IORING_SETUP_CLAMP
))
10359 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
10361 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
10362 if (p
->cq_entries
< p
->sq_entries
)
10365 p
->cq_entries
= 2 * p
->sq_entries
;
10368 ctx
= io_ring_ctx_alloc(p
);
10371 ctx
->compat
= in_compat_syscall();
10372 if (!capable(CAP_IPC_LOCK
))
10373 ctx
->user
= get_uid(current_user());
10376 * This is just grabbed for accounting purposes. When a process exits,
10377 * the mm is exited and dropped before the files, hence we need to hang
10378 * on to this mm purely for the purposes of being able to unaccount
10379 * memory (locked/pinned vm). It's not used for anything else.
10381 mmgrab(current
->mm
);
10382 ctx
->mm_account
= current
->mm
;
10384 ret
= io_allocate_scq_urings(ctx
, p
);
10388 ret
= io_sq_offload_create(ctx
, p
);
10391 /* always set a rsrc node */
10392 ret
= io_rsrc_node_switch_start(ctx
);
10395 io_rsrc_node_switch(ctx
, NULL
);
10397 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
10398 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
10399 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
10400 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
10401 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
10402 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
10403 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
10404 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
10406 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
10407 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
10408 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
10409 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
10410 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
10411 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
10412 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
10413 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
10415 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
10416 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
10417 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
10418 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
10419 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
10420 IORING_FEAT_RSRC_TAGS
;
10422 if (copy_to_user(params
, p
, sizeof(*p
))) {
10427 file
= io_uring_get_file(ctx
);
10428 if (IS_ERR(file
)) {
10429 ret
= PTR_ERR(file
);
10434 * Install ring fd as the very last thing, so we don't risk someone
10435 * having closed it before we finish setup
10437 ret
= io_uring_install_fd(ctx
, file
);
10439 /* fput will clean it up */
10444 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
10447 io_ring_ctx_wait_and_kill(ctx
);
10452 * Sets up an aio uring context, and returns the fd. Applications asks for a
10453 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10454 * params structure passed in.
10456 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
10458 struct io_uring_params p
;
10461 if (copy_from_user(&p
, params
, sizeof(p
)))
10463 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
10468 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
10469 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
10470 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
10471 IORING_SETUP_R_DISABLED
))
10474 return io_uring_create(entries
, &p
, params
);
10477 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
10478 struct io_uring_params __user
*, params
)
10480 return io_uring_setup(entries
, params
);
10483 static int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
, unsigned nr_args
)
10485 struct io_uring_probe
*p
;
10489 size
= struct_size(p
, ops
, nr_args
);
10490 if (size
== SIZE_MAX
)
10492 p
= kzalloc(size
, GFP_KERNEL
);
10497 if (copy_from_user(p
, arg
, size
))
10500 if (memchr_inv(p
, 0, size
))
10503 p
->last_op
= IORING_OP_LAST
- 1;
10504 if (nr_args
> IORING_OP_LAST
)
10505 nr_args
= IORING_OP_LAST
;
10507 for (i
= 0; i
< nr_args
; i
++) {
10509 if (!io_op_defs
[i
].not_supported
)
10510 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
10515 if (copy_to_user(arg
, p
, size
))
10522 static int io_register_personality(struct io_ring_ctx
*ctx
)
10524 const struct cred
*creds
;
10528 creds
= get_current_cred();
10530 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
10531 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
10539 static int io_register_restrictions(struct io_ring_ctx
*ctx
, void __user
*arg
,
10540 unsigned int nr_args
)
10542 struct io_uring_restriction
*res
;
10546 /* Restrictions allowed only if rings started disabled */
10547 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10550 /* We allow only a single restrictions registration */
10551 if (ctx
->restrictions
.registered
)
10554 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
10557 size
= array_size(nr_args
, sizeof(*res
));
10558 if (size
== SIZE_MAX
)
10561 res
= memdup_user(arg
, size
);
10563 return PTR_ERR(res
);
10567 for (i
= 0; i
< nr_args
; i
++) {
10568 switch (res
[i
].opcode
) {
10569 case IORING_RESTRICTION_REGISTER_OP
:
10570 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
10575 __set_bit(res
[i
].register_op
,
10576 ctx
->restrictions
.register_op
);
10578 case IORING_RESTRICTION_SQE_OP
:
10579 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
10584 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
10586 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
10587 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
10589 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
10590 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
10599 /* Reset all restrictions if an error happened */
10601 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
10603 ctx
->restrictions
.registered
= true;
10609 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
10611 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
10614 if (ctx
->restrictions
.registered
)
10615 ctx
->restricted
= 1;
10617 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
10618 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
10619 wake_up(&ctx
->sq_data
->wait
);
10623 static int __io_register_rsrc_update(struct io_ring_ctx
*ctx
, unsigned type
,
10624 struct io_uring_rsrc_update2
*up
,
10630 if (check_add_overflow(up
->offset
, nr_args
, &tmp
))
10632 err
= io_rsrc_node_switch_start(ctx
);
10637 case IORING_RSRC_FILE
:
10638 return __io_sqe_files_update(ctx
, up
, nr_args
);
10639 case IORING_RSRC_BUFFER
:
10640 return __io_sqe_buffers_update(ctx
, up
, nr_args
);
10645 static int io_register_files_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10648 struct io_uring_rsrc_update2 up
;
10652 memset(&up
, 0, sizeof(up
));
10653 if (copy_from_user(&up
, arg
, sizeof(struct io_uring_rsrc_update
)))
10655 if (up
.resv
|| up
.resv2
)
10657 return __io_register_rsrc_update(ctx
, IORING_RSRC_FILE
, &up
, nr_args
);
10660 static int io_register_rsrc_update(struct io_ring_ctx
*ctx
, void __user
*arg
,
10661 unsigned size
, unsigned type
)
10663 struct io_uring_rsrc_update2 up
;
10665 if (size
!= sizeof(up
))
10667 if (copy_from_user(&up
, arg
, sizeof(up
)))
10669 if (!up
.nr
|| up
.resv
|| up
.resv2
)
10671 return __io_register_rsrc_update(ctx
, type
, &up
, up
.nr
);
10674 static int io_register_rsrc(struct io_ring_ctx
*ctx
, void __user
*arg
,
10675 unsigned int size
, unsigned int type
)
10677 struct io_uring_rsrc_register rr
;
10679 /* keep it extendible */
10680 if (size
!= sizeof(rr
))
10683 memset(&rr
, 0, sizeof(rr
));
10684 if (copy_from_user(&rr
, arg
, size
))
10686 if (!rr
.nr
|| rr
.resv
|| rr
.resv2
)
10690 case IORING_RSRC_FILE
:
10691 return io_sqe_files_register(ctx
, u64_to_user_ptr(rr
.data
),
10692 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10693 case IORING_RSRC_BUFFER
:
10694 return io_sqe_buffers_register(ctx
, u64_to_user_ptr(rr
.data
),
10695 rr
.nr
, u64_to_user_ptr(rr
.tags
));
10700 static int io_register_iowq_aff(struct io_ring_ctx
*ctx
, void __user
*arg
,
10703 struct io_uring_task
*tctx
= current
->io_uring
;
10704 cpumask_var_t new_mask
;
10707 if (!tctx
|| !tctx
->io_wq
)
10710 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
10713 cpumask_clear(new_mask
);
10714 if (len
> cpumask_size())
10715 len
= cpumask_size();
10717 if (in_compat_syscall()) {
10718 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
10719 (const compat_ulong_t __user
*)arg
,
10720 len
* 8 /* CHAR_BIT */);
10722 ret
= copy_from_user(new_mask
, arg
, len
);
10726 free_cpumask_var(new_mask
);
10730 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
10731 free_cpumask_var(new_mask
);
10735 static int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
10737 struct io_uring_task
*tctx
= current
->io_uring
;
10739 if (!tctx
|| !tctx
->io_wq
)
10742 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
10745 static int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
10747 __must_hold(&ctx
->uring_lock
)
10749 struct io_tctx_node
*node
;
10750 struct io_uring_task
*tctx
= NULL
;
10751 struct io_sq_data
*sqd
= NULL
;
10752 __u32 new_count
[2];
10755 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
10757 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10758 if (new_count
[i
] > INT_MAX
)
10761 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
10762 sqd
= ctx
->sq_data
;
10765 * Observe the correct sqd->lock -> ctx->uring_lock
10766 * ordering. Fine to drop uring_lock here, we hold
10767 * a ref to the ctx.
10769 refcount_inc(&sqd
->refs
);
10770 mutex_unlock(&ctx
->uring_lock
);
10771 mutex_lock(&sqd
->lock
);
10772 mutex_lock(&ctx
->uring_lock
);
10774 tctx
= sqd
->thread
->io_uring
;
10777 tctx
= current
->io_uring
;
10780 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
10782 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10784 ctx
->iowq_limits
[i
] = new_count
[i
];
10785 ctx
->iowq_limits_set
= true;
10788 if (tctx
&& tctx
->io_wq
) {
10789 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
10793 memset(new_count
, 0, sizeof(new_count
));
10797 mutex_unlock(&sqd
->lock
);
10798 io_put_sq_data(sqd
);
10801 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
10804 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10808 /* now propagate the restriction to all registered users */
10809 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
10810 struct io_uring_task
*tctx
= node
->task
->io_uring
;
10812 if (WARN_ON_ONCE(!tctx
->io_wq
))
10815 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
10816 new_count
[i
] = ctx
->iowq_limits
[i
];
10817 /* ignore errors, it always returns zero anyway */
10818 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
10823 mutex_unlock(&sqd
->lock
);
10824 io_put_sq_data(sqd
);
10829 static bool io_register_op_must_quiesce(int op
)
10832 case IORING_REGISTER_BUFFERS
:
10833 case IORING_UNREGISTER_BUFFERS
:
10834 case IORING_REGISTER_FILES
:
10835 case IORING_UNREGISTER_FILES
:
10836 case IORING_REGISTER_FILES_UPDATE
:
10837 case IORING_REGISTER_PROBE
:
10838 case IORING_REGISTER_PERSONALITY
:
10839 case IORING_UNREGISTER_PERSONALITY
:
10840 case IORING_REGISTER_FILES2
:
10841 case IORING_REGISTER_FILES_UPDATE2
:
10842 case IORING_REGISTER_BUFFERS2
:
10843 case IORING_REGISTER_BUFFERS_UPDATE
:
10844 case IORING_REGISTER_IOWQ_AFF
:
10845 case IORING_UNREGISTER_IOWQ_AFF
:
10846 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
10853 static int io_ctx_quiesce(struct io_ring_ctx
*ctx
)
10857 percpu_ref_kill(&ctx
->refs
);
10860 * Drop uring mutex before waiting for references to exit. If another
10861 * thread is currently inside io_uring_enter() it might need to grab the
10862 * uring_lock to make progress. If we hold it here across the drain
10863 * wait, then we can deadlock. It's safe to drop the mutex here, since
10864 * no new references will come in after we've killed the percpu ref.
10866 mutex_unlock(&ctx
->uring_lock
);
10868 ret
= wait_for_completion_interruptible(&ctx
->ref_comp
);
10871 ret
= io_run_task_work_sig();
10872 } while (ret
>= 0);
10873 mutex_lock(&ctx
->uring_lock
);
10876 io_refs_resurrect(&ctx
->refs
, &ctx
->ref_comp
);
10880 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
10881 void __user
*arg
, unsigned nr_args
)
10882 __releases(ctx
->uring_lock
)
10883 __acquires(ctx
->uring_lock
)
10888 * We're inside the ring mutex, if the ref is already dying, then
10889 * someone else killed the ctx or is already going through
10890 * io_uring_register().
10892 if (percpu_ref_is_dying(&ctx
->refs
))
10895 if (ctx
->restricted
) {
10896 if (opcode
>= IORING_REGISTER_LAST
)
10898 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
10899 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
10903 if (io_register_op_must_quiesce(opcode
)) {
10904 ret
= io_ctx_quiesce(ctx
);
10910 case IORING_REGISTER_BUFFERS
:
10911 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
10913 case IORING_UNREGISTER_BUFFERS
:
10915 if (arg
|| nr_args
)
10917 ret
= io_sqe_buffers_unregister(ctx
);
10919 case IORING_REGISTER_FILES
:
10920 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
10922 case IORING_UNREGISTER_FILES
:
10924 if (arg
|| nr_args
)
10926 ret
= io_sqe_files_unregister(ctx
);
10928 case IORING_REGISTER_FILES_UPDATE
:
10929 ret
= io_register_files_update(ctx
, arg
, nr_args
);
10931 case IORING_REGISTER_EVENTFD
:
10932 case IORING_REGISTER_EVENTFD_ASYNC
:
10936 ret
= io_eventfd_register(ctx
, arg
);
10939 if (opcode
== IORING_REGISTER_EVENTFD_ASYNC
)
10940 ctx
->eventfd_async
= 1;
10942 ctx
->eventfd_async
= 0;
10944 case IORING_UNREGISTER_EVENTFD
:
10946 if (arg
|| nr_args
)
10948 ret
= io_eventfd_unregister(ctx
);
10950 case IORING_REGISTER_PROBE
:
10952 if (!arg
|| nr_args
> 256)
10954 ret
= io_probe(ctx
, arg
, nr_args
);
10956 case IORING_REGISTER_PERSONALITY
:
10958 if (arg
|| nr_args
)
10960 ret
= io_register_personality(ctx
);
10962 case IORING_UNREGISTER_PERSONALITY
:
10966 ret
= io_unregister_personality(ctx
, nr_args
);
10968 case IORING_REGISTER_ENABLE_RINGS
:
10970 if (arg
|| nr_args
)
10972 ret
= io_register_enable_rings(ctx
);
10974 case IORING_REGISTER_RESTRICTIONS
:
10975 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
10977 case IORING_REGISTER_FILES2
:
10978 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
10980 case IORING_REGISTER_FILES_UPDATE2
:
10981 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10984 case IORING_REGISTER_BUFFERS2
:
10985 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
10987 case IORING_REGISTER_BUFFERS_UPDATE
:
10988 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
10989 IORING_RSRC_BUFFER
);
10991 case IORING_REGISTER_IOWQ_AFF
:
10993 if (!arg
|| !nr_args
)
10995 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
10997 case IORING_UNREGISTER_IOWQ_AFF
:
10999 if (arg
|| nr_args
)
11001 ret
= io_unregister_iowq_aff(ctx
);
11003 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
11005 if (!arg
|| nr_args
!= 2)
11007 ret
= io_register_iowq_max_workers(ctx
, arg
);
11014 if (io_register_op_must_quiesce(opcode
)) {
11015 /* bring the ctx back to life */
11016 percpu_ref_reinit(&ctx
->refs
);
11017 reinit_completion(&ctx
->ref_comp
);
11022 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
11023 void __user
*, arg
, unsigned int, nr_args
)
11025 struct io_ring_ctx
*ctx
;
11034 if (f
.file
->f_op
!= &io_uring_fops
)
11037 ctx
= f
.file
->private_data
;
11039 io_run_task_work();
11041 mutex_lock(&ctx
->uring_lock
);
11042 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
11043 mutex_unlock(&ctx
->uring_lock
);
11044 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
,
11045 ctx
->cq_ev_fd
!= NULL
, ret
);
11051 static int __init
io_uring_init(void)
11053 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11054 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11055 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11058 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11059 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11060 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
11061 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
11062 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
11063 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
11064 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
11065 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
11066 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
11067 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
11068 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
11069 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
11070 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
11071 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
11072 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
11073 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
11074 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
11075 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
11076 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
11077 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
11078 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
11079 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
11080 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
11081 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
11082 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
11083 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
11084 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
11085 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
11086 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
11087 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
11088 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
11089 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
11090 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
11092 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
11093 sizeof(struct io_uring_rsrc_update
));
11094 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
11095 sizeof(struct io_uring_rsrc_update2
));
11097 /* ->buf_index is u16 */
11098 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS
>= (1u << 16));
11100 /* should fit into one byte */
11101 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
11103 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs
) != IORING_OP_LAST
);
11104 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
11106 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
11110 __initcall(io_uring_init
);